|
Sample Design for the 1990 Bay Area Household Travel Survey - Working Paper 1
Click HERE for graphic. Table of Contents Executive Summary . . . . . . . . . . . . . . . . . . . . . . i I. Background: Need for Travel Survey Information. . . . . . . . . 1 A. The Regional Transportation Database . . . . . . . . . . . 1 B. The Role of MTC in Collecting Travel Behavior Information. 2 C. Contents of the Sample Design Working Paper. . . . . . . . 4 II. Bay Area Demographics - 1980 vs 1990 . . . . . . . . . . . . . 5 III. Sample Design and Survey Design Concepts. . . . . . . . . . .12 A. Sampling Concepts and Terminology. . . . . . . . . . . . .12 B. Criteria of Sample Design. . . . . . . . . . . . . . . . .15 C. The 1990 Travel Survey Sampling Frame. . . . . . . . . . .16 IV. Estimating Adequate Sample Size. . . . . . . . . . . . . . . .20 A. Balancing Policy and Statistical Objectives. . . . . . . .20 B. Sample Size Determination: Trip Rate Analysis. . . . . . .21 C. Sample Size Determination: Trip Distribution Analysis. . .25 D. Sample Size Determination: Mode Choice Analysis. . . . . .26 V. Estimating Project Cost . . . . . . . . . . . . . . . . . . . .33 A. Comparative Metropolitan Household Travel Surveys. . . . .33 B. The 1990 Survey: Target Size and Target Cost . . . . . . .35 VI. Estimating Project Schedule. . . . . . . . . . . . . . . . . .44 V-H. Bibliography. . . . . . . . . . . . . . . . . . . . . . . . .47 Executive Summary The San Francisco Bay Area's Metropolitan Transportation Commission (MTC), the regional transportation planning agency for the nine- county Bay Area, is planning to conduct a regional household travel survey to coincide with the 1990 Census. This working paper discusses the sample design prerequisites for developing a sample survey design. Included in this paper are seven sections: 1) the need for travel survey information; 2) Bay Area demographic patterns; 3) sample design and survey design concepts and terminology; 4) statistics and procedures for estimating adequate sample size; 5) estimating project cost; 6) estimating project schedule; and 7) a bibliography on sample surveying in general and household travel surveys in particular. The purpose of this working paper is to describe the sample design for the 1990 Survey in terms of survey objectives, desired precision levels, target sample sizes and projected costs. This work basically serves as rationale for future work in developing the survey design. The survey design considers actual survey content, that is, format of the questions to be asked of all or some of the survey respondents. This paper is not intended to address the detailed sample design questions of what methodology should be used in household sample selection; or the practicality of one geo-coding scheme versus another. These questions will be addressed by MTC's Travel Survey Consultant. Elements of the survey design such as questionnaire content and use of a panel survey will be addressed in future MTC reports. Why collect travel behavior data? Why not just rely on transit and highway counts and the decennial Census? The basic purpose for collecting travel behavior information from household travel surveys is for the development of travel demand models to simulate the various components of travel behavior: auto ownership, trip frequency (trip generation), and the distribution of trips between origins and destinations by travel mode. The basic sampling frame intended for the 1990 Travel Survey is the residential telephone exchange numbers for Bay Area county households. This sampling frame is essentially the same as that for the 1981 MTC Travel Survey, where random digit dialing procedures were successfully used in selecting residential telephone numbers. Random digit dialing should eliminate potential bias of directory-based dialing procedures, where unlisted numbers are not included. Any type of i telephone survey, however, will contain a bias against households without telephones. The use of random digit dialing to access residential telephone exchanges is a suitable and appropriate sample frame for the 1990 Household Travel Survey. The Survey is intended to focus on Bay Area households, purposely excluding group quarter population (2.5% of total population). Additionally, telephone surveys deliberately (though undesirably) exclude households without telephones (an additional ñ3.6% of the Bay Area population). Correcting the sample frame to offset the zero-telephone bias may not be cost- effective. Determining an adequate sample size for conducting household travel surveys is a blend of both policy/political and statistical objectives. The policy objective may be to have an economical/cost-effective data collection program with geographic equity in terms of selecting household samples from various parts of the region. One underlying policy issue is how much money is available for data collection as opposed to other vital funding needs. The basic assumption for statistical adequacy is at what geographic scale do we want (need) "accurate" survey results: regional level or county-level accuracy? Regional-level accuracy. requires many fewer sam,oles than a county-level accurate survey, as discussed below. Money does buy accuracy, but a tradeoff between data quality and survey cost allows flexibility in determining how large a sample to survey. The table entitled "How Many Household Samples Are Required?" (Table S-1) indicates that to gain county-level accuracy for total trips per household requires nine times as many household samples (9,764 versus 1,108) as a regionally-accurate sample. Superdistrict-level accuracy would require 33 times as many samples as a regionally-accurate sample. With the objective in mind of developing county-level accurate trip rate estimates, and adequate sample sizes for developing good trip distribution and mode choice models, a sample size target of 9,900 regional households was determined (Table S-2). Relative to the 1981 MTC Travel Survey, the sample sizes will increase in all counties except San Francisco, where we intend to collect 1,300 samples instead of the previous 3,000. The sampling rate will range from a low of 0.38% in Santa Clara to a high of 1.16% in Napa County (Table S-3). The regional sampling rate is projected at 0.44 percent of all households, or 1 in 229 Bay Area households. ii A regional survey of 9,000 to 10,000 sample households will provide the Bay Area with accurate regional total trips per household within ñ2 percent; accurate county trip rates within ñ5 percent; and accurate superdistrict-level total trips per household at within ñ10 percent. A 9,000 to 10,000 sample survey will yield an accurate estimate of regional transit trips per household within ñ5 percent. A basic component of sample design is the ultimate cost to the sponsoring agency. Cost information on a per sample household basis has been collected and compared for the Bay Area and other metropolitan areas. Costs range from $20 per household to over $200 per household. Costs generally include data collection costs (costs associated with survey data collection and geo-coding of responses), and may or may not include post-survey processing costs. Post-survey processing costs, that is the "value-added" provided by public agency staff, could run into the hundreds of thousands of dollars (the typical case being MTC involvement in the 1981 Travel Survey in terms of trip linking, sample expansion and data reporting). A target cost per household of $70 has been determined by inflating the cost of the 1981 Travel Survey to current dollar levels. One of the more critical elements of developing and conducting a major household travel survey is the schedule. Basically, the schedule for the 1981 MTC Travel Survey is translated into comparable dates for the 1990 Survey (Table S-5). This is assuming the project will be funded and conducted in the Spring of 1990. Given this schedule, the final survey consultant report on the 1990 Survey should be completed by June 30, 1990. Next steps for the 1990 MTC Travel Survey includes the presentation of this sample design analysis to MTC staff and a Travel Survey Advisory Panel for review and comment. At these review sessions, issues and ideas should be discussed and debated: - Use of Panel Surveys in collecting longitudinal data; - Use of Attitudinal as well as the standard Behavioral questions; - Collection of Weekday vs Weekend trip records; - Reducing non-response by providing payment to participants ($5-$15); - Importance of non-household travel survey data collection (Figure S-1): - Commercial /Trucks / Goods Movement Inventory; - Visitor Surveys; - Non-Bay Area residents commuting to Bay Area jobs surveys; - Cordon Line Surveys (at regional boundaries); -iii- - Transit Operator On-Board Surveys; - Traffic Counting Programs; - Highway Speed surveys; - Use of Data at appropriate geographic level in travel model development; - Use of computer-assisted geo-coding technologies; - Use of computer-assisted telephone list development techniques; - Problem with too many surveys being conducted during the same time as the 1990 Census; and - Problem of too long a survey and non-response / refuse to complete survey. The list of potential items to discuss and debate will continue to grow once we embark on our first round of "brainstorming" sessions. The basic sample design and survey design are well under way, and the 1990 Household Travel Survey should provide Bay Area transportation planners and decision-makers with the database for planning our transportation systems well into the twenty-first century- -iv- Table S-1 How Many Household Samples Are Required? Allowable Error ñ10% ñ5% ñ2% Regional-Level Accuracy Total Trips/HH 277 1,108 6,925 Vehicle Trips/HH 352 1,409 8,805 Transit Trips/HH 2,447 9,787 61,170 County-Level Accuracy Total Trips/HH 2,441 9,764 61,027 Vehicle Trips/HH 3,026 12,106 75,660 Transit Trips/HH 38,176 152,709 954,430 Superdistrict-Level Accuracy Total Trips/HH 9,039 36,141 225,885 95% Confidence Level Table S-2 Comparative Sample Size 1981 vs 1990- Household Surveys Household Samples 1981 1990 County Survey Survey San Francisco 2,996 1,300 San Mateo 415 1,100 Santa Clara 855 2,000 Alameda 846 2,000 Contra Costa 483 1,300 Solano 158 600 Napa 75 500 Sonoma 234 600 Marin 147 500 Region 6,209 9,900 Table S-3 1990 Household Travel Survey Sampling Rates 1990 Sample Sampling County Households Households Rate San Francisco 313,600 1,300 0.41% San Mateo 245,000 1,100 0.45% Santa Clara 530,200 2,000 0.38% Alameda 480,900 2,000 0.42% Contra Costa 298,400 1,300 0.44% Solano 109,300 600 0.55% Napa 43,000 500 1.16% Sonoma 145,900 600 0.41% Marin 97,800 500 0.51% Region 2,264,100 9,900 0.44% Table S-4 How Accurate? 1981 vs 1990 Household Surveys Total Trips / HH ñ error% 1981 1990 County Survey Survey San Francisco 3.1% 4.7% San Mateo 7.8% 4.8% Santa Clara 5.5% 3.5% Alameda 5.9% 3.8% Contra Costa 7.5% 4.6% Solano 12.9% 6.6% Napa 17.8% 6.8% Sonoma 12.2% 7.6% Marin 13.3% 7.2% Region 2.1% 1.7% Avg/county 9.6% 5.5% 95% Confidence Level Table S-5 1981 Travel Survey Schedule Proposed 1990 Travel Survey Schedule 1981 Survey 1990 Survey Work Task Schedule Schedule Draft RFP Reviewed 10/6/80 10/6/89 RFP Mailed 11/14/80 11/14/89 Proposal Closing 12/12/80 12/12/89 Review of Proposals 12/18/80 12/18/89 Consultant Selected 12/23/80. 