Activity-Based Travel Demand Models: A Primer (2014)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

69 3 ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) 3.1 FOUNDATIONS 3.1.1 Demand Models and Supply Models Activity-based models are travel demand models. Travel demand models may estimate the demand for travel by regional residents, the demand for travel for commercial purposes, or the demand for travel for special purposes or destinations such as special events or airports. This demand is typically characterized by in- formation about origins, destinations, timing, and modes of travel. Supply models predict the performance of the transportation system, given a set of input travel demand. These performance indicators include measures such as O-D travel times and costs by time of day and mode, and estimated link volumes. The time, costs, and other net- work impedance estimates produced by the supply model are then fed back as input to the demand model. Activity-based models forecast the demand for travel for regional residents: the purpose and number of activities to participate in, the amount and type of travel required to fulfi ll these activities, the destinations of these activi- ties, the mode of travel used to access activity locations, and the timing of this travel. This demand is primarily infl uenced by household and individual characteristics and by the per- formance of the transportation system as re- fl ected in travel times, costs, and accessibilities. The household and individual characteristics are input to the activity-based model based on exogenous sources of base-year and future-year demographic information. The network perfor- mance inputs to the activity-based model are produced by a network supply model, which is typically executed sequentially and iteratively with the activity-based model. 3.1.2 Aggregate Versus Disaggregate One of the distinguishing features of activity- based models is that they are typically imple- mented using a disaggregate microsimulation framework, in which the choices are pre- dicted at various decision-making levels, such as households, persons, tours, and trips. In a tradi tional trip-based model, aggregate esti- mates of demand are predicted fi rst. Then each subsequent step in the model system further disaggregates the overall aggregate estimates of demand. For example, the total number of trip productions and attractions by purpose are fi rst predicted during the trip generation step

70 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER for each TAZ. These total trip productions and attractions are then disaggregated to O-D pairs in the trip distribution step. Within each O-D pair, the number of trips is then further dis- aggregated to estimate the number or trips by mode during the trip mode choice step. Once all these disaggregation steps have been per- formed, the demand can be input to the net- work assignment model. In contrast, in an activity-based model system, disaggregate estimates of demand are predicted first, and then these estimates are aggregated by geography, time of day, and market segment for input in the network as- signment model. Producing disaggregate esti- mates of demand helps reduce bias in the estimates of demand and provide greater flex- ibility to analyze the impacts of policies and investments. However, the use of this disaggre- gate data must be informed by the analysis con- text. Although the activity-based model pro- duces precise disaggregate estimates of demand at detailed spatial and temporal resolutions for many market segments, the accuracy of these estimates at fine levels of disaggregation should be carefully considered in application. 3.1.3 Discrete Choice Models Activity-based models simulate the activity- travel decisions of households and individuals that collectively result in the activity patterns researchers observe. These activity patterns are composed of many smaller, often-related deci- sions, such as what time to depart home for work in the morning, what mode to take, whether to make an extra stop for groceries on the way home, and where to make that stop. Other longer-term decisions also have a bear- ing on activity and travel, such as the choices of where to work, where to live, how many cars to own, and whether or not to participate in an employer’s transit pass program. Most travel demand modelers are already familiar with mode choice models. In Figure 3.1 the left diagram depicts the choice from among 5 different mode choice options. Choice models, however, are also used to model other types of choices, such as the choice of destination and time of day depicted by the center and right- most diagrams, respectively. Choices made in space and time, respectively, are in reality continuous dimensions, which modelers parse into discrete units for analytical and computa- tional convenience. Some of the early activity- based models employed regression models and hazard-based duration models to predict activ- ity durations or ending times, given a starting point; however, in practice discrete choice for- mulations such as the ones shown in Figure 3.1 have proved to be easier to calibrate and inte- grate with other model structures and to incor- porate sensitivity to travel conditions that vary by time of day. For destination choices, the fundamental unit of analysis might be a zone, or it could be smaller units, such as a gridcell, microzone, or Figure 3.1. Choice structures applied to activity-travel dimensions of mode, space, and time. 2014.11.1 systems i which sp these dat time inte [Insert Fi [Caption Figur <H3>3.1 Activity- Figure 3. activity-b example, would pu the numb decisions depend o 8 C46 Primer n use to dat ace and time a are proces rvals, and fo gure 3.1] ] e 3.1. Choic .3.1 Choice based mode 2<FIG3.2> ased model as a TAZ), rchase a tra er of autom might be so n where ind FINAL for co e have used are transfo sed in surve r the m inte e structure Horizons ls consider d illustrates a ing systems the number nsit pass. W obiles to ow mewhat int ividual hous mposition.do 60-, 30-, and rmed into di ys, for the cr nance of ge s applied to ifferent tim few of the c . These inclu of househol hile workpla n is an exam erdependent ehold memb cx even 15-m screte interv eation of ne o-databases activity-tra time. e horizons in ommon long de the choic d automobil ce location ple of a hou . For examp ers work. In inute decisi als has imp tworks load . vel dimens addition to -term choic es of workp es to own, a and transit p sehold-leve le, the numb the case of on intervals ortant impli ing points, f ions of mod different un e models th lace locatio nd whether ass are indi l choice. In er of cars o some part-t . The ways i cations for h or assignme e, space, an its of analy at appear in n (defined, an individua vidual choic real life, the wned might ime worker 134 n ow nt d sis. for l es, se 3 s,

71 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) even parcel. Likewise, modelers parse time into intervals and choose a starting time interval for an activity. In practice, activity-based modeling systems in use to date have used 60-, 30-, and even 15-minute decision intervals. The ways in which space and time are transformed into dis- crete intervals has important implications for how these data are processed in surveys, for the creation of networks loading points, for assign- ment time intervals, and for the maintenance of geo-databases. 3.1.3.1 Choice Horizons Activity-based models consider different time horizons in addition to different units of anal- ysis. Figure 3.2 illustrates a few of the com- mon long-term choice models that appear in activity-based modeling systems. These include the choices of workplace location (defined, for example, as a TAZ), the number of household automobiles to own, and whether an individual would purchase a transit pass. While work- place location and transit pass are individual choices, the number of automobiles to own is an example of a household-level choice. In real life, these three decisions might be somewhat interdependent. For example, the number of cars owned might depend on where individ- ual household members work. In the case of some part-time workers, however, the direc- tion of causality might be reversed. In addition, whether someone bought a transit pass might also depend on workplace and the availability of automobiles. The sequence in which these decisions are represented in activity-based modeling systems is part of the model design. In addition, there have been models developed in a research setting that attempt to integrate these decisions into a more complex, single multidimensional choice. The set of choices shown in Figure 3.3 are simplified depictions of models aimed at daily tour pattern generation and certain stop de- tails. The left diagram represents a model that predicts the number of shopping tours in a daily pattern, given the existence of at least 1. This is an example of using a choice mode ap- proach to predict the frequency of occurrences of something like tours, where the number is likely to be small (e.g., 0, 1, 2, 3+). Figure 3.2. Examples of long-term and mobility choices. 2014.11.1 however, transit pa which the design. In integrate [Insert Fi [Caption T aimed at model th 1. This is somethin [Insert Fi [Caption 8 C46 Primer the directio ss might als se decision addition, th these decisi gure 3.2] ] Fi he set of cho daily tour p at predicts th an example g like tours, gure 3.3] ] FINAL for co n of causalit o depend on s are represe ere have be ons into a m gure 3.2. E ices shown attern genera e number o of using a c where the n mposition.do y might be workplace nted in activ en models d ore complex xamples of in Figure 3. tion and ce f shopping t hoice mode umber is lik cx reversed. In and the avai ity-based m eveloped in , single mu long-term a 3<FIG3.3> rtain stop de ours in a da approach to ely to be sm addition, wh lability of a odeling sys a research ltidimension nd mobility are simplifi tails. The le ily pattern, g predict the all (e.g., 0, ether so e utomobiles. tems is part setting that a al choice. choices. ed depiction ft diagram r iven the ex frequency o 1, 2, 3+). one bough The sequen of the mode ttempt to s of models epresents a istence of at f occurrenc 135 a ce in l least es of 2014.11.1 Figu T intermed efficient applied m first inter additiona ordered c simplest <H3>3.1 Because these inte described example, represent it is poss does not 8 C46 Primer re 3.3. Exam he center an iate stop bef for predictin ultiple time mediate stop l stops. The hoices; how model struct .3.2 Joint an many choice rdependenc as separate choosing be s the joint c ible to enum necessarily FINAL for co ple choice d right diagr ore or after g a “yes-no s. For exam and then re re are other ever, in pra ures work th d Conditio s are interd ies to the ex choices, su tween comb hoice of tou erate every lead to more mposition.do structures ams in Figu the primary ” type of res ple, it may b -evaluated ways to repr ctice, activit e best over nal Choices ependent, ac tent practica ch as mode inations of r primary de combination accuracy an cx for daily ac re 3.3 illust destination ponse. This e applied o to predict w esent this d y-based mod a wide rang tivity-based l. Most mod and destinat mode and d stination an of destinat d is less pra tivity and s rate binary c stop. Binary type of “add nce to predi hether there ecision proc elers have e of input c modeling s elers recog ion, are real estination. F d mode in a ion and mod ctical in ter top genera hoices of w choice stru stop” mod ct an initial is room in t ess, includin found that s ases. ystem desig nize that wh ly bundled c igure 3.4<F hierarchical e, all on the ms of mode tion models hether to ad ctures are el might be insertion of he schedule g linked an ome of the ns try to cap at have been hoices, for IG3.4> manner. W same level, l estimation 136 . d an a for d ture hile that and Figure 3.3. Example choice structures for daily activity nd stop gen ration models.

72 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER The center and right diagrams in Figure 3.3 illustrate binary choices of whether to add an intermediate stop before or after the primary destination stop. Binary choice structures are efficient for predicting a “yes-no” type of re- sponse. This type of “add stop” model might be applied multiple times. For example, it may be applied once to predict an initial insertion of a first intermediate stop and then re-evaluated to predict whether there is room in the sched- ule for additional stops. There are other ways to represent this decision process, including linked and ordered choices; however, in prac- tice, activity-based modelers have found that some of the simplest model structures work the best over a wide range of input cases. 3.1.3.2 Joint and Conditional Choices Because many choices are interdependent, activity-based modeling system designs try to capture these interdependencies to the extent practical. Most modelers recognize that what have been described as separate choices, such as mode and destination, are really bundled choices, for example, choosing between combi- nations of mode and destination. Figure 3.4 rep- resents the joint choice of tour primary destina- tion and mode in a hierarchical manner. While it is possible to enumerate every combination of destination and mode, all on the same level, that does not necessarily lead to more accuracy and is less practical in terms of model estima- tion and application. It is worth noting that this example shows the choice of destination first, and conditional on destination, the choice of mode. In some model systems, particularly in Europe, this ordering is reversed; however, the way it is shown here is more familiar to most U.S. modelers. The important takeaway is that the choice of destination conditions the choice of mode, and that the composite travel times and costs of the modes available to travel to each destination alternative affect the choice of the destination. Another example of a joint or conditional choice would be the choice of tour starting and ending times, as shown in Figure 3.5. Here there is an obvious logical constraint being enforced in which the tour ending time intervals must be later than the tour starting time intervals. Im- plicit in this choice of starting and ending times is the tour duration. The choice hierarchy may seem incontrovertible due to temporal ordering of starting times before ending times; however, the utility of ending time and duration may in- fluence the tour starting time. Yet another example of a conditional choice would be the choice of trip mode, conditional on tour mode. This would seem to be a rather obvious hierarchical relationship in which the mode chosen for the whole tour dictates what is available for individual trips on the tour. As Figure 3.4. Tour destination conditioning tour mode choice. Figure 3.5. Tour starting time conditioning tour ending time choice. 2014.11.1 applicatio condition this order modelers mode, an destinatio [Insert Fi [Caption A ending ti enforced intervals hierarchy times; ho [Insert Fi 8 C46 Primer n. It is wort al on destin ing is rever . The impor d that the co n alternativ gure 3.4] ] Fig nother exam mes, as show in which the . Implicit in may seem wever, the u gure 3.5] FINAL for co h noting tha ation, the ch sed; howeve tant takeawa mposite trav e affect the ure 3.4. Tou ple of a join n in Figure tour ending this choice o incontrovert tility of end mposition.do t this examp oice of mod r, the way it y is that the el times an choice of th r destinatio t or conditi 3.5<FIG3.5 time interv f starting an ible due to t ing time and cx le shows th e. In some m is shown he choice of d d costs f th e destination n condition onal choice >. Here the als must be d ending tim emporal ord duration m e choice of odel system re is more f stination co e modes ava . ing tour m would be th re is an obvi later than th es is the to ering of star ay influenc destination f s, particula amiliar to m nditions the ilable to tra ode choice. e choice of t ous logical e tour starti ur duration. ting times b e the tour sta irst, and rly in Europ ost U.S. choice of vel to each our starting constraint b ng time The choice efore endin rting time. 137 e, and eing g 2014.11.1 [Caption Y condition which the tour. As s and bike however, trip mode start and choices o certainly place ups these con [Insert Fi [Caption 8 C46 Primer ] Figure 3 et another e al on tour m mode chos hown in Fig unavailable there may b choice. For end times, t f destination conditional tream. It is ditional rela gure 3.6] ] FINAL for co .5. Tour sta xample of a ode. This w en for the w ure 3.6<FIG for the subs e one or sev example, a he decision s for those on tour mod part of the m tionships co mposition.do rting time conditional ould seem t hole tour di 3.6>, a per equent trip m eral other c fter choosin of whether a stops. As a c e, it is also odel design nsistently th cx conditionin choice wou o be a rather ctates what son chooses ode choice hoice decisi g the tour m nd how man onsequence conditional , and a chall roughout th g tour endi ld be the cho obvious hi is available to walk for s. In an acti ons that take ode, there m y intermed , while the c on a handfu enge in its i e model sys ng time cho ice of trip m erarchical re for individu the tour mo vity-based m place in be ay be the ch iate stops to hoice of trip l of other ch mplementat tem. ice. ode, lationship in al trips on th de, leaving S odeling sys tween tour a oices of tou insert, and t mode is oices that ta ion, to main 138 e OV tem, nd r he ke tain

73 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) shown in Figure 3.6, a person chooses to walk for the tour mode, leaving SOV and bike un- available for the subsequent trip mode choices. In an activity-based modeling system, however, there may be one or several other choice deci- sions that take place in between tour and trip mode choice. For example, after choosing the tour mode, there may be the choices of tour start and end times, the decision of whether and how many intermediate stops to insert, and the choices of destinations for those stops. As a consequence, while the choice of trip mode is certainly conditional on tour mode, it is also conditional on a handful of other choices that take place upstream. It is part of the model de- sign, and a challenge in its implementation, to maintain these conditional relationships con- sistently throughout the model system. 3.1.3.3 Utility Maximization Although there are a number of ways that one could go about attempting to model these choices, the travel demand modeling profession has come to accept the theoretical premise that these choices are not simply random. Rather, they are part of a deliberate process in which an individual trades off the worth of one al- ternative course of action versus another and chooses the alternative that is most likely to maximize his or her welfare. Accordingly, dis- crete choice models based on the principal of random utility maximization (RUM) have be- come the primary method for modeling activity and travel choices. While there are a number of competing theories of decision making that might work better than RUM for certain deci- sion contexts—such as elimination by aspects, regret minimization, and prospect theory— RUM has proved to be robust over a wide range of decision contexts. The assumption is that people choose the alternative that provides them with the highest utility among available alternatives. RUM has been found to be robust over a wide range of decision making and choice contexts. While it carries with it certain assumptions, it is applied probabilistically in model formulations, which allows modelers to account for measurement error and random heterogeneity in the popu- lation. Some of RUM’s less realistic assump- tions include that the decision maker has full knowledge of the attributes of each alterna- tive and pays equal attention to all available alternatives. 3.1.3.4 Random Utility Theory Because most travel demand modeling profes- sionals are at least familiar with mode choice models, this section highlights certain impor- tant aspects of discrete choice models that are central to their use in activity-based travel model ing and important terminology. Of par- ticular concern are the roles of choice sets and composite utility (or logsums) and how models are applied in a simulation environment. We assume that the decision maker selects the alternative that is perceived to offer the Figure 3.6. Tour mode choice conditioning trip mode choice, with intervening models. 2014.11.1 Figur <H3>3.1 Although the travel choices a individua alternativ models b method f of decisio eliminati over a wi 8 C46 Primer e 3.6. Tour .3.3 Utility M there are a demand mo re not simpl l trades off e that is mo ased on the or modeling n making th on by aspect de range of FINAL for co mode choic aximizatio number of w deling prof y random. R the worth of st likely to m principal of activity and at might wo s, regret mi decision con mposition.do e condition n ays that on ession has c ather, they one alterna aximize hi random util travel choi rk better th nimization, texts. cx ing trip mo e could go a ome to acce are part of a tive course o s or her welf ity maximiz ces. While t an RUM for and prospec de choice, w bout attemp pt the theore deliberate p f action ver are. Accord ation (RUM here are a nu certain dec t theory—R ith interve ting to mod tical premis rocess in w sus another ingly, discre ) have beco mber of com ision contex UM has pro ning model el these cho e that these hich an and chooses te choice me the prim peting the ts—such as ved to be ro 139 s. ices, the ary ories bust

