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.
B-1 B.1 Trip Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-1 B.2 Trip Distribution/Destination Choice . . . . . . . . . . . B-5 B.3 Mode Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-7 B.4 Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .B-11 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .B-11 In preparing this report, a literature review of transfer- ability of model parameters was undertaken. This appendix presents the results of this review, which are mixed regarding the validity of transferring model parameters in many cases. The purpose of this appendix is not to warn practitioners against transferring parameters but to provide background information on research findings regarding transferability and information that may be helpful in areas where some data may be available for model estimation but not enough to estimate a complete set of model components. It is recognized, however, that many areas do not have enough data for model estimation and must use transferred parameters such as those presented in Chapter 4. The literature review found that while transferability was valid in some studies, its validity could not be demonstrated in others. In general, transferability was demonstrated for trip generation and mode choice in some cases but not others, while the literature on transferability of other param- eters, including trip distribution, time of day, and freight/ truck modeling, was insufficient to draw any conclusions. More research into model transferability, the conditions under which transferability is most likely to be valid, and ways in which the validity of transferred parameters could be improved is needed. This appendix includes several references that describe methods for scaling that could be used if limited model estimation data (possibly from a small household activity/travel survey or NHTS samples in the model region) are available. B.1 Trip Generation B.1.1 Spatial Transferability Several studies in the literature have examined spatial transferability in the context of trip generation, as discussed in the following paragraph. Caldwell and Demetsky (1980) evaluated spatial transfer- ability of linear regression models of household-level trip generation and zonal-level trip generation, using data from three cities in Virginia: Roanoke, Harrisonburg, and Winchester. In the household-level model, they considered two explanatory variables (auto ownership and household size) and used total trip productions per household as the dependent variable. In the zonal-level model, they used a single explanatory vari- able (zonal-level number of cars), with total zonal trip produc- tions classified by home-based work, home-based nonwork, and nonhome-based productions as the dependent variable. Overall, the results of the study suggest that trip generation models can be transferred between cities, at least as long as care is taken in selecting âsimilarâ cities. âSimilarâ cities are implicitly defined in the study as those with similar house- hold size, household auto ownership levels, and per capita income. Gunn et al. (1985) examined the transfer scaling approach for spatial transferability using two adjacent urban regions of the Netherlands: one located around Rotterdam and The Hague and the other located around Utrecht. The transfer- ability analysis was based on data collected at each of the two urban regions, though the data were collected at the two locations at different points in time, as well as at differ- ent times of year. To accommodate the intrinsic differences in background variables across the two spatial contexts due to different times of data collection and different periods within the year of data collection, the authors used a nation- wide travel survey as a control data set and then examined the spatial transferability of a daily shopping trip generation A p p e n d i x B Review of Literature on Transferability Studies
B-2 model as well as a personal business trip generation model (which are parts of a linked disaggregate-level nested logit system of mode-destination and trip generation specific to each trip purpose). The overall empirical results indicate that a simple uniform scaling of the coefficients between the joint model components of the base area and the transfer area is quite adequate relative to separate locally estimated models for the two areas, both from a statistical log-likelihood ratio fit perspective as well as from a prediction perspective on a suite of predefined market segments. This is quite interesting, given that the specifications adopted in these joint models are not particularly comprehensive in trip determinant variables. Specifically, the independent variables included level-of-service variables, demographic variables (cars per licensed driver, gender, and a central business district destination dummy variable), and an intrazonal trip dummy variable. Koppelman and Rose (1983) indicated that aggregate models are not likely to be spatially and temporally transferable, even in cases where the underlying disaggregate-level behavioral process is similar. This is because of differences in the dis- tribution of variables within aggregate population groups in the estimation and application contexts. In their empirical analysis, the authors, among other things, examined the intra- regional transferability of household-level linear regression trip generation models between two sectors of each of three urban areas, Baltimore, Minneapolis-St. Paul, and Washington, D.C. The dependent variables in the analysis included number of stops and number of tours. The results indicate large differ- ences in parameter estimates of the trip generation model between sectors in each urban region. However, the authors found reasonable predictive ability of the transferred models based on typical goodness-of-fit and prediction measure comparisons between the transferred models and locally esti- mated models. At the same time, their statistical tests reject transferability, despite the closeness of goodness-of-fit and prediction errors. Wilmot (1995) also examined the transferability of household-based linear regression trip generation models. He used total trips per household as the dependent variable and considered household size and number of workers as the independent variables. He examined transferability within cities, between areas in a city, and between several cities in South Africa. His results suggest that model specification does influence the level of transferability, as does the difference in average income between the estimation and application contexts. Wilmot also emphasized the need to have quality data in the application context to evaluate transferability. In his study, he found a substantial improvement in transferability when the constant in the linear regression model is updated based on application context data. Agyemang-Duah and Hall (1997) built upon the earlier research in two ways. First, they used an ordered-response model that respects the discrete and ordinal nature of number of trips and includes built-in upper limits for trip rates as the values of the explanatory variables increase. Second, they included variables related to cost of travel and accessibility in evaluating spatial transferability. The research focused on weekday home-based shopping trips made by households with one or more vehicles in the metropolitan Toronto area, based on a 1986 travel survey. The independent variables included household size, number of children less than 16 years old, number of vehicles in the household, number of full-time employed individuals working outside the home, number of part-time employed individuals working outside the home, number of individuals employed at home, number of unemployed individuals, and accessibility to shopping opportunities. Spatial transferability was examined by evaluat- ing models estimated on a core area (estimation area) to pre- dict trip generation in a periphery area (application context). Similarly, spatial transferability also was examined between the eastern and western parts of the metropolitan area, and among three pairs of municipalities. The transferability was assessed for a simple transfer scheme as well as a transfer updating scheme where factors (or scales) are applied to the latent index contribution of socioeconomic variables and the accessibility variable (the model coefficients used here are as obtained in the estimation context). Transferability was eval- uated using a transferred pseudo R2 measure (or the fraction of the constants-only log-likelihood ratio value in the pre- diction context explained by the model coefficients obtained from the estimation context), comparison of predicted versus observed aggregate shares, weighted root mean square error (the average relative error in the aggregate predicted shares weighted by the predicted shares), and two other related measures. The results indicate that the simple transfer mech- anism works quite well for model transfer, though the transfer updating procedure substantially improves the predictive ability of the transferred model. Kawamoto (2003) examined the spatial transferability of a linear regression model of total home-based trip produc- tions at the person level between two urban areas in Brazil: Sao Paulo and Bauru. They used a standardized form of the regression model, where the dependent and independent variables are represented in standardized form and are unit free. This procedure requires the values of the mean and stan- dard deviation of each model variable in the application area, and represents a transfer updating scheme where the scaling is done on a variable-by-variable basis. Transferability was evaluated based on a Wald test statistic of parameter equality in the regression models in the estimation and application contexts after accommodations for variance differences in the two contexts. The variables considered in the analysis included relationship with householder, educational attain- ment, number of cars in household, student status, employment
