Research Methods for Understanding Aircraft Noise Annoyances and Sleep Disturbance (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.

Appendix G. Suggested Annoyance Survey Research Protocol The primary goal of this proposed new national survey of aircraft noise annoyance in the United States is to update previous estimated dose-response relationships and provide a best estimate of the relationship between aircraft noise exposure and the self-reported annoyance of residents for the nation as a whole. The “Schultz Curve” has been a cornerstone of aircraft noise and land use compatibility policy for the past 30 years. Yet, the data providing the basis for that relationship are out-of-date, drawn from multiple transportation modes, and generally from non-US surveys. It is important to note that the reactions are to be surveyed as distinct from reactions that are manifest as complaints. There may be some correlation between the two forms of personal reactions, but surveyed results are not biased by such factors as knowledge of how and where to complain. Highly annoyed was chosen by Schultz because it is the response of those who have “attended to the outdoor noise,” and can be thought of as exhibiting “a definite and conscious response to it.” (Schultz, 1978). They are also the ones more likely to view noise as a problem that should be dealt with. Note that the U.S. EPA (1974) also used “highly annoyed” as a part of the basis for its recommendation of levels “requisite to protect public health and welfare with an adequate margin of safety.” Projected noise exposure is the annual exposure consistent with the recommendation of the U.S. EPA – annual average day-night sound level, DNL or Ldn (U.S. EPA, 1974). The following sections discuss our approach to each Phase of the proposed work. Phase 1 – Test Plan Summary The test plan will consist of six basic components:  Selecting airports  Developing noise exposure contours  Sampling respondents  Surveying respondents  Determining noise exposures for respondents  Analysis of results In brief, we are suggesting selecting 16 airports based on precision considerations (see “Stage 1 Sampling: Selection of airports” in Phase 3 discussion). For these 16, we will first determine DNL contours so that sampling of households may be stratified by noise exposure (for sampling of households, see “Stage 2 Sampling: Selection of Households,” Phase 3). Individuals within the sampled households will be selected (see “Stage 3 Sampling: Respondent Selection,” Phase 3), and surveyed (see “Surveying Methods,” Phase 3). G-1

For completed interviews, specific values of DNL and other noise metrics will be determined, (for our proposed methods for determining the noise contours and respondent noise metrics, see “Determining Noise Exposure,” Phase 3). Analysis will then examine three methods for developing dose-response relationships: logistic regression; an alternative based on human judgments of loudness; and Schultz’ original cubic polynomial function (see “Proposed Analysis Plan” in Phase 4 discussion). Airport Coordination Once the survey airports are identified, we will contact each and coordinate a visit. Each airport should be kept informed of the eventual survey so that staff can respond appropriately to questions that may be raised by citizens or the press.16 We propose making the upcoming survey widely reported and known because it is almost certain that airport communities will become aware of it as the identification of survey subjects takes place. We will visit the airports to discuss the survey and provide information about its conduct. We also expect to collect specific information about the airport: recent controversies related to noise, last completed FAR Part 150 Study, if any, community outreach programs, if any, etc. Phase 2 – Survey Instrument Survey Content The fundamental annoyance question will be based on one recommended by Fields (2000): “Thinking about the last 12 months or so, when you are here at home, how much does the noise from aircraft bother, disturb or annoy you: not at all, slightly, moderately, very, or extremely?” There are some variables that previous studies have found to be either insignificantly or significantly correlated with surveyed annoyance. Most personal variables (gender, level of education, occupation, household size, etc.) have been investigated and shown to have little influence on reported annoyance (Miedema and Vos, 1999). However, there are a few significant factors, such as fear of aircraft crashes, and reported sensitivity to noise that could be of importance in understanding individual annoyance reports (Miedema and Vos, 1999). Our effort proposed here is to produce results that guide national policy, and are based on national level results. For example, the geographic distribution of individual noise sensitivities, fear, etc. is likely random and would not, in any case, aid in national policy formation. Two factors that have been found to correlate with annoyance and may do so in our national survey results are differences in annoyance at different airports or in different communities (Fields 2000), and differences in reported annoyance in different climates (Miedema 2005), though the latter is less significant. 16 We will discuss whether and how this information should be disseminated with FAA. G-2