12/23/89 Survey Pre-Test 2/3-2/18/80 2/3-2/18/90 Survey Conducted 3/81 - 6/81 3/90 - 6/90 Final Survey Report 6/30/81 6/30/90 Click HERE for graphic. I. Background: Need for Travel Survey Information A. The Regional Transportation Database Maintaining a current, up-to-date regional transportation database is an important component of any region's overall transportation planning program. Collecting and synthesizing the region's information database into a framework suitable for analysis by planners and policy-makers is an important part of "defining the transportation problem". As problems are defined, potential solutions are devised, implemented, and subsequently monitored via a new cycle of database update. Regional transportation databases may be broadly defined to include the following components: - Highway and transit facility inventory and usage patterns; - Transit operator on-board surveys and ridership statistics by line; - Highway volume, speed and level-of-service characteristics; - Demographic / Land Use / Economic Databases, including: - Census data, including journey-to-Work commuter characteristics; - Projections databases (ABAG, State, Other); - Financial Databases, including: - Government revenues and expenditures: past, present and projected; - Projected financial needs and revenue shortfalls; - Cost Of transportation and non-transportation goods and services, including: - Consumer Price Indices; - Cost of Construction Indices; - Cost of Gasoline; - Parking charges for on-street and off-street parking; - Travel Behavior Inventory, including: - Trip frequency (trips generated); - Trip distribution (trip length); - Choice of Mode; - Socio-economic characteristics of household: - Auto ownership and auto availability; and, - Household income. Various agencies provide various pieces of the transportation database puzzle. Transit operators monitor their own route's performance, conduct marketing surveys, and conduct on-board surveys to determine travel behavior characteristics -1- of their riders. Highway "operators" (CalTrans, local public works departments) monitor highway facilities by counting vehicles, and by conducting special vehicle classification vehicle speed, and vehicle occupancy surveys. In comparison to Bay Region-specific data collection efforts, the federal Bureau of the Census conducts the decennial Census every ten years. The Census serves as the most comprehensive demographic / economic benchmark against which most other databases are compared. In particular, the Journey-to-Work component of the decennial Census is a rich, detailed database developed for metropolitan areas, providing detailed commuting patterns and commuter characteristics. The Census Journey-to-Work "packages" are an indispensable and invaluable component of the Bay Area's regional transportation database. The large sample size provides statistically valid information at a geographic scale (and cost) unavailable elsewhere. B. The Role of MTC in Collecting Travel Behavior Information. MTC, as the metropolitan planning organization for the nine-county Bay Area, collects data and conducts surveys of regional and corridor significance, cutting across modal distinctions and county lines. Certain of the databases, such as consumer price indices, gas prices, and construction cost indices are readily available from other sources. MTC collects and compiles this information into a format more appropriate for Bay Area data users. The transportation finance database is an integral component of MTC's mission in terms of providing relevant estimates of funding priorities, and -revenue and cost forecasts. MTC and its predecessor agency, the Bay Area Transportation Study Commission (BATSC), have a long-standing commitment to regional transportation databases in the form of Travel Behavior Surveys. The first survey, the 1965 BATS Home Interview Survey, collected data on travel patterns for more than 30,000 Bay Area households. This survey was updated by MTC with the 1981 Household Travel Survey, querying over 7,000 Bay Area households as to their daily travel patterns and socioeconomic background characteristics. The 1965 Survey was conducted as a home interview study, that is, surveyors actually visited sampled households at their homes. The 1981 Survey, in comparison, was a telephone survey, and included mail-out / mail-back travel diaries. -2- Household travel surveys were a common database task for metropolitan areas in the 1950's and 1960's. These massive scale origin-destination surveys often sampled five percent of a region's population. Given their cost (approximately $200 per household in today's dollars) and magnitude (30,000 Bay Area households), household surveys fell out of fashion until the late 1970's to early 1980's, when telephone surveys became the primary data collection technique. The telephone survey has become, and will continue to be, the method of choice for cost-effectively collecting travel behavior data. Why collect travel behavior data? Why not just rely on transit and highway counts and the decennial Census? The basic purpose for collecting travel behavior information from household travel surveys is for the development of travel demand models to simulate the various components of travel behavior: auto ownership, trip frequency (trip generation), and the distribution of trips between origins and destinations by travel mode. Behavioral travel models cannot be developed solely on the basis of traffic and transit counts. Travel demand models are based on the socioeconomic characteristic of a household (income, auto ownership, family size and lifestyle) compared to the actual travel choices made by each particular household and household members. Travel surveys answer the basic questions of: have trip rates (trip frequencies) so structurally changed such that changes in demographic variables cannot explain the variation in trip-making? Are average trip lengths stable over time? How does the value of travel time change with relative changes in household income and price levels? What is an adequate (accurate?) sample size for these household travel surveys? Metropolitan transportation planners from the 1950's and 1960's depended on large scale 5 percent sample surveys to determine travel patterns using zonal "aggregate" models.1 The 1970's saw the emergence of the new "disaggregate"2 breed of travel demand models, which were thought to require many fewer household ___________________________ 1. "Aggregate" models typically use "gravity" models in trip distribution and "diversion curves" in mode choice. 2 "Disaggregate" models typically use logit" choice models in determining choice probabilities (whether to own 0, 1, or 2+ autos; whether to take auto or transit to work, etc.) Disaggregate models use the person or the individual household as the basic unit of trip-making, e.g., trips per household or mode choice of a worker. Aggregate models use the travel analysis zone (TAZ) as the basic unit, e.g., trips per zone, trips between zones. Disaggregate models allow inclusion of behavioral variables such as auto ownership, workers per household, household size, income, etc. -3- samples to develop statistically adequate travel models. In practice', however, the larger the sample size, the more likely we are to capture more of the variation in trip-making characteristics. The rule-of-thumb for number of samples for use in disaggregate travel demand model development is 1,100 sample households per geographic area of interest. This rule-of-thumb will be further tested in a subsequent section of this paper. C. Contents of the Sample Design Working Paper Following this background section are seven sections detailing proposals and plans for the 1990 MTC Household Travel Survey. Section H discusses changes in Bay Area demographics from 1980 to 1990, based on ABAG's Projections '87 forecasts for the year 1990. Section III is a review of sample design and survey design concepts and terminology as they relate to the 1990 Travel Survey. The fourth section describes statistical methods used in determining "adequate" and "accurate" sample-sizes for the 1990 Survey. Following this is a discussion of proposed project costs, partly based on the cost of the 1981 MTC Travel Survey in comparison with other metropolitan area survey efforts. Section VI provides a brief discussion on the proposed schedule for the 1990 Survey. Finally, a bibliography of sample survey and household travel survey resources is included as Section VII. The purpose of this working paper is to describe the sample design for the 1990 Survey in terms of survey objectives, desired precision levels, target sample sizes and projected costs. This work basically serves as rationale for future work in developing the survey design. The survey design considers actual survey content, that is, format of the questions to be asked of all or some of the survey respondents. This paper is not intended to address the detailed sample design questions of what methodology should be used in household sample selection; or the practicality of one geo-coding scheme versus another. These questions will be addressed by MTC's Travel Survey Consultant. Elements of the survey design such as questionnaire content and use of a panel survey will be addressed in future MTC reports. -4- II. Bay Area Demographics - 1980 vs 1999 This section of the sample design working paper provides a basic review of demographic changes in the Bay Area between 1980 and 1990. ABAG's Projections '87 forecasts for 1990 are compared against 1980 Census/ABAG estimates of .households, population and employment. Comparisons of Bay Area demographic characteristics, by county, are shown in Tables 2.1. 2.2 and 2.3. Regionally, ABAG is expecting a 15 percent growth in the number of households, a 14 percent growth in household population, and a 24 percent growth in the number of employed residents and total Bay Area jobs. This is equivalent to 290 thousand new housing units, 700 thousand new Bay Area residents, 626 thousand new workers, and 600 thousand new jobs regionally, 1980 to 1990. Bay Area total population (Table 2.3) is expected to increase from 5.2 million residents in 1980 to 5.9 million by 1990. The difference between household population (Table 2.1) and total population (Table 2.3) is group quarters population. Group quarters population includes nonhousehold persons living in college dormitories, correctional institutions, nursing homes, military quarters, emergency shelters, etc. The 121 thousand group quarters population in the Bay Area in 1980 is expected to increase to 148 thousand persons by the year 1990 (Table 2.3). On a county by county basis, Solano County is by far the fastest growing Bay Area county in terms of percent growth in households and population. Solano County is expected to gain 29 thousand housing units and over 76 thousand new residents over this ten year period, at 36 percent over this period (a 3.1% annual growth rate). The resident commuter market in Solano County is expected to grow by over 48 thousand workers (48%); whereas employment opportunities in Solano are projected at only 29 thousand new jobs, 1980 to 1990 (32%). We should be expecting our 1990 Travel Survey and the 1990 Census Journey-to-Work data to show an increasingly long-distance commute for Solano County resident workers, particularly to jobs in Contra Costa and Alameda Counties. San Francisco is expected to be one of the slowest growing Bay Area counties in terms of percent growth in households and population . The county will gain 14,000 housing units and add 64 thousand new residents. Average household size in San Francisco will increase from 2.19 persons per household in 1980, to 2.29 persons per household by 1990. Employed resident growth in San Francisco (61 thousand new -5- workers) is expected to exceed new job formation (35 thousand new jobs). Given this phenomena of faster growth in workers relative to job formation, we should expect the 1990 Survey and 1990 Census to show a lower proportion of San Franciscans living-and-working in San Francisco, relative to 1980 (i.e., more San Franciscans reverse commuting" to jobs in the East Bay and the Peninsula). In terms of greatest absolute growth, Santa Clara County is expected to gain the most new housing units, population, workers