B-3 status, and if the individual is a child younger than 11 years. The results indicate that the standardized regression models are indeed transferable between the two cities, though the unstandardized versions are not. This is interesting, especially given that the Sao Paulo data was collected in 1987, while the Bauru data was collected in 1998. Cotrus et al. (2005) examined the spatial transferability of linear regression and Tobit models of person-level trip generation models, using data from Tel Aviv and Haifa in Israel. The data were drawn from the 1984 and 1996/1997 Israeli National Travel Habits Survey. The models included age, car availability, possession of a driverâs license, employ- ment status, education level, and whether the individual defines herself/himself as the head of the household. The results indicate that the Tobit models fit the data better, but that equality of coefficients in the two areas is rejected for both the regression and Tobit models on the basis of statistical tests. In particular, the coefficients on the license holding and age variables are statistically different, while those of other coefficients are not. However, the trans- ferred models appear to do quite well in terms of aggregate predictions. Greaves and Stopher (2000) employed the data transfer- ability approach to transfer trip production models. Specifically, they used the 1995 Nationwide Personal Transportation Survey (NPTS) data and clustered households into relatively homogenous groups for each of six trip purposes: home- work, home-school, home-shop, home-other, other-work, and other-other. A classification and regression tree method, combined with the standard analysis of variance procedure, was adopted to determine the clusters. The number of clusters varied from six groups for the home-work, home-school, and work-other purposes to 16 groups for the remaining purposes. The clustering variables included household size, number of workers, number of vehicles, and number of children and adults by age group. Within each cluster for each trip purpose, a cumulative frequency distribution was developed for number of trips produced. They then applied the cluster scheme to predict the trip productions for a survey sample of households in the Baton Rouge MPO region. For this process, they applied the clustering scheme to the add-on sample as developed earlier from the main NPTS sample, and then drew a random realization from the cumulative trip production frequency distribution for each purpose and each Baton Rouge region sample household based on the cluster to which the sample household is assigned. Next, they compared the trip pro- duction predictions from their method and from a borrowed model that is based only on household size as the indepen- dent variable, using the survey-collected trip productions as âground reality.â They found that their approach does better than the borrowed model, a result that is not surprising given that the borrowed model is based only on a single household size variable, while the authorsâ approach effectively uses several independent variables. They also compared the model estimates obtained from estimating trip production models using their synthesized trip production data and the actual survey trip production data, and concluded that the trip production models for âhome-work and home-school are well estimated, home-shop and work-other are acceptably estimated, and home-other and other-other are marginally well estimated.â Stopher et al. (2003) undertook a similar analysis as Greaves and Stopher, except that they examined the effectiveness of their approach in application areas (Dallas and Salt Lake City) where household travel surveys may not be based on the same survey collection methodology as NPTS (the Baton Rouge household travel survey used earlier was patterned after the 1995 NPTS). Specifically, the household travel surveys were collected over the fall or spring of a year, rather than the year-round data collection of NPTS, and were based on an activity survey rather than the trip-based survey of NPTS. The study also examined if the travel characteristics are a function of city characteristics in addition to demographic attributes that formed the clustering basis in the earlier work. Their results show that the simulation does not work well for the Dallas and Salt Lake City areas, though this result may simply be an artifact of the way the survey questions were worded and interpreted by respondents. They also conclude that city characteristics do matter in trip production estimates, and they recommend using contextual variables such as city population size and transit service quality. In addition, they suggest the use of a Bayesian updating of the travel character- istics for the clusters using small samples from the application context. Reuscher et al. (2002) also pursued a data transferabil- ity analysis of vehicle trips per household, vehicle miles of travel (VMT), person trips per household, and person miles of travel (PMT) rates. They used a combination of cluster/ regression analysis, judgment, and well-established relation- ships between VMT and area type and demographics. In particular, they first classified the census tracts in the United States into nine groups defined by area type (urban, suburban, and rural) and income (very low, very high, and other). Next, they developed household size-specific, number of vehicles- specific, and census tract (CT) cluster-specific vehicle trip, VMT, person trips, and PMT rate estimates (and standard error of estimates) using the 1995 NPTS data. Based on this initial classification, they subsequently undertook a clustering analysis procedure to determine the final clusters based on a combination of household size, number of vehicles, and the initial CT clusters. Once this clustering was established, the travel characteristics for any CT tract in the United States could be determined based on the cluster to which it belongs. The authors assessed their approach using data from Baton
B-4 Rouge and three NPTS add-on samples from New York, Massachusetts, and Oklahoma, and found their approach to be better than other approaches that cluster CT tracts based on metropolitan statistical area (MSA) size, census region, and census division. Mohammadian and Zhang (2007) used methods similar to the earlier data transferability studies but considered a more comprehensive set of variables to cluster households on, including demographics, pedestrian-friendly environ- ment characteristics (such as intersection density, road density, and block size), transit usage, and congestion factors (the Urban Mobility Index measure, total number of road users divided by road density, and the percentage of workers driving to work divided by road density). A combination of principal component analysis and cluster analysis was under- taken to define a total of 11 relatively homogenous groups of household types using the 2001 NHTS. This clustering scheme was then transferred to the NHTS add-on samples from New York, Wisconsin, Texas, Kentucky, and Hawaii. The transferred travel characteristics from the original NHTS sur- vey were then compared to the actual travel characteristics directly collected in add-on samples, as a way of assessing the performance of transferability. They found reasonable transfer- ability on such travel characteristics as person/vehicle trips and tours by purpose. Zhang and Mohammadian (2008a) applied the data transferability approach by generating a synthetic population for the application context using well-established population generation methods. Their application context corresponded to the New York region. They classified the generated popula- tion using the approach in Mohammadian and Zhang (2007) and compared the mean values of trips per person and trip distance per person from the simulated data with the mean values from corresponding