Our survey instrument will, nevertheless, include some questions about personal variables to confirm that they still remain, for the most part, insignificant. On the other hand, our analysis will test for differences in annoyance in different communities, at different airports and in different climates. Testing for the significance of different airports or different climates is straight-forward in terms of categorizing respondents: it is simply a matter of which airport they live near. Testing for the significance of different communities is more of a challenge. What is a community? How is it defined geographically? We will test characterizing respondents by the predominant type of aircraft operation that produces their noise exposure: departure, arrival or sideline (start of takeoff, reverse thrust) and associated runway used. These variables, at a given exposure level, should characterize relatively small geographic areas, possibly communities. At a minimum, we should learn whether annoyance correlates with type of operation – possibly useful information for land use compatibility, at the local level, if not the Federal level. We will draft the questionnaire for review. OMB Review OMB approval of the instrument will be necessary. The approval is of more than the instrument, but will also require a “Supporting Statement” and, because our information collection will use statistical methods, details about the methods must be provided including: 1. Describe (including a numerical estimate) the potential respondent universe and any sampling or other respondent selection methods to be used. 2. Describe the procedures for the collection of information including:  Statistical methodology for stratification and sample selection,  Estimation procedure,  Degree of accuracy needed for the purpose described in the justification,  Unusual problems requiring specialized sampling procedures, and  Any use of periodic (less frequent than annual) data collection cycles to reduce burden. 3. Describe methods to maximize response rates and to deal with issues of nonresponse. 4. Describe any tests of procedures or methods to be undertaken. 5. Provide the name and telephone number of individuals consulted on statistical aspects of the design and the name of the agency unit, contractor(s), grantee(s), or other person(s) who will actually collect and/or analyze the information for the agency.17 The OMB recommends that agencies need to allow at least 120 days for consideration of initial public comments, the second public comment period and OMB review, plus additional time for preparation of the Information Collection Request (ICR), as well as time lags for publication of Federal Register notices (OMB, 2006). We will coordinate with OMB to alert OMB of the pending submission. Once finalized, the completed OMB Form 81-I and Supporting Statement will be submitted. 17 OMB Form 83-I, available: http://www.whitehouse.gov/sites/default/files/omb/inforeg/83i-fill.pdf G-3

During the time of the OMB review, we will simultaneously conduct work that falls conceptually under Data Collection, Phase 3. We propose to conduct these efforts – mainly noise exposure computations – so that once the OMB review is complete and our approach approved, we can immediately start identifying possible respondents and quantifying their noise exposures and be ready to commence surveying immediately after the first year. We recognize that Phases 3 and 4 are intended to occur after the first year, but we propose this Phase 3 work for the first year so that a reasonable schedule is maintained. Phase 3 – Data Collection No nationally representative survey of aircraft noise has been conducted in the US. The largest previous coordinated studies in the US were conducted at nine airports from 1967 to 1970 (Tracor Inc., 1971). This proposed survey provides an opportunity to not only update the “Schultz Curve” on a national basis, but to do so with improved understanding of what variables are likely to be important in affecting annoyance, with a widely accepted form of survey instrument, with improved statistical analysis capabilities and with much improved methods for determining noise exposure metrics. Sampling Respondents The sampling plan is developed to meet the study’s primary goals of estimating the relationships between aircraft noise exposure and the responses of annoyance and sleep disturbance. Additionally, the plan will allow exploration of variations in that relationship across different airports, including variations that may be due to climate, airport size, location, and other factors. A three-stage sampling plan is proposed, in which the first stage is a sample of airports to represent a national range of locations and airport types, the second stage is a probability sample of addresses with diverse noise exposure in the selected airports, and the third stage is selection of an adult at each sampled address to take the survey. Stage 1 Sampling: Selection of airports We propose surveying sixteen airports, with two airports selected from each of the eight FAA Regions. The airports sampled in the ACRP 02-35 project will be excluded. The list of airports eligible for sampling will be compiled in conjunction with the sponsor’s needs, and the airports will be selected in collaboration with the sponsor so that they represent a wide range of climates, urban development, size, number of runways, and fleet mix. Sixteen airports are proposed to obtain an acceptable precision for national estimates as described below. The airports are to be sampled first one, then three at a time as specified in the RFP. Information from the first airport and possibly from other early sampled airports may be used to change the sample allocation (see Stage 2 Sampling, below) for airports to be sampled later. For example, it may be discovered in the initial surveys that stratification by community characteristics does not increase precision, so later airport surveys would not need to use that stratification. Such changes in the sampling design could be accounted for in the analysis although some ability to compare airports would be lost if the changes were too great. An alternative that could be considered would be to sample all airports concurrently, but with fieldwork spread over a 12-month period for each airport. This alternative would accelerate study G-4