74 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER maximum utility from a set of alternatives that are mutually exclusive, which we call the choice set. The observer does not know the true utilities; however, they may be inferred from the choices made. Sources of error include missing variables, unobserved taste variation (preferences), measurement error (actual versus perceived travel time), and using the incorrect functional form (e.g., linear, nonlinear, hierar- chical). We treat these errors as random and additive, such that Ui = Vi + ei. Total utility, Ui, is composed of a systematic portion, Vi, which we account for through the variables in our models, and a random compo- nent symbolized by the error term, ei. Using a mode choice example to illustrate how a utility function is formulated, we can write UtilityTransit = a * in-vehicle time + b * fare + c * (access time + egress time) + d * wait time + mode-specific constant Utility equals the weighted sum of the at- tributes of the alternative. The weights in the model are known as model parameters, shown here as a, b, c, and d. These parameters can be estimated from survey data, borrowed from another model, or asserted based on experience. The parameters convert the modal attributes in various units, such as minutes and cents, to a general value called a “util” (because they mea- sure utility). This has important implications for how the weights can be compared to one another. There is an additional term called a mode-specific (or alternative-specific) constant, which represents the value (in utils) of all of the attributes of the alternative that are not explic- itly listed in the utility equation. In the case of transit, this could include difficult-to-measure factors such as transit reliability, transit safety, and the influence of weather on the choice of transit. 3.1.3.5 Utility Expressions and Choice Probabilities The probability of choosing an alternative i from a set of choice alternatives C may be ex- pressed probabilistically as P(i : C) = Prob(Ui ≥ Uj, ∀j ⊂ C) = Prob(Vi + ei ≥ Vj + ej, ∀j ⊂ C) General assumptions for the distribution of the error term, following a Gumbel distribution, lead to the multinomial logit model: ∑ )( )()( = ∀ P i C V V : exp exp i j j This is the model that is used so often to represent mode choices. It carries with it the important simplifying assumption that error terms are independently and identically dis- tributed (IID). The IID property is important, because it assumes that a change in the utility (e.g., level of service or cost) of one alternative will have an equal proportional effect on the probabilities of choosing all of the other alter- natives, all else being equal. This assumption may not necessarily hold in all choice contexts. 3.1.3.6 Nested Models and Composite Utilities (Logsums) One reason for relaxing the IID assumption is to account for correlation between alternatives that may share unobserved similarities. For ex- ample, travelers may view two types of transit alternatives, such as bus and light rail, as being more similar (i.e., closer substitutes) than other model alternatives. Thus, a change in the level of service of bus should have a greater effect on light rail than it would on, for example, auto mobile alternatives. A common alternative form for this is the nested logit model. As pre-

75 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) viously discussed, activity-based modeling sys- tems make extensive use of such hierarchical or nested choices for more than just mode choice. Consider the example of the hierarchical choice of tour ending time, conditional on tour starting time, as shown in Figure 3.7. We can represent the conditional prob- ability of a choice that appears in a lower-level nest on the choice made in the upper-level nest in the following formulas. The probability of mode i is conditional on nest n: P(i) = P(i | n) * P(n) P i V V V V V V exp exp exp ln exp exp ln exp i n n j n n j n n n j n n j n m m j m m j mm ∑ ∑ ∑∑ ( ) ( ) ( ) ( ) ( ) = θ θ ∗ + θ θ           + θ θ                ∈ ∈ ∈∀ where qm is a dispersion parameter specific to each nest and reflects the correlation between alternatives in the same nest. To be consistent with utility maximization parameters, these parameters must have values greater than zero and less than or equal to one. Logsum terms represent composite utility of lower-level nested alternatives. The left term in the equation represents the conditional choice of mode, and the right term represents the unconditional choice of destina- tion zone. In the equation such a logsum vari- able is the term Vln exp j n n j n ∑ ( )θ  ∈ This is the natural log of the denominator of the lower-level nest and represents the com- posite utility of the nested alternative (modes). Note that the denominator for the lower-level choice (mode) appears in the utility expres- sion of the upper-level choice of destination zone. Because we take the natural log of this sum, this term is commonly referred to as the “logsum.” In choice theory, the logsum repre- sents the maximum expected utility that may be derived from the lower-level choice, which in this case is mode. In the choice of a desti- nation, the logsum term represents the mode- weighted accessibility for travel to each zone alternative. The equation also includes another portion of the utility of the zone alternatives V m that represents other attributes of the zone, such as attraction variables. Thus, it is com- mon in activity-based models to use composite accessibilities, such as mode choice logsums, to account for travel times and costs by all avail- able modes when choosing a destination. The assumption is that the destination is chosen first; however, this conditional ordering could be reversed. 3.1.3.7 The Importance of Choice Sets The choice set is the group of alternatives con- sidered to be available to the chooser in a given choice context. The role of choice set forma- tion and restrictions are important in activity- Figure 3.7. Use of logsum of the lower-nest mode choice alternatives in the upper nest. 2014.11.1 this is the extensive C starting t [Insert Fi [Caption Figur W on the ch is conditi <EQ> <EQ> )( PiP = ∑ ∈ = j iP )( 8 C46 Primer nested logi use of such onsider the ime, as show gure 3.7] ] e 3.7. Use o e can repre oice made i onal on nest )()|( nPni ∗ n nj nni V V exp( )exp( | | θ θ FINAL for co t model. As hierarchica example of t n in Figure f logsum of sent the con the upper- n: ∑ ∀ ⎢ ⎢ ⎣ ⎡ ⎢⎢⎣ ⎡ ⎢⎢⎣ ⎡ ∗ m V exp exp ) mposition.do previously d l or nested c he hierarchi 3.7<FIG3.7 the lower-n ditional prob level nest in ∑ ∑ ∈ ⎜⎜⎝ ⎛+ ⎜⎜⎝ ⎛+ j mm j nn V ln ln θ θ cx iscussed, a hoices for m cal choice o >. est mode c ability of a the followin ∈m mj n nnj V V exp( exp( | | θ θ ctivity-based ore than ju f tour endin hoice altern choice that g formulas ⎥ ⎥ ⎦ ⎤ ⎥⎥⎦ ⎤ ⎟⎟⎠ ⎞ ⎥⎥⎦ ⎤ ⎟⎟⎠ ⎞ m ) ) modeling s st mode cho g time, cond atives in th appears in a . The probab ystems mak ice. itional on t e upper nes lower-level ility of mod 143 e our t. nest e i

76 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER based modeling systems. This is particularly true for nested choices and conditional choice relationships. In conditional choice contexts, the upstream model choice will in many cases condition the availability of alternatives down- stream. For example, the choice of tour mode conditions the availability of certain trip modes. Generally speaking, if a person does not choose to drive for the tour mode, then we would not expect driving to be available for any trip on the tour. Similarly, we might expect that bicycle would be available only to persons who left home with a bicycle; however, in household surveys it is common to find exceptions to these assumptions as people sometimes leave cars and bikes behind, used rental or company vehicles, or have access to car-sharing or bike- sharing services. In addition, the presence or absence of an alternative in a lower-level choice may greatly affect the composite utility of the upper level choice. In a policy context, if one were to model the addition of a new transit service to the region that would greatly improve travel time by certain zones, then this addition of a new alternative to serve those zone pairs would make those destinations more attractive. This change in accessibility would be reflected in the mode choice logsums that would be used by an upstream destination-choice model and possi- bly even long-term workplace and automobile ownership choice models. 3.1.3.8 Other Considerations The multinomial and nested logit models de- scribed are by far the most common model forms used in practice in activity-based model- ing system. They owe their ubiquity to being relatively easy to comprehend and to imple- ment over a wide range of choice contexts. Use of different model forms is, of course, possible. Most variations focus on different methods for handling the error terms in the models to better account for heterogeneity across users and cor- relation across choice alternatives. Some choice model variations specify different forms of the dependent variable that may be more appli- cable in certain modeling contexts, such as or- dered choices and combinations of discrete and continuous choices. Still other models deviate from the RUM paradigm in an attempt to cap- ture different decision-making theories, such as risk aversion. Computational complexity, run time, and the ability to explain results to end users are the chief challenges of adopting the more advanced model forms in practice. Inter- ested readers may want to consult recent re- search in the area of discrete choice and related econometric models. Excellent basic references on discrete choice models include the works of Ben-Akiva and Lerman (1985); Hensher, Rose, and Greene (2005); Koppelman and Bhat (2006); and Train (2009). 3.1.4 Activity Pattern Structure An activity-based travel model differs from a trip-based model by modeling decisions to par- ticipate in activities. The central focus of the models is whether, when, and where to par- ticipate in activities and for how long. Travel is a derived demand resulting from the need for people to engage in activities outside the home. Trips are a means of traveling between activity locations and decisions related to trip scheduling, such as mode and departure time, are made to accommodate desired arrival and departure times from activity sites. In some activity-travel modeling systems, these deci- sions are coordinated between members of the same household. Activity-based travel models also are characterized by their disaggregate representation of individuals and households, which typically using simulation methods. This enables modelers to track these individuals and to effectively use their demographic character- istics in analysis.

77 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) 3.1.4.1 Activities Versus Trips Activity modeling does bear some resemblance to trip-based modeling in terms of generating activities, distributing them to locations and time periods, and choosing travel modes for them. Some activity purposes, such as work and school, have similar labels in the trip- based world; however, we actually model the trips within an activity-based modeling system as separate entities that allow persons to travel between activity locations. Activities have a duration, which we model, that has intrinsic value to the participant. People derive satisfaction from participating in activities, and we assume that the amounts of time that we observe people participating in ac- tivities reflect the utility the participants derive from it. Therefore, when we model the sched- ule of activities and travel, we take into account the expected amounts of time that individuals will spend in each activity; how they prioritize their time between work or school, and shop- ping and recreational activities; and how much time they are willing to devote to travel. Modeling activities also means allowing for the possibility of in-home substitutions and trade-offs, such as telecommuting from home, at-home leisure, eating, and other activities. This is important for modeling future scenarios in which gasoline prices are higher or predict- ing the impacts of online commerce and social media. One response that people may choose in reaction to high travel costs is to undertake ac- tivities at home. In addition, in-home activities of other household members are important. For example, many parents of young children time their work departure times and forego some discretionary activities out of the home so they can be at home for their children. Some of the most advanced activity-based modeling systems try to capture this dynamic. 3.1.4.2 Tours and Half-Tours A key aspect of activity-based travel models is that travel is organized around tours. A tour is a series of trips beginning and ending at home or work anchor location. By model- ing decisions on a tour basis, there is enforced consistency between the outbound and return portions of the tour, so that a mode chosen to go to work conditions the mode available for the return home. Common to tour-based activity modeling is the identification of a primary destination on each tour and the insertion of intermediate stops either before or after the primary destina- tion. In addition, there may be subtours within a tour. Figure 3.8 shows a home-based tour in which work is the primary activity/destination. 3.1.4.3 Primary Stops on Tours How to determine which stop in the tour is the primary destination is one key design deci- sion. While it is possible in recent tour-based household surveys to ask the primary purpose of the tour, this has not always been the case and is certainly not true in all surveys, particu- larly older ones. Using a hierarchical typology based on activity purposes is one method, which works well for work, school, and col- lege purposes, but for other purposes primacy is less clear. Other tie-breaking rules include the first stop on the tour, the stop farthest from the home anchor point, and the stop with the longest duration. These have important impli- cations for the construction of tour schedules, Figure 3.8. Home-based tour. 2014.11.18 C46 Primer FINAL for composition.docx 148 household members are important. For example, many parents of young children time their work departure times and forego some discretionary activities out of the home so they can be at home for their children. Some of the most advanced activity-based modeling systems try to capture this dynamic. <H3>3.1.4.2 Tours and Half-Tours A key aspect of activity-based travel models is that travel is organized around tours. A tour is a series of trips beginning and ending at home or work anchor location. By modeling decisions on a tour basis, there is enforced consist ncy b t een the outbound and return portions of the tour, so that a mode chosen to go to work conditions the mode available for the return home. C mmon to tour-based activity modeling is the identification of a primary destination on each tour and the insertion of intermediate stops either before or after the primary destination. In addition, e may be subtours within a tour. Figure 3.8<FIG3.8> shows a home-based tour in which work is the primary activity/destination. [Insert Figure 3.8] [Caption] Figure 3.8. Home-based tour. <H3>3.1.4.3 Primary Stops on Tours shop workhome

78 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER since time-window availability criteria for the insertion of intermediate stops would be influ- enced by both activity duration and travel time to the primary destination. 3.1.4.4 Intermediate Stops on Tours There can be zero or more intermediate stops on the tour, which are stops made between the anchor location and the primary destination. Some activity-based modeling systems refer to sequence of one or more stops between the anchor location and the primary destination as the first half of the tour, or outbound half, and the sequence of one or more stops between the primary destination and the anchor location as the second half of the tour, or return half. In Figure 3.8, there is one intermediate stop on the return half of the tour; the stop is between work and home. There are no stops on the out- bound half, which is between home and work. Whether to model stops on tours using this half-tour schema, or a more sequential method, is a design decision. 3.1.4.5 Work-Based Tours (or Subtours) As shown in Figure 3.9, the sequence of trips between work and lunch is referred to as a work-based tour (or subtour). In this case, the anchor location for the tour is the workplace, and the primary destination is lunch, and there are no intermediate stops on this subtour. In trip-based practice, both of these trips would have been cast as nonhome-based work trips. Although it is possible to allow nonwork loca- tions to be anchors for subtours, nonwork sub- tours are observed less frequently in survey and are difficult to identify accurately; therefore, this has generally not been done in practice. Because of the frequency of work-based sub- tours, however, these are typically generated as part of a daily activity pattern. 3.1.5 Activity Types In general, disaggregation of travel purposes by activity types makes activity-based models more sensitive to variations in travel behavior than trip-based models and allows them to be more accurate when matching person types with activity locations and times of day. While the labeling of activity types will vary from place to place, the following list of activity types is generally found in most activity-based modeling schemes: • At home; • Work at home; • Work (at workplace); • School (K–12); • University/college; • Personal business/medical; • Shopping; • Eat meal; • Social/recreational; and • Escort passenger. Two schemes for classifying these activi- ties are important to model specification be- cause they contextualize decision making. One scheme is based on the relative fixity in place and time of an activity, and the other scheme is related to whether activities and travel involve coordination with other household members. Figure 3.9. Home-based tour with a work- based subtour. 2014.11.18 C46 Primer FINAL for composition.docx 150 As shown in Figure 3.9<FIG3.9>, the sequence of trips between work and lunch is referred to as a work-based tour (or subtour). In this case, the anchor location for the tour is the workplace, and the primary destination is lunch, and there are no intermediate stops on this subtour. In trip-based practice, both of these trips would have been cast as nonhome-based work trips. Although it is possible to allow nonwork locations to be anchors for subtours, nonwork subtours are observed less frequently in survey and are difficult to identify accurately; therefore, this has generally not been done in practice. Becaus of the freque cy of work-based subtours, however, these are typically generated as part of a daily ac vity pa t rn. [Insert Figure 3.9] [Caption] Figure 3.9. Home-based tour with a work-based subtour. <H2>3.1.5 Activity Types In general, disaggregation of travel purposes by activity types makes activity-based models more sensitive to variations in travel behavior than trip-based models and allows them to be more accurate when matching person types with activity locations and times of day. While the labeling of activity types will vary from place to place, the following list of activity types are generally found in most activity-based modeling schemes: <BL> lunch shop workhome