clusters from the actual observed survey data (from the New York NHTS add-on sample). The results show good fit of the simulated and observed travel characteristics. Zhang and Mohammadian (2008b) further improved upon Zhang and Mohammadian (2008a) by fitting a gamma dis- tribution for the trip rate per person and trip distance per person for each cluster using the main NHTS survey, and next updated the parameters of this distribution using a small sample randomly selected from the NHTS add on for New York (as suggested by Stopher et al., 2003). The authors used a Bayesian approach to updating and compared the parameters of the updated gamma distribution within each cluster with the equivalent best fit gamma distribution parameters from the corresponding cluster of households from the entire New York add-on sample. The authors note that the parameters of the updated gamma distribution are closer to those from the New York add-on sample compared to the unupdated gamma distribution parameters. B.1.2 Temporal Transferability There have been relatively few studies of temporal trans- ferability in the context of trip generation. Ashford and Holloway (1971) employed data from the Pittsburgh area collected in 1958 and 1967 to examine the temporal stabil- ity of parameters from a zonal-level linear regression model as well as a household-level linear regression model (more specifically, a cross-classification model).1 The authors found substantial differences in the estimated coefficients between the regression models for the 2 years and concluded that trip generation projections over long-term planning horizons are likely to be unreliable other than for gross level-of-magnitude estimates. Kannel and Heathington (1972) performed a similar analysis of stability of parameters for a household-level linear regres- sion model using data from Indianapolis in 1964 and 1971. The independent variables considered in this analysis were household size and auto ownership. The study found sub- stantial and statistically significant differences in estimated coefficients of the linear regression models estimated in 1964 and 1971, reinforcing the finding from Ashford and Holloway (1971). Doubleday (1976) evaluated the temporal transferability of a linear regression model of the cross-classification type using employment status and profession, presence and age of children, and household car ownership as determinant variables of individual-level trip generation by trip purpose. The data were drawn from the Reading area in England from 1962 and 1971. The results indicate, among other things, that the trip generation models provide good predictive results for employed males, but not so for retired individuals, home- makers, and employed females. The inclusion of the presence and age distribution of children appeared to provide more stable results over time. Badoe and Steuart (1997) studied the temporal transfer- ability of linear regression home-based trip generation models at the household level with a simple transfer method and using data from the Greater Toronto area from 1964 and 1986. Specifically, they examined model parameter stability and the predictive ability of models estimated from the 1964 data to explain household-level trip generation in 1986. The independent variables used were household size, number of vehicles owned by household, number of licensed drivers in the household, and number of employed individuals. 1Cross-classification is but a form of linear regression where the effects of independent variables (such as car ownership, household size, etc.) are allowed to have a general non-linear effect. An equivalent linear regression formulation would have appropriately defined dummy vari- ables to represent the effect of each combination value of the independent variables.
B-5 The empirical results indicate generally large differences in the sensitivity to explanatory variables of total home-based trips, home-based work trips, home-based shopping trips, home- based social and recreational trips, and home-based personal business trips. Badoe and Steuart then evaluated predictive ability using a transfer R2 measure (i.e., the R2 measure as computed using the 1964 linear regression models on the 1986 trip generation data without any adjustments of the 1964 regression results), a transferability index (the ratio of the transfer R2 measure and the R2 measure from the 1986 linear regressions), the transfer root mean square error (RMSE) of the predictions using the 1964 models for the 1986 data, and a measure of relative RMSE (the ratio of the transfer RMSE and the RMSE from the 1986 linear regression models). These results indicate, as expected, that the transferred measures are not as good as the prediction measures based on the 1986 linear regression models though the differences are rather marginal. The differences in the transfer and 1986 model predictive abilities narrow further when the linear regression predictions are aggregated to obtain zonal-level trip ends. This is, of course, because of compensating errors and the loss of variation in the aggregation of trips to the zonal level. But the results do show statistically significant biases (overpredictions) in using the 1964 model to predict zonal-level trip ends in 1986. Overall, the authors find good temporal transferability of the 1964 models for total home-based trips and home-based work trips, but quite poor forecast performance for the home- based nonwork trip categories. However, they also note that the poor forecast performance for the nonwork categories can be partly attributed to the generally low ability to explain nonwork trips using the explanatory variables they used as well as ignoring trip chaining behavior. Cotrus et al. (2005), in their study as discussed under spatial transferability, also examined temporal transferability of trip generation models in Haifa and Tel Aviv over time. Their results indicate statistically significant differences in coefficients in each urban area over time, rejecting temporal stability in the behavioral relationship characterizing trip generation. However, the authors acknowledge that their result may be an artifact of not considering several other explanatory variables in the models, including income, land use variables, spatial structure attributes, the economic conditions, and the trans- portation system characteristics. In addition, the results may also be affected by the different survey designs, periods of data collection, and variable definitions used in the 1984 and 1996/1997 Israeli Travel Habits Surveys. B.1.3 Summary The results of studies of the spatial and temporal trans- ferability of trip generation models have been rather mixed. Unfortunately, it is difficult to synthesize the results from the various efforts to provide any conclusive guidelines for trans- ferability because of the different variable specifications used, the different dependent variables adopted (some of which are at the person level and some at the household level), the different trip purposes considered, the different geographic and temporal periods of the studies, the different model forms employed, and the different independent variable specifications in the models. Besides, most of the trip generation studies have not controlled for land use, accessibility, and transportation system characteristics when studying spatial and temporal transferability. A study by Lin and Long (2007) highlights this issue and suggests that including these additional variables can enhance spatial transferability. However, the study by Lin and Long focuses only on household auto work trips and not on other kinds of trips that are likely to exhibit more variation in trip generation relationships across space and time. In general, however, it appears safe to say that trip gen- eration transferability will be improved with better variable specifications, a disaggregate-level analysis at the household or person level rather than at an aggregate zonal level, a model structure that reflects the ordinal and discrete nature of trips, and a transfer approach that involves transfer scaling of coefficients. In the context of transfer scaling, it should be pointed out that most trip generation analyses of transferability have focused on a simple transfer approach, rather than on a transfer approach that combines some limited information from the application context to update the estimation context relationships for use in the application area. Another important issue to note in the earlier trip genera- tion studies is that they have all been trip based and do not consider trip chaining and the more general interdependence among trips of individuals. Thus, separate models for home- based trips and nonhome-based trips are developed, without any consideration of the dependence between these categories of trips. Consequently, differences in trip chaining tendencies from one area to another, or from one time period to another, could immediately result in findings of poor trip generation transferability, even if models of the number of stops (out-of- home activity participations) have good transferability. This issue needs careful attention in the future and suggests the need for transferability analysis in the context of tour-based and activity-based frameworks for travel demand modeling. B.2 Trip Distribution/ Destination Choice B.2.1 Temporal Transferability The literature on transferability of trip distribution/ destination choice is relatively limited and has been focused on temporal transferability, not spatial transferability. Volet and Hutchinson (1986) evaluated the ability of growth factor-based