completion, would allow more efficient use of fieldwork, and would ensure that all airports are surveyed under comparable conditions to allow ready comparison and pure estimates of airport heterogeneity. It would also allow seasonality to be considered for each airport, and would decrease the airport-to-airport variability, since the seasonality effects would be removed from the variability. This alternative would result in a more precise estimate of the overall dose- response relationship as well as decreased costs. We can discuss the strengths of this approach with the sponsor. Anticipated precision for national estimates: Previous surveys have exhibited a great deal of variability in the dose-response relationship among different airports. As occurs in many studies with multiple levels of sampling, the airport-to-airport variability is the driving factor for the anticipated precision for the national estimate of the relationship between noise exposure and %HA (see Lohr, 1995; Jenney and Lohr, 2009). Using data from Fidell and Salvati (2004) and Fidell et al. (2011) to estimate the heterogeneity among airports, we anticipate that with 16 airports and 700 households sampled per airport, the margin of error for the slope and intercept in the logistic model will be approximately 0.04 and 1.8, respectively. Consequently, the anticipated margins of error for estimating the percentage of persons who are highly annoyed (%HA) for DNL values between 55 dB and 65 dB are between 4 and 5 percentage points. Of course, if the historical relationships do not hold or the variability among airports has changed, the precision from this study will differ from the anticipated values. Because of the high anticipated variability among airports, as estimated from the historical data, the only way to obtain more precision for estimating a national overall dose-response curve is to increase the number of airports surveyed; increasing the sample sizes at each airport beyond 700 will have minimal effect on the precision of the national estimates. If ten airports were sampled, the margins of error would be about 35% larger than from the sample of 16 airports we propose. Surveying more than 16 airports would reduce the margins of error commensurately. Stage 2 Sampling: Selection of Households For each selected airport, HMMH will provide a map of DNL contours (see Determining Noise Exposure, below). These contours will be used to stratify addresses into groups based on the DNL exposure, for example, 5 strata with DNL exposure 50-55 dB, 55-60 dB, 60-65 dB, 65-70 dB, 70+ dB. Alternative DNL noise stratum will be constructed for airports with unusual noise exposure profiles, for example, airports where few households are exposed to noise greater than 65 dB. Additional stratification may be considered based on types of noise exposure conditions (for example, sideline or under flight path) and/or community characteristics within DNL noise stratum. The stratification will guarantee that the sample at each airport contains households with diverse values of noise exposure. To implement the noise stratification, noise levels will be computed for each census block, with each block and all its addresses then being assigned to the appropriate stratum. As discussed below, we are considering the addresses on the US Postal Service computerized Delivery Sequence File (DSF) as the likely household sampling frame. These addresses can be geo-coded to the appropriate census blocks with a relatively small degree of geocoding error that will be of no real importance for the allocation to strata. In general, since the proportions of households in the highest DNL noise stratum are expected to be small, these strata will be sampled at higher G-5