and jobs, 1980-1990. (The larger base in Santa Clara relative to all other Bay Area counties yields the rather modest percentage changes.) Santa Clara is expected to gain 71 thousand additional housing units, 177 thousand new residents, 156 thousand new resident workers and a strong 204 thousand growth in employment opportunities. Household sizes are expected to decline slightly from 2.77 in 1980 to 2.73 persons per household by 1990. We are expecting the 1990 Survey and 1990 Census to show a hefty increase in incommuting to Santa Clara County from the Peninsula, East Bay, as well as Santa Cruz and San Benito Counties. Alameda County, the second largest Bay Area county, is expected to gain 55 thousand new housing units, 1980 to 1990; 141 thousand new residents; 146 thousand new workers; and 111 thousand new jobs. Household size is projected to remain steady at 2.53 persons per household. The 1990 Census should show a strong percentage increase in the number of daily commuters from the Central Valley (San Joaquin, Stanislaus Counties) to fill Alameda County jobs. We will probably also see a strong growth in out-commuting from Alameda County to Santa Clara, San Mateo and San Francisco Counties. The third largest Bay Area county is Contra Costa. The 241 thousand households in Contra Costa in 1980 should grow to 298 thousand by 1990 (15 thousand less than San Francisco). The population in Contra Costa should exceed 767 thousand by 1990 (50 thousand more than San Francisco). ABAG is expecting a 96 thousand increase in the number of Contra Costa resident workers matched with a- 85 thousand growth in new jobs. Contra Costa is expected to show the fastest percentage growth in new employment, at 42.2% over the ten year period, or a strong 3.6%/year annual growth rate. (Marin County is showing the second fastest percent growth in jobs, 1980-1990, at 41.2%.) Given the increased in-commuting from Solano County residents to fill Contra Costa County jobs, we should expect increased out-commuting from Contra Costa to jobsites in Alameda and San Francisco Counties'. -6- San Mateo County is projected as the second slowest growing county in the Bay Area in terms of percent increase in housing units and population. ABAG is projecting a 20 thousand increase in the number of San Mateo housing units; a 40 thousand increase in population; a 40 thousand increase in the number of employed residents; and a 58 thousand increase in local jobs. Workers per household is expected to increase from 1.40 in 1980 to 1.44 by 1990. (Santa Clara County has the highest workers per household in the Bay Area at 1.45 in 1980 increasing to 1.55 by 1990.) Strong growth in San Mateo job openings (22.4% over ten years), coupled with the slowest Bay Area growth in total labor force (12.6% over - ten years), is expected to yield major changes in San Mateo commute patterns. The 1990 Census and 1990 Survey may wind up showing an absolute decline in the number of San Mateo workers commuting to San Francisco. We should expect a heavy increase in incommuting from Alameda to San Mateo, as well as a strong increase in the number of San Mateo to Santa Clara County commuters. Sonoma County is second to Solano County in terms of fastest percent growth in households and population. Sonoma is the largest (in terms of population, households, etc.) of the four north bay counties. ABAG is showing Sonoma to grow by 36% in households (+31 thousand); by 24% in population (+69 thousand); by 39% in number of workers (+50 thousand); and by 35% in number of jobs (+36 thousand). Sonoma County has the fewest workers per household of the nine Bay Area counties, increasing from 1.14 workers per household in 1980 to 1.24 by 1990 (where the regional average of workers per household is expected to reach 1.40). The "surplus" commuters from Sonoma County are expected to fill jobs primarily in Marin and San Francisco Counties. Marin County is the slowest growing Bay Area county in terms of population growth, adding only 8,000 new residents between 1980 and 1990. On the other hand, household production in Marin will add 9,000 new units (a 10% increase). Household sizes in Marin are expected to drop from 2.43 in 1980 to 2.30 person per household by 1990. Marin is expected to increase by 18 thousand new workers and 31 thousand new jobs. Marin will still remain a premier bedroom community with jobs in Marin County held by Marin and Sonoma County residents, and Marin resident workers working predominantly in Marin and San Francisco. -7- Napa County is the smallest of the nine Bay Area counties and is projected to remain the smallest. Napa is expecting a gain of 7,000 new housing units (+17%), 11,000 new residents (+11%), 9,000 additional employed residents (+22%), and 11,000 new jobs (+31%) in the 1980's decade. These workers/jobs forecasts for Napa indicate a net stability in the number of Napa out-commuters (9,700 in 1980). Interestingly enough, the number of Santa Cruz to central Bay Area county commuters in 1980 (14,700) exceeds the number of Napa to central Bay Area county commuters (9,700). In summary, the 1980 Census and the Association of Bay Area Governments Projections '87 demographic forecasts provide the necessary preview of the 1990 Census year with respect to growth trends of the 1980's. Basically, we will use ABAG's geographic distribution of households within the Bay Area to better determine the fair share of household samples selected by county and sub- county unit. The commuter interpretations for 1990 are offered to pre-sage the near-term commute future (and the 1990 Census Journey- to-Work). -8- Click HERE for graphic. Click HERE for graphic. -9- Click HERE for graphic. -10- Click HERE for graphic. III. Sample Design and Survey Design Concepts Sample design and survey design concepts and terminology are reviewed in this section. These survey terminologies are then related specifically to the proposed 1990 Household Travel Survey. All of the concepts and definitions are gleaned from several survey sampling textbooks, notably those by Kish (1965), Cochrane (1963), Babbie (1973) and Stopher and Meyburg (1979). A. Sampling Concepts and Terminology. Technical terms, in this case sample survey jargon, need to be carefully defined to avoid confusion between the survey scientist and the transportation planner. Similarly, transportation planning jargon ("home-based work trip","trip", "linked trip") should be defined for the mutual benefit of transportation planners and survey scientists. The following definitions are paraphrased from the Babbie or Kish textbooks. 1. An element is the basic unit of analysis in survey research. If doing corporate research, an element might be an individual corporation. Typically, elements are persons, or, perhaps, persons in households. For the 1990 Survey, the basic element will be persons in households.. 2. A universe is a hypothetical aggregation of all survey elements. For example, "all persons in the San Francisco Bay Area". No time or spatial descriptions are used in defining universe, such that this is essentially a useless term in survey practice. 3. The term population refers to the aggregation of all elements. "The population must be defined in terms of (1) content, (2) units, (3) extent, and (4) time." For the 1990 Bay Area Travel Survey we can define the population as (1) all persons age 5 or greater, (2) in family and non-family households, (3) residing in the nine-county San Francisco Bay Area, (4) as of April 1, 1990. This statement of the population is a practical definition, yet excludes non-household (group quarters) population and children younger than five years of age. This population definition would also exclude non-residents (commuters and visitors) present in the nine county Bay Area on the April 1, 1990 "Census Day". 4. A subset of the population is the survey population. This is defined as the "aggregation of elements from which the sample is actually selected." For telephone surveys the survey population would exclude persons living in households without telephones. If the telephone survey is based on telephone directories, then the survey population would also exclude persons with unlisted telephone numbers. -12- The 1981 Travel Survey used random digit dialing procedures, such that households with unlisted numbers could be sampled. 5. The term sampling unit is used to specify the "elements or set of elements considered for selection in some stage of sampling". For example, each of the nine Bay Area counties could be considered a sampling unit, as could sub-county units such as superdistricts, travel analysis zones or census tracts. Households and persons within households would also be defined as a sampling unit stage. The terms "primary sampling units","secondary sampling units","tertiary sampling units", and "final sampling units" could be applied in the 1990 Travel Survey to Counties, Census Tracts, Households, and Persons in Households, respectively. The final sampling units are equivalent to the elements (e.g., persons in households). 6. The sampling frame is an actual list or roster of sampling units from which selections are made. The sampling frame may be a hard copy paper printout of all households living in the population area, or it could be a computer file of residential telephone number exchanges stratified by each of the sampling units (e.g., by county). Survey researchers typically define the sampling frame and the survey population jointly. "The researcher often begins with a ... population in mind for his study; then he searches for possible sampling frames. The frames available for his use are examined and evaluated, and the researcher decides which frame represents a survey population most appropriate to his needs." For the 1990 MTC Travel Survey the issue is what is the most cost-effective and comprehensive method to obtain listings of residential telephone exchanges. 7. Population parameters are summary descriptors of given variables within a population. For example, the mean number of trips per household for all Bay Area households is a population parameter. This term is sometimes used interchangeably with population value or true value. 8. Sample statistics are summary descriptors of given variables within a survey sample. Similar to the above example, the mean trips per household based on a sample survey is a statistic. Sample statistics are used in estimating population parameters. 9. Confidence levels and confidence intervals are commonly used terms to describe sample statistic accuracy or precision. For example, the researcher may claim that he is "95 percent confident" that a sample statistic (i.e., regional trips per household) is within plus or minus 2 percent of the "true" population parameter. As the confidence interval expands to, say, 99 percent confidence, the accuracy might fall within ñ5 percent of the parameter. 10. Basic types of survey designs include cross-sectional and longitudinal surveys. Cross-sectional survey data is collected at one selected point in time for describing some population. Longitudinal surveys are where the data are collected -13- over a period of time to permit analysis of changes in behavior or attitudes over time. Household travel surveys are typically considered cross-sectional even though the sample data is collected over a period of months. 