B-6 and gravity-based trip distribution models for commuting trips estimated in the Toronto region in 1971 to predict the spatial distribution of commuting trips in 1981. They devel- oped models for three different spatial resolutions of the traf- fic zone system in the Toronto region: a 38-zone system, a 77-zone system, and a 124-zone system. The overall conclu- sion of this study is that the growth factor model outperforms the gravity model in predicting the 1981 spatial patterns, though both the growth factor and gravity models have dif- ficulty in replicating commute trend shifts due to changes in the urban spatial structure of employment centers and residential locations. Duffus et al. (1987) conducted a similar temporal transferability analysis with gravity-type trip distri- bution models using data from Winnipeg in the years 1962, 1971, 1976, and 1981. The authors used a rather coarse spa- tial resolution, partitioning the Winnipeg planning area into 36 âsuperâ zones. The results indicate that transferability in terms of zone-to-zone forecast errors deteriorates with the length of time of the temporal transferability period and with the inclusion of K-factors in the estimation phase. Elmi et al. (1999) examined the temporal transferability of entropy-type aggregate trip distribution models for commute trips based on data collected in the Toronto region in 1964, 1986, and 1996. The number of zones was 815 in 1964, and 1,404 in 1986 and 1996 (it is not clear how the authors reconciled this difference in zone systems in their empirical analysis). The authors also examined the influence of an improved model specification on transferability through the stratification of the trip data into two spatial markets (the Toronto Central area and the rest of the Greater Toronto area), and segmentation based on gender, auto ownership level, driverâs license status, and worker occupation. Their results show that the coefficient on the impedance parameter (represented as the auto travel time between zones) is not temporally stable, though the trans- ferred model forecasts are comparable to those obtained from locally (in time) estimated models. In addition, the extent of transferability deteriorates with an increase in time span between the estimation and application years, as also found by Duffus et al. Further, the authors observe that improved model specifications through the trip data stratifications enhance transferability significantly as measured by the disaggregate transfer log-likelihood value fits. However, this result did not carry over to transferability as measured by the zone-level root mean square forecast errors. Overall, the authors conclude that, from a pragmatic perspective, a simple model devoid of any stratification is adequate in forecast performance. The above studies have used an aggregate trip distri- bution model, with auto travel time as the only measure of travel impedance. In contrast, Karasmaa and Pursula (1997) examined temporal transferability in the context of a dis- aggregate nested logit trip destination-mode choice model, which effectively considers travel time and cost characteris- tics by multiple modes (walk, car, and public transport) in destination choice decisions. However, like the earlier trip distribution models, Karasmaa and Pursula also confined their attention to home-based work trips in the paper. The research is based on data from the Helsinki metropolitan area, collected in 1981 (estimation context) and 1988 (transfer context). The authors examined the effects of model specification by using travel time and travel-cost variables only, and then adding the number of cars per household as an additional socio- economic variable. Four transfer approaches were evaluated: transfer scaling, Bayesian updating, combined transfer, and joint context estimation. The influence of the size of the appli- cation context data on transferability was also examined by using five different samples. The authors found no substan- tial differences in disaggregate transfer predictive fit across different sample sizes and different updating methods. All sample sizes and transfer methods did well in disaggregate predictive fit compared to the locally estimated joint choice model (i.e., the model directly estimated using 1988 data). However, the implied money value of time was quite different based on estimation sample size and transfer updating proce- dure (the research restricted the implied money value of time to the same across modes and across the mode and destination choice dimensions). Also, the transferred modelâs predictions of changes in behavior due to an across-the-board 30 percent increase in public transport travel time varied substantially based on sample size and transfer updating method. The authors made some tentative conclusions about the effective- ness of the alternative transfer methods based on the modelâs predictions of behavioral changes, including the superiority of the transfer scaling approach for simple models and large transfer biases (i.e., large differences in the locally estimated parameter values in the estimation and application contexts), and the better performance of the combined transfer approach when the sample size in the application context is large and the transfer bias is small. Gunn et al. (1985) also examined destination choice model transferability, as part of their joint system of mode, destina- tion, and trip generation system. B.2.2 Summary There has been little previous research on studying trans- ferability of trip distribution and destination choice models. Further, the earlier studies in this area have been confined to temporal transferability of work trips. Within this restricted context, the results from earlier studies suggest that trip distribution/destination choice models transfer reasonably well over time in terms of predictive fit and forecast errors, though the behavioral parameters do show temporal instabil- ity. However, there seems to be no clear indication of which type of updating method would be best suited for what type
B-7 of transfer context. Of course, the trip-based nature of earlier studies completely ignores issues of destination linkages of stops and identifies the need for transferability analysis in the context of tour-based and activity-based frameworks for travel demand modeling. B.3 Mode Choice B.3.1 Spatial Transferability Watson and Westin (1975) studied the spatial transfer- ability of binary logit intercity mode choice models among different subareas in the Edinburgh-Glasgow area of Scotland. Specifically, they identified six travel âcorridorsâ in the Edinburgh-Glasgow area based on whether the origin and destination ends were in the central city, the suburbs, or periph- eral to the urban area. The modes considered were the auto- mobile and train. They included level-of-service variables and a mode-specific constant, but no socioeconomic char- acteristics of the travelers. The models estimated in the six travel corridors were then compared for similarity in model coefficients, and each model also was transferred to the other five corridors to evaluate modal split predictions. Their find- ings indicate that there is a high level of model transferability between the three models estimated in the corridors with a trip-end in the central city. However, this is not the case for the models estimated in the remaining three corridors that did not have a trip-end in the central city. Atherton and Ben-Akiva (1976) examined the spatial trans- ferability of a home-to-work trip mode choice model estimated on data collected in Washington, D.C., in 1968 to New Bedford, Massachusetts, and Los Angeles. Data from 1963 in New Bed- ford and 1967 in Los Angeles were available to test the extent of transferability of the multinomial logit model estimated from Washington, D.C. The alternatives considered in the mode choice model included driving alone, sharing a ride, and public transit. The authors conclude, based on statistical tests of parameter equality and predictive ability in the transfer contexts, that the Washington, D.C. model is transferable to the other two application areas. They further examined the benefit of updating approaches that (1) update the constants only based on aggregate shares of the alternative modes in the application area, (2) update the constants as well as