sampling rates (generally termed “oversampled”) in order to generate sufficient sample sizes for model fitting. The greatest precision for estimating the airport-specific relationship between noise exposure and annoyance will be obtained if addresses are selected randomly within each stratum. For each address selected into the sample, the precise latitude and longitude will be determined, and an accurate value of noise exposure and of other noise-based metrics will be assigned for each specific address. Sample sizes and allocation at Stage 2: The anticipated precision for estimating quantities of interest for each airport depends on the sample size and the allocation of sampled points to strata; additionally, since the major models considered are nonlinear, the anticipated precision depends on the model quantities themselves. Many surveys have as their primary goal estimation of a population mean or proportion such as the percentage of persons who are unemployed, and specify a survey design that provides high efficiency for estimating such a quantity. This survey is different: the primary goals are estimating the relationship between DNL and %HA and estimating the heterogeneity of that relationship at different locations. The standard sampling designs used to estimate population means efficiently will not necessarily be the most cost-effective for estimating the regression or covariance parameters of interest. Instead, the desired allocation of observations to strata will give high precision for estimating model parameters and for estimating the predicted value of %HA at desired noise levels. The measures of information and optimal experimental designs studied by Abdelbasit and Plackett (1983) and Chaloner and Larntz (1989) provide useful guidance for the allocation of sampled addresses to strata. We used data from Fidell and Salvati (2004) and Fidell et al. (2011) as a basis for estimating the anticipated precision with different allocations and sample sizes, using a logistic dose-response model. Anticipated precisions were similar when other models were used. Westat statisticians have developed computer programs for calculating the anticipated precision of model parameter estimates and predictions under different models, sample sizes, and allocations of observations to strata. If the historical relations hold, and each of five DNL noise strata (55-60 dB, 60-65 dB, 65-70 dB, 70-75 dB, 75+ dB) is allocated one-fifth of the observations, it is anticipated that a sample size of 700 households near an airport would give a margin of error of approximately 4 percentage points for predicting %HA at DNL levels between 55 and 65 dB. The anticipated margins of error for the slope and intercept for a single airport are 0.03 and 2, respectively (somewhat different from the values estimated previously for all 16 airports). Some airports may not have noise exposures across the full range; if, for example, households are sampled only at DNL levels between 55 dB and 65 dB the anticipated margin of error remains at 4 percentage points for predicting %HA at DNL 55 dB or 60 dB, but increases to 7 percentage points for predicting %HA at DNL = 65 dB. For most airports we expect the anticipated margin of error for predicting %HA at DNL=65 dB to be less than 5 percentage points when a sample of size 700 is taken. Stage 3 Sampling: Respondent Selection G-6

The final sampling stage is to select at each selected address a sample of eligible adults to take the survey. One possibility is to select all persons at sampled addresses. However, while this procedure avoids the need for another stage of sampling, it has the disadvantage that the responses of people in the same household are likely to be similar to one another, making the sample results less precise than they would be if the same sized sample were more widely spread across households. For these reasons, we propose administering the survey instrument to only one adult per household, to be randomly selected from the eligible adults using the Westat- developed Rizzo method (Rizzo et al., 2004). Development of survey weights: For analyses involving “population” characteristics such as the demographic characteristics of persons living in the sampling region, we propose to construct weights for the data as is done for most representative samples. The first step applies the reciprocal of the sampling rate within each airport (the higher the probability of being sampled, the lower the weight and vice versa), the second step incorporates the reciprocal of the sampling rate within each household, and the third step adjusts for nonresponse. Determining Noise Exposure Noise exposure determinations will be made with the FAA’s INM. These computations will serve two purposes. First, noise contours will permit selection of potential survey respondents by DNL noise exposure band: 50-55 dB, 55-60 dB, 60-65 dB, 65-70 dB, and 70-75 dB, see Stage 2 Sampling, above. Second, it will permit computation of specific noise exposure (DNL) values at each respondent location. We plan also to compute additional noise-related metrics for each respondent, including number of aircraft noise events that exceed a specified level, referred to as “number above” or NA. Often this metric is NA70, meaning number of aircraft events that produce a Sound Exposure Level (SEL) louder than 70dB at the location.18 Other noise related metrics such as probability of awakening (ANSI, 2008)19, arrival, departure or sideline noise predominance will also be considered. We propose to use the HMMH proprietary software RealContours™. RealContours™ automates the preparation of INM inputs directly from flight track data to permit modeling of the full diversity of activity as precisely as possible, at a cost equivalent to the more simplified and less accurate manual approach. RealContours™ improves the precision of modeling by utilizing operations monitoring results in four key areas:  It directly converts the flight track trace for every identified aircraft operation to an INM track, rather than assigning all operations to a limited number of prototypical tracks.  It models each operation on the specific runway that it actually used, rather than applying a generalized distribution to broad ranges of aircraft types.  It can use each aircraft’s actual climb performance on departure to select the “best-fitting” climb profile for that aircraft type in the INM database. 18 SEL is a sound energy integrated metric. The SEL value for most jet aircraft operations is about 7 to 10 dB higher than the maximum level. Outdoors, speech interference commences when background levels exceed approximately 60 dBA. Consequently NA70 is a rough measure of how many times speech interference occurs outdoors. 19 HMMH was instrumental in developing the standard and it is based on Anderson and Miller (2007). G-7