11. "The primary longitudinal designs are trend studies, cohort studies, and panel studies." Trend studies are where a general population is sampled and studied at different points in time. Two cross-sectional studies, such as the 1965 BATS Home Interview Survey and the 1981 MTC Household Travel Survey, could be considered trend studies given the similarity in questionnaire content. Another good example of (shorter-range) trend studies are Gallup Polls conducted over a period of months to track shifts in political opinions. The Bay Area Council's annual survey of Bay Area residents is another example of a longitudinal trend study. Cohort studies, in contrast, follow a specific sample group (rather than the trend studies general group) over a period of time. For example, a survey in 1975 of recent high school graduates could be compared with that of high school graduates at their ten year reunion in 1985 and at their fifteen year reunion in 1990. With cohort studies, the actual sample elements studied may be different in each survey period. Panel studies involve the collection of data over time from the same sample of respondents. The original survey respondents are re-interviewed each time the survey is conducted, gathering valuable information about changes in behavior or attitudes from the exact same set of individuals. Problems with panel studies include the cost in continually tracking panel members; panel attrition, where respondents may be unwilling or unable to participate in subsequent survey cycles; and basic complications in data analysis due to changing composition of the household due to divorce, household re-location, new family members, and the basic survey statistics that are desired from such studies. Panel surveys are the least frequently conducted of the survey designs mentioned, but represent the "most sophisticated survey design for most explanatory purposes". Panel surveys remain a serious option for the 1990 Household Travel Survey in terms of selecting a sub- sample of the main sample in order to re-survey the sub-sample on, say, a two-year cycle. This way we could more efficiently monitor changes in travel behavior due to such exogenous variables as gas price and economic conditions or to internal household decisions such as lifestyle arrangements, changing work locations, or new/fewer household members. -14- B. Criteria of Sample Design. Kish suggests that sample design covers two domains: a selection process, where the rules, standards and procedures are developed to determine the sampling frame and survey population; and an estimation process, where sample statistics can be used in determining satisfactory sample sizes. Further, Kish states that a "good sample design requires the judicious balancing of four broad criteria:" - Goal orientation; - Measurability; - Practicality; and - Economy. Goal orientation is meant to consider both the selection process and the estimation process in terms of meeting research objectives. The research objective for the 1990 Travel Survey is the determination of travel behavior patterns of a cross-section of Bay Area residents in the spring of 1990, coincident with the federal 1990 Census. The research must address detailed characteristics of the household, each person in the household, and the trips each person will make on a pre-selected day. Our objective is to capture a representative cross-section of the Bay Area population in terms of auto ownership, income, race and ethnicity, and differing household structures (family, non-family, households by household size, retired households, multi-worker households, etc.) Measurability considers the ability to effectively quantify the variation in the sample statistics to provide a better handle in determining characteristics of the population at large. Kish suggests this is the "scientific bridge" between sample survey statistics and the true population parameters. Our ability to measure the standard deviation around a mean and identifying standard errors of the mean is an important component in sample design. The intent of measurability is to understand the probability of sample selection; and to reduce unmeasurable biases through better sample design. 'Practicality refers to problems in accomplishing the design essentially as intended." In other words, the most theoretically pristine sample selection procedure is worthless if the office/field instructions are not "simple, clear, practical and complete." For the 1981 Travel Survey this meant the development of clear, coherent instructions for interviewers, ample pre-testing of the survey -15- instrument, as well as flexibility in adapting the survey instrument to provide more efficient survey responses. Practicality means being able to adapt to whatever irregularities occur during the conduct of the survey, especially with regards to proper development of sampling frame survey lists. Economy refers to the cost-effectiveness of the survey in fulfilling the research objectives. Inadequate pre-testing of the survey instrument will likely end up with too many poorly worded questions. In this case, more time will be spent by interviewers and respondents in answering these questions than desired. Poorly designed surveys will increase the boredom factor for survey respondents, resulting in interviews being terminated by irritated respondents. The philosophy is that we will be trying for the most amount of usable information at the least per unit cost. C. The 1990 Travel Survey Sampling Frame The previous section on Bay Area demographics touched on projected number of Bay Area households, workers, persons and jobs by the year 1990. This section has introduced sample survey concepts and related them to 1990 Survey objectives. In this sub-section, the sampling frame for the proposed 1990 Survey is identified, and potential biases are described. The basic sampling frame intended for the 1990 Travel Survey is the residential telephone exchange numbers for Bay Area households. This sampling frame is essentially the same as that for the 1981 MTC Travel Survey, where random digit dialing procedures were successfully used in selecting residential telephone numbers. Random digit dialing should eliminate potential bias of directory- based dialing procedures, where unlisted numbers are not included. Any type of telephone survey, however, will contain a bias against households without telephones. Approximately 3.8 percent (74 thousand) of Bay Area households (1.97 million) did not have a telephone, according to the 1980 Census of Housing (Table 3-1). The "zero-telephone" market ranged from 1.5 percent in Marin County to 6.3 percent of San Francisco households. Alameda County had the greatest number (19,500) of zero-telephone households in the Bay Area. Interestingly, on a state-wide basis, 5.3% of California's 8.63 million households (455 thousand) did not have a telephone. -16- Racial and ethnic minority households are more likely not to own telephones than Bay Area households in general (Table 3.2). The 1980 Census shows 9.6 percent of the Bay Area's black households and 7.0 percent of the spanish origin households without telephones. Elderly households, on the other hand, have a higher propensity to own telephones than the population in general, with only 2.7 percent of elderly households (where the "householder" or head of household is elderly) without a residential telephone. The proportion of racial and ethnic minority households without a phone is a definite concern in terms of sample frame bias. Unfortunately, the cost to identify and include zero telephone households within our Travel Survey sampling frame may be too prohibitive to consider. These households would have to be interviewed within a home interview format, which could conceivably cost $200 per household interview. The issue of how to circumvent this sample frame bias should be addressed by MTC's Travel Survey Consultant. The 1990 Census will provide us with similar information on zero- telephone households. The household telephone item, question #H12 on the 1990 Census "Long-Form" questionnaire, is identical to the 1980 Census question. Another perhaps unmeasurable bias may be in the number of households with multiple phone listings per dwelling unit. More affluent households (or households with teenagers) may have an increased probability of being chosen from the sample frame due to multiple telephone lines. In addition to biases due to residential telephone ownership patterns, the 1990 Travel Survey sampling frame will exclude group quarters population. Group quarters, as defined previously, includes persons living in college dormitories, convents, monasteries, group homes, nursing homes, emergency shelters, inmates in correctional institutions, as well as "street people". We could assume that certain groups (e.g., inmates in correctional institutions, nursing homes) will have limited mobility outside their "residences". Exclusion of college dormitories is more problemsome, given the greater mobility afforded to college students. The sample frame will typically exclude group quarters population (approximately 2.5 percent of the Bay Area population by 1990) due to practicality: difficulty in identifying members (elements) of the group quarter, difficulty in ascertaining trip records from individuals within the group quarter, and difficulty in collecting "household" type information (e.g., autos per household, household income) from group quarters residents. The relevancy of household-based trip rate characteristics would quite -17- likely be skewed by including group quarter data. In summary, the use of random digit dialing to access residential telephone exchanges is a suitable and appropriate sample frame for the 1990 Household Travel Survey. The Survey is intended to focus on Bay Area households, purposely excluding group quarter population (2.5% of total population). Additionally, telephone surveys deliberately (though undesirably) exclude households without telephones (an additional ñ3.6% of the Bay Area population). Correcting the sample frame to offset the zero- telephone bias may not be cost-effective. -18- Click HERE for graphic. Click HERE for graphic. Notes: Elderly households are where householder (i.e., head of household) is age 65+. Black and Spanish Origin households are where householders are of the specified race. Persons of Spanish Origin may be of any race. Source: 1980 Census of Housing, Detailed Housing Characteristics, HC80-1-B6, Tabs 94,96. -19- IV. Estimating Adequate Sample Size A. Balancing Policy and Statistical Objectives Determining an adequate sample size for conducting household travel surveys is a blend of both political and statistical objectives realities. The policy / political objective may be to have a household travel survey which fairly well represents the travel behavior for residents of each individual Ba y Area County. A basic assumption for adequacy3 is at what geographic scale do we want "accurate" survey results: regional-level or county-level accuracy? Regional-level accuracy requires many fewer samples than a county-level accurate survey, as discussed below. Political (fiscal) realities concern the availability of funding. The simple question: "how many samples do we need?" could be answered (not necessarily facetiously) with: "how much money is available?" Basically, money buys accuracy, that is, increased sample sizes should decrease the sample error by a measurable amount, given proper sample frame and sample selection design. To develop statistically superior trip rate estimates for sub- county geographic units at, say, a 99 percent confidence level with trip rate errors at ñ1 percent, would require a sample survey costing in the millions of dollars. On the other hand, a statistically adequate trip rate for regional-level accuracy at a more appropriate 90 percent confidence level with trip rate error at ñ5 percent, would require about 775 regional household samples (at $50,000 - $75,000 total survey cost). This same 775 sample household survey would, however, provide inferior if not intolerable (unacceptable) errors at a county and sub-county level. The key to determining adequate sample size involves a tradeoff analysis of the fiscal realities of government financing in the 1980's and 1990's with the need for better, more accurate