79 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) 3.1.5.1 Mandatory, Maintenance, Discretionary, and At-Home Activities Activities are sometimes grouped into four general categories according to priority in the daily activity pattern schedule. The categories are mandatory, maintenance, discretionary, and at-home. Mandatory activities consist of work and school. They are the least flexible in terms of generation and scheduling and are the basic building blocks of activity schedules for workers and students. Some model systems differentiate between work-at-home (telecommuting) and work-out-of-home activities. Some models also categorize school activities by grade level. Maintenance activities include escort, shopping, and other maintenance (e.g., doctor’s visits). Some modeling systems model certain purposes explicitly, while others combine them into more general categories, like “other.” This is a design decision that should depend on lo- cal modeling needs. For example, in areas with a large contingent of senior citizens, explicit modeling of a medical activity purpose may be desirable. Many of these maintenance activi- ties are performed on behalf of the household, such as picking up or dropping off household members or going grocery shopping. In model systems that represent joint travel explicitly, the escort purpose may be replaced by more detailed descriptions. Discretionary activities include eating out, visiting, and other recreational activities. They are the most flexible in terms of generation and scheduling and are often substituted for in-home activities, particularly for households with poor accessibilities to recreational oppor- tunities. In some activity-based modeling systems, maintenance activities are grouped under the discretionary activities in recogni- tion of the fact that they often have similar scheduling flexibility and are often found on the same tour. Most activity-based models used in prac- tice classify at-home activities into working at home and other at-home activities. The reasons for this lack of further stratification are partially due to lack of survey informa- tion on in-home activities. In addition, how- ever, where in-home activity data have been collected analysts have found it difficult and perhaps unnecessary to effectively distinguish between nonwork at-home activities. In addition, some modeling systems also differentiate activities on work-based subtours from those belonging to the main home-based tour. One reason for this is because subtours tend to be more constrained in terms of time; therefore, activities on work-based subtours are likely to have significantly shorter average durations and travel distances. 3.1.5.2 Independent, Joint, and Escort Trips The definitions of independent, joint, and es- cort trips depend on the level of involvement between individuals, as can be determined from available household diary data. In older trip-based surveys, determining whether two household members participated in an activ- ity together or shared a ride could be difficult, because older survey methods did not stress consistency across individuals when report- ing events. In more recent activity-based sur- veys, survey firms have been more vigilant and systematic in ensuring consistency, making it somewhat easier to identify joint activities and travel, albeit not without some challenges in identifying information. As depicted in Figure 3.10, a fully joint tour is one in which two or more household members travel together to some out-of-home location at which they participate in an activity together. In this case the activity is a leisure ac- tivity, but generally any nonmandatory out-of- home activity type would qualify. Usually two or more household members who commute to

80 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER work or school together are considered to be not engaged in a joint activity, the assumption being that they are engaging in independent work or school activities, even if in close prox- imity. In such cases, there is joint travel on the tour, but this is simply represented as a shared- ride mode choice for both persons, not as an instance of joint activity participation. If, how- ever, they were to stop on the way home from work for a meal, the meal event would indeed represent joint activity participation, making this a partially joint tour. Figure 3.11 represents two variations on a partially joint tour. In the left diagram, two household members travel together to run an errand, a joint activity, after which one goes to work while the other goes shopping, and they both return home separately. In the right dia- gram, one household member goes to work at the beginning of a tour; the second household member meets the first at a store to go shop- ping together; they then continue on to dinner, and travel home together. In practice, fully joint tours are more common than partially joint tours. Figure 3.12 represents joint travel but not joint activity participation. In this example, an adult drops off a child at school on her way to work, and picks up the child from school Figure 3.10. Fully joint tour between two household members. Figure 3.11. Two variations on a partially joint tour. Figure 3.12. An escort tour. 2014.11.18 C46 Primer FINAL for composition.docx 153 to be more constrained in terms of time; therefore, activities on work-based subtours are likely to have significantly shorter average durations and travel distances. <H3>3.1.5.2 Independent, Joint, and Escort Trips The definitions of independent, joint, and escort trips depend on the level of involvement between individuals, as can be determined from available household diary data. In older trip- based surveys, determining whether two household members participated in an activity together or shared a ride could be difficult, because older survey methods did not stress consistency across individuals when reporting events. In more recent activity-based surveys, survey firms have been more vigilant and systematic in ensuring consistency, making it somewhat easier to identify joint activities and travel, albeit not without some challenges in identifying information. [Insert Figure 3.10] [Caption] Figure 3.10. Fully joint tour between t o household members. As depicted in Figure 3.10<FIG. 3.10>, a fully joint tour is one in which two or more household members travel together to some out-of-home location at which they participate in an activity together. In this case the activity is a leisure activity, but generally any nonmandatory home joint leisure 2014.11.18 C46 Primer FINAL for composition.docx 154 out-of-home activity type would qualify. Usually two or more household members who commute to work or school together are considered to be not engaged in a joint activity, the assumption being that they are engaging in independent work or school activities, even if in close proximity. In such cases, there is joint travel on the tour, but this is simply represented as a shared-ride mode choice for both persons, not as an instance of joint activity participation. If, however, they were to stop on the way home from work for a meal, the meal event would indeed represent joint activity participation, making this a partially joint tour. Figure 3.11<FIG3.11> represents two variations on a partially joint tour. In the left diagram, two household members travel together to run an errand, a joint activity, after which one goes to work while the other goes shopping, and they both return home separately. In the right diagram, one household member goes to work at the beginning of a tour; the second household member meets the first at a store to go shopping together; they then continue on to dinner, and travel home together. In p actice, fully joint tours are more common than partially joint tours. [Insert Figure 3.11] [Caption] . Figure 3.11. Two variations on a partially joint tour. home joint errand indep. shop work work joint shopping joint dinner home 2014.11.18 C46 Primer FINAL for composition.docx 155 [Insert Figure 3.12] [Caption] Figure 3.12. An escort tour. Figure 3.12<FIG. 3.12> represents joint travel but not joint activity participation. In this example, an adult drops off a child at school on her way to work, and picks up the child from school on her way home from work. The school and work activities are considered to be independent activities; however, the drop off and pick up events are generally referred to as escort activities for the driver. For the person being escorted (the child in this case), it is simply an independent school activity with a shared-ride mode. In practice, some activity-based modeling systems explicitly distinguish between activities performed with other household members, while others use a looser correlation specification. For example, it is possible to generate escort trips for an individual adult in the household, and the propensity to undertake escort trips will be correlated with household structure, particularly the presence and ages of children. This would be an example of a correlated model. An explicitly coordinated activity-based model system would generate some home drop off child pick up child work school

81 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) on her way home from work. The school and work activities are considered to be indepen- dent activities; however, the drop off and pick up events are generally referred to as escort activities for the driver. For the person being escorted (the child in this case), it is simply an independent school activity with a shared-ride mode. In practice, some activity-based modeling systems explicitly distinguish between activi- ties performed with other household members, while others use a looser correlation specifi- cation. For example, it is possible to generate escort trips for an individual adult in the house- hold, and the propensity to undertake escort trips will be correlated with household struc- ture, particularly the presence and ages of chil- dren. This would be an example of a correlated model. An explicitly coordinated activity-based model system would generate some escort trips as outcomes of a joint decision model in which children (or other household members) were directly matched with other household mem- bers as part of tour mode choice process. 3.1.6 Daily Activity Patterns There are many different ways in which choice elements can be represented and integrated into an activity-based modeling system. Differences in activity-based model design are expressed in how certain choices are represented structur- ally, as well as in their sequencing. Figure 3.13 depicts the representation of the choice of an overarching day pattern for an individual using two different schemes for representing daily ac- tivity patterns. The first model (Model A) represents day patterns as combination of tour types. Given a large number of combinations of different type tours, there could be thousands of individu- ally defined alternatives. Although practically this type of day-pattern model would eliminate those that are observed rarely and group cer- tain alternatives. The exact number of tours of each type would be chosen in a subsequent se- ries of models. The second model (Model B) defines day- pattern alternatives differently by characteriz- ing the day patterns as being either mandatory or nonmandatory, with a secondary choice of whether to include joint activities with other household members. As used here, a manda- tory pattern is defined as a pattern involving work, school, or college activities, but may include other, discretionary activities, such as eating out, shopping, and social/recreational. A nonmandatory day pattern would include only discretionary activities. Joint activities with Figure 3.13. Two different ways of representing daily activity pattern choices. 2014.11.1 escort tri members process. <H2>3.1 There are an activit how certa 3.13<FIG individua [Insert Fi [Caption F 8 C46 Primer ps as outcom ) were direc .6 Daily Ac many diffe y-based mo in choices a 3.13> depic l using two gure 3.13] ] igure 3.13. FINAL for co es of a join tly matched tivity Patte rent ways in deling syste re represent ts the repre different sch Two differe mposition.do t decision m with other h rns which choi m. Differenc ed structura sentation of emes for re nt ways of cx odel in whic ousehold m ce elements es in activit lly, as well the choice o presenting d representin h children ( embers as p can be repr y-based mo as in their se f an overarc aily activity g daily acti or other hou art of tour m esented and del design a quencing. F hing day pa patterns. vity pattern sehold ode choice integrated in re expressed igure ttern for an choices. 156 to in N on m an da to ry N on m an da to ry

82 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER other household members are an extra dimen- sion that could be added to either a manda- tory or nonmandatory day pattern. The exact number of mandatory, nonmandatory activities and tours, as well as joint activity participation would be determined in downstream models. 3.1.7 Implementation Framework There are many different ways in which devel- opers of activity-based models have structured the sequencing and information flow between model components. Figure 3.14 is a general- ized representation of the major model steps that are common in activity-based models used in practice. As the model system progresses, travelers make decisions: whether to travel, where to go, how many stops to make, what mode to choose, and so on. Earlier decisions influence and con- strain the decisions made later. For example, the number of vehicles owned, modeled in the automobile ownership (mobility) model, influ- ences the number of tours and the mode used on each tour. The mode used for the tour then influences the location of stops on the tour, and so on. This conditioning effect is referred to as “downward vertical integrity.” Activity-based models also use information from models that are lower in the model chain to inform the choices made by decision makers in upper-level models. This information typi- cally takes the form of accessibility variables, which are formed from the composite utility of a lower-level choice. For example, a mode choice logsum, which reflects accessibility by all modes of transport, can be used to inform the choice of destination for the tour or stop. This representation of the composite utility represents the maximum expected utility that the decision maker can expect to receive from a lower-level choice, before making that choice. This flow of composite utilities back up the Figure 3.14. Major steps and information flow in an activity-based modeling system. Figure 3.14. Major steps and information flow in an activity-based modeling system. Model Inputs Synthetic Population Long-term Choices Mobility Choices Daily Activity Patterns Tour & Trip Details Trip Assignment Model Outputs Upward Integrity: Expected utility (accessibility) of choice alternatives in lower models affects choices made in higher models Downward Integrity: Choices made in higher models affect choices made in lower models

83 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) model stream is referred to as “upward verti- cal integrity.” Together the downward and up- ward vertical integrity built into activity-based model system designs help to ensure a high degree of internal consistency among space, time, and mode dimensions and to recognize their interdependence. 3.1.8 Disaggregation One of the chief strengths of disaggregate modeling methods is elimination of aggrega- tion bias that may threaten the validity of forecasted responses to transportation system and policy changes. Activity-based models do this by simulating activity-travel patterns of individual travelers and their interactions with other household members, which has led activity-based modeling practice to use micro- simulation techniques, typically Monte Carlo methods. As an outcome, this creates forecasts in which each trip is represented as a whole number (integer), a stark departure from con- ventional trip-based modeling practice. Four- step models generate interval values for trips and then allocate fractions of trips to TAZs, modes, and time periods. This difference be- tween whole trips and fractions has important implications for both computational efficiency and forecast variance. 3.1.8.1 Logit Models and Aggregation Bias Figure 3.15 illustrates the issue of aggregation bias using logit models. The horizontal axis represents the cost of a choice for two differ- ent decision makers, Person A and Person B. The vertical axis represents the resulting logit- calculated choice probabilities. The probabili- ties for each individual, as predicted by the model, are quite different, and their average (circled) is halfway in between. If we were to aggregate these two individuals and take their average cost, then we would obtain a different probability, following the lighter-gray dashed lines. The probability of this average cost is dif- ferent from the average probability we obtain when we calculate each person’s probability. Figure 3.15. Aggregation bias in application of logit models. 2014.11.1 Figure 3. axis repre The verti for eac halfway i then we w probabili calculate [Insert Fi [Caption 8 C46 Primer 15<FIG3.15 sents the co cal axis repr individual, n be ween. ould obtain ty of this av each person gure 3.15] ] Figu FINAL for co > illustrates st of a choic esents the re s predicted If we were t a different erage cost i ’s probabili re 3.15. Ag mposition.do the issue of e for two di sulting logi by the mod o aggregat probability, different fr ty. gregation b cx aggregation fferent deci t-calculated l, are quite d these two in following th om the aver ias in appli bias using sion makers choice prob ifferent, an dividuals an e lighter-gr age probabi cation of lo logit model , Person A a abilities. Th d their avera d take their ay dashed li lity we obtai git models. s. The horiz nd Person B e probabilit ge (circled) average cos nes. The n when we 160 ontal . ies is t,

84 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER Another aspect of aggregation bias is shown in Figure 3.16. The average impact of a change is not equal to the impact calculated at the average of the explanatory variables; this is symbolized by the tangents to the curve, representing slopes at each point. Again, the black lines represent the individual outcomes, and the gray line represents the slope corre- sponding to the averaged outcome. Because of the sigmoid (S-shape of the curve), the logit model is most sensitive (elastic) to change in inputs at its center region and is relatively less sensitive ( inelastic) to changes in inputs at its top and bottom ends. This is one reason why some aggre gate models predict larger shifts in response to scenario inputs changes than dis- aggregate models. A real-life example might be a mode shift in response to a new toll charge. Imagine the perceived cost of the toll being affected by per- sonal values of time, where Person A has a high willingness to pay (so perceived cost is not that onerous) and Person B has a low willingness to pay (so perceived cost is considered to be very onerous). Because both persons are already at the far ends of the distribution, they are less likely to react to a cost change by changing their baseline choices. By grouping travelers under a single average value of time, however, the per- ceived cost represents an average condition, the gray slope, and has the potential to over estimate the elasticity of response to the toll. 3.1.8.2 Computational Efficiency of Disaggregate Data Structures In an aggregate model framework, used in most trip-based models, it is necessary to create and maintain a separate trip table for each sociode- mographic segment that one wants to use as an explanatory variable in a model. For example, to represent household automobile ownership levels or income group affiliation in a mode choice model, there needs to be separate trip tables for each level of each variable. To add more variables, such as trip purpose or house- hold type, requires creating separate trip tables Figure 3.16. Aggregation bias in application of logit models. 2014.11.1 A impact of variables Again, th correspon model is sensitive aggregate models. [Insert Fi [Caption 8 C46 Primer nother aspe a change is ; this is sym e black line ding to the most sensiti (inelastic) t models pre gure 3.16] ] Figu FINAL for co ct f aggreg not equal to bolized by t s represent t averaged ou ve (elastic) o changes in dict larger s re 3.16. Ag mposition.do tion bias is the impact he tangents he individua tcome. Bec to change in inputs at its hifts in resp gregation b cx shown in Fi calculated a to the curve l outcomes, ause of the s inputs at its top and bo onse t scen ias in appli gure 3.16<F t the averag , representin and the gra igmoid (S-s center regi tto ends. T ario inputs cation of lo IG3.16>. T e of the exp g slopes at e y line repres hape of the on and is rel his is one re changes than git models. he average lanatory ach point. ents the slo curve), the l atively less ason why s disaggrega 161 pe ogit ome te

85 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) for each level of each variable, a step that can be inefficient if not intractable for large num- bers of variables. Each additional trip table leads to further fractional allocations of trips, increasingly sparse matrices, and proportional increases in computer memory requirements and hard-disk space. In contrast, disaggregate models main- tain representation in a list-table format. An example of the trip-list table format is shown in Figure 3.17, depicting a day’s worth of ac- tivities for one individual. Each record repre- sents a trip, and each column can represent an explanatory variable to be used in the model. To specify an additional variable in a model is simply a matter of adding a column to the trip- list table; the column addition has a relatively small impact on computational resources, com- pared with aggregate methods. The trip-list table structure is also conve- nient for querying model outputs. The table format in the figure lends itself well to creat- ing new variables and grouping outcomes by household or person attributes, geographic unit, activity purpose, trip or tour mode, and potentially other variables. For example, we might want to calculate activity and trip dura- tion and add gender as an explanatory vari- able, so that we can summarize the amount of time spent shopping or commuting and make a comparison between women and men. 3.1.8.3 Monte Carlo Simulation Prediction using simulation methods is an- other important difference between activity- based modeling systems and trip-based models. Table 3.1 uses mode choice models as an ex- ample since this is the one place where discrete Hhid Perid Dayno Tourno Tripno Activity OTAZ DTAZ Depart Mode Age Inc 626 1 2 1 1 Escort 39 82 7:00 HOV2 55 4 626 1 2 1 2 Work 82 1290 7:10 SOV 55 4 626 1 2 1 3 HHbus 1290 160 15:25 SOV 55 4 626 1 2 1 4 Shopping 160 96 16:10 SOV 55 4 626 1 2 0 5 Home 96 39 17:00 SOV 55 4 626 1 2 2 6 Jnt shop 39 87 19:00 HOV2 55 4 626 1 2 0 7 Home 87 39 21:00 HOV2 55 4 Figure 3.17. Example of activity-based model trip-list table structure. TABLE 3.1. COMPARISON OF 4-STEP AND MICROSIMULATION IMPLEMENTATIONS Conventional 4-Step Model-Mode Choice Activity/Tour-Based/Simulation-Mode Choice For each market segment, defined by trip purpose and household demographic group, predict the probability of each mode for each O-D pair. Predict probability of each simulated chooser selecting each mode for a specific O-D pair and purpose. Allocate the number of trips for each market segment and O-D pair to modes in proportion to their predicted probabilities. Use Monte Carlo random draws to predict a single mode choice. Sum over market segments to form trip tables. Sum over choosers and purposes, grouped by O-D pair, to form trip tables for network assignment.