 It selects the specific airframe and engine combination to model on an operation-by-operation basis, resulting in a far more detailed and truly representative fleet mix. RealContours™ does not modify any of the noise and performance data in the INM, nor does it modify the computational algorithms. The FAA has reviewed RealContours™ and has stated that it is does not require any special approvals for applying the INM because of the aforementioned characteristics. RealContours™ was used for the approved Part 150 at Baltimore Washington International Airport, has been used for the Environmental Impact Statement at Providence Rhode Island International Airport, and is used for annual updates of DNL contours at Boston Logan International Airport. Application of RealContours™ using departures from runway 27, Boston International Airport: It depicts a collection of departures plotted over a photo of Boston and vicinity. The contours that follow the radar traces are DNL noise exposure levels computed using HMMH’s RealContours™ to compute annual DNL contours directly from radar systems. The green is the 65 dB DNL contour; the blues are in 5-dB increments down to 50 dB DNL. The RFP mentions possible use of monitoring data as part of determining noise exposure. We use monitoring data on a site-by-site basis to compare with INM computed levels. Our experience comparing monitored levels with those computed using RealContours™ for INM input preparation has shown how accurate the INM can be when realistic data are input. Table 15 Comparison of Measured and Modeled Annual DNL Values gives a comparison of measured and modeled DNL values for BWI at 19 Remote Monitoring Stations. Though levels at four sites differ by more than 3dB, the average difference is less than 2 dB, with measured being both lower than and higher than modeled. We regard this agreement as excellent. We also note that determining reasons for differences between measured and modeled is a very time consuming task, and beyond the scope or need of this project. Sometimes, understanding these differences requires a site visit to identify the exact situation: e.g., monitor shielded from direct view of aircraft operations by building; monitor close to road with heavy truck traffic. Table 15 Comparison of Measured and Modeled Annual DNL Values Remote Monitoring Station Measured Modeled Measured minus Modeled RMS01 50.4 52.3 -1.9 RMS02 54.9 55.6 -0.7 RMS03 65.5 63.9 1.6 G-8

Remote Monitoring Station Measured Modeled Measured minus Modeled RMS05 53.5 53.6 -0.1 RMS06 53.4 53.9 -0.5 RMS07 61.2 59.1 2.1 RMS08 56 55.8 0.2 RMS09 59.2 62.5 -3.3 RMS10 51.8 51.1 0.7 RMS12 62.8 63.9 -1.1 RMS13 51.1 51.3 -0.2 RMS14 62.6 65.1 -2.5 RMS15 68.9 75.3 -6.4 RMS17 50 54.2 -4.2 RMS19 65 68 -3 RMS20 70.5 70.4 0.1 RMS21 62.3 63.8 -1.5 RMS22 57.8 57 0.8 RMS23 61.2 57.9 3.3 Use of this approach linked to use of a single survey instrument and consistent interviewing techniques will likely minimize if not eliminate methodological differences that traditionally affect comparisons of airport-to-airport surveys. We will acquire flight track data for each airport. We expect to use up to one year of data for each when there is no cost for the data. In our pricing, we have assumed that access to these data, either from the airport or through FAA will have no cost, other than our labor to import and standardize format. Surveying Methods We suggest the survey be conducted by telephone, using an address-based sample. A telephone survey offers a number of advantages over other modes of data collection. An in-person collection would offer higher quality data with respect to response rate and coverage, but would be 6-8 times the cost of data collection. A more viable alternative to a telephone survey might be one that contacted the respondent by mail and asked the respondent to fill out the survey on paper or by the web. This method is likely to be less expensive and may provide data that is less subject to social desirability bias (Tourangeau and Yan, 2007). However, without interviewer involvement, it is up to the individuals in the household to follow a respondent selection rule (e.g., Battaglia, et al, 2008; Hicks, et al., 2012). There is some evidence that mail survey results may show higher estimated levels of annoyance (Janssen, Vos, van Kempen, Breugelmans, and Miedema, 2011; Yamada, Kaku, Yokota, Namba, and Ogata, 2008). The suspicion is that the person who is most concerned about noise will respond. There are other advantages to a mail survey approach. For example it has fewer coverage issues when compared to a telephone frame (see discussion below). A mail survey is likely to have a higher response rate when compared to a telephone survey (e.g., Cantor, et al., 2007). Pending further discussions with the sponsor, we have assumed a methodology that collects the annoyance data by telephone. However at project award, we can review available evidence and re-consider this decision if deemed appropriate. G-9