statistics on urban resident travel patterns. The science of determining accurate sample sizes is balanced with the art of budgeting scarce financial resources. ___________________________ 3 The term adequacy is introduced to suggest blending of statistical and policy objectives. This is not a statistical term as precise as accuracy which can be expressed in terms of percent error. Adequacy is used in the context of first identifying political / institutional objectives (e.g., county-level accurate trip rates) and accuracy levels (e.g., 95 percent confidence level with ñ5 percent error), then evaluating sample sizes as to whether they are adequate or inadequate with respect to these objectives. -20- B. Sample Size Determination: Trip Rate Analysis Most of the work in the literature on determining adequate sample sizes focusses on estimation of accurate trip rates4 (trip generation), most notably the work by Smith (1979, 1980) and Stopher (1982). Smith's work investigated the use of coefficient of variations in determining adequate samples for cross- classifications5 types of trip generation models. Stopher's work tested Smith's hypotheses for use in the 1981 Detroit Household Travel Survey. Basically, both agree that approximately 900 to 1,200 households are required to provide accurate region-level trip rate estimates. The coefficient of variation is an important statistical measure in this field of assessing adequate sample sizes. It is defined as the ratio of the sample standard deviation to the sample mean. For example, the 1980 regional mean trips per households in the Bay Area was 8.71 trips per household. The standard deviation on this mean trip rate is 7.40 trips per household.6 Thus, the coefficient of variation for Bay Area total trips per household is 0.85 (7.40 + 8.71). Similarly, if the mean household income in an area is $30,000 and the standard deviation is $5,000, then the coefficient of- variation is 0.17 ($5,000 + $30,000). Coefficients of variation (C.V.) are essentially unitless measures, such that the variability in mean statistics are directly comparable even ff the basic units of measurement are dissimilar (trip rates versus income in dollars). Larger coefficients of variation represent a larger variability in the statistic being observed. Transit trips per household, being essentially a "statistically rare event", ___________________________ 4 Trip rates as defined here include trips produced per household. 5 Cross-classification trip generation models typically stratify trips per household by socio-economic class variables: auto ownership (0, 1, 2+ autos/household), income. (low, medium, high) and/or household size (0, 1, 2, 3, 4+ persons/household). MTC's trip generation models are of a linear regression form, such that auto ownership, income and household size are rated as continual, rather than categorical variables. The difficulty in using cross- classification models in travel forecasting is in adequately segmenting zonal households into the appropriate market segment cells. 6 Standard deviations on trip rates are typically quite large, sometimes exceeding the mean trip rate by several hundred percent. The 7.41 standard deviation relative to the 8.71 trip rate can be interpreted to mean that 90 percent of all households will have from -3.4 to +20.9 trips per household. No households can have negative trips such that the minimum trip rate is effectively zero trips/hh. The standard error of the mean is a better measure of the trip rate variation in determining the population mean trip rate. The same example suggests that the true population parameter is somewhere between 8.56 and 8.86 trips per household (90% confidence interval). The standard error of the mean is defined as the standard deviation divided by the square root of the sample size (s.d. + (N^-2)). -21- will have correspondingly higher C.V.s than a more common event such as total trips per households or vehicle trips per household. Mean trips rates, their companion standard deviations and coefficients of variation, are presented in Tables 4.1 through 4.3. These trip rates, standard deviations, etc., were developed recently by MTC staff and are based on the 1981 Travel Survey. The paper by Kollo and Purvis (1984) provides more discussion on the statistical variation of 1965 Home Interview and 1981 Household Travel Survey trip rates. The sample size determination formula, as used by Smith and Stopher, is shown in Table 4.1. The formula solves for n (sample size), as follows: n = (C.V.^2* Z^2)+E^2, where C.V. = coefficient of variation, Z = Z-score for normal distribution (=1.64 for 90% confidence level) (=1.96 for 95% confidence level) (=2.58 for 99% confidence level) E = tolerable error (expressed as 0.10, 0.05, 0.02, etc.)7 Later in this paper we solve for the projected error based on assumed sample sizes, basically an algebraic transformation of the above, as follows: E = sqrt ((C.V.^2 * Z^2) /n). Regionally-accurate sample size requirements are reported in Table 4.1. The regional sample sizes required are calculated for a ñ10 percent, a ñ5 percent and a ñ2 percent tolerable error, at a 95% confidence in the trip rates8. Trip rates are shown for total trips per household, vehicle trips per household, and transit trips per household. Rates are also shown for trips by trip purpose - home-based work, home-based shop, home-based social/ recreation, home-based school, and non-home-based trips. The Bay Area's C.V. for total trips per household (0.85) is very similar to those reported in the literature (Smith, 1979). Transit trips per household are statistically ___________________________ 7 The ^ symbol is the exponentiation operator (e.g., A 2, "raised to the second power" or "squared"). 8 A more lenient 90% or more stringent 99% confidence interval could be calculated instead of the 95% level. To convert sample sizes into the 90% confidence interval, multiply the shown samples by 0.70. To convert to the 99% confidence interval the sample should be expanded by 1.73. -22- the rarest of trips and have the highest coefficients of variation (2.53). Given the regional sample size of 6,209 households from the 1981 Survey (n=6209), we solved for required sample size at the various error levels. As tolerable error decreases (becomes more strict), more samples are required. A sample of 1,108 households will provide us with regionally-accurate total trips per household at a ñ5 percent error level (95% C.L.). To reduce the potential error to ñ2 percent would require a total of 6,926 sample households. To further reduce the error to ñ1 percent at a 95 percent confidence level would require 27,700 sample households. A "very accurate" (99 percent confidence, ñ1 percent) regional travel survey would require 48,000 sample households. (At $100 per household, this would cost 4.8 million dollars). Vehicle trips are slightly less common than total trips, thus the higher coefficient of variation and corresponding sample size requirements. Approximately 1,400 sample households will yield a regionally accurate estimate of vehicle trips per household within plus or minus five percent. Transit trips per household are the least common trip we would typically measure in trip rate analysis. Given the higher probability of not surveying households who may take transit trips on a daily basis, household travel surveys are rarely used to determine regional transit trip-making rates. Transit operator on- boar.d surveys provide much of the necessary information on total transit trip-making, trips by trip purpose, by route, and by time period. Nevertheless, a 9,800 sample household travel survey will produce the necessary coverage to give us 95% confidence that our transit trips per household estimate is within plus or minus five percent of the true population parameter. Thus, a 9,800 sample household survey will yield accurate" estimates of transit trips per household. Trips by trip purpose further stratify trip-making into base components. Each of the component trip rates are statistically "rarer" than the sum-of-the-parts trip rate. Some days, some households will not take home-based work or home-based shop trips; other days, they will make several of these types of trips. The statistical variability is greater and, therefore, the greater the number of samples are required to produce similar accuracy levels. At a 95 percent confidence level, for a ñ5 percent tolerable error, we would require 1,500 sample regional households for accurately measuring home-based work trips / household; 2,300 samples for home-based shop trips; 4,000 samples for home-based social/recreation trips; 6,000 household samples for home-based school trips; and 3,200 samples to accurately measure non-home- -23- based trips per household. Sample size requirements for county-level accurate trip rates are shown in Table 4.2. Trip rates for total trips per household, vehicle trips per households and transit trips per household are shown for the nine Bay Area Counties. Standard deviations and coefficients of variation are reported along with sample size requirements at three tolerable error levels. The coefficient of variation for total trips per household ranges from a low of 0.78 in Napa to 0.95 in Sonoma County. Thus, Napa County will require fewer household samples (943) relative to Sonoma County (1,375) given an accuracy standard of plus or minus five percent, 95 percent confidence interval. The important thing to note is that to have trips rates at ñ5 percent for each of the nine counties will require 1,084 sample households per county, summing to 9,764 required samples. At the same ñ5 percent error level, we would require 12,100 household samples to attain county-accurate vehicle trip per household. County-level-accuracy for transit trips per household would require a phenomenal 152,700 household samples. Clearly, accurate transit trips per household below a regional level is prohibitively expensive . Sample size requirements for superdistrict-level-accuracy for total trips is shown in Table 4.3. MTC's 34 superdistricts (Figure 1) are commonly used by MTC staff in reporting sub-county level travel characteristics: commuter matrices, trip tables, demographic forecasts, etc. To attain the ñ5 percent accuracy level would require approximately 1,000 samples per each of the MTC 34 superdistricts, or 36,000 samples ($3.6 million). However, a 9,000 sample survey should yield accurate superdistrict level trips per household within plus or minus 10 percent - a laudable and achievable accuracy standard. Similar superdistrict summaries for vehicle trip and transit trips per household are not included due to the unrealistic sample size requirements stretching well beyond our duty (and fiscal ability) to incorporate economic sample designs. In summary, approximately 9,000 to 10,000 sample households will provide the Bay Area with accurate regional total trips per household within ñ2 percent; accurate county trip rates within ñ5 percent; and accurate superdistrict-level total trips per household at within ñ10 percent. A 9,000 to 10,000 sample survey will yield an accurate estimate of regional transit trips per household within ñ5 percent. -24- C. Sample Size Determination: Trip Distribution Analysis is that it is "not The basic conclusion of Smith (1979) regarding trip distribution feasible, therefore, to produce an accurate O-D trip table from any reasonably sized home-interview survey. Even the large surveys conducted in the past [1950's,1960's] had no hope of reproducing interchange volumes at the zonal level within a reasonable degree of accuracy." Trip tables built from travel survey trip records are typically considered "lumpy", having many zonal interchanges with zero trips and other interchanges showing thousands of trips. Trip table "lumpiness" diminishes as expanded trips