estimate a single factor that scales the other coefficients, and (3) use a Bayesian update method based on the inverse of the variance-covariance matrices of the coefficient estimates from the estimation context and the application context as weighting factors. The results indicate that the Bayesian update approach works best, especially when the disaggregate sample available from the application context is small in size and the original estimation context choice model is well specified. However, there is little difference in the extent of transferability between the model with no updating and that with even the Bayesian update. Talvitie and Kirshner (1978), in their study of urban com- mute mode choice model transferability between Washington, Minneapolis-St. Paul, and San Francisco, used the same vari- able specification as that in Atherton and Ben-Akiva. The modal alternatives are drive alone, shared ride, and bus with walk access (the individuals choosing the Bay Area Rapid Transit System in the San Francisco Bay area were removed from the analysis). The authors examined transferability both within each region and between regions. The within-region transferability was examined by partitioning the sample from each region in three ways: (1) urban travel versus suburban travel (not done for the San Francisco sample), (2) central busi- ness district (CBD) travel versus non-CBD travel, and (3) a random split of the sample into two sub samples. Overall, the results of statistical tests of parameter equality between the samples within each region were mixed and inconclu- sive although there was more evidence of nonequality of parameters than equality of parameters. The between-region transferability in terms of model parameter equality also was statistically rejected with a high level of confidence. These results are clearly different from the results of Atherton and Ben-Akiva. The authors suggest that several factors may have played a role in their findings, including variations in net- work coding routines and differential trimming of outlying data points across the data sets. Galbraith and Hensher (1982) emphasized the need to consider both level-of-service variables as well as a reasonably extensive set of socioeconomic and contextual characteristics in mode choice models before evaluating transferability. They also identified the need to use consistent data (i.e., same measure- ment procedures, sampling procedures, variable definitions, questionnaire wording, etc.) in the estimation and application contexts to engage in any meaningful debates about the extent of model transferability. Their empirical analysis of the spatial transferability of mode choice models involved examining the intra-urban transferability of commute binary mode choice coefficients from two suburban areas in Sydney. The alterna- tives included car and rail. In addition to the usual level- of-service variables, the final specification used in the paper included variables representing gross annual individual income, number of licensed drivers in the household, and number of cars in the household. Their statistical tests reject parameter equality of the logit models in the two suburban regions though they find that a specification that normal- ized travel cost by income transferred relatively better than a specification that used a non-normalized travel-cost variable. However, in an evaluation of predictive ability at the mode share level, the simple transferred models without any updat- ing performed quite adequately relative to the locally estimated model. They find a Bayesian transfer update approach to
B-8 perform somewhat better than the approach without any updating and the approach that updates the constants/scale. Koppelman and Wilmot (1982) focused on the intra- regional transferability of a commute mode choice model for bread winners who work in the central business district of Washington, D.C. They caution against the sole use of model parameter equality as an indicator of whether a model is transferable or not, indicating that model parameter equality is a symmetric property between two contexts, while transferability is a directional property. In their empirical analysis, they used disaggregate measures of transferability (transfer log-likelihood ratio, transfer log-likelihood index, and the transfer rho-squared) as well as aggregate measures of transferability (root mean square error and relative root mean square error). The data sample was partitioned into three groups based on three predetermined geographic sec- tors in the Washington, D.C. area, and model transferability was studied between the resulting three pairs of sectors. The alternatives included drive alone, shared ride, and transit, and the variables included in the specification are level-of-service variables, income, vehicles per driver, a government worker dummy variable, and the number of workers in the household. The results reject parameter equality across the models for the three pairs of sectors. Further, the disaggregate measures of transferability reject the hypothesis of intraurban trans- ferability, even if the modal constants are adjusted to match the application area modal shares. However, the transferred models provide close to 80 percent of the information provided by local models, indicating that the extent of transferability is not bad at all from a nonstatistical perspective. Further, the transferred models perform quite well compared on the basis of aggregate modal share predictions. This seeming inconsistency between statistical tests and transfer errors is not uncommon, and the authors recommend that âalthough statistical tests can be used to alert the planner or analyst to differences between models, they must be considered with reference to the magnitude of errors that are acceptable in each application context.â Koppelman and Rose (1983) studied the intraregional transferability of a multinomial work mode choice model by partitioning the Baltimore region into a North sector and a South sector. The modal alternatives were drive alone, shared ride, and transit, while the independent variables included level-of-service variables as well as socioeconomic variables such as income and cars per driver. The results reject trans- ferability based on parameter equality, disaggregate measures of transferability, and aggregate measures of transferability, though there is substantial improvement in the aggregate measures of prediction when the estimated model constants are adjusted based on the aggregate modal shares in the applicant region. The authors conducted a similar analysis of intraregional transferability of mode choice models from the Washington, D.C. area and Minneapolis-St. Paul, and found that the transfer performance is much better in these other urban areas relative to Baltimore. However, even in these other areas, intraregional transferability is rejected based on statistical tests. Koppelman et al. (1985) examined the effectiveness of model updating using limited data from the application context on intraregional and interregional work travel mode choice transferability. Specifically, they studied the effect of updat- ing alternative specific constants and the scale of the model. The data used for the intraregional transferability analysis were from Washington, D.C., with the same use of three sectors as defined in Koppelman and Wilmot (1982). The data used for interregional transferability were from Washington, D.C., Minneapolis-St. Paul, and Baltimore. The independent variables used included three level-of-service variables, a car per driver variable specific to the drive-alone and shared-ride alternatives, and modal constants. The same transferability measures as developed in Koppelman and Wilmot (1982) were used in evaluating transfer effectiveness. The results indicate that transferability is improved substantially when the constants are updated, and even more so when the con- stant and scale are updated. However, the returns from updating the constant and scale are not as high as with updat- ing the constant only. This holds for both interregional and intraregional transferability. Gunn et al. (1985) conducted a similar evaluation of the effect of model updating as Koppelman et al. (1985), using a joint system of mode, destination, and trip generation system (see discussion of this paper under Section B.1.1). Their results corroborate the findings of Koppelman et al. (1985) that updating constants and the scale leads to improved model transferability. McComb (1986) assessed spatial transferability using data from a single âhigh-qualityâ data source (the transportation supplement of the Canadian Labor Force Data) for 10 cities in Canada. He used the same uniform model specification and consistent data collection and preparation across the cities and examined socioeconomic moderating effects of sensitivities to level-of-service variables. The work trip mode choice model developed for the City of Winnipeg was used as the estimation context, while the other cities were considered as the application contexts. Four modal alternatives were considered: drive alone, shared ride (driver and passenger), transit, and walk/other. The independent variables included level-of-service-variables, sex of individual, family income, age, work trip distance, and peak versus off-peak work start time. The author found that coefficient equality cannot be rejected between cities of similar socioeconomic make-up, size, and transportation system quality (such as Edmonton and Winnipeg, and Calgary and Winnipeg, at the time). However, coefficient equality was rejected for cities that are