The data collection approach follows the sequential administration of the surveys as specified in response to questions to the RFP. Sampled households at a single airport will be administered the survey using the procedures specified below. This initial collection will be used to evaluate the data collection methods. These methods would then be modified based on the results and applied at subsequent administrations. The RFP states that data would be collected from no more than 3 airports during a single point in time. Our budget and procedures assumes this sequential approach, based on the rationale for the sequential design specified in the RFP. However, as discussed in the sample design section, this procedure may not yield the most efficient method. At the initiation of the project, we suggest discussing this design relative to alternatives that may be less expensive and/or provide greater analytic power. Proposed Survey Data Collection Procedures As specified in the sample design section, we are proposing an address-based sample. For a telephone survey, an alternative method of sampling would be a random digit dial frame (RDD). We suggest an address frame because it provides much more precision with respect to targeting the community surrounding the airports, as well as the ability to stratify according to the noise contours of interest. An RDD sample has significant disadvantages with respect to targeting at this level of geography which we will present upon request. An address-based sample does require finding the telephone numbers for the specific address. This will initially be done by matching the address to existing ‘reverse directory’ data-bases that associate phone numbers with addresses, primarily from numbers listed in the phone book. They also include information from other data-bases that the vendor of phone numbers accesses (e.g., warranties; subscriptions). When initially making a call into the household, the protocol will verify the address with the respondent. If it is not the sampled address, the unit will be moved to the mail-survey stage of the process (see Mail Survey below). Prior experience with this method indicates that approximately 40% of the addresses will yield a telephone number that represents the sampled household. For those addresses that have a correct telephone number, the annoyance survey will be conducted directly. For those addresses where a telephone number is not available, a short mail survey will be sent to the sampled address. This survey will ask recipients to provide their telephone numbers for the telephone survey. Once a telephone number is obtained from a returned survey, the telephone interview will be administered for the sampled address. When initiating the telephone collection, we will use procedures to maximize the response rate (Dillman, et al., 2009): 1. Send a notification to explain the survey to all households with a telephone number. 2. Attempt to complete the interview (making follow-up calls to contact the respondent). 3. Follow-up with those households that do not express a specific objection. Step 1 will include a token incentive ($2). The primary purpose is to draw attention to the notification letter and its contents. This methodology has been shown to significantly increase the response from households (Cantor, et al., 2008). With respect to step 3, the vast majority of the households that refuse the interview will do so without listening to the initial explanation of the survey. In many cases, calling these households back results in gaining cooperation once the respondent understands the legitimacy of the study. G-10

The proposed protocol for the mail survey that is used to collect the telephone numbers will: 1. Send the survey to the sampled address 2. Send a ‘thank you/reminder’ postcard to all addresses 3. Send a follow-up survey to addresses that do not respond 4. Send a third survey to addresses that have not responded As with the telephone procedures, an incentive of $2 will be sent in the initial mailing. Once a phone number is obtained from the mail survey, an attempt to complete a telephone interview will be made following steps 2 and 3 listed above for the telephone interview. No notification letter will be sent to these households, as they are already aware of the study and have proactively provided their telephone number. We anticipate the entire process will take approximately 17 weeks to complete. Phase 4 – Data Analysis & Final Report Proposed Analysis Plan The analysis plan is developed here for the primary goals of (1) modeling the national dose- response relationships between noise exposure (measured by DNL, day-night average sound level) and the effects on persons in surrounding communities such as percentage of persons who are highly annoyed (%HA) and percentage of persons reporting sleep disturbance, and (2) investigating the heterogeneity of the dose-response relationships at different airports, and relating deviations from the national models to airport-specific factors. For brevity, the following discussion focuses on the response %HA; analogous models are considered for other responses. Schultz (1978) modeled %HA as a cubic polynomial function of DNL. FICON (1992) recommended continued use of the DNL metric for noise levels, and described other models for relating %HA to DNL including a logistic regression model (U.S. EPA, 1982): Logistic model: %𝐻𝐴 = 1001 + exp (−𝛽0 − 𝛽1𝐷𝑁𝐿) More recently, Fidell et al. (2011) proposed an alternative model: Alternative model: %𝐻𝐴 = 100 exp (− 𝛼 𝑚 ) where 𝑚 = [10𝐷𝑁𝐿10 ]0.3. The data are used to estimate β0 and β1 for the logistic model, or α for the alternative model. The logistic model has the advantage that additional household-level covariates (for example whether the household is under a flight path) can be included to supplement the basic relationship. The alternative model may also be modified to allow household-level covariates with some adaptations of the theory. The Schultz (1978) model does not allow individual household-level covariates to be used. We expect the fits from the three types of models to be similar, but will explore all the models. Estimates of the dose-response relationship are sensitive to the method used to combine information from the sampled airports. Combining all observations across airports in a single model (called here a combined analysis), as has been done in several studies, gives airports with higher sample sizes or certain allocation patterns higher influence in determining the national G-11