are reported at more aggregate levels, e.g., MTC 34 superdistricts or nine county level. Rather than accept trip table values from travel surveys as "gospel", transportation planners use travel demand models to simulate trip volume interchanges, resulting in smoother, less lumpy zonal distributions. Typically, "gravity models" are used in simulating trip distribution patterns. MTC's trip distribution models include gravity models stratified by all trip purposes as well as "logit destination choice" models to simulate work trip distribution patterns. The calibration of gravity models and logit destination choice models are based on trip length frequency distributions (TLFD) and average trip lengths (ATL). Regional TLFDs from "observed" travel survey records are matched in model "simulations" using gravity "friction factors . The research by Smith and others notes that the sample size requirements for accurate trip rate analysis are much more stringent than those for determining regional trip length frequency distributions. In comparison to trip rate analysis, where the basic unit of analysis is the household, in trip distribution analysis the basic unit is the trip. Smith quotes others as suggesting that as few as 600 sample trips are needed for estimating adequate trip length frequency distributions. Given two sample trips per household per a specific trip purpose, then 300 households would suffice to provide adequate trip length data for model calibration purposes. The opening statement of this sub-section suggests the . impossibility in developing "accurate" zone-to-zone trip tables from household travel surveys. On the other hand, county-to-county trip interchange accuracy can be estimated using procedures described by Smith (1980)9. The 1981 MTC Travel Survey is quite accurate in terms of estimating intra-county trips, ranging from ñ1.1% error (90% confidence level) in San Francisco to ñ6.1% error for intra-Napa County trips. The largest inter-county ___________________________ 9 Smith's 1980 paper was published by FHWA. The 1979 reference is an abridgement of the work included in a Transportation Research Record (TRB). Appendix C to the 1980 paper includes statistical formulae for analyzing trip distribution accuracy. -25- trip interchanges, for example, San Mateo to San Francisco and Alameda to San Francisco, are accurate within plus or minus 8.3 to 9.5 percent at a 90% confidence level. Trip interchanges greater than 100,000 trips will typically be within ñ10 percent in the Bay Area 10,000 trips)10. A trip interchange of 10,000 total trips will typically be within 40 percent (plus or minus 4,000 trips). Lastly, a trip interchange with only 1,000 trips can probably be bracketed to somewhere between 0 and 2,000 trips (ñ100%). In conclusion, no economically-designed sample will be sufficient to satisfy accuracy demands for trip interchanges with small volumes (less than 100,000 trips). Basically, household surveys are good estimators of large trip interchanges, and have relatively large percentage errors for small trip interchanges. Almost any size household travel -survey, however, will provide sufficient detail on trip length frequency distributions for purposes of calibrating trip distribution models. Trip interchange accuracy can be increased by increasing sample size. D. Sample Size Determination: Mode Choice Analysis In estimating adequate sample sizes for developing accurate trip generation and trip distribution characteristics, standard statistical formula can be adapted to provide ready answers. In the case of sample size adequacy for mode choice model development, Smith (1979) relies on the reported experiences of others, rather than statistical measures. Smith writes that: "although the required sample size for logit modeling is difficult to derive theoretically, a reasonable range of required samples can be determined from past research in model calibration. In mode-choice modeling with data bases that contain trips (rather than households) as the primary observation unit, about 100-400 samples have been used to calibrate adequate models. Other logit models have been successfully calibrated by using data from about 500-1300 households. Thus, it seems reasonable to be able to produce an adequate model by using the 887 households required to develop [trip generation) production models." It appears that Smith's basic conclusion is that trip generation models require more detailed information than that required for logit mode choice model development. We have doubts whether this is really the case. Research at MTC indicates that mode choice models require as many samples as are affordable. Other researchers will, of course, claim success in adequately calibrating mode choice models. The question is whether their samples can be further stratified to have one sub-sample ___________________________ 10 Accuracy on a 100,000 trip interchange (out of 17 million total trips) is based on a sampling rate of 0.20% (1 in 500). Increasing the sampling rate from 0.20% to 0.40% would reduce the error from ñ11.6% to ñ8.2%, 90% confidence level. -26- of the main sample serve as an independent check on another sub- sample. In other words, can a sample be randomly separated into two estimation datasets and produce similar logit statistics? How variable are the model coefficients given this sub-sampling procedure? A brief exposure to MTC's model development research with the 1981 Travel Survey is in order to further explain the potential problems of too small a sample. MTC staff research in logit model estimation and validation suggested the need to "split" the main work trip samples from the 1981 Survey into what we called the "estimation" dataset and the "validation" dataset. The "estimation" dataset was used in estimating several logit model specifications, producing several sets of model coefficients. This is the process of disaggregate model estimation. Following estimation is the process of disaggregate model validation. This is where the coefficients derived from the "estimation" dataset were applied to the independent "validation" dataset to show prediction errors and model biases. Given the dissimilarities between the estimation and validation dataset trip-making characteristics, regardless of splitting techniques, it was virtually impossible to have coefficients from one dataset adequately replicate the travel choice of the independent, validation dataset trips. Discouraged with these results, we- "pooled" the estimation and validation datasets back together into the total sample for all subsequent model estimation and validation work. Our conclusion is that given the variability in trip-making characteristics, the larger the database, the better the opportunity to capture the variability. The 1981 MTC Travel Survey, even though the largest of its kind in the 1980's, was still too small to permit adequate testing of sub- sample model specifications. Future research at MTC and elsewhere should re-address this "minimum sample size" issue by pursuing this sub-sampling technique. If 500 trip samples are adequate for estimating stable model coefficients, then this can be rigorously tested by repeatedly drawing 500 random trip samples from a large, regional household travel survey and repeatedly applying the same model specification, comparing coefficient and logit statistic results. If the results are stable across the tests, then a possible conclusion is that 500 sample trips are indeed adequate for estimating good mode choice models. If, on the other hand, the coefficients are unstable, bounce around uncontrollably, flip-flop from positive to negative, then we can conclude that 500 samples are probably inadequate. -27- In summary, the more trip samples we have, the better are our chances for estimating good, stable, adequate mode choice models. In comparison to trip generation or trip distribution analysis, there are no readily available statistical tests to apply to determine adequate sample sizes. -28- Click HERE for graphic. Click HERE for graphic. Note: Sample size, n, determined with the formula: n = C.V.^2 * Z^2 / E^2 where n = sample size C.V. = coefficient of variation Z = Z-score for normal distribution (=1.96 for 95% confidence level) (=1.64 for 90% confidence level) E = tolerable error (expressed as 0.10, 0.05, 0.02) -30- Click HERE for graphic. Click HERE for graphic. -32- V. Estimating Project cost A basic component of sample design is the ultimate cost to the sponsoring agency. Cost information on a per sample household basis has been collected and compared for the Bay Area and other metropolitan areas. Costs range from $20 per household to over $200 per household. Costs generally include data collection costs (costs associated with survey data collection and geo-coding of responses), and may or may not include post-survey processing costs. Post-survey processing costs, that is the "value-added" provided by public agency staff, could run into the hundreds of thousands of dollars (the typical case being MTC involvement in the 1981 Travel Survey in terms of trip sampling, sample expansion and data reporting). This section summarizes cost information on Bay Area and other metropolitan area household travel surveys, sampling rates for California and other area's surveys, and proposed "target" sample costs and "target" sample sizes. This section concludes with a target sample size per county, including projected accuracy for trip rates by county and region. A. Comparative Metropolitan Household Travel Surveys Comparative costs of metropolitan area travel surveys are shown in Table 5.1. Most of these surveys are of the telephone survey format,.rather than the home interview format. (These cost values are tentative and have not been reviewed for accuracy with the respective metropolitan planning organizations.) Costs range from $20 per household in the Portland and Atlanta surveys to over $200 in Houston. Further research will be necessary to make the appropriate comparisons between similar survey instruments (telephone vs. home-interview) and consultant versus in-house processing costs. The scale of the survey would seem to make a difference given the lower per unit cost of large (+4,000 sample).surveys relative to small (<2,500 sample household) surveys. The marginal cost for adding one more sample to a large 9,000 sample survey should logically be less expensive than the marginal cost for a small 1,100 sample survey. The second half of Table 5.1 compares costs from the 1965 Bay Area Home Interview Survey and the 1981 Bay Area Telephone Survey. Bay Area consumer prices indices for "all items" are used to convert costs into various years constant dollars.11 The ___________________________ 11 Bay Area consumer price index for 1965 annual average was 94.7; for 1981, 279.0. The January 1989 Bay Area consumer price index was 381.3. Inflation averaged 7.0 percent per year for the 1965 to 1981 period and 4.4 percent per year for 1981 to 1988. 