86 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER choice models are consistently used in trip- based modeling systems. In a trip-based model, market segments are defined by trip purpose and household demographic groups, and the model predicts the probability of each mode for each O-D pair. The model then allocates the fraction of trips for each segment and O-D pair to modes in proportion to their predicted prob- abilities. This is an aggregate prediction, which is then summed over all market segments to form trip tables. In an activity-based model using simula- tion, the model predicts the probability of each simulated chooser selecting a mode for a specific O-D pair and purpose and then uses Monte Carlo random draws to predict a single mode choice, represented in integer format. To form trip tables for network assignment, the model aggregates over all of the individual trip records, grouped by O-D pair. There are three basic steps in Monte Carlo prediction. Here mode choice is used to illus- trate an example, but the same applies to any of the choice models discussed thus far. 1. Predict the probability and cumulative probability for each alternative outcome as shown in Table 3.2. 2. Draw a random number from a uniform distribution on the unit interval (0…1): for example, Rand() = 0.76. 3. Create lower and upper bounds for each alternative. Select the alternative with the range on the cumulative probability array that includes the random draw, as shown (bold and shaded) in Table 3.3. Monte Carlo simulation has advan- tages and disadvantages compared with ex- pected values used in trip-based models. The key advantage of Monte Carlo simulation is that explanatory variables can be included in models with little computational overhead (as opposed to aggregate models, in which each market segment increases the number of calcu- lations exponentially). Outcomes of previous model components can be used as explanatory variables in subsequent components. 3.1.8.4 Simulation Variance The key disadvantage of Monte Carlo simula- tion is that multiple runs are required in order to determine the expected values, or average results, for certain model outputs. This has im- plications for forecasting, but there are ways to compensate for the disadvantages that are em- ployed in most practical activity-based models. The amount of variability in model results depends on the number of agents making the choice decision and the size of the probabil- ity of the choice. For example, lower probability choices have more variability in their outcomes TABLE 3.3. EXAMPLE OF MONTE CARLO LOWER AND UPPER BOUNDS AND ALTERNATIVE SELECTION Monte Carlo SOV HOV Bus LRT Walk Bike Lower Bound 0.00 0.57 0.85 0.88 0.96 0.97 Upper Bound 0.56 0.84 0.87 0.95 0.96 1.00 TABLE 3.2. EXAMPLE OF MONTE CARLO PROBABILITY AND CUMULATIVE PROBABILITY Monte Carlo SOV HOV Bus LRT Walk Bike Probability 0.56 0.28 0.03 0.08 0.01 0.04 Cumulative Probability 0.56 0.84 0.87 0.95 0.96 1.00 Note: LRT = light rail transit.

87 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) than higher probability choices. Because most outputs from activity-based models are aggre- gations of choices, one run of the model can be a sufficient indication of the expected outcome from an area-wide policy. Following the law of large numbers, as the number of model runs gets very large, aggregate outcome averages will converge to a consistent estimate. In terms of variance, regional VMT, vehicle hours of travel (VHT), district-level tour flows, tours and trips by mode, and higher facility- type link estimates and transit line boardings tend to be very stable from run to run. For more dis aggregate analysis, such as TAZ-level origins and destinations, lower facility-type link loadings, and lower ridership transit routes, can have more variation and therefore multiple runs of the model system may be required, where results are averaged across the runs. There are ways of compensating for Monte Carlo variability. One way is to fix the ran- dom number seed in the functions used by the program to generate random numbers. While this can be a bit complex for a large number of choices, this results in the program generat- ing the same sequence of random numbers for successive runs, which means that outcomes will only vary according to changes in inputs. This ensures stability from run to run but at the cost of representing only one possible outcome from the model. For some applications, it may be preferable to do many runs and average the aggregate results to obtain an expected value. The ability to control the random seeds and sequences in model application is useful because it provides confidence that the model imple- mentation runs consistently and that it produces the same outputs, given the same inputs. Once the random numbers are controlled, users are able to exploit this model feature to run the model system more efficiently and to produce better performance measures. Activity-based models should be run multiple times in order to account for simulation variation (also referred to as simulation error). The dis aggregate outputs can be used to produce distributions and con- fidence intervals for core activity-based model measures, in addition to average values that can be used as inputs to tradi tional static assignment models. If regional-scale DTA and simulation models are adopted, use of dis aggregate outputs to produce multiple network simulations may be desirable. Empirical testing of the number of runs required to produce results with confidence is also desirable, although only a limited num- ber of regions have implemented this in practice. The number of runs is dependent on the spatial, temporal, or typological detail that is of interest. Analysis of smaller spatial, temporal, or typo- logical segments requires more runs. An additional important issue is how these performance indicators are transmitted and explained to decision makers. Because tradi- tional 4-step models do not produce distribu- tions of outcomes given fixed inputs, decision makers are most familiar with the single-point forecasts generated by these models. Thus, the communication and interpretation of activity- based models that include ranges of potential outcomes presents new opportunities and chal- lenges. Distributions or ranges of outcomes provide the advantage of illustrating the degree of uncertainty around different outcomes, but may be misinterpreted by decision makers if not properly presented. 3.1.8.5 Convergence and Equilibration As described in Chapter 2, linked demand- and-supply model systems such as activity- based model systems typically include pro- cesses of iterative feedback. These feedback processes are implemented in the network supply model as well as between the net- work supply model and the activity-based model components. Iterative feedback is used to ensure that the models are achieving con- vergence to an equilibrium, or at least a stable,

88 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER condition. Convergence is important within the context of activity-based model systems because it provides confidence in the integrity of the model system and helps ensure that the model will be a useful analytic tool. To be use- ful in an application context, the model sys- tem must produce similar outputs when seeded with similar outputs, so that the analyst can have confidence that changes in model outputs can be attributed to changes in model inputs and not to “noise” resulting from the configu- ration of the model system. The focus of this discussion is on the sys- tematic and iterative exchange of informa- tion implemented between the activity-based demand model and a network assignment model in order to converge to a stable solution. The most general description of this process is that the activity-based demand model produces estimates of demand that are then used as input to the network assignment model. The network assignment model assigns this travel demand to network paths, producing estimates of link volumes and speeds, and O-D travel times and costs by travel mode, time of day, and possibly user class. This process is iteratively repeated in order to achieve stable estimates of travel demand, link volumes, speeds, and travel times and costs. The simplest, naïve way to configure the model system to iterate between the demand-and-supply components is to feed the estimates of demand from the activity- based model directly into the network assign- ment model and then to feed the estimates of travel times and costs from the network model directly into the activity-based model. However, using this direct feedback ap- proach may necessitate iterating between the activity-based model and the network assign- ment model many times before an acceptable level of stability in model system outputs is achieved. Given the time required to execute the entire model system, lengthy run times may result that can compromise the usefulness of the model in an application context. To help the model system achieve an ac- ceptable level of stability more quickly, a number of convergence strategies have been developed. First, enforcement strategies are often employed. These enforcement strategies are similar to those used in traditional trip- based models and include methods such as the averaging of travel demand across successive iterations before network assignment, the aver- aging of network skims or impedances before demand simulation, and the averaging of link- level volume or impedances before the genera- tion of updated skims. Figure 3.18 illustrates two commonly used enforcement methods. Note that it is recommended to use both strate- gies simultaneously. Figure 3.18. Model system enforcement strategies. 2014.11.18 C46 Primer FINAL for composition.docx 170 and the averaging of link-level volume or impedances before the generation of updated skims. Figure 3.18 <FIG3.18> illustrates two commonly used enforcement methods. Note that it is recommended to use both strategies simultaneously. [Insert Figure 3.18] [Caption] Figure 3.18. Model system enforcement strategies. Second, the activity-based model sample rates are often varied. Because the activity- based model is implemented in a disaggregate Monte Carlo simulation framework, it is possible to run the model using only subsamples of the population; this option can significantly reduce model run times. Many activity-based model systems have been configured so that earlier iterations of the model run employ small subsamples, such as 10% or 25% of the regional population, while later iterations of the model run use a full 100% sample. Third, the overall model system can be configured to run a fixed number of iterations or to run until a prespecified convergence gap criterion has been achieved. The gap criterion is typically based on changes in demand or changes in travel times. Note that if a convergence criterion is used in conjunction with an enforcement strategy, then the criterion must not use the Activity-Based Model Network Model Average Demand Activity-Based Model Network Model Average Skims Direct Demand Direct Skims

89 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) Second, the activity-based model sample rates are often varied. Because the activity- based model is implemented in a disaggregate Monte Carlo simulation framework, it is pos- sible to run the model using only subsamples of the population; this option can significantly reduce model run times. Many activity-based model systems have been configured so that earlier iterations of the model run employ small subsamples, such as 10% or 25% of the re- gional population, while later iterations of the model run use a full 100% sample. Third, the overall model system can be configured to run a fixed number of iterations or to run until a prespecified convergence gap criterion has been achieved. The gap criterion is typically based on changes in demand or changes in travel times. Note that if a conver- gence criterion is used in conjunction with an enforcement strategy, then the criterion must not use the enforcement metric. For example, if demand is averaged successively across itera- tions, changes in demand cannot be used as the convergence criterion. For model system con- vergence, it is more common in practice to assert a fixed number of iterations, typically between three and ten, than to use a convergence crite- rion. Use of a fixed number of iterations should be based on an empirical investi gation that identifies the degree of convergence associated with different configurations. Finally, note that different levels of con- vergence, and by extension different numbers and types of iterative execution of the model system, are required for different application contexts. For example, analyses of detailed geographic, spatial, modal, or demographic segments require higher levels of convergence in order to ensure that difference between alter- natives are attributable to policy or investment being tested, and not because of the way the model system is configured. 3.2 DESIGN This section presents concepts relevant to activity-based model system design and pro- vides an overview of an overall approach to activity-based model design. The following sections first consider concepts related to spa- tial scale, temporal scale, and typological or market segmentation detail. As described in Chapter 2, the resolution, or level of detail, associated with these key dimensions is a criti- cal consideration, and activity-based models are distinguished from trip-based models in two important ways. First, in contrast to the zone-based looping structure of most trip- based models, adding more zones, more time periods, more demographic segments, or more trip purposes does not greatly increase the run time of activity-based model components. Add- ing detail along all of those dimensions has not been practical in the past because the run time and data storage requirements in trip-based models are proportional to the square of the number of zones, times the number of popula- tion segments, times the number of trip pur- poses, times the number of time periods. In an activity-based demand microsimulation, however, the run time depends primarily on the number of different households and per- sons simulated. The amount of detail used in the various dimensions may add to memory re- quirements but will not substantially influence run times. This fundamental difference is what has made it possible to include more detail in activity-based models. 3.2.1 Spatial Scale Activity-based models and, more generally, many travel demand models are intrinsically spatial models. The locations of households, employment, tour and trip origins and destina- tions, and many other model inputs and out- puts are spatial. A number of core components

90 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER of activity-based models systems, such as usual work and school location models, tour destina- tion location models, and stop location models predict travelers’ choices of locations. Spatial scale refers to the resolution or level of detail used to define the spatial units that collectively make up the region and that are used to characterize key model inputs, out- puts, and sensitivities. There is no correct spa- tial scale. Rather, when seeking to identify the appro priate level of detail to incorporate into the model system design, it is necessary to con- sider critical issues such as the types of sensi- tivities to policies and projects that the model system is required to have, in conjunction with considering the type and availability of spatial data needed to implement and apply the model. Key model system data inputs for which spatial information is required include • Household and population totals, poten- tially incorporating key demographic seg- mentations such as income, age, or other attributes; • Employment information, often in the form of employment totals by industrial sector; • School enrollment by grade; • Parking supply and cost information; • Hotel rooms and open space, and other urban form buffer or proximity-based measures; • Skims of travel times and costs by mode and time of day; and • Accessibility indicators. Generally speaking, use of fewer, larger spatial units reduces model run times and also reduces the level of data preparation burden. But larger spatial units introduce aggregation bias and reduce the sensitivity of the model to effects such as the local land use mix or the distance to transit. Conversely, use of a greater number of smaller spatial units often increases model run times but also reduces aggregation bias and increases the sensitivity to small-scale land use and transportation system effects. Multiple spatial scales may be used within a model system. For example, microzone or parcel geographies may be used to represent the locations of employment and population for measuring the attractiveness of locations, while TAZ geographies may be used to repre- sent automobile travel times and transit access points used to represent transit travel times. It may be useful to use very small spatial units when estimating walk mode travel times and distances, either for entire trips or to access transit, but this may not be necessary when estimating the travel times and distances for long automobile trips. The spatial scale of the activity-based model should also be defined in coordination with the spatial scale used for the network assignment model. The following sec- tions describe some of the spatial scales that are commonly used in activity-based model systems. 3.2.1.1 Zones Travel analysis zones are used in most travel demand model systems. The term TAZ is generic and does not imply or refer to any spe- cific scale. However, TAZs are often defined so that they are similar to or consistent with an existing geographic system such as a region’s Census tracts or Census block groups. The number of TAZs in a region typically ranges from 500 to 5,000. However, within a re- gion, there may be a fair amount of variation in size among the TAZs. TAZs defined for the purposes of trip-based models can be readily used in activity-based model development. As a result, implementing an activity-based model using a traditional TAZ-level of spatial of detail is relatively straightforward because TAZ-level information developed to support trip-based

91 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) model implementation can be used directly, or easily translated to, the activity-based model implementation. 3.2.1.2 Microzones Like TAZ, the term microzone is a generic term that does not refer to a specific scale. Instead this term is intended to describe a geographic system that incorporates more spatial de- tail than a typical TAZ system. In a number of regions, microzones have been defined at a resolution that is similar to that of Census blocks, although the block geography is usually modified to ensure that the individual micro- zones will be meaningful within the model sys- tem. For example, blocks that represent water features such as rivers or lakes may be com- bined with adjacent microzones. Developing microzone-level spatial information, especially for future-year scenarios, can be more involved than developing TAZ-level spatial information. However, there are a number of spatially de- tailed, publically available datasets that can be used to create these microzone-level assump- tions. A typical model might include 30,000– 150,000 microzones, an order of magnitude more than the typical number of TAZs but also less than a typical number of parcels in a region. 3.2.1.3 Parcels Parcels have a more specific definition than TAZs or microzones. Parcel geographies are most often defined by local-level municipal and county tax assessors’ offices. Parcels are usually extremely fine-grained, with each spa- tial unit often corresponding to the geography asso ciated with a single building. However, as with TAZs and microzones, there is significant variation in parcel sizes. For example, large in- stitutions that contain diversity of buildings, employment, and uses may be represented by a single parcel. Using a parcel-level spatial scale can provide the greatest ability to incorporate local-level, smaller-scale land use and trans- portation system attributes, such as the mix of employment within a short walking distance or the distance to the nearest actual transit stop. Developing, maintaining, and forecasting parcel-level attributes requires more effort than developing similar TAZ-level or microzone- level attributes, especially on the employment side. There are often inconsistencies and errors in the base-year or observed data sources, and developing future-year parcels requires careful consideration of the sources for detailed future population and employment assumptions and potentially methods and practices for splitting parcels as development occurs. 3.2.2 Temporal Scale and Scheduling The explicit representation of the time-of- day and scheduling choices, and the relation- ship between these schedule choices and other choices of tour and trip destinations and travel modes, is one of the primary features that dis- tinguishes activity-based model systems from trip-based models. In many trip-based models, travel demand is estimated at a daily level, and a set of fixed factors may be applied to dis- aggregate this daily demand to time periods in order to generate time-period-specific estimates of network performance, such a peak hour or peak period traffic volumes and speeds. But use of fixed factors renders the model system insen- sitive to many influences on travelers’ choices of travel time, such as accessibilities and the number or type of other household and person activities. The following sections describe some of the issues of temporal scale, then address no- tions of activity scheduling, and finally explain how scheduling and time-of-day choice relate to other elements of the activity-based model system.