relationship. An alternative method of estimation is to fit the model to each airport separately, then average the coefficients across the airports. We propose using mixed models (Demidenko, 2004; Stiratelli et al., 1984) to estimate the overall relationship and investigate variations in the relationship among airports. Mixed models capture the best features of the combined analysis method and the averaging coefficients method. In a mixed logistic model, each airport is allowed to have its own slope and intercept, and these are estimated along with the slope and intercept that best describe the overall relationship among all airports. The models may be used to assess whether all airports have the same dose-response relationship as the overall curve. If there is heterogeneity among airports, the mixed models can estimate the degree of heterogeneity as well as investigate airport characteristics that are associated with divergence from the overall model. Mixed logistic models have been successfully used in other settings in which relationships are thought to vary across localities; see, for example, Kaufman et al. (2003). Westat statisticians have formulated a mixed model version of the alternative model that allows airports to have different levels of α, and have developed computer code for the SAS® statistical software package (SAS Institute, 2011) that can be used to fit and evaluate the mixed logistic and alternative models. The figure to the right shows differences in model fit between the combined and mixed model approaches on the data in Fidell and Salvati (2004). Because the data set contains airports with unusual dose-response relationships, the estimates of the overall relationship between DNL and %HA depend on how the information from airports is combined. The combined fits (dashed lines), which use all the data in a single model, allow airports with unusual relationships and large sample sizes to unduly influence the estimates and result in a curve for the logistic model which appears too flat. The fits for the logistic and alternative models are very similar when the mixed model formulation (solid lines) is used; the mixed models provide a better description of the overall relationship. Alternative Regression Methods for Computing Annoyance Dose-Response Curves Two approaches will be taken to estimate the quantities in the models. First, each model will be fit directly. Second, the survey design will be incorporated into the model fit to account for the effects of unequal weighting. Recently developed statistical methods in Rabe-Hesketh and G-12

Skrondal (2006) and Rao et al. (2010) will be adapted for incorporating the survey design when estimating model quantities. The fits of the models will be assessed and compared by analyzing residuals from the models as well as through goodness-of-fit tests (Vonesh et al., 1996; Hosmer and Lemeshow, 2000). The representative sample of airports, uniform sampling design at selected airports, and use of modern statistical methods for analysis will give an updated version of the dose-response relationship between noise and annoyance that will likely give reliable information for setting aviation policy. The models proposed here allow investigation of household, airport and land use factors that may be associated with differences in the degree of annoyance in response to airport noise. Reporting The final report will include the elements in the RFP, in two broad categories: (1) A full documentation of the entire survey process, including sampling, data collection and data processing, weighting, and variance estimation. This includes selection of airports and rationale, development of noise contours, survey instrument, sampling of persons living around the selected airports, data collection approach, and procedures for processing the collected data, as well as detail on preparing weights. Response rates will be reported in accordance with established practice given in the guidelines of the American Association of Public Opinion Research. (2) Results from the data analysis. Summary statistics on noise exposure, annoyance, and other responses of interest will be provided for each airport and for the national sample. The report will contain technical details of the dose-response models fitted and their properties and implications. The results will be given for separate airports as well as the aggregated sample, and will include investigations into factors that may be associated with possible heterogeneity among airports or neighborhoods. The report will discuss the significance of additional noise-related metrics, such as arrival, departure, or sideline dominance, NA70, or probability of awakening for predicting annoyance. G-13