B-9 very different in character (such as Toronto and Winnipeg and Ottawa and Winnipeg). Koppelman and Wilmot (1986) reported an analytic and empirical investigation of omission of variables on the spatial transferability of mode choice models using the same data set and procedures in Koppelman and Wilmot (1982). Three different specifications were considered to evaluate omitted variable effects on transferability, with each subsequent speci- fication, including the variables in the earlier specification and new variables as follows: (1) three level-of-service variables and modal constants, (2) addition of cars per driver variables specific to drive alone and shared ride, and (3) addition of a government worker dummy variable and a number of workers in the household variable, both specific to the shared- ride mode. The results indicate substantial improvement in transferability with improved specifications, and with modal constant updating based on the aggregate share in the appli- cation context. The authors also indicate that models with only level-of-service variables and constants are unlikely to achieve adequate levels of transferability for practical use. Koppelman and Pas (1986) also examined spatial transfer- ability of a mode choice model using the Washington, D.C. data, but added a multidimensional element to the analysis. The main focus was on whether a nested logit model of auto ownership and mode choice is more or less transferable than a simpler joint multinomial logit model of auto ownership and mode choice. The nested logit model was estimated using a two-step sequential estimation approach, which can lead to a loss of efficiency. In the empirical analysis, the nested logit modelâs logsum parameter is not statistically significantly different from 1 at the 0.05 level of significance. The results show that the transferred models without updating are able to capture more than 85 percent of the information obtained from locally estimated models for both the multinomial and nested logit models, indicating that both these models are transferable across three sectors in the Washington, D.C. area. The multinomial logit model has a small advantage in the extent of transferability though this improvement over the nested logit model is marginal. However, this result is likely to be specific to the empirical context in the study, because the nested logit specification essentially collapsed to the multinomial logit specification for all the three sectors in the Washington, D.C. area. Further analysis is needed to examine the effect of model structure on transferability. Abdelwahab (1991) examined spatial transferability of intercity mode choice models between two regions in Canada encompassing travel between 23 major metropolitan areas. He used the 1984 Canadian Travel Survey (CTS) in the analy- sis and geographically divided the 23 metropolitan areas into two regions: an eastern region, including Thunder Bay and cities east of Thunder Bay, and a western region, including Winnipeg and cities west of Winnipeg. The intercity travel in each of these regions was categorized based on trip length (short trips less than 600 miles and long trips) and purpose (recreational and business). The author used two transfer updating methods, one being the constant-only update scheme and the second being the Bayesian update method that updates all model coefficients. The independent variables used in the analysis are not provided in the paper. The results indicate that the transferred models explain about 50 to 93 percent of the information (i.e., the difference between the log-likelihood value at convergence and the log-likelihood value at market shares) provided by the locally estimated models. Overall, the findings indicate poor transferability, as measured by dis- aggregate predictive fit and aggregate error, for both updating methods considered. Karasmaa (2001) explored the spatial transferability of work trip mode choice models in the Helsinki and Turku regions of Finland. The Helsinki region was used as the esti- mation context, and the Turku as the transfer context. Four transfer approaches were evaluated: transfer scaling with re-estimation of alternative-specific constants and the scale, Bayesian updating, combined transfer, and joint context estimation. The influence of the size of the estimation con- text data on transferability also was examined by using four different sample sizes for estimation of the Helsinki mode choice model using a 1995 mobility survey. The results show that the joint context estimation is generally the best method of transfer, especially when the estimated coefficients of the locally estimated models are quite different between the estimation and application contexts. The combined transfer estimation approach is best when there is a large estimation sample and the transfer bias is small between the estimation and application contexts. All the above transferability studies were focused on a devel- oped country setting. In contrast, Santoso and Tsunokawa (2005) examined spatial transferability in a developing country. Travel survey data from Ho Chi Minh City in Vietnam is used as the case study. A work trip mode choice model with three modes (walking, bicycling, and motorcycles) was estimated for the urban area of the city, and its transferability to the suburban area was assessed. The independent variables included level-of-service variables, sex of the individual, and the ratio of number of vehicles to the number of workers. They considered four updating procedures: updating of only the constants, updating of the constants and scale, Bayesian updating, and the combined transfer approach. The transfer- ability results indicate that the Bayesian updating approach does not provide any tangible improvement over the simple transfer model (with no updating at all), while the other three methods do provide improvements. This result holds up for even small sizes of disaggregate data from the transfer context and is in contrast to the finding of Atherton and Ben-Akiva (1976). Among the remaining three approaches,
B-10 the approaches involving updating of the constants and scale and the combined transfer approach are particularly effective. Interestingly, while Koppelman et al. (1985) find that the gain from updating the constants and scale is not as high as with updating the constants only, the current study finds substantial gains from updating both the constants and scale, with rela- tively small gains (compared to the simple transfer approach) when only the constants are updated. B.3.2 Temporal Transferability McCarthy (1982) examined the temporal transferability of work trip mode choice models in the San Francisco Bay area using before and after data sets associated with the Bay Area Rapid Transit (BART) study. The research was confined to only those individuals who did not change residences and employment locations in the pre-BART and post-BART sam- ples. Data collected from November 1973 to April 1974 were used to develop a pre-BART sample (with only car and bus as the modes) as well as an immediate post-BART sample (BART was a viable mode). In addition, another short-run post-BART sample was collected in the fall of 1975 after the entire BART system became operational. The explan- atory variables used in the analysis are the usual generic level-of-service variables as well as alternative-specific vari- ables for family income, number of vehicles per driver, and a San Francisco employment dummy variable. The results show that the pre-BART binary choice model coefficients are stable in the post-BART data context. Next, a model with the pre-BART coefficients for generic variables, the car-specific coefficients from the pre-BART estimation, and freely esti- mated alternative-specific coefficients for the BART mode was developed from the immediate post-BART sample, and the transferability of this updated model to the sample from the fall of 1975 was examined. The results indicate that the coefficients are all stable, and a statistical test of coefficient equality can be marginally rejected at the 0.05 level of sig- nificance, but not at the 0.01 level of significance. Predictive success indices confirm the good temporal transferability of the updated mode choice model to the post-BART period. Badoe and Miller (1995a) examined the temporal transfer- ability