92 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER 3.2.2.1 Temporal Scale People experience time as a continuous phe- nomenon, with one moment seamlessly transi- tioning to the next; in activity-based model systems time is broken down into discrete intervals or time periods. As with the represen- tation of space, the resolution used to define these time periods can vary from one activity- based model to another, and different temporal resolutions may be used even within the same activity-based model system. The temporal scale used in the models defines the alternatives that can be used in the time-of-day and scheduling models. The earliest activity-based models tended to use a relatively coarse temporal scale of four or five broad time periods, typically defined consis- tently with the five or six time periods used in the network assignment model component of the overall model system. Shortly thereafter, more detailed temporal scales were incorpo- rated into activity-based model systems, in- cluding hours and half-hours. Some activity- based model systems have incorporated the use of quasi continuous time. These different tem- poral scales result in different model sensitivi- ties, run times, and complexity. Because these time periods are often used in combination with each other (for example, jointly predict- ing the time a traveler leaves his or her home in the morning and the time he or she returns home in the evening), use of more detailed tem- poral resolutions leads to a multiplicative in- crease in the number of time-period combina- tions that the model system needs to consider. However, the potential benefit of this increased temporal detail is improved model sensitivity. Often, different temporal scales are used within the same overall model system. Spe- cifically, it is common practice to use detailed time periods such as hours or half-hours to support activity scheduling, while using only broad time periods when assigning travel demand in static network assignment models. Most activity-based models used in practice are linked to static network assignment models in order to generate an estimate of network impedances required for input to the model components such as the scheduling models. Static assignment models cannot reasonably be applied to large regions for very short time intervals because the time required to complete a given trip can exceed the boundaries of the static network assignment time period. As a result, many activity-based model systems are linked to static network assignments that use five to eight time periods in order to portray the main differences in network performance by time-of-day periods. 3.2.2.2 Activity Scheduling Scheduling, or time-of-day, model components are included in activity-based models to rep- resent the important fundamental dimension of time in activity and travel choice. At the simplest level, time-of-day models are used to predict when activities start and end, as well as their duration. In most activity-based model systems, there are two levels at which schedule choice is considered: the tour level and the trip level. Trip-level scheduling is constrained by tour-level scheduling. Activity-based models have generally em- ployed one of two primary approaches for representing the scheduling process. Most activity-based models used in practice have im- plemented a scheduling hierarchy to schedule tours first. Using this hierarchy, mandatory pur- pose activities such as work and school tours are scheduled first, followed by maintenance purpose activities, and finally discretionary purpose activities such as social/ recreational tours. Some activity-based models also incor- porate intra-household interactions, and these models typically schedule any joint travel made by members of the same household after sched-

93 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) uling individual mandatory activities but be- fore scheduling any individual maintenance or discretionary activities. An alternative to the hierarchy-based approach is to build the sched- ule chronologically through the day, perhaps starting with an initial basic schedule. It should be noted that even within those models that use a scheduling hierarchy, there are different approaches. For example, some activity-based models assume an initial over- all daily framework describing when travelers leave their homes, when they return, and then populate the travelers’ entire day by adding details such as stops and departure times. In contrast, other activity-based models define an initial set of primary activities and then build out the schedule for the day, ultimately result- ing in information about when travelers leave home and when they return. 3.2.2.3 Time Constraints and Time Windows An important aspect of activity-based models is properly representing the effect of time con- straints on people’s activity and travel choices. Some early activity-based models did not rigor- ously do this, allowing tours to be scheduled during overlapping time periods. However, more advanced activity-based models carefully account for time constraints to ensure that all tours and trips are made consistently, and also to more accurately incorporate the effect of time constraints on activity and travel choices. After an activity is scheduled, the time periods used are made unavailable for scheduling other activities. This blocking out of time may also incorporate the travel time expected when transitioning from one activity to the next. The remaining time periods are referred to as avail- able time windows, and are available for other activities. This logic ensures that no person can be in more than one place at one time. Use of time windows also extends to the scheduling of joint activities in activity-based models with intra-household interactions. These models take into account the schedules of multiple persons within the household, scheduling joint activities only when all participants have suffi- cient time for the activity and associated travel. 3.2.2.4 Sensitivities The scheduling and time-of-day models in- cluded in activity-based models system are sensitive to a broad range of factors, including person and household characteristics, trip and tour characteristics, accessibilities, and indi- vidual activity patterns and scheduling pres- sure. For example, these models reflect the fact that higher income workers may work longer hours, but that they tend not to work very early or late. These models can also show how, as more activities are scheduled during a day, the duration of these activities is reduced. One distinguishing feature of the scheduling and time-of-day models included in most activity- based models used in practice is the use of shift variables. Shift variables are used to concisely represent rescheduling sensitivities based on ac- tivity purpose, traveler, and other attributes in- cluding, importantly, travel time and cost. Shift variables allow for a single variable to affect the entire temporal distribution. For example, use of shift variables can capture the effect that longer travel times tend to lengthen the dura- tion of work tour, shift departures from home to work earlier, and shift arrivals back home later. They also capture the tendency to shift travel out of the most heavily congested time periods or to avoid peak period tolls. 3.2.2.5 Linkages with Other Models In most activity-based model systems used in practice, certain types of choices are made at both the tour level and the trip level, including choices of destination, travel mode, and time of day. The tour-level choices of destination, mode, and time of day are usually executed

94 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER sequentially, followed by these same choices being made at the trip level. There is no single correct placement of scheduling and time-of- day models in relation to other models. At the tour level, some activity-based models place the scheduling model in sequence after tour desti- nation choice but before tour mode choice, while other activity-based models incorporate two-stage tour scheduling, with a preliminary tour time of day selected before both tour desti- nation choice and tour mode choice, and a final tour time of day selected after destination and mode choice. In most activity-based models, trip-level scheduling choices always follow trip stop location choices, although sometimes trip- level scheduling precedes, and other times fol- lows, trip mode choice. 3.2.3 Sociodemographics and Population Synthesis 3.2.3.1 Sociodemographics Sociodemographics refers to a set of attri- butes that characterize individual households and persons in a population. The household sociodemographic attributes that usually are of greatest interest in activity-based models are household size, number of workers, pres- ence of children, age of the head of household, and household income, although many other household-level attributes are also used. The person-level attributes that are often used in activity-based models include age, gender, and worker or student status, although many other person-level attributes are also used. Activity-based models are used to make predictions of whether, when, where, and how to participate in activities and to provide infor- mation about the travel required in order to en- gage in these activities. The socio demographic attributes included in a travel demand model should provide information that helps ex- plain how different households and persons make different activity-related and travel- related choices. Information about household size, household income, person age, and other socio demographic attributes are included in activity-based models because they have been shown to provide meaningful explanatory power regarding these choices. Other variables such as housing type and own/rent status may become more commonly used in the future in cases where activity-based models are inte- grated with a land use model that predicts such outcomes. 3.2.3.2 Market Segmentation In travel demand forecasting, market segmen- tation refers to the structuring of different decision-making units and different choice contexts into smaller groups in order to avoid issues of aggregation bias and to provide more accurate model sensitivities. Market segmenta- tion is different in an activity-based model than in a traditional trip-based model. In a trip-based model, detailed market segments are defined at the beginning of the model stream, and this market segmentation is either held constant, or perhaps is simplified, in each subsequent step of the model system. For example, a trip-based model may include segmentation by automobile ownership level that reflects the fact that, rela- tive to households that own private vehicles, households with zero vehicles may generally make fewer trips, may choose their trip desti- nations differently, and may choose different travel modes. Separate matrices representing zero-vehicle and nonzero-vehicle households are needed throughout the entire model system. 3.2.3.3 Synthetic Population Market segmentation is also used in activity- based models but with a greater degree of flexibility. Activity-based models usually use a synthetic population that is essentially a list of all of the households and people within the

95 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) modeled area and that includes detailed infor- mation related to key explanatory variables. Because the synthetic population is in a dis- aggregate list-based format and includes de- tailed sociodemographic information, there is greater flexibility with respect to the definition and use of market segments in activity-based models. Market segmentation is used in a num- ber of ways in activity-based models. First, the synthetic population process that generates this key input to the activity- based model is guided by a set of demographic “marginal controls” that represent the distri- bution of important attributes that impact travel demand choices of the population, such as household size and income. Second, once the synthetic population has been created ac- cording to the market segmentation implied by the control variables, it is common practice to calculate additional market segmentation attributes, such as person types. This person- type segmentation, which may include values such as “worker,” “student,” or “nonworking adult” is used to structure choices in the model system. For example, nonworking adults will not make choices related to usual workplaces or work tours. Finally, the specifications of the individual activity-based model components may include additional market segmentation. For example, although the synthetic popula- tion used in the model may include continu- ous variables such as those related to income and age, it is often better to group income and age into categories in order to achieve a better model fit. A key design question is determining how these categories should be defined. Note that the market segmentation used in any given individual model component does not neces- sarily need to correspond to the segmentation used to define the marginal controls, although it is good practice to align these segments to the greatest extent possible. A more detailed discussion of the development of a synthetic population can be found in Section 3.3. 3.2.4 Long-Term and Mobility Choices Some of the important choices that influence day-to-day travel behavior are not made on a daily basis, but are made on a less frequent, longer-term basis. Examples of such choices in- clude decisions of where to work (for workers) and where to go to school (for students). Other longer-term choices are related to the specific mobility options people and households decide to use. This can include owning automobiles, driving licenses, bicycles, transit passes, and toll transponders. Workers can also decide to follow specific types of work schedules and may or may not have a free or subsidized park- ing space available at the workplace. All of these mobility decisions can significantly influ- ence the availability and attractiveness of dif- ferent location, mode, and scheduling choices that create daily activity and travel patterns. 3.2.5 Activity Purposes and Joint Travel Early activity-based models tended to include only 3 or 4 distinct activity purposes, such as work, school, other, maintenance, and discre- tionary. Recently, as many as 7 to 10 activity purposes have been included in activity-based models. Escort activities (also referred to as “chauffeuring” or “serve passenger”) tend to have different characteristics from other ac- tivities, particularly in terms of mode choice, since they tend to involve automobile shared- ride tours. Meal activities can usefully be separated from other types of maintenance ac- tivities, because they tend to take place during certain periods of the day at locations where food service employment is located. Shop- ping is another type of maintenance activity that can be tied to specific attraction vari- ables, such as retail employment, and tends to happen during store opening hours. Medical visits are another activity purpose that can be tied to a specific attraction variable (medical employment).

96 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER On the discretionary side, it can be useful to separate social visits as a separate activity purpose, as they often occur at residential loca- tions, outside working hours. Outdoor recre- ation can be another useful activity category, as it can be tied to open space/parks/sport fields, and so forth, as attraction variables. In general, if the land use data have sufficient detail that will allow one to predict where specific types of activities are likely to take place, then the data can also be useful to distinguish those types of activities in the activity-based model compo- nents. In this way, the model can better pre- dict which types of people tend to visit certain types of locations during certain periods of the day and can more accurately predict changes in behavior when the distribution of land uses or demographic characteristics of the population change. 3.2.6 Travel Modes As described in Chapter 2, the set of modes used in an activity-based model is similar to the set that would be used in a trip-based model. These modes include automobile modes such as drive alone (DA), and shared ride (SR2, SR3), as well as transit modes and non motorized modes. However, the activity-based model is not limited to a simple representation of modal alternatives but can also include detailed sub- modal alternatives such as managed lanes, bus rapid transit, and commuter rail. The representation of travel modes is re- lated to the structure of the activity-based model, in which mode choices are made both at the tour level and the trip level. The tour mode is defined as the primary mode for the entire sequence of trips that make up the tour. How- ever, the tour mode is not necessarily used for all the trips on a tour and is not even explicitly reported by travelers in a household travel sur- vey. The tour mode is defined in the model de- sign and is determined based on the nodes that are used for the trips on a tour. Tour modes may be defined at a relatively aggregate level. For example, a typical tour mode choice model might include the following alternatives: • Drive alone (DA) • Shared ride 2 (SR2) • Shared ride 3+ (SR3) • Walk (WK) • Bike (BI) • Walk-to-transit (W-TRN) • Drive-to-transit (D-TRN) The inputs to the tour mode choice model in an activity-based model are generally similar to the types of inputs to a traditional trip-based mode choice model. These common inputs in- clude information about purpose, time of day, automobile ownership, and household income. However, there are some features that distin- guish tour mode choice models. Perhaps the most significant difference is that tour mode choice models consider the travel times and costs for the entire round trip, including both the journey to the tour destination and the re- turn from the destination to home. In order to accurately capture these times and costs the model uses time-period- and direction-specific multimodal network skims. Another significant difference is that, because the activity-based model is implemented using a disaggregate micro simulation framework, tour mode choice models often include detailed household- and person-level variables, such as person type, age, and parking subsidies that may significantly improve the explanatory power of the model relative to a trip-based mode choice model. Similar to trip-based mode choices, tour mode choice models are typically segmented by pur- pose and may include land use and other urban form variables. The trip mode is the travel model that is used for each individual trip on the tour. Trip

97 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) mode choice models may contain more detail, such as transit submodes or tolling alternatives. Consistency between the tour mode and the trip mode is essential, although this does not mean that all trips on a tour use the same mode. Rather, it requires a logical consistency across all of the choices made by an individual. Tour mode is defined in relationship to trip mode. In the simplest cases, all trips on a tour have the same mode. For example, if a traveler drives alone from home to work and back, there are two DA trips, and thus this tour can be eas- ily classified as a DA tour. However, it is also common for travelers to use multiple different modes on the same tour. Common examples of this include tours in which the vehicle occu- pancy on a traveler’s tour changes as a result of picking up or dropping off passengers, or in which a person uses both transit and walk modes on the same tour. For tours where mul- tiple trip modes are used, a hierarchy is used to identify the tour mode. For example, if a person’s tour includes both shared-ride trips and DA trips, the tour would be classified as a shared-ride tour. Similarly, if a tour includes both walk-to-transit trips and walk trips, the tour would be classified as a transit tour. These classifications are then used in model applica- tion to ensure that the predicted trip modes are consistent with the predicted tour modes. Table 3.4 illustrates the availability of trip modes by tour mode. 3.2.6.1 Accessibilities Accessibility measures are critical to ensur- ing reasonable policy sensitivity at the various levels of the model to changes in infrastructure or land use, or both. In general, four types of accessibility variables are included in the models: 1. Direct measures of travel times, distances, and costs from modeled network paths; 2. Detailed logsums calculated across alterna- tives of models that include direct measures; 3. Aggregate (approximate) logsums calcu- lated across alternatives of models that in- clude direct measures; and 4. Buffer measures representing the activity opportunities and urban design surround- ing each parcel or microzone (e.g., Census block). The direct measures are used in all mode, destination, and time-of-day choice models wherever possible. Often, however, the model TABLE 3.4. TRIP AND TOUR MODE AVAILABILITY Trip Mode Tour Mode DA SR2 SR3 Walk Bike W-TRN D-TRN DA X X X X SR2 X X X X SR3 X X X Walk X X X X X X X Bike X X X W-Bus X X W-Rail X X D-Bus X D-Rail X

98 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER hierarchy makes it impossible to use a direct measure because it depends on a yet- unmodeled outcome. This would be the case, for example, for travel time in a destination-choice model that is higher in the hierarchy than mode and/ or time-of-day choice, since in order to measure travel time directly it is necessary to know the mode and time of day. In such cases, detailed logsums can be calculated from the lower-level choice models and used instead of direct mea- sures in the upper-level model. A typical exam- ple in practical activity-based models is the use of tour mode choice model logsums in higher- level models such as tour time-of-day choice, tour destination choice, and workplace loca- tion choice. There are cases when it is not practical to use the most fully detailed versions of the log- sums that are calculated on the fly during the simulation every time one is needed. To address this issue, a common approach is to precalcu- late more aggregate accessibility logsums to be used in models where using the more impracti- cal ones would not be computationally or con- ceptually feasible. For example, some model systems use aggregate accessibility logsums cal- culated from each origin TAZ or microzone, to all possible destinations, via all possible modes, with the different modes and destinations weighted approximately as they would be in a fully detailed logsum across a tour mode and destination-choice model. Aggregate logsums are typically calculated for each combination of up to 4 or 5 critical dimensions, including • Origin TAZ or microzone; • Tour purpose; • Household income group or VOT group; • Household automobile sufficiency (auto- mobiles owned compared with driving-age adults); and • Household residence distance from transit service. These few dimensions are typically chosen because they tend to be the most critical vari- ables in mode choice models and will thus help determine how much influence the different available modes will have on the logsum mea- sures. For example, the accessibility logsum for the zero-vehicle household segments depends critically on how accessible destinations are by nonautomobile modes from the given ori- gin, while the logsums for the lowest income (or VOT) group will be most sensitive to travel costs such as tolls and transit fares. Aggregate measures are used most often in the day-level models and some of the longer- term models, where the model is not yet con- sidering a tour to a specific destination, but is considering, for example, how many tours to make for a given purpose from the home loca- tion during the day. In that case, the overall ac- cessibility from the residence for each purpose can have an effect. It is through these types of variables that activity-based models can rep- resent true induced trip and suppressed trip effects (as opposed to simply shifting destina- tions or modes). Although it is important to be able to rep- resent such effects, they are, both in reality and in the models, small relative to other types of choice responses that occur at the other levels of the model system. So, although it is impor- tant to include these effects, they may not be so substantial that it would be worthwhile try- ing to implement fully detailed logsums for all levels of the model system, a process that could increase model complexity and run times by an order of magnitude. For that same reason, it is typical to use less spatial and temporal detail in the aggregate logsums than is used in the fully detailed logsums. For example, if the model uses parcels or Census blocks for the basic spatial unit, it may use only TAZ-level detail for the aggregate log- sums. One reason is that they mainly represent accessibility over all distances for the entire re-

99 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) gion, while other measures such as buffer-based measures can be better at measuring very local accessibility over short distances. As another example, in many cases time of day is not in- cluded as an explicit dimension in the aggre- gate logsums. Instead, the model design might use the most typical times of day for each travel purpose to specify which time-of-day– related congestion levels to use in calculating the mea- sures. It would be possible to use time of day as an added dimension, and that is another case where model designers trade off whether add- ing detail that is likely to have a small influence on the results would be worth the increase in computation time. Finally, buffered measures represent the accessibility to very nearby destinations, as could be made by walk, bike, or very short car trips. The typical measures that are buffered include • The number of nearby households; • The number of nearby jobs of various types (as proxies for activity locations); • The number of nearby school enrollment places of various school types; • The number of nearby paid parking places, and their average price level; • The number and average size of nearby parks and open space (for recreation); • The number of nearby street intersections of various types (e.g., T-junctions, 4+ links); • The number of nearby dead-ends and cul- de-sacs; and • The number of nearby transit stops. As seen from these variables, the buffer measures represent the neighborhood charac- teristics in terms of land use and urban design along a number of dimensions. The two differ- ent types of measures for street system design can be used to represent the positive acces- sibility of having a dense network versus the negative accessibility of a layout with many dead-ends and cul-de-sacs. The traditional way to calculate buffer measures has been to use a simple radius, such as a quarter-mile or half-mile, and count up every thing within that radius, giving everything an equal weight. This method provides a sim- ple density measure within a circle. Clearly, these measures are most relevant when the spatial units themselves are much smaller than the radius of the buffer area. Thus, using buffer-based measures is really only useful when the spatial unit of the model is the parcel or (at the largest) the Census block. Also, the measures are more important to include in the models in such cases, because they represent the neighborhood effects around the spatial alternatives of interest. For example, one may be more likely to shop at a parcel (or block) where there are additional shopping (or meal) opportunities located nearby. In models using fairly large TAZs as the main spatial unit, these neighborhood effects are already included to some extent, but in a less consistent way, be- cause the relevant nearby attractions may not be in the same TAZ. One way to make the buffer measures more accurate and relevant is to use on-street shortest-path distance to measure the distance to the edge of the buffer, rather than using straight line (as the crow flies) or Euclidean distances. In this way, the buffers represent the effects of possible obstacles such as rivers, free- ways, rail yards, as well as street layouts with poor connectivity. A second way to make the measures more relevant is to use distance-decay weighted measures rather than simply count- ing up everything within a certain distance and weighting equally. This makes the measures more behaviorally relevant, with nearer at- tractions being weighted more highly, and also helps to avoid some of the boundary effects and arbitrariness of using a single fixed radius.