33 1965 Travel Survey cost 1.5 million in 1965 dollars. This is equivalent to 6.1 million in current year 1989 dollars. The 1981 Survey cost MTC 337 thousand dollars (1981 dollars). This is equivalent to $16 per household in 1965 dollars and $65 dollars per household sample in today's dollars. The '81 Survey would cost MTC approximately $460 thousand if it were held today with the same sample size (7,091). The 1965 BATS Survey collected travel day information for both weekdays and weekend days. Of the 30,686 households sampled, 20,486 were surveyed as to their weekday travel patterns and 10,200 were surveyed as to their weekend travel. For the '81 Survey, 6,209 were sampled for weekday travel habits and 888 for their weekend patterns (7,091 completed, total household samples). A key issue for the 1990 Bay Area Travel Survey is whether or not to include a weekend component within the total sample. Should the survey be oriented to collecting weekday travel only? Various metropolitan areas have embarked on several major data collection efforts over the past thirty years (Table 5.2). Sample sizes have declined dramatically for surveys conducted in the 1970's and 1980's and the total number of households in 12 each of the metropolitan areas has increased appreciably between survey years.12 Sampling rates in the 1980's have hovered between 0.15% and 0.36%, including the: - 1986 San Diego survey (0.32%); - 1984 Dallas-Ft. Worth survey (0.25%); - 1982 Minneapolis-St. Paul survey (0.34%); - 1981 MTC Bay Area survey (1.17% in San Francisco; - 0.21% in other eight counties for total sampling rate of 0.36%); and the - 1980 Detroit survey (0.15%). Shown in Table 5.2 is the proposed Bay Area 1990 sample of 9,900 households, with a projected 2.264 million 1990 households and a 0.44% sampling rate. (This survey will probably be the largest of its kind in the United States, given Bay Area funding commitments for data collection in concert with the 1990 Census.) The California State Department of Transportation (CalTrans) embarked on a major series of household travel surveys between 1976 and 1980 (Table 5.3). The largest component was a 6,900 household sample survey for the Los Angeles region, ___________________________ 12 Table 5.2 is in draft stage and will be updated as more information becomes available. The TRB Transportation Planning Applications Task Force is sponsoring a background survey on past and future metropolitan area household travel surveys. Data from this survey and other research will be incorporated into this Working Paper as the research progresses. -34- surveying approximately 0.17 percent of that region's households. The second largest component was a Bay Area survey of 2,200 households13. Smaller California metropolitan areas such as Sacramento, Fresno. Stockton, Santa Barbara, Salinas-Monterey, Modesto, and Santa Cruz, were surveyed for 700 sample households each, resulting in fairly substantial sampling rates from 0.24% in Sacramento to 0.97% in Santa Cruz. CalTrans is currently pursuing funding (1.5 million) to update the statewide household travel database with household surveys to be conducted Fall, 1990. B. The 1990 Survey: Target Size and Target Cost Various sample size designs at a regional and county level are summarized in Tables 5.4.1 through 5.4.4. The first two tables show five different sampling "options" using a "conservative" $100 per household and a "target" $70 per household cost. The $100 figure is being used by others in estimating costs for 1990 data collection efforts. The $70 figure is a realistic expectation for the MTC survey given a survey design similar to the 1981 Survey and the impact of inflation (4.5% per year). Given the 1981 Survey cost MTC $49 per household in 1981 dollars, this same survey instrument could probably be used in today's dollars at 65$ per household. Applying a $5.00 per household cushion should give us somewhere around $70 per household. Sampling Option "A" uses the 1981 Survey sampling rate of 0.36% (7,091 / 1970549). Sample sizes would range from 155 in Napa County to 1,909 in Santa Clara County for a regional total of 8,151 household samples. Option "A" would cost around $571 thousand, upwards to a conservative $815 thousand. The most overly-optimistic sampling rate, Option "B", would be the 1965 sampling rate of 2.21 percent of Bay Area households (30,686 / 1387000). This would yield around 50,000 sample households region- wide, peaking at 11,700 samples for Santa Clara County alone. This survey would cost in the neighborhood of $3.5 to $5.0 million. This sampling option should not be taken too seriously. Money does buy accuracy, but beyond a certain point the marginal increase in accuracy just isn't worth the marginal increase in cost. What would a million dollar survey mean in terms of data reliability and accuracy? Sampling Option "C" suggests that we could acquire anywhere from 10,000 to 14,300 ___________________________ 13 Unfortunately, due to survey and sample design discrepancies between the 1980 CalTrans Survey and the 1981 MTC Survey, the two surveys were never merged into one, comprehensive regional transportation database. -35- sample households for a million dollars. Using the target $70 per household cost would yield 3,300 sample households in Santa Clara County, 3,000 sample households in Alameda County, down to 300 sample households in Napa County (Table 5.4.2). The million dollar Option "C" survey would provide a county-level accuracy (trip rates at ñ5%, 95% confidence) for the five largest Bay Area counties. Sampling Option "D" would use the '81 Survey sampling rate for San Francisco of 1.17%. This is another too costly option, priced at $1.8 to $2.6 million dollars. This option would yield 26,000 region-wide sample households. Sampling Option "E" is the cheapest of the sampling plans, using the non-San Francisco sampling rate from the '81 Survey (0.21%). This survey would be priced at around $333 to $476 thousand dollars. Sample statistics would be adequate for the region and Santa Clara and Alameda Counties. Approximately 4,800 region-wide household samples would be collected under this plan. The most "do-able" of the five sampling options is Option "C", the million dollar survey. The 10,000 household samples would provide not only regionally-accurate travel data, but county-level-accuracy as well. Using a conservative $100 per sample, the survey would cost $1,000,000 (Table 5.4.3). Using a more Bay Area-Specific cost of $70 per household would cost $700,000 (Table 5.4.4). The sampling rate for this survey would be targeted at a 0.44% regional average, or 1 in 229 households. Three "comparative sampling frames" are shown in Tables 5.4.3 and 5.4-4. The eleven" sampling frame apportions the 10,000 household samples equally among the nine Bay Area Counties according to projected 1990 households. The "minimum" sampling frame increases the number of sampled households in the smaller Bay Area counties - Solano, Napa, and Marin - to a reasonable 500 sample minimum. In this "minimum" sampling frame, the sampling rates are reduced proportionately among the remaining six counties. The "target" sampling frame in Tables 5.4.3 and 5.4.4 are MTC staff attempts at rounding off the target county samples to appropriate minimum and maximum values. The objective is to have statistically accurate trip rates (ñ5 percent, 95% confidence) in the five largest Bay Area counties - Santa Clara, Alameda, San Francisco, Contra Costa, San Mateo - with at least 1,100 household samples in each of these counties. The north bay counties would be divided into two super-counties, each with 1,100 sample households. Marin and Sonoma counties combined would have 1,100 sample households as would the Napa-Solano combination. The resulting sampling rates would range from 0.38% of households sampled in Santa -36- Clara County (2,000 total samples), to 1.16% of households in Napa (500 total households). This "target" sampling frame appears to be a more logical blending of geographical equity in collecting samples balanced with the desired statistical accuracy at county- level. The target county samples are tested to determine the projected accuracy of county and regional trip rates in Table 5.4.5. Coefficients of variation (C.V.) and household samples (n) serve as the dependent variables. The formula now solves for percent error (E): E = sqrt (C.V.2 * Z2 / N) The 9,900 target regional sample for the 1990 Travel Survey will result in a regional trip rate accuracy at ñ1.7% for total trips, ñ1.9% for vehicle trips, and ñ5.0% for regional transit trips per household (all at a 95 percent confidence level). County level accuracy for total trips per household should range from ñ3.5% in Santa Clara to ñ7.6% in Sonoma. The five largest counties will each be within plus or minus five percent. The four north bay counties will individually be between ñ6.6 to ñ7.6 percent. As two-county groups, the north bay counties will be within the plus or minus five percent objective, as well. Vehicle trips per household accuracy will range from ñ3.7 percent in Santa Clara to ñ7.4% in San Francisco and Sonoma Counties. The San Francisco projected error is high due to the high coefficient of variation on vehicle trips made per San Francisco resident household, a "statistically rarer" event in San Francisco than in the other Bay Area counties. We should not expect to get very accurate county-level transit trips from a 9,900 sample survey. Transit trips per households will be measurable within ñ7.7 percent in San Francisco to ñ44.6 percent in Napa (though the absolute error in Napa should be substantially smaller than the larger Bay Area counties). As stated previously, the various transit operators' on-board surveys will serve as a necessary sub-regional component to augment the Bay Area's regional database. In conclusion, various metropolitan areas have embarked on household travel surveys at varying price levels. The Bay Area can be reasonably confident in establishing target cost estimates per household based on past survey efforts. An appropriate balance is needed between sub-regional statistical accuracy and geographic equity in the distribution of household samples. -37- Table 5.1 Comparative Metropolitan Household Travel Surveys Cost per Sample Household Estimates Region Year Samples Total Cost Cost/HH Bay Area 1981 7,091 $336,762 $47.49 Portland 1985/88 7,000 $140,000 $20.00 Washington, D.C. 1987/88 7,000 $350,000 $50.00 Atlanta 1980 4,900 $100,000 $20.41 Pittsburgh 1978/81 3,200 $184,992 $57.81 San Diego 1986 2,754 $150,000 $54.47 Philadelphia 1986,88 2,500 $260,000 $104.00 Dallas 1984 2,471 $320,000 $129.50 Minneapolis-St.Paul 1982 2,460 $180,500 $73.37 Baltimore 1988 2,000 $20,000 $10.00 Denver 1985 1,645 $60,000 $36.47 Houston 1984 1,500 $400,000 $266.67 Phoenix 1983/84 1,461 $35,000 $23.96 Washington, D.C. 1980 1,218 $30,450 $25.00 Vancouver, Wash 1985 1,031 $37,900 $36.76 Snohomish County, Wash. 1987 880 $45,000 $51.14 Pierce County, Wash. 1987 800 $45,000 $56-25 Bremerton, Wash. 1985 734 $74,500 $101.50 Cincinnati 1978 412 $30,000 $72.82 Bay Area Travel Surveys 1965 BATSC Home Interview 1965$ 30,686 $1,533,500 $ 49.97 1965 BATSC Home Interview 1981$ 30,686 $4,517,914 $147.23 1965 BATSC Home Interview 1989 $30,686 $6,174,483 $201.21 1981 MTC Telephone Survey 1965 $7,091 $114,306 $16.12 1981 MTC Telephone Survey 1981 $7,091 $336,762 $47.49 1981 MTC Telephone Survey 1989 $7,091 $460,241 $64.91 -38- Table 5.2 Comparative Metropolitan Household Travel Surveys Household Sampling Rates Household Total Sampling Region Year Samples Households Rate Bay Area 1965 30,686 1,387,000 2.21% Bay Area 1981 7,091 1,970,549 0.36% Bay Area 1990 9,900 2,264,100 0.44% Los Angeles 1967 30,800 3,061,600 1.01% Los Angeles 1976 6,947 4,027,894 0.17% San Diego 1966 San Diego 1977 1,399 609,544 0.23% San Diego 1986 2,500 775,000 0.32% Washington, DC 1955 26,100 450,680 5.79% Washington, DC 1968 14,918 547,224 2.73% Washington, DC 1980 1,218 Minneapolis- St.Paul 1958 18,634 366,511 5.08% Minneapolis- St.Paul 1970 3,871 433,460 0.89% Minneapolis- St.Paul 1982 2,430 721,000 0.34% Detroit 1968 40,000 Detroit 1980 2,500 1,650,000 0.15% Dallas- Ft. Worth 1964 Dallas- Ft. Worth 1984 2,471 1,000,000 0.25% Chicago 1956 Chicago 1970 Honolulu 1960 Honolulu 1981 Atlanta 1961 Atlanta 1972 Denver 1961 Denver 1971 Denver 1985 -39- Click HERE for graphic. -40- Click HERE for graphic. Click HERE for graphic. 