of a morning peak work trip logit mode choice model in Toronto over the long-transfer period from 1964 to 1986. They also assessed if transferability was related with variable specification. The alternative modes in the analysis were auto driver, transit, and walk. The independent variables included level-of-service-variables as well as spatial, personal, and household characteristics of the commuter. In addi- tion to a single pooled model, the authors also formulated 10 models to represent 10 mutually exclusive and homo- geneous (in sensitivity to level-of-service variables) segments. Overall, statistical tests reject the hypothesis of equality of coefficients between the 1964 and 1986 estimations for all the pooled and market-segmented specifications. However, from a pragmatic perspective, the transferred models provide useful information in the application context. Specifically, the pooled models that were transferred provide at least as much as 76 percent of the log-likelihood improvement (over the constants-only log-likelihood) provided by locally estimated 1986 models. Updating the constants and scale increases this percentage to 84 percent. Improved model specifications, in general, provide better transferability, though the seg- mented model with 10 market segments did not perform well (suggesting overfitting in the estimation context). Badoe and Miller (1995b) used the same data and approach as in Badoe and Miller (1995a) but focused on comparing the performance of alternative transfer updating schemes for different sample sizes of disaggregate data availability in the application (transfer) context and different model speci- fications. The joint context estimation and the combined transfer estimation procedure provide the best transferability results. If the estimation data sample is available, the authors recommend the joint context estimation over the combined transfer approach. The simple transfer scaling approach also provides a reasonable method for model transfer. However, the authors state that âthe Bayesian approach cannot be recommended as an updating procedure.â Finally, model specification improvements led to a substantial improvement in transferability. Karasmaa and Pursula (1997) also examined the temporal transferability of mode choice models, but within the context of a joint mode-destination choice model. They found the transfer scaling approach to be best for simple models and large transfer biases (i.e., large differences in the locally estimated parameter values in the estimation and application contexts), the combined transfer approach to be best when the sample size in the application context is large and the transfer bias is small, and the joint context estimation and Bayesian update approaches to be best with small sample sizes in the applica- tion context. B.3.3 Summary There is substantial literature on work trip mode choice model transferability although much of it is focused on spa- tial transferability rather than temporal transferability. There does not appear to be any published literature on transfer- ability for non-work mode choice. The literature on work mode choice transferability in space and time is mixed. However, some general conclusions are as follows: â¢ Coefficient equality between the estimation and applica- tion contexts should not be used as the sole yardstick for
B-11 assessing transferability; rather disaggregate and aggregate prediction measures that provide an assessment of the amount of information provided by the transferred model also should be considered. â¢ Transferability improves with improved variable speci- fication. â¢ Model updating leads to a substantial improvement in transferability relative to a simple model transfer, even if the updating is simply a constants-only updating to reflect the aggregate mode shares in the application context. â¢ There is no consensus regarding which update method is best, and it would behoove the analyst to consider all of the updating procedures that are possible in order to assess which performs best in any given context. It is interesting to note that most of the mode choice trans- ferability studies have been undertaken in the 1970s and 1980s, with significantly fewer studies undertaken recently. Also, while there has been substantial focus on tour-based mode choice and activity-based modeling in general in the past two decades, there does not appear to be any analysis of transfer- ability in the context of tour-based mode choice modeling. B.4 Conclusions Overall, the literature provides mixed results regarding the effectiveness and validity of transferability though there also is a clear indication that transferability improves with a better variable specification and with a disaggregate-level model (at the individual or household level) in the estimation context (thus capturing more behavioral determinants that effectively get controlled for in the application context). The results also emphasize that, whenever possible, some level of model updating should be undertaken using local data collected in the application context. While the collection of a small disaggregate-level data set in the application context would allow model updating using any of the methods iden- tified earlier in this document (and the analyst can compare alternative updating methods), the synthesis suggests that even simple updating procedures such as a constants-only updating scheme using aggregate travel data in the application context typically provide superior results than the simple (no-update) transfer approach. However, it is recognized that even aggregate travel data may not be available in some application contexts, and there may not be resources available to collect such data prior to model transfer. In such instances, the results suggest that the simple transfer scheme should be accompanied by a careful selection of the âestimationâ city, so that the âestimationâ city is similar to the application city in terms of such factors as the distributions of household size, household auto ownership levels, employed individuals, household income, and popu- lation density. Further, it would be best to estimate travel models at a disaggregate level in the estimation context, and then apply the disaggregate-level model parameters using explanatory variable data from the application context to forecast travel. If this is not possible, an alternative approach suggested by Hu et al. (2007) may be considered, which is based on using census tracts as the unit for transfer. Specifically, Hu et al. classify all census tracts in the country into one of 11 clusters based on a combination of household income, household buying power, geo-economic nature of tract (rural/suburban/ urban/mega-urban/extreme-poverty), employment rates, life-cycle status, and number of household vehicles. For each cluster, a model is developed using households from the NHTS that are identified as belonging to that cluster. In application, each census tract of the application city is first classified into one of the 11 clusters. Then, for each census tract in the application city, the corresponding model esti- mated using the NHTS data is applied, with the exogenous variables for the tract extracted from census data. Travel sta- tistics at the tract level (number of person trips by purpose per household, number of vehicle trips per household, PMT per household and VMT per household) are then converted to a traffic zone level using blocks as a linking mechanism. It should be noted, however, that this method does not pro- vide spatial information on trips (origins and destinations of trips) and so may be of limited use for travel modeling. Further, these authors also emphasize the importance of local data collection in the application context. References Abdelwahab, W. M. (1991). âTransferability of Intercity Disaggregate Mode Choice Models in Canada.â Canadian Journal of Civil Engi- neering, 18, pp. 20â26. Agyemang-Duah, K. and F. L. Hall (1997). âSpatial Transferability of an Ordered Response Model of Trip Generation.â Transportation Research 31A, pp. 389â402. Ashford, N. and Holloway, F. (1971). âThe Permanence of Trip Gen- eration Equations.â Report prepared for the Urban Mass Trans- portation Administration, Florida State University: Tallahassee, Florida. Atherton, T. J. and M. E. Ben-Akiva (1976). âTransferability and Updating of Disaggregate Travel Demand Models.â Transportation Research Record 610, pp. 12â18. Badoe, D. A. and E. J. Miller (1995a). âAnalysis of Temporal Transferabil- ity of Disaggregate Work Trip Mode Choice Models.â Transportation Research Record 1493, pp. 1â11. Badoe, D. A. and E. J. Miller (1995b). âComparison of Alternative Methods for Updating Disaggregate Logit Mode Choice Models.â Transportation Research Record 1493, pp. 90â100. Badoe, D. A. and G. N. Steuart (1997). âUrban and Travel Changes in the Greater Toronto Area and the Transferability of Trip- Generation Models.â Transportation Planning and Technology, Vol. 20, pp. 267â290.