100 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER 3.3 COMPONENTS This section introduces the major types of com- ponents included in most activity-based model systems. The various types of components are then described in greater detail in the following subsections. Although the exact structure of implementations in specific regions or software packages is not described in detail, the similari- ties in the design of those implementations are highlighted. In certain cases where important differences can be found among the existing U.S. implementations, the design differences and main options are introduced and discussed. If the reader wishes to find more detailed infor- mation regarding specific model implementa- tions, the authors recommend consulting the references. The material from the Travel Model Improvement Program (TMIP) activity-based model webinar series (Resource Systems Group 2012a) can also be a useful source of detailed information, as many of the webinar series topics follow the same sequence as the model component subsections that follow. Figure 3.19 depicts the main component sections of applied activity-based model sys- tems. Between the model inputs and outputs, the figure shows • Longer-term choices; • Mobility choices; • Day activity patterns (DAPs); • Tour and trip details; and • Trip assignment. Going from top to bottom in the figure, each model component is conditional on the choices simulated in the higher components. This is termed “downward integrity,” meaning that the choices are consistent with previously predicted choices. For example, the number of tours for each purpose predicted in the DAP Figure 3.19. Downward and upward model integrity. Figure 3.19. Downward and upward model integrity. Model Inputs Synthetic Population Long-term Choices Mobility Choices Daily Activity Patterns Tour & Trip Details Trip Assignment Model Outputs Upward Integrity: Expected utility (accessibility) of choice alternatives in lower models affects choices made in higher models Downward Integrity: Choices made in higher models affect choices made in lower models

101 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) model components determines which tours will be simulated in the “Tour & Trip” details components. In well-designed model systems, information also flows from the lower com- ponents to the higher components. The term “upward integrity” refers to accessibility infor- mation from the available choices in the lower- level model components affecting the choices simulated in the higher-level models. Upward integrity is achieved mainly through the use of various types of accessibility measures that approximate the expected utilities (logsums) from the lower-level models. Next, each of the various types of components in Figure 3.19 is described in greater detail. 3.3.1 Population Synthesis 3.3.1.1 Purpose In the activity-based model system, households and persons are used as the core decision- making units, making choices about key con- siderations like the number of vehicles a house- hold chooses to own, the type and amount of activities that occur, and the locations of key destinations such as work and school. Activity- based model systems typically employ micro- simulation, in which these choices are repre- sented at the level of individual household or individual person. Population synthesis is used to create the lists of households and persons, or synthetic population, that are the basis for simulating these choices. The choice models that make up the activity-based model system should be specified to use only demographic variables that are available in the synthetic pop- ulation, and the synthetic population should include all of the demographic attributes that are used in the choice models that compose the model system. Creating a synthetic population is the first step of running the activity-based model system. 3.3.1.2 Design The first step in creating a synthetic population is designing its structure. This design process involves • Selecting the sociodemographic variables that are going to be controlled (the mar- ginal controls). • Identifying sources of information for these variables. • Determining the categories used to classify these variable. • Specifying the geography that will be used. • Identifying the source of the household and person data that will be sampled to create the synthetic population. The selection of sociodemographic vari- ables can be influenced by known relationships between these variables and travel demand choices as well as by anticipated policy appli- cation and analysis needs of the model system. Typical variables that are controlled in house- hold population synthesis include • Household size; • Household income; • Age of householder; • Number of household workers; and • Presence of children. Some population synthesis tools also in- clude the ability to simultaneously control person-level attributes, most frequently age and gender. Information describing the distribution of these attributes in the population can be de- rived from a number of sources. Base-year data are often derived from the decennial Census, the ACS, or other TAZ-level control totals used for trip-based modeling or other regional plan- ning efforts. Future-year data may be derived from regional socioeconomic forecasts, land

102 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER use models, or other tools. It is common that agencies may have only limited or no forecast information on some of the marginal controls included in the population synthesis. In these cases, base-year distributions may be assumed to remain fixed, although such an assumption will influence the distribution of other marginal controls. A key concern when selecting marginal controls for population synthesis is to choose a sufficient number of variables that are not highly correlated with each other. Using too few control variables may produce a synthetic population that doesn’t accurately reflect the true population. Conversely, using too many control variables may result in too many sparse cells in the multidimensional distribution. Note that this multidimension distribution is created not only regionally but also ideally at smaller geographic levels such as TAZs. However, it is possible to specify control attributes at multi- ple spatial resolutions or geographic units, pro- vided that the smaller geographies nest within the larger geography. An additional concern with specifying marginal control distributions is to minimize the use of control variables or categories that may result in a sparse matrix for certain cells. A sparse matrix may make it difficult to find samples in the PUMS data or in the household survey data that match these rare combina- tions. And even if some samples are found, the relative rarity of these samples may mean that they are repeatedly drawn into the synthetic population an unreasonable number of times. In general, it is desirable to use the small- est geography for which data are available, although in some instances this geography may be larger than the base geography used in the model system. For example, the activity-based model may use parcels as the basic spatial unit, but the population synthesis may be performed at the TAZ level. The synthetic population may then be allocated to parcels using a set of rules or assumptions. It is also necessary to acquire household and person data files that will be sampled to create the synthetic population. In the United States, this disaggregate sample has typically been derived from Census PUMS data. It is also possible to use household survey data as a data source for this sample. 3.3.1.3 Implementation After the synthetic population design has been established and all required input data col- lected and prepared, a population synthesis software tool is typically used to actually cre- ate the synthetic population. There are a num- ber of available tools that are differentiated by their unique features. However, most all of these perform two basic functions: (1) creating the joint multidimensional sampling distribu- tion from the set of independent marginal con- trols, and (2) drawing samples of households and their associated persons into the popula- tion in such a way as to match this sampling distribution. The first step of creating the multi- dimensional sampling distribution involves fit- ting or balancing the multiple control dimen- sions of the design. Most frequently, this is achieved through the use of an iterative pro- portional fitting (IPF) process, although other approaches exist. This process produces, at the level of geography specified in the design, the set of shares of households within each sam- pling type, based on the marginal controls. Alternative approaches to IPF exist and have advantages such as the ability to control both household-level and person-level marginal con- trols simultaneously. In addition, these methods can avoid some of the limitations and distor- tions that can be an outcome of a simple IPF process. For example, when there are signifi- cant structural changes in the population be- tween a base year and a forecast year, a simple IPF procedure can distort the expected distri-

103 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) bution. An IPF also can produce distorted dis- tributions when there are a number of empty cells in the seed marginal distributions. The second step involves drawing the PUMS or household survey-based samples according to these shares so that they match the targets and marginal distributions both in aggregate and at the more detailed spatial geographies. Before running this sampling, the estimated shares are applied to the input totals and con- verted to integer values using rounding in order to generate the targets required for sampling. Selection probabilities are calculated for all samples based on the target distribution, and random Monte Carlo simulation is then usually used to draw samples in order to create the syn- thetic population. There are a number of other methods for drawing samples to create a syn- thetic population that may involve more com- plex algorithms but also provide more features. After the population synthesis tool has produced the population, the results are usu- ally validated against the marginal distribu- tions that were used as input to the process to ensure that the results are reasonable. It is also common for the synthetic population to be postprocessed in order to calculate new vari- ables (e.g., person type) that will be used in the activity-based model system. 3.3.2 Long-Term Models Some of the important choices that influence day-to-day travel behavior are not made on a daily basis but are made on a less frequent, longer-term basis. One such choice is the decision of where to live. In most activity- based model systems, residential choice is implicit in the population synthesis process, described in the preceding section. Related choices are the decisions of where to work (for workers) and where to go to school (for stu- dents). Other longer-term choices are related to the specific mobility options that people and households decide to use. These can include owning automobiles, driving licenses, bicycles, transit passes, and toll transponders. Workers also can decide to follow specific types of work schedules, have usual modes for work travel, and may or may not have a free or subsidized parking space available at the workplace. All of these mobility decisions can significantly influ- ence the availability and attractiveness of dif- ferent location, mode, and scheduling choices that create daily activity and travel patterns. Figure 3.20 provides a schematic overview of how the longer-term and mobility choices fit into an activity-based model system. These choices are simulated for each household and person in the synthetic population. Then, con- ditional on these predicted choices, a travel day is simulated by running the day-pattern, tour- level, and trip-level models. After the simulated trips are assigned to the networks, the travel times and accessibility measures can be recalcu- lated for another iteration of the model system, Figure 3.20. Longer-term and mobility choice models in an activity-based model. 2014.11.18 C46 Primer FINAL for composition.docx 200 attractiveness of dierent location, mode, an scheduling c oices that create daily activity and travel patterns. [Insert Figure 3.20] [Caption] Figure 3.20. Longer-term and mobility choice models in an activity-based model. Figure 3.20<FIG. 3.20> provides a schematic overview of how the longer-term and mobility choices t into an activity-based model system. These choices are simulated for each household and person in the synthetic population. Then, conditional on these predicted choices, a travel day is simulated by running the day-pattern, tour-level, and trip-level models. After the simulated trips are assigned to the networks, the travel times and accessibility measures can be Population Synthesizer Longer Term & Mobility Models Simulation of Day Patterns, Tours & Trips (conditional upon longer term & mobility choices) Highway and transit assignments Usual work and school location Auto ownership / availability Free parking eligibility / reimbursement Transit pass ownership

104 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER including the longer-term and mobility models (see Table 3.5). 3.3.2.1 Usual Location Most workers and students have a usual place where they go to work or study. Although they may go to a different location on some days (e.g., a business meeting away from the office or a field-trip away from school), they are most likely to visit the usual location on any given day. As a result, this location does not depend on the activity pattern followed on a specific day, and thus it can be modeled at the upper level, before modeling the day activity pattern or subsequent choices. Also, the usual work and school locations can be important anchor points for other travel-related choices, such as where one typically goes to shop or do errands, or the likelihood that household members will carpool together to work and/ or school. One possibility for the usual work location is to work from home on a regular ba- sis (as is done by more than 5% of workers in the United States, with the percentage growing over time). Models of usual work and school location are quite similar to tour or trip destination- choice models in that they predict the choice of a single location from among many alternative destinations. Table 3.6 lists the types of variables that tend to influence location choice models. Key variables are the accessibility variables, includ- ing the mode choice logsum measuring the TABLE 3.6. TYPICAL VARIABLES IN A LOCATION CHOICE MODEL Households Persons Land Use Accessibility • Income • Size • Children • Seniors • Automobiles • Worker status • Occupation • Driver • Gender • Telecommuter • Employment density by type • Household density • Student enrollment • Mixed use • Parking density • Intersection density • Agglomeration and competition effects • Distance or distance- decay functions • Mode choice logsum • Mode/destination logsum TABLE 3.5. A CLASSIFICATION OF LONGER-TERM AND MOBILITY CHOICE MODELS Model Household Decision Person Decision Worker Decision Student Decision Location Models • Work location • Work at home • School location Vehicle Models • Auto ownership • Auto type • Bike ownership • Toll transponder • Auto allocation • Driver’s license Personal Mobility • Transit pass Worker Mobility • Work schedule type • Pay to park at work

105 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) attractiveness of traveling between home and a potential work location across all available modes of travel. Although a change in acces- sibility may not cause a person to change his or her work location from one day to the next, over the longer term it will influence the distri- bution of O-D commute patterns. In contrast to aggregate zone-based 4-step models, it is not necessary to include all zones as choice alternatives for each case. Because many thousands of different workers, students, or tours are being simulated, it is most effi- cient to use only a subsample of the possible locations as choice alternatives in the model. This is particularly true when modeling at the micro zone or parcel level of spatial resolution, in which case there may be many thousands, or even millions of different choice alternatives in the region. An efficient form of selecting a set of alternative locations for the model (in both model estimation and application) is im- portance sampling of alternatives, as described in Ben-Akiva and Lerman (1985). The general concept is to use a simple weighting function to determine the sampling probabilities that approximates the choice probabilities in the model itself. In this case, a typical sampling weight is calculated by using a fairly simple at- traction function and impedance function for each alternative, resembling the functions used in a gravity model (but typically much simpler than the utility functions in the location choice model itself). One important aspect of work location models is that the number of jobs available in any zone or geographic district is typically an input to the model system, so the total predicted number of people with their usual workplace in a given area should be approximately equal to the number of jobs available in the area. In other words, the usual work location model should be doubly constrained, both at the home end and the work end. In practice, this is achieved in activity-based models by using an iterative “shadow price” method, whereby the utility of each possible work location area is varied between iterations so that the total de- mand for jobs will converge to total available supply. A shadow price procedure can also be used for usual school destinations, particularly for types of schools that attract longer distance commutes, such as colleges and universities. One final point that can be very important in some regions: When balancing demand and supply for jobs, it may be important to take ac- count of workers who commute from outside the modeled region, as well as workers who commute from inside the region to jobs outside the region. In locations near regional bound- aries, the proportion of such workers can be quite substantial. In those cases, a common ap- proach is to estimate the number or percentage of internal-external (IX) and external-internal (XI) commute trips from other sources (e.g., a more aggregate model of external trips), and use those input data to modify the simulation so that a certain percentage of jobs in each area are prefilled by external workers, and a certain percentage of workers from each area are simu- lated to have a workplace outside the region. 3.3.2.2 Automobile Availability All activity-based models used in practice have a model to predict how many automo- biles are owned and available for use by each household. The number of vehicles available, relative to the number of adults or workers in the household, tends to be one of the most im- portant and significant variables in subsequent models, such as those of tour generation and mode choice. The most significant variables in automobile ownership models tend to be household size, composition, and income. Accessibility from the residence location to des- tinations of various types by automobile versus nonautomobile modes is also an important in- put, as people who live in areas where a wide