41- Table 5.4.3 1990 Bay Area Household Travel Survey Target Sample Design - Conservative $100/Household Cost ABAG Pro '87 Projected 1990 Comparative Sampling Frames Households Even Minimum Target San Francisco 313,615 1,385 1,324 1,300 0.41% San Mateo 245,006 1,082 1,034 1,100 0.45% Santa Clara 530,172 2,342 2,238 2,000 0.38% Alameda 480,873 2,124 2,030 2,000 0.42% Contra Costa 298,415 1,318 1,259 1,300 0.44% Solano 109,344 483 500 600 0.55% Napa 43,009 190 500 500 1.16% Sonoma 145,906 644 616 600 0.41% Marin 97,768 432 500 500 0.51% Region 2,264,108 10,000 10,001 9,900 Sampling Rate 0.44% 0.44% 0.44% Cost per Interview $100 $100 $100 Total Cost $1,000,000 $1,000,100 $990,000 Table 5.4.4 1990 Bay Area Household Travel Survey Target Sample Design - Target $70/Household Cost ABAG Proj.'87 Projected 1990 Comparative Sampling Frames Households Even Minimum Target San Francisco 313,615 1,385 1,324 1,300 0.41% San Mateo 245,006 1,082 1,034 1,100 0.45% Santa Clara 530,172 2,342 2,238 2,000 0.38% Alameda 480,873 2,124 2,030 2,000 0.42% Contra Costa 298,415 1,318 1,259 1,300 0.44% Solano 109,344 483 500 600 0.55% Napa 43,009 190 500 500 1.16% Sonoma 145,906 644 616 600 0.41% Marin 97,768 432 500 500 0.51% Region 2,264,108 10,000 10,001 9,900 - sampling Rate 0.44% 0.44% 0.44% Cost per Interview $70 $70 $70 Total Cost $700,000 $700,070 $693,000 -42- Click HERE for graphic. -43- VI. Estimating Project Schedule This last section considers the simplest, yet most critical element of developing and conducting a major household travel survey - the schedule. Basically, the schedule for the 1981 MTC Travel Survey is translated into comparable dates for the 1990 Survey. This is assuming the project will be funded and conducted in the Spring of 1990. Given this schedule, the final survey consultant report on the 1990 Survey should be completed by June 30,1990. The 1980 Travel Survey and proposed 1990 Travel Survey schedule is summarized in Table 6.1. The first key milestone for the 1990 Survey will be the release of a Request for Proposals (RFP) by mid- November, 1989. The consultant should be selected and "on-board" by the end of December, 1989. Survey pre-tests, if any, will be conducted the first few weeks in February. The Survey will be conducted over a period of three months - March to May 1990. (Census Day will be April 1,, 1990. Data for the Journey-to-Work will be based on the last week in March, 1990.) The survey consultant responsibilities will include: developing the sampling frame (list of eligible telephone numbers), hiring and training of interviewers, conducting the telephone interviews, and coding and editing of responses (including geo-coding of origin- destination, residence and work locations). The RFP will fully outline the responsibilities of MTC staff and consultant staff. The spring, summer and fall of 1989 will be devoted to: - Comparative analysis of metropolitan area travel surveys: - Cost; - Sampling Rates; - Innovative Survey Designs; - Preparation of a Request for Proposal; - Travel Survey Advisory Panel The Travel Survey Advisory Panel will be composed of members chiefly from the academic community in the fields of city planning, survey research, and transportation planning/engineering. The mission of the Travel Survey Advisory Panel will be to advise MTC staff on the content and form of the 1990 Travel Survey sample design and survey design. Relevant questions will be debated and options outlined. The Advisory Panel will serve as a "brainstorming" center to gauge demand for travel survey information. -44- Next steps for the 1990 MTC Travel Survey includes the presentation of this sample design analysis to MTC staff and the Travel Survey Advisory Panel for review and comment. At these review sessions, issues and ideas should be discussed and debated: - Use of Panel Surveys in collecting longitudinal data; - Collection of weekday vs weekend trip records; - Use of Attitudinal as well as the standard Behavioral questions; - Reducing non-response by providing payment to participants ($5-$15); - Importance of non-household travel survey data collection: - Commercial /Trucks / Goods Movement Inventory; - Visitor Surveys; - Non-Bay Area residents commuting to Bay Area jobs surveys; - Cordon Line Surveys (at regional boundaries); - Transit Operator On-Board Surveys; - Traffic Counting Programs; - Highway Speed surveys; - Use of Data at appropriate geographic level in travel model development; - Use of computer-assisted geo-coding technologies; - Use of computer-assisted telephone list development techniques; - Problem with too many surveys being conducted during the same time as the 1990 Census; and - Problem of too long a survey and non-response / refuse to complete survey. The list of potential items to discuss and debate will continue to grow once we embark on our first round of "brainstorming" sessions. The basic sample design and survey design are well under way, and the 1990 Household Travel Survey should provide Bay Area transportation planners and decision-makers with the database for planning our transportation systems well into the twenty-first century. -45- Table 6.1 1981 Travel Survey Schedule Compared to proposed 1990 Travel Survey Schedule 1981 Survey 1990 Survey Work Task Due Date Due Date Draft RFP Reviewed 10/6/80 10/6/89 RFP Mailed 11/14/80 11/14/89 Proposal Closing Date 12/12/80 12/12/89 Review of Proposals 12/18/80 12/18/89 Consultant Selected 12/23/80 12/23/89 Survey Pre-Test 2/3-2/18/80 2/3-2/18/90 Survey Conducted March -May '81 March -May '90 Final Survey Report 6/30/81 6/30/90 VIII. Bibliography Household Travel Surveys A. General Reference Works Elizabeth S. Ampt, Anthony J. Richardson and Werner Brog. New Survey Methods in Transport. VNU Science Press, Utrecht, The Netherlands, 1985. Earl R. Babbie. Survey Research Methods. Wadsworth Publishing Co., Inc., Belmont, CA, 1973. William G. Cochrane. Sampling Techniques. John Wiley & Sons, Inc., New York, N.Y., Second Edition, 1963. Leslie Kish. Survey Sampling. John Wiley & Sons, Inc., New York, N.Y., 1965. Donald A. Krueckeberg and Arthur L. Silvers. Urban Planning Analysis: Methods and Models. John Wiley & Sons, Inc., New York, N.Y., 1974, Anthony J. Richardson and Arnim H. Meyburg. Survey Methodology in Transportation Planning. 1989. Peter R. Stopher and Arnim H. Meyburg. Survey Sampling and Multivariate Analysis for Social Scientists and Engineers. Lexington Books, Lexington, MA, 1979. Donald P. Warwick and Charles A. Lininger. The Sample Survey: Theory and Practice. McGraw-Hill Book Co., New York, N.Y., 1975. Thomas R. Willemain. Statistical Methods for Planners. MIT Press, Cambridge, MA, 1980. B. Travel Survey References John F. Anderson, Marsha A. Niebuhr, Ann Braden, and Stephen R. Alderson. Telephone Interviews: Cost-Effective Method for Accurate Travel Surveys. In Transportation Research Record 1097, TRB, National Research Council, Washington, D.C., 1986, pp. 4-6. Werner Brog and Elizabeth Ampt. State of the Art in the Collection of Travel Behavior Data. In Special Report 201, TRB, National Research Council, Washington, D.C., 1983, pp. 48-62. -47- Alan C. Clark and Celia Goldstucker. Mail-Out/Mail-Back Travel Survey in Houston, Texas. In Transportation Research Record 1097, @B, National Research Council, Washington, D.C., 1986, pp. 13-19. John F. DiRenzo, Robert A. Ferlis, and Philip I. Hazen. Sampling Procedures for Designing Household Travel Surveys for Statewide Transportation Planning. In Transportation Research Record ???, IRB, National Research Council, Washington, D.C., ???, pp. 37-43. Federal Highway Administration. Assessing the Need for New Survey Data for Travel Models. FHWA, Washington, D.C., February, 1989 (Draft). Robert A. Ferlis. Field Data Collection and Sampling Procedures for Measuring Regional Vehicle Classification and Occupancy. In Transportation Research Record 701, TRB, National Research Council, Washington, D.C., 1479, pp. 1-6. Household Survey Manual. Bureau of the Budget, Executive Office of the President. Washington, D.C., 1969. Hanna P.H. Kollo and Charles L. Purvis. Changes in Regional Travel Characteristics in the San Francisco Bay Area: 1960-1981. In Transportation Research Record 987, TRB, National Research Council, Washington, D.C., 1984, pp. 57-66. David L. Kurth. A Small Sample Mail-Out/Telephone Collection Travel Survey. In Transportation Research Record 1097, TRB, National Research Council, Washington, D.C., 1986, pp. 7-12. Susan Liss. Nationwide Personal Transportation Study: Experiences with Previous Surveys and Options for the Future. In Transportation Research Record 1097, TRB, National Research Council, Washington, D.C., 1986. pp. 3133. Eric J. Miller and David F. Crowley. A Panel Survey Approach to Measuring Transit Route Service Elasticity of Demand. Paper submitted to the Transportation Research Board Annual Meeting, January 1989. Elaine Murakami and Donald R. Pethick. Puget Sound Council of Governments Origin-Destination Travel Survey, 1985. In Transportation Research Record 1097, TRB, National Research Council, Washington, D.C., 1986. pp. 23-31. Bahar B. Norris and Gordon A. Shunk. Special-Purpose Travel Surveys. In Transportation Research Record 1097, TRB, National Research Council, Washington, D.C., 1986. pp. 20-22. -48- Bahar B. Norris and Gordon A. Shunk. The 1984 Home Interview Survey in the Dallas-Fort Worth Area: Changes in Travel Patterns, 1964-198-4. In Transportation Research Record 1134, TRB, National Research Council, Washington, D.C., 1987, pp. 1-9. William L. Nicholls H. Designing Telephone Surveys for the Greater Bay Area. Metropolitan Transportation Commission, Berkeley, CA, June 1977. Don L. Ochoa and George M. Ramsey. The 1976-1980 Statewide Travel Survey. California Department of Transportation, Sacramento, CA, December 1981. Rai Parvateneni, Peter Stopher and Cleveland Brown. Origin- Destination Travel Survey for Southeast Michigan. In Transportation Research Record 886, TRB, National Research Council, Washington, D.C., 1982, pp. 1-8. Emily B. Peterson and John R. Hamburg. Travel Surveys: Current Options. In Transportation Research Record 1097, TRB, National Research Council, Washington, D.C., 1986, pp. 1-3. Marilyn M. Reynolds, Sydwell M. Flynn, and David B. Reinke. 1981 San Francisco Bay Area Travel Survey. In Transportation Research Record 877, TRB, National Research Council, Washington, D.C., 1982, pp. 51-58. Michael E. Smith. Design of Small-Sample Home-Interview Travel Surveys. In Transportation Research Record 701, TRB, National Research Council, Washington, D.C, 1979, pp. 29-35. Michael E. Smith. Design of Small-Sample Home-Interview Travel Surveys. FHWA, Washington, D.C., August 1980. Peter R. Stopher and Ira M Sheskin. Toward Improved Collection of 24-H Travel Records. In Transportation Research Record 891, TRB, National Research Council, Washington, D.C., 1982, pp. 10-17. Peter R. Stopher. Small-Sample Home-Interview Travel Surveys: Application and Suggested Modifications. In Transportation Research Record 886, TRB, National Research Council, Washington, D.C., 1982, pp. 41-47. Peter R. Stopher. Data Needs and Data Collection - State of the Practice. In Special Report 201, TRB, National Research Council, Washington, D.C., 1983, pp. 63-71 -49- Peter R. Stopher. Sampling Strategies for Small Sample Surveys. Paper presented to the Planning Applications Conference, Orlando, Florida, February 1987. Roger B. Trent and Cecil R Pollard. Individual Responses to Rising Gasoline Prices: A Panel Approach. In Transportation Research Record 935, TRB, National Research Council, Washington, D.C, 1983, Urban Origin-Destination Surveys. Federal Highway Administration, U.S. Department of Transportation,'July 1975. Kenneth R. Yunker. Tests of the Temporal Stability of Travel Simulation Models in Southeastern Wisconsin. In Transportation Research Record 610, TRB, National Research Council, Washington, D.C., 1976, pp. 1-5. Yacov Zahavi and James M. Ryan. Stability of Travel Components over Time. In Transportation Research Record 750, TRB, National Research Council, Washington, D.C., 1980, pp. 19-26. -50-