B-12 Caldwell, L. C., III, and M. J. Demetsky (1980). âTransferability of Trip Generation Models.â Transportation Research Record 751, pp. 56â62. Cotrus, A. V., J. N. Prashker, and Y. Shiftan (2005). âSpatial and Temporal Transferability of Trip Generation Demand Models in Israel.â Journal of Transportation and Statistics, 8, pp. 1â25. Doubleday, C. (1976). âSome Studies of the Temporal Stability of Person Trip Generation Models.â Transportation Research 11 (4), pp. 255â264. Duffus, L. N., A. S. Alfa, and A. H. Soliman (1987). âThe Reliability of Using the Gravity Model for Forecasting Trip Distribution.â Transportation, Vol. 14, No. 1, pp. 175â192. Elmi, A. M., D. A. Badoe, and E. J. Miller (1999). âTransferability Analysis of Work-Trip-Distribution Models.â Transportation Research Record 1676, pp. 169â176. Galbraith, R. A. and D. A. Hensher (1982). âIntra-Metropolitan Trans- ferability of Mode Choice Models.â Journal of Transport Economics and Policy, XVI, pp. 7â29. Greaves, S. P. and P. R. Stopher (2000). âCreating a Synthetic Household Travel and Activity Survey: Rationale and Feasibility Analysis.â Transportation Research Record 1706, pp. 82â91. Gunn, R. F., M. E. Ben-Akiva, and M. Bradley (1985). âTests of the Scaling Approach to Transferring Disaggregate Travel Demand Models.â Transportation Research Record 1037, pp. 21â30. Hu, P. S., T. Reuscher, and R. L. Schmoyer (2007). âTransferring 2001 National Household Travel Survey.â Prepared for the Federal Highway Administration, U.S. Department of Transportation by Oak Ridge National Laboratory. Kannel, E., and K. Heathington (1972). âThe Temporal Stability of Trip Generation Relationships.â Technical paper prepared as part of an investigation conducted by Joint Highway Research Project, Engineering Experiment Station, Purdue University in Cooperation with the Indiana State Highway Commission and the U.S. Department of Transportation, Federal Highway Administration. Karasmaa, N. (2001). âThe Spatial Transferability of the Helsinki Metropolitan Area Mode Choice Models.â Presented at the 5th Workshop of the Nordic Research Network on Modeling Transport, Land Use, and the Environment, pp. 1â24. Karasmaa, N. and M. Pursula (1997). âEmpirical Studies of Transfer- ability of Helsinki Metropolitan Area Travel Forecasting Models.â Transportation Research Record 1607, pp. 38â44. Kawamoto, E. (2003). âTransferability of Standardized Regression Model Applied to Person-Based Trip Generation.â Transportation Planning and Technology, Vol. 26, No. 4, pp. 331â359. Koppelman, F. S., G-K. Kuah, and C. G. Wilmot (1985). âTransfer Model Updating with Disaggregate Data.â Transportation Research Record 1037, pp. 102â107. Koppelman, F. S. and E. I. Pas (1986). âMultidimensional Choice Model Transferability.â Transportation Research-B, Vol. 20B, No. 4, pp. 321â330. Koppelman, F. S. and J. Rose (1983). âGeographic Transfer of Travel Choice Models: Evaluations and Procedures.â Optimization and Discrete Choice in Urban Systems, pp. 272â309. Koppelman, F. S. and C. G. Wilmot (1982). âTransferability Analysis of Disaggregate Choice Models.â Transportation Research Record 895, pp. 18â24. Koppelman, F. S. and C. G. Wilmot (1986). âThe Effect of Omission Vari- ables on Choice Model Transferability.â Transportation Research-B, Vol. 20B, No. 3, pp. 205â213. Lin, J. and L. Long (2007). âTransferability of Household Travel Data across Neighborhood Types and Geographic Areas Using NHTS.â Technical Report, Civil and Materials Engineering Department, University of Illinois at Chicago. McCarthy, P. S. (1982). âFurther Evidence on the Temporal Stability of Disaggregate Travel Demand Models.â Transportation Research-B, Vol. 168, No. 4, pp. 263â278. McComb, L. A. (1986). âAnalysis of the Transferability of Disaggre- gate Demand Models among Ten Canadian Cities.â Tribune Des Transports, Vol. 3-1, pp. 19â32. Mohammadian, A. and Y. Zhang (2007). âInvestigating the Transfer- ability of National Household Travel Survey Data.â Transportation Research Record 1993, pp. 67â79. Reuscher, T. R., R. L. Schmoyer, and P. S. Hu (2002). âTransferability of Nationwide Personal Transportation Survey Data to Regional and Local Scales.â Transportation Research Record 1817, pp. 25â32. Santoso, D. S. and K. Tsunokawa (2005). âSpatial Transferability and Updating Analysis of Mode Choice Models in Developing Countries.â Transportation Planning and Technology, 28, pp. 341â358. Stopher, P. R., P. Bullock, and S. Greaves (2003). âSimulating Household Travel Survey Data: Application to Two Urban Areas.â Presented at the 82nd Annual Meeting of the Transportation Research Board, Washington, D.C. Talvitie, A. and D. Kirshner (1978). âSpecification, Transferability and the Effect of Data Outliers in Modeling the Choice of Mode in Urban Travel.â Transportation 7, pp. 311â331. Volet, P. and B. G. Hutchinson (1986). âExplanatory and Forecasting Capabilities of Trip Distribution Models.â Canadian Journal of Civil Engineering, 13, pp. 666â673. Watson, P. L. and R. B. Westin (1975). âTransferability of Disaggregate Mode Choice Models.â Regional Science and Urban Economics, 5, pp. 227â249. Wilmot, C. G. (1995). âEvidence on Transferability of Trip Generation Models.â Journal of Transportation Engineering, Vol. 121, No. 5, American Society of Civil Engineers, pp. 405â410. Zhang, Y. and A. Mohammadian (2008a). âMicrosimulation of House- hold Travel Survey Data.â Presented at the 87th Annual Meeting of the Transportation Research Board, Washington, D.C. Zhang, Y. and A. Mohammadian (2008b). âBayesian Updating of Transferred Household Travel Survey Data Using MCMC Simu- lation with Gibbs Sampler.â Presented at the 87th Annual Meeting of the Transportation Research Board, Washington, D.C.