106 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER variety of activities can be reached without an automobile tend to own fewer automobiles, all else equal. The automobile and nonautomobile accessibility to the usual work locations of any workers in the household is also a key variable, and that is a key reason for predicting automo- bile ownership conditional on the usual work and school locations. An implicit assumption in this structure is that it will be possible for most households to increase their automobile ownership level if doing so will allow them to take advantage of better employment opportu- nities. If automobile ownership costs were to increase to the point where that is not the case, then a different model hierarchy might be more appropriate. Table 3.5 lists a few other possible types of vehicle-related models including • Automobile type choice (e.g., body type and/or fuel type); • Toll transponder ownership; • Driving license ownership; • Automobile allocation among household drivers; and • Bicycle ownership. Although it is possible to include all of those models within an activity-based model structure, most of those choices are not rep- resented in most of the applied activity-based models in the United States. Regarding bicycle ownership, the cost of owning a bicycle is low enough that the cost of ownership is not one of the more significant factors discouraging greater bicycle use in the United States. There are more important aspects to bicycle use, such as provision of safe infrastructure, that are more crucial to predicting bicycle use. A similar argument has been made for modeling ownership of a driving license— almost anyone who wants to use an automobile can easily get a license, so it is not necessary to model license-holding to explain automobile use. Recently, however, it has been noted that a lower percentage of teenagers have obtained a license, so there may be a renewed interest in modeling license-holding to help explain how travel preferences may be shifting across generations. Ownership of a transponder for electronic toll collection (ETC) may be useful for predict- ing which types of households are most likely to use tolled facilities like high-occupancy toll (HOT) lanes and express lanes, at least in the short term. In the longer term, it may be that ETC technologies become so ubiquitous that they will no longer be relevant as a predictive variable. The two vehicle-related models that may be most valuable to enhance activity-based model systems are models of vehicle type choice and vehicle allocation among household members. The type of vehicle(s) that a household owns and the travel pattern of the person in the household who uses each one can have a significant influ- ence on the pollutant emissions generated by travel, including greenhouse gas emissions. With a great deal of progress being made in traffic simulation models and related vehicle emissions models, it could be valuable for activity-based models to be able to predict the type of vehicle used to make each simulated trip. 3.3.2.3 Other Long-Term Models As previously shown in Section 3.3.2, Table 3.5 lists three other potential types of mobility models that may be valuable to in- clude as longer-to-medium-term choices. Two of these models are included in several of the activity-based model systems used in practice. Transit pass ownership: This is typically modeled as a binary yes-or-no choice, although it would also be possible to model the type of transit pass that a person owns. The key aspect of owning a transit pass is that once a pass is purchased the marginal cost of using transit be-

107 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) comes zero. Thus, a person who buys a transit pass or receives a subsidized transit pass for com- muting to work might also be more likely to use transit for other purposes as well. When model- ing transit pass ownership, it is also important to simulate the price of purchasing a pass and the effect of that price on pass ownership, so that pass owners are not represented as being totally unresponsive to transit fare policies. Availability of free parking: Parking price data input to activity-based models typically represent the average price of paid parking spaces within a certain zone or area. In reality, however, not everyone who parks within the area will need to pay for a parking space. That is particularly true for workers, many of whom received free or subsidized parking places at or near the workplace. A simple way to incorpo- rate this within an activity-based model struc- ture is to include a model that predicts whether or not each worker has a free parking space available at work, or whether they are subject to paying the market price for a parking space. This type of model also provides a way of sim- ulating employer-related parking policies. A final type of mobility model that could be valuable in activity-based model systems is to predict the type of work schedule that each worker has, including the variability and flex- ibility of the work schedule from one day to the next. People with different types of work schedules may exhibit quite different peak- spreading sensitivity to changes in conges- tion profiles across different times of day, and modeling this aspect of work behavior would provide a means to more accurately model employer-based policies that allow or encour- age different types of work schedules. 3.3.3 Day-Pattern and Tour- and Trip-Level Models Figure 3.21 shows a somewhat different over- view of activity-based model components, expanding on what is predicted below the Figure 3.21. Typical activity-based model structures. 2014.11.1 <H3>3.3 The DAP widely ac designs, h of the ex number o purposes importan <BL> • M <BBL> 8 C46 Primer .3.1 Day-Pa part of acti ross practic owever, is act sequence f tours that . Often, the ce and prior andatory pu FINAL for co Figure 3.21 ttern and T vity-based m al implemen that the main and specifi each individ purposes are ity in structu rposes mposition.do . Typical ac our Genera odel system tations in th focus of th cation of ch ual makes f grouped in ring the day cx tivity-based tion s is where t e United St e day-patter oices that ar or each of a to three diff ’s activity a model stru he model sy ates. The co n models is e simulated number of d erent types, nd travel pa ctures. stem design mmon featu tour genera , the main ou ifferent act indicating th ttern: s vary most re of all of t tion. Regard tput is the e ivity and tou eir general 210 hose less xact r

108 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER longer-term and mobility models. In the “ Basic Activity-Based Model Structure” column at the left, the “Tour & Trip Details” from Figure 3.19 are shown as split into a number of differ- ent subcomponents. The tour-level models of primary destination choice, mode choice, and scheduling are shown in one box, with the models of intermediate stop generation and lo- cation below the tour level, and the trip-level models of mode choice and departure time choice as the lower level. This general sequence of modeling the different types of choices is fol- lowed in virtually all practical activity-based model systems in the United States. In many activity-based model systems, the day activity patterns are predicted separately for each individual in the household, predicting the number of tours each person makes in the day for each activity purpose (and possibly some other aspects of the full-day pattern as well, de- pending on the specific model system design). Although there is some influence of household size and composition through the use of house- hold characteristics as exogenous variables in the day-pattern models, there are no explicit linkages or interactions simulated between the day patterns of different household members. The column at the right indicates that there is a second class of models that does in- clude explicit simulation of intra-household interactions. Such models contain extra sub- components in the day-pattern part of the model system, first predicting DAP types at the household level and then predicting joint travel activities and tours involving multiple house- hold members, and using those predictions to condition the individual DAP for each person in the household. The lower tour- and trip- level models then take into account the fact that some tours involve multiple household members, while others do not. Such distinction between models systems with and without ex- plicit intra-household interactions is discussed in more detail in the following section. 3.3.3.1 Day-Pattern and Tour Generation The DAP part of activity-based model systems is where the model system designs vary most widely across practical implementations in the United States. The common feature of all of those designs, however, is that the main focus of the day-pattern models is tour generation. Regardless of the exact sequence and specifica- tion of choices that are simulated, the main out- put is the exact number of tours that each indi- vidual makes for each of a number of different activity and tour purposes. Often, the purposes are grouped into three different types, indicat- ing their general importance and priority in structuring the day’s activity and travel pattern: • Mandatory purposes — Work — School • Maintenance purposes — Escort (pick up, drop off, accompany others) — Medical — Shopping — Personal business (e.g., errands, civic activities) • Discretionary purposes — Meal (eating out) — Social visit — Recreation The exact grouping and number of dif- ferent purposes considered in the models may vary somewhat from one example to the next, but all of them tend to treat mandatory (work and school) tours as the highest priority ac- tivities and tours around which any remaining tours in the day are arranged and scheduled. It is also important to remember that any tour may contain multiple activity stops for different purposes and that a tour is classified according

109 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) to the activity at the primary destination of the tour. The primary destination is typically deter- mined by prioritizing all stops based on some combination of the activity purpose (manda- tory purposes highest priority, discretionary purposes lowest) and the duration of stay at the destination (activities of longer duration hav- ing higher priority). While the common feature of DAP models is that they generate tours for different pur- poses, one can find a number of variations in the way the models are specified and arranged. Following is a summary of some of the key dif- ferences found between different models used in practice. Inclusion of intermediate stops In some cases, the DAP models also predict some aspect of extra (intermediate) stops made during the day. For example, the model can predict if there are any additional stops made for each different activity purpose. A reason for including these details at the day-pattern level is that there may be substitution between mak- ing additional home-based tours versus making addi tional stops to be chained into a smaller number of tours. People who live in dense urban areas nearby many activities tend to be able to return home more easily between ac- tivities and have less need for chaining multiple activities into tours, all else equal. This type of trade-off can be captured most explicitly by in- cluding some aspect of intermediate stops in the day-level choices. The alternative is to generate all intermediate stops in the tour- and trip-level models, as indicated by the “Stop Generation” box in Figure 3.21. Note that all activity-based model systems include models of stop genera- tion and allocation to tours at that lower level. Modeling some aspects of stop generation at the higher day-pattern level simply conditions those lower-level models so that they already have some information about certain activities that need to be allocated to tours. Scheduling of mandatory tours In some model systems, work and school tours are generated and scheduled before the genera- tion of additional tours for nonmandatory pur- poses. As discussed further in the next section, the rationale for this structure is that manda- tory activities tend to be of long duration and block out much of the time in the day so that it is not available for scheduling other activi- ties. Thus, all else equal, those who make man- datory tours of the longest durations are less likely to make additional tours during the same day. Explicit modeling of intra-household interactions This feature is probably the most substantial distinction among the model systems used in practice, because it adds quite a bit of informa- tion as well as complexity to the model system. The models simulate joint household choices of various types including the following: • Joint household DAP types. Often, house- hold members will tend to coordinate their overall travel patterns, for example, whether they stay at home all day, go to work or school, or leave the house for some other nonmandatory purpose. For example, if a child stays home from school because of illness, it is more likely that a parent will stay home from work that day as well. Modeling pattern coordination can be important, because it can increase the possibilities for joint travel across house- hold members. • Fully joint tours. A fully joint tour is one in which two or more household members leave home together, travel together to all of the same locations along the tour, and return home together as well. This type of tour accounts for a substantial percentage of the observed tours for nonmandatory purposes and for a substantial percentage

110 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER of observed multioccupant vehicle tours (ridesharing) as well. The fact that a tour is jointly made by multiple household mem- bers can be important to condition the tour scheduling and tour mode choice models [i.e., the drive alone (SOV) mode choice alternative is not available for such tours]. • Joint half-tours to work or school. There are a number of possibilities for house- hold members to coordinate their work or school commute travel. Examples are one adult dropping another adult off at work, or two children traveling together to the same school. The most common example is a parent dropping children off and/or pick- ing them up at school. These types of coor- dination are typically modeled as half-tours rather than full tours, because the partici- pants do not necessarily travel together for the full home-based tour. For example, a parent may drop a child at school and then return home or drive on to their workplace, and then the other parent may pick up the child at school in the afternoon. Because of the many different possibilities for this type of travel, the model specification tends to be quite complex, and the recent examples found in practice use somewhat different model structures. • Allocation of household maintenance tours. Some model systems have modeled the generation of maintenance tours (such as shopping, escort, and personal business) at the household level and then allocated each tour to a specific household member. However, this type of model has not gained widespread use in practice. Possible rea- sons are that it is not apparent that includ- ing such a model substantially affects the model forecasts, and also that the coding of activity purposes in most survey data is not precise enough to know when someone is performing an activity on behalf of the household versus for the individual (e.g., grocery shopping versus shopping for per- sonal items). From a conceptual standpoint, the inclusion of intra-household interactions adds aspects of behavioral realism to the models. There has not yet been a direct comparison of forecasts from models based on the same data with and without such interactions, so it is not yet clear how much the model predictions tend to be influenced by adding this additional level of modeling. 3.3.3.2 Scheduling The topics of this section and the two follow- ing sections—scheduling, location choice, and mode choice—are relevant at both the tour level and the trip level. There is no standard practice for the exact hierarchy to be used be- tween the three choice dimensions, and, in fact, all three of the structures shown in Figure 3.22 can be found in the models in practice in the United States. The common feature of these three structures is that mode choice is esti- mated conditional on destination choice, which is standard practice in the United States in both trip-based and activity-based modeling (but is not always standard practice in other coun- tries). Time-of-day choice, however, has been modeled in all three possible positions relative to the other tour-level choices. In reality, there is some degree of simultane- ity across all three of these choice dimensions, and there is no obviously correct way to model it. The best structure statistically may vary by tour purpose, although there are yet few models in practice that allow different model structures for different tour purposes. The best structure also will tend to depend somewhat on what level of temporal detail is used to model time of day and scheduling decisions. As dis- cussed previously, a number of different levels of temporal detail have been used in practice in activity-based models including the following:

111 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) • Four, five, or six broad time periods of the day (e.g., early, a.m. peak, midday, p.m. peak, and late); • One-hour periods (24 periods in the day); • Half-hour periods (48 periods in the day); • Ten-minute periods (144 periods in the day); • Five-minute periods (288 periods in the day); and • Continuous time (e.g., 1,440 one-minute periods in the day). The earliest activity-based model systems tended to use four or five time periods in the day, similar to 4-step model systems. In general terms, the broader the time periods, the higher in the tour hierarchy the time-of-day choice should be modeled, since the choice will tend to be less sensitive to shifts in the input variables. So, the models that place time of day above destination choice tend to be the activity-based models that use the broad time periods. The more recent trend has been to use greater tem- poral detail and to model time of day below destination choice, either above or below mode choice. An impediment to using many time periods at the tour level is that most tour-level time-of- day models simultaneously predict the time that the tour (or the primary tour activity) begins, as well as the time that it ends. This simultaneous choice means that instead of the model having N choice alternatives, where N is the number of choice periods, it has N * (N + 1)/2 choice alter- natives. So, a model that uses 48 half-hour time periods in the day has 1,176 possible scheduling alternatives. This dimensionality issue tends to make it impractical to use periods much shorter than 30 minutes at the tour level. At the trip level, however, the choice is only one- dimensional, predicting the trip departure time conditional on what has already been predicted at the tour level and for any preceding trips in the tour. So, it can be practical to use periods as short as 5 or 10 minutes to predict time of day at the trip level. Of course, it is unlikely that separate highway travel time skims will be available for each different period when one uses such short periods. Nevertheless, using short periods still has advantages in terms of modeling the sched- uling of travel and activities across the day, as discussed below. Also, with increased interest in using activity-based models together with DTA and/or network micro simulation, it can be very useful to be able to predict trip departure times to a high level of detail. An additional advantage of using shorter time periods is that it allows a more continu- ous treatment of time, approaching a duration 2014.11.1 F of behavi models b much the <H3>3.3 The topic choice— exact hie structure States. Th on destin activity-b choice, h choices. [Insert Fi [Caption 8 C46 Primer rom a conce oral realism ased on the model pred .3.2 Schedu s of this sec are relevant rarchy to be s shown in F e common ation choice ased model owever, has gure 3.22] ] FINAL for co ptual standp to the mode same data w ictions tend ling tion and the at both the t used betwe igure 3.22< feature of th , which is st ing (but is n been model mposition.do oint, the inc ls. There ha ith and with to be influe two follow our level an en the three FIG3.22> c ese three str andard prac ot always st ed in all thr cx lusion of in s not yet be out such int nced by add ing sections— d the trip le choice dime an be found uctures is th tice in the U andard pract ee possible p tra-househo en a direct c eractions, so ing this add schedulin vel. There is nsions, and in the mode at mode cho nited States ice in other ositions rel ld interactio omparison it is not ye itional level g, location c no standard , in fact, all ls in practic ice is estim in both trip countries). ative to the ns adds aspe of forecasts t clear how of modeling hoice, and m practice fo three of the e in the Uni ated conditi -based and Time-of-day other tour-le 215 cts from . ode r the ted onal vel Figure 3.22. Tour-level odel hierarchies found in practice in the United States.

112 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER model but using discrete time. So, instead of simply using alternative-specific constants for each time period, one can model factors that tend to shift departure times earlier or later, or shift activity durations shorter or longer, using continuous variables in the model specification periods (Vovsha et al. 2004). An important concept in tour and trip scheduling models is that of the time window. The time window is the time period in which no other travel or activities have yet been scheduled in the simulated day. Figure 3.23 illustrates this concept. Initially, there is no tour scheduled for a person, so the available time window is a 24-hour period stretching from 3 a.m. to 3 a.m. (the most typical bound- ary from one day to the next used in travel surveys and activity-based models). Tours are scheduled in priority order, with mandatory (work and school tours) typically scheduled first, then joint tours scheduled in any remain- ing time windows, and then finally individual nonmandatory tours scheduled last. The figure shows a work tour scheduled first from 8 a.m. to 5 p.m., and a joint tour including multiple household members scheduled from 8 p.m. to 11 p.m. If another tour is generated during the day, there are three remaining time-window segments possible to schedule it in: from 3 a.m. to 8 a.m., 5 p.m. to 8 p.m., and 11 p.m. to 3 a.m. For most tour purposes, tours that are scheduled very early in the morning or very late at night are relatively rare, so this tour would be mostly likely to be scheduled in the period from 5 p.m. to 8 p.m. and would not neces- sarily need to span that entire 3-hour window. The time-window constraint affects not only the timing of this additional tour, but also its destination and mode; small time windows rule out, or make less attractive, destination–mode combinations that require a lot of travel time. Note that one of the complexities of modeling joint tours is that the available time window must then consider all of the household mem- bers participating in the tour, not just a single individual. One important difference found in prac- tice is the manner of scheduling any addi- tional stops and travel time within the tour. Figure 3.24 illus trates two different approaches for a home-based work tour. In Approach 1, the tour scheduling (time-of-day) model predicts the times arriving at work (8 a.m.) and the time departing from work (5 p.m.), and then the trip-level models simulate the additional tour details outward, generating a shopping stop on the way home with a duration of 70 minutes. When the travel times between home and the Figure 3.23. Scheduling of tours using time windows. 2014.11.18 C46 Primer FINAL for composition.docx 219 Figure 3.23. Scheduling of tours using time windows. [Insert Figure 3.24] [Caption] Figure 3.24. Two different approaches for scheduling travel tours. One important difference found in practice is the manner of scheduling any additional stops and travel time within the tour. Figure 3.24<FIG. 3.24> illustrates two different approaches Shop 3am 6am 9am Noon 3pm 6pm 9pm Midnight 3am Work Eat Windows available for making additional tour Shop 3am 6am 9am Noon 3pm 6pm 9pm Midnight 3am Work Eat Shop 3am 6am 9am Noon 3pm 6pm 9pm Midnight 3am Work Eat Approach 1: Start with activity at primary destination and simulate tour details “outward” Approach 2: Start with entire tour duration and simulate tour details “inward”

113 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) two tour destinations are added in the result is that the person leaves home at 7:30 a.m. and arrives back home at 7 p.m. Approach 2 works in the opposite direction, first predict- ing the time leaving home and the time arriving back home and then simulating the tour details inward from those times, toward the primary activity at work. These two approaches have relative ad- vantages and disadvantages for modeling. The rationale for Approach 1 is that the duration of the activity at the primary destination, es- pecially for work and school tours, is of high priority and should condition the participa- tion and timing of activities on intermediate stops. Approach 2 may fit better with a model structure that includes some aspect of sched- uling at the day-pattern level but generates all intermediate stops at the lower level after the tour duration has been simulated, since the full duration of the tour already includes the time needed to schedule any intermediate stops. With either approach, simulating feasible ac- tivity schedules can be a complex process, as it may not always be possible to travel to a fea- sible stop location and then on to the next des- tination within the available time window. For this reason, most activity-based model software includes capabilities to re-simulate day patterns if a feasible schedule is not simulated the first time. This issue can be prevented to some ex- tent by incorporating time/space constraints in the model specification to the greatest extent possible, and thus not including choice alterna- tives that cannot be physically reached within the remaining time window available. 3.3.3.3 Location Destination choice is perhaps the most dif- ficult choice dimension to explain adequately in travel demand model systems. This is partly because there are so many possible alternative locations for any activity; our input data tend to tell us fairly little about those locations. As for using data at the parcel level, there may be a fair amount of quantitative information about the land use on the parcel, but even that does not relate the qualitative information Figure 3.24. Two different approaches for scheduling travel tours. 2014.11.18 C46 Primer FINAL for comp sition.docx 219 Figure 3.23. Scheduling of tours using time windows. [Insert Figure 3.24] [Caption] Figure 3.24. Two different approaches for scheduling travel tours. One important difference found in practice is the manner of scheduling any additional stops and trav l time within the tour. Figure 3.24<FIG. 3.24> illustrates two ifferent approaches Shop 3am 6am 9am Noon 3pm 6pm 9pm Midnight 3am Work Eat Windows available for making additional tour Shop 3am 6am 9am Noon 3pm 6pm 9pm Midnight 3am Work Eat Shop 3am 6am 9am Noon 3pm 6pm 9pm Midnight 3am Work Eat Approach 1: Start with activity at primary destination and simulate tour details “outward” Approach 2: Start with entire tour duration and simulate tour details “inward”

114 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER that may be most important in actual travel choices. The main variables to include in models of tour primary destination choice are the same as listed earlier in Table 3.6. For tour locations, however, there may be additional information, such as the tour purpose, the remaining time window available for the tour after other tours have been scheduled, and whether or not the tour includes multiple household members. For models of work and school tour destination choice, we also know the person’s usual work and/or school location, so those are included as a special alternative, which is the most likely one to be chosen. (In some model systems, it is assumed that all school tours go to the usual school location, so there is no separate destination-choice model for school tours.) Another type of location model is the model to predict the location of any intermedi- ate stops along the tour. This model is more complex because it must consider accessibility relative to both the tour origin and the tour des- tination (or the location of the previously simu- lated intermediate stop, in cases where there is more than one stop on a half-tour). The loca- tion choice models use the generalized cost of going from the tour origin to the stop and then to the primary destination (or the other way around if the stop is on the way home from the primary destination). If an intermediate stop has already been simulated, then the location choice models use the generalized cost of going from Stop 1 to the primary destination. Be- cause the tour mode choice model has already been applied by this point, the generalized cost is typically calculated assuming that all trips along the tour are by the main tour mode (even though the trip mode choice model, which is run below this model, might predict that the trip mode is different from the tour mode for a small percentage of cases). See Figure 3.25. The fact that an intermediate stop location must be predicted relative to two other loca- tions is a major reason that the tour-based ap- proach is so difficult to implement within the aggregate zone-based software framework typically used for 4-step models. The stop loca- tion model would need to be run for every O-D pair in the region. This is not such an issue in a stochastic microsimulation model, where the model only needs to be run once for each inter- mediate stop that is simulated. When applying destination-choice models in practice, it is often noticed that the fact that an O-D pair crosses particular types of bound- aries, such as river crossings or county or state borders, appears to have an inhibitory influ- ence on destination choice that was not cap- tured in the models. Thus, it is best to include Figure 3.25. Determining generalized cost for intermediate stop locations. 2014.11.1 <H3>3.3 Mode ch are very tour. Bec is logical represent occupanc such as p park-and T predeterm estimatio dependin O particular 8 C46 Primer Figure 3.25 .3.4 Mode oice is mode similar to th ause people to model th the infrequ y along tour ark-and-ride -ride lot to r our mode ch ined nestin n data decid g on activity ne fairly typ structure h FINAL for co . Determin led at both t eir trip-level tend to use is as a singl ent cases of s as the driv are also mo etrieve their oice models g structure ( e which nes purpose, av ical mode c as nesting on mposition.do ing general he tour leve counterpart the same mo e, tour-level multimodal er may pick deled at the cars on the typically u e.g., from p ting structur ailable mod hoice struct the shared cx ized cost fo l and the tri s, but they c de for an en decision. Su tours, as we up or drop tour level, way home. sed nested lo revious state e performs es, or other ure is shown -ride, transit r intermedi p level. Tou onsider all tire tour in bsequent tr ll as cases o off passeng because use git modelin d preferenc best. The ne local charac in Figure 3 , and nonmo ate stop loc r-level mode segments of the large ma ip-level mod f changing a ers at stops. rs have to re g, either usi e research) o sting structu teristics. .26<FIG3.2 torized alte ations. choice mod the roundtr jority of cas els are used utomobile Alternatives turn to the s ng a r letting the re may vary 6>. This rnatives, 223 els ip es, it to ame

115 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) such effects in the original model specification, but without going to the extreme of adding a variable (k-factor) for every district pair. (Even observed trip O-D matrices are typically de- rived from survey data and may include sig- nificant measurement and/or sampling error, so should only be used as an indication of actual O-D flows rather than absolute targets.) 3.3.3.4 Mode Mode choice is modeled at both the tour level and the trip level. Tour-level mode choice models are very similar to their trip-level counter parts, but they consider all segments of the roundtrip tour. Because people tend to use the same mode for an entire tour in the large majority of cases, it is logical to model this as a single, tour-level decision. Subsequent trip-level models are used to represent the infrequent cases of multimodal tours, as well as cases of changing automobile occupancy along tours as the driver may pick up or drop off passengers at stops. Alternatives such as park-and-ride are also modeled at the tour level, because users have to return to the same park-and-ride lot to retrieve their cars on the way home. Tour mode choice models typically used nested logit modeling, either using a pre- determined nesting structure (e.g., from previ- ous stated preference research) or letting the estimation data decide which nesting struc- ture performs best. The nesting structure may vary depending on activity purpose, available modes, or other local characteristics. One fairly typical mode choice structure is shown in Figure 3.26. This particular structure has nesting on the shared-ride, transit, and non- motorized alternatives, although other models may have different structures. Also of interest are the path type alternatives below the main mode alternatives. For the automobile alterna- tives, these alternatives can include whether or not the path includes a tolled facility. For transit, the path type can indicate which tran- Figure 3.26. Typical mode choice alternatives and nesting structure. 2014.11.1 T Tr 8 C46 Primer Figure 3 able 3.7. Va aditional V FINAL for co .26. Typica riables Co ariables mposition.do l mode cho mmonly Us Vari cx ice alternat ed in Activi ables Possib Sim ives and ne ty-Based M le with Dis ulation sting struct ode Choice aggregate ure. Models 225

116 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER sit submode(s) are used along the transit path. (Some regions may have more types of transit than are indicated in this figure.) There is no clear consensus whether it is better to include these path type choices in the activity-based model mode choice structure or to let the net- work software path-finding process find the best path along all possible path types and just pass one set of skims for the best alternative across all path types. One can find examples of both approaches in practice. Including the path type choices within the activity-based mode choice structure has the advantage of giving the user more control over the types of variables used and amount of disaggregation in the util- ity equations, and also allows the logsum effect of a mode being more attractive overall when there is a choice among two or more attractive available path type alternatives. Table 3.7 provides a list of variables found in mode choice components of activity-based model systems. The variables at the left are traditional variables that are included in most model systems, including aggregate trip-based models. The variables at the right are addi- tional explanatory variables that can be used in disaggregate activity-based models. Several of these variables are choice outcomes predicted by higher-level models at the longer-term/ mobility and day-pattern levels. Including these variables can greatly increase the explanatory power and policy-relevance of the mode choice models. 3.3.4 Component Linkages Activity-based models include a number of subcomponent models that interact and are intended to provide behavioral realism by ad- dressing numerous choice dimensions such as activity generation, destination choice, mode choice, and time-of-day choice. These sub- component models are linked and executed in a manner that is intended to realistically represent the interaction of the various im- portant dimensions of choice that individuals and households face in carrying out their daily activities and travel, as discussed in an earlier section (Bowman 1998). Typically a set of mul- tinomial logit and nested logit choice models is estimated and implemented (Bowman 1998). The activity-based model components do not equilibrate explicitly, although measures of accessibility from lower modes such as mode choice are fed back up to higher-level models such as automobile ownership. Most activity- based models are implemented using Monte Carlo simulation, which means that they are TABLE 3.7. VARIABLES COMMONLY USED IN ACTIVITY-BASED MODE CHOICE MODELS Traditional Variables Variables Possible with Disaggregate Simulation • Purpose and time of day • Travel time and cost • Car ownership/sufficiency • Household income • Household size • Urban density • Pedestrian friendliness • Tour complexity • Travel party • Escorting arrangement • Transit pass • Free parking eligibility • Toll transponder • Person type • Age • Gender • Daily schedule, time pressure

117 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) subject to some degree of simulation variation. In some regions, multiple activity-based model simulations are executed and averaged before being used as input to the network assignment model. It should be noted that all choices modeled in an activity-based system are interdependent to some degree, but it would not be possible to estimate a model along all dimensions simulta- neously. So then one of the key aspects that dif- ferentiate different model designs—even within the same “family” of designs—is which model components are modeled and applied jointly versus which ones are modeled and applied sequentially. 3.3.5 Execution Sequencing In most U.S. model systems, work (and school) locations are predicted before automobile owner ship on the assumption that one can more easily buy or sell an automobile to fit one’s commuting needs than one can find a dif- ferent job to match one’s car ownership level. In reality, these two choices are interdependent and have sometimes been modeled that way. An important design option at the tour level is the hierarchy to use for the three types of models: destination choice, mode choice, and time of day. Nearly all activity-based models used in the United States have included destination choice above mode choice, mean- ing that tour mode choice is conditional on the chosen tour destination. (There are one or two exceptions that use joint mode and des- tination models.) This is similar to the 4-step model hierarchy in which mode choice is mod- eled after trip distribution. For tour purposes with specific, unique destinations such as work, school, medical visits, social visits to friends or relatives, and so forth, it makes sense behavior- ally to model destination choice above mode choice, using the logsum across all modes as an accessibility variable for each destination. (See the section on accessibility variables.) For other tour purposes where many different destina- tions may be reasonable alternatives, such as grocery shopping, doing errands, it might make more sense, behaviorally, to model destination choice conditional on which tour mode a per- son prefers, as the choice of mode may be more constrained than the choice of destination. In the theory of nested discrete choice models, however, the best hierarchy depends on the unexplained variance in the data used for model estimation. Typically, we have a more substantial range and accuracy of the variables in our models to explain the attrac- tiveness of travel modes than we do to explain the attractiveness of particular destinations, particularly when those destinations are aggre- gate zones consisting of many different pos- sible activity locations. This means that the un- explained error in the destination choice tends to be greater than the unexplained error in the mode choice, and so the statistical error terms will tend to be more highly correlated across alternative destinations than across alternative modes. When estimating a nested logit model, either simultaneously or sequentially, this will tend to result in estimation results that indi- cate that destination choice should be modeled below mode choice. In fact, that is the nesting order that is typically used in applied models in several European countries, such as the United Kingdom. However, common practice in the United States has remained to model des- tination choice above mode choice. In nested models, it is typically thought that parameters on accessibility logsum terms should be in the range between 0 and 1 in order to obtain models that give reasonable policy responses in practice. So, in many applied activity-based models, the parameters on the mode choice log- sum variables in the destination-choice models have been constrained to values not exceeding 1.0, after values have been estimated that are above 1.0. (Additional approaches have been

118 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER used to obtain valid logsum parameters, such as including distance impedance variables in the destination-choice models in addition to the mode choice logsum variables.) What potential activity-based model users should take away from this discussion is that the best way to model destination choice in re- lation to or in combination with mode choice remains an issue that could benefit from further research and testing in practice. It is likely that future model designs will include more flex- ibility where tour mode and destination choice are modeled simultaneously in nested (or cross- nested) logit models, with the data deciding which order of nesting (mode choice above or below destination choice) determined primarily by what the estimation results indicate is the best structure for each tour purpose. There is no clear consensus on where in this hierarchy the tour time of day models should be placed. In the earlier activity-based models that only used four or five broad time periods across the day, tour time-of-day choice was typically modeled above both destination choice and mode choice. When there are dif- ferent automobile and transit level-of-service skim variables for different time periods, this hierarchy has the advantage that the time-of- day models indicate which time-of-day-related level-of- service measures to use in the mode choice and destination-choice models (i.e., use a.m. peak congestion levels or midday conges- tion levels to model the home-to-destination half of the tour?). Research and practice have indicated that the more detailed the time periods used in the time-of-day models, the more likely that trav- elers are to shift time periods, and thus the lower down in the choice hierarchy that the time-of-day models should be. More recent activity-based models have moved toward more continuous time models with periods of 15, 30, or 60 minutes in length, and these model systems have tended to place tour time- of-day choice either below destination choice and above mode choice or below both destina- tion choice and mode choice. A third option is to estimate simultaneous nested tour-mode/ time-period models, and let the estimation data indicate which is the best nesting structure to use for each tour purpose. However, each level of nesting added to the estimation stage makes the model estimation much more complex, par- ticularly when destination sampling must be used. For example, there is still no example of a joint simultaneously nested tour destination/ mode/time-of-day model, including all three levels of tour choices. An alternative to estimating complex multi- dimensional nested models is to assume a nest- ing structure and estimate and apply the models as sequential nested models. This may mean es- timating a mode/time-of-day logsum across all possible mode and time-of-day combinations for use in the destination-choice model. The advantage of calculating all the accessibility logsum variables across all times of day is that the upper-level models can be made sensitive to policy changes that vary by time of day. The disadvantage is that it can greatly increase the runtime of the model, particularly if there are many different time periods used. These design considerations affect the vertical integrity of the models—the idea that although each choice in the hierarchy is condi- tional on the choices simulated above it, each choice alternative also receives information about the expected utility across all of the re- maining choices alternatives at all levels below it, if it were to be chosen. The more different types of choices and choice levels that are in- corporated into the design, the more difficult it has become to maintain the ideal vertical integ- rity of a fully branched tree of nested models from top to bottom. Just as at the tour level, the relative order- ing of the trip mode and departure time models can vary from one model design to the other. In

119 Chapter 3: ACTIVITY-BASED MODEL CONCEPTS AND ALGORITHMS (FOR MODELERS) this case, however, it is not as critical because the trip-level models do not have as much in- fluence on the model results as the tour-level models—they simply provide some more detail based on the tour-level choices. For example, most model systems use a hierarchy to determine the main mode of a tour, and that in turn constrains which modes can be used for any trips in the tour. A typical hierar- chy from lowest to highest is • Walk; • Bike; • Drive alone; • Shared ride; and • Transit. This would mean that a walk tour could only contain walk trips, a bike tour could only contain bike and walk trips, and so forth, while a transit tour can also contain trips by any other mode. In practice, the most common variations in mode along a tour are variations in car occupancy along the tour as a result of picking up or dropping off passengers. For ex- ample, the mode can shift from shared ride 2 to drive alone if a passenger is dropped off at the destination, or can change from shared ride 2 to shared ride 3 if a passenger is picked up. One of the main tasks of the trip mode choice model is to get the right vehicle occupancy for the various trips along the tour, depending on factors such as whether the trip is leaving or re- turning home, whether it is leaving or going to an escort (serve passenger) activity, and so on. Other wise, the most common type of mixed tour is one where a person uses transit for one half of a tour but gets a ride by car (or walks) for the other half of a tour. For model systems where the tour time-of- day model already models the main tour arrival and departure times at the most detailed level (e.g., 30 or 60 minutes), then the trip-level de- parture time model only needs to be used to model the departure time from any intermedi- ate stops, typically at that same level of tempo- ral detail. In some recent model systems, how- ever, even more detailed time periods have been used at the trip level, with periods as detailed as 5 or 10 minutes. Since the trip-level choice is only one-dimensional (compared to the two- dimensional tour time of day choice), it is more feasible to use smaller time periods at the trip level. Also, the choice is constrained by the pre- vious choices at the tour level, so there may not be many available alternatives to choose from in any case.