Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
P· _ · - rolect ~InalngE OVERVIEW The chapter presents the main findings of the study. It includes the results of each of the models that were estimated using the dataset described in Chapter 2. We first present the results of the analysis of the cross-sectional dataset of major metropolitan areas, and then describe the results of the time-series analysis for New York City. CROSS-SECTIONAL REGRESSION MODELS In estimating the cross-sectional models we first developed a general theoretical mode! for the each of the modes and market segments of interest, postulating (in general terms) what we thought should be the most important determinants of the dependent variables. This process can be outlined as follows: The increases in single occupant vehicle use we have observed in recent decades are likely to represent a combination of new commuters, disproportionately choosing the private vehicle mode, and existing commuters who have changed" modes. In either case, these changes can be hypothesized to be a function of the following basic factors: · The private vehicle has become an option for more people, or more household members, through increased vehicle availability; · The solo driving option has become relatively more attractive in terms of cost (either because vehicle ownership and operating costs have declined or because incomes have increased, making the costs easier to afford); and · Driving alone has become relatively more attractive in terms of other characteristics (either because of direct changes in transportation alternatives such as highway capacity or transit fare and service levels, or because reorientation of jobs andJor housing have made ndeshanng or pubic transit less feasible or convenient)
A general theoretical mode! for explaining the changes in mode share and private vehicle occupancy can therefore be stated as follows: AOccupancypv = F(^household composition,Avehicle availability,Ajobs/housing location, ArelaVve vehiclecosts) AMode Share = F(A vehicle availability, /\relative vehicle costs, Jobs I housing location, Alevels of see/ice) Given these theoretical models, we then determined how we might best measure these determinants, and then selected the most closely corresponding explanatory variables from among the available data. Table ~ ~ provides an example of how we selected variables for testing in the mode! estimation from among the data available in each general category. Table 11. Example of Variable Selection : . ~ ~ ~ . - ~ A. .~ ,~ ~ ~ .- .~--. i- - .~ -- i- ~ -, ~ ~ .~ ~ ~ ~ ~ i- i~ ,~, ,. ~ -, . :: : I:- I: Ca~g:o~::: I: ~ : I:::; :Examp:le Variables T s I:: I: Act ................ ............ ~. ~....... ~ ~ . ~...... ~ ~ .~. ~. ~.~. ~.~ it. ~ ~ ~.~.~ ~.~.~.~ ~.~.~ ~ Measures of housing and workplace · °/O suburban households "dispersion" · °/O suburban population · °/O suburban employment · Total central city employment · Age of housing stock Vehicle availability | · vehicles per household · vehicles per capita · vehicles per worker · vehicles per worker not working at home · TO zero vehicle households · °/O 2+ vehicles households Relative vehicle costs · Income · Gasoline prices · State gasoline taxes · Parking costs Levels of service · Average commute time · °/O of commutes >45 minutes · Average transit fare · Peak transit vehicles · Peak transit vehicle miles · Transit vehicle per worker · Transit vehicle miles per worker · HOV lane miles Other characteristics · Household size · Workers per household 36
This table is not a comprehensive listing of the data assembled, but is intended to represent the types of variables tested in the mode} estimation procedure. A full list of the variables collected and the corresponding data is provided in Appendix A. Before estimating the models, we produced correlation matrices of the explanatory variables to identify any potential multicollineanty problems. If a high degree of correlation was shown to exist among variables in our theoretical model, we then reformulated the mode! to include similar but less correlated variables. This mode] was estimated and then refined further by substituting other explanatory variables (based on the preliminary results). The final models were selected based on goodness of fit, the reasonableness and statistical significance of the estimated coefficients, and how well the mode! confirmed the theoretical "priors" described above. The final models are presented in the following sections. Static mociels for ~ 990 The static models relate observed mode shares and occupancies in 1990 for each of the metro areas to various socioeconomic and transportation related data also for that year. Separate models were estimated for the suburb to central city and within central city commute markets, the two markets historically of primary importance to public transit systems in the US. Single Occupant Vehicle Share Models Table 12 presents the mode] of 1990 single-occupant vehicle shares for the suburb to central city commute. The mode] contains four parameters: downtown parking cost, suburban private vehicles available per suburban household, transit vehicle hours per worker not working at home, and central city employment. Downtown parking cost represents the "prevailing" cost of parking in the core of the urban area (rather than an estimate of the average parking cost actually paid by all commuters). It is meant to convey the "price" of parking. The coefficient has a negative sign and is statistically significant at better than the i% level, indicating that parking prices are an important determinant of SOV shares, with higher prices causing lower shares. 37
Table 12.1990 SOV Share Model for Suburb to Central City Commute ~ . . ~ ~ . ~ . ~ ~ I:: ?: ~ . ~. ~ . ~: ~ :: ~ Variable: : ~ ~ ~ Parame:ter: tiStatistic: Elasticity ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~.~ ~ ~.~ ~.~ ~ ~.~ ~ ~ -it ~ .~ ~ ~ ~.~.~ ~.~ ~ .. ~i. ~ ~ ~ ~ ~ ~ . ~ ~ ~ ~ ~ ~ .~ i. i. ~ . i ~ ~ . ~ ~ of. ~ Downtown parking cost -0.0127 -3.290 r Private vehicles available per household in 0.1157 1.497 suburbs Transit vehicle hours per worker not working -0.0575 -3.893 at home Central city employed labor force 0.0538 2.069 N Adjusted R2 I -0.09 +0.27 -0.12 +0.03 33 0.53 Source: Charles River Associates, 1997. The transit vehicle hours per worker not working at home variable is a measure of the level of available transit service. It generally describes both the available capacity per commuter, as well as the total size or scope of the transit system. The coefficient is negative and significant at better than the i% level, indicating that the availability of good transit service has a strongly negative effect on SOV mode shares. The private vehicles per household in suburbs variable is a measure of vehicle availability that is specific to the suburbs. The coefficient is positive, as expected, but statistically significant at only the 15% level. A possible interpretation of this result is that vehicle availability in the suburbs is uniformly very high, particularly compared to the central city, and as such is not a large determinant of the SOV mode share for the suburb to central city commute. Finally, the central city labor force variable is statistically significant and indicates a positive relationship between employment and SOV shares for this commute market. Given the discussion of the importance of central city density to transit in Chapter i, we refight expect lower SOV shares to be associated with higher concentrations of central city employment. But the variable in this mode} measures the absolute number of employed central city residents, rather than the density," so the interpretation of this result is not as clear. It could reflect the definitional issues associated with central cities described in Chapter 2; for example, total! central city employment in Los Angeles is quite comparable to that of New York, and in Dallas it is comparable to that of Philadelphia, despite much different patterns of density and mode share. On the other hand, larger central city Here and in similar subsequent tables, the elasticity estimate is computed at the mean values of the influencing variables. t However, a satisfactory model could not be fit with an employment density variable, perhaps because of its very high correlation with the downtown parking cost variable. 38
labor force could be could be associated with central cities that are larger in area. To the extent that public transit will be less convenient for the distribution of trips in a larger downtown area (because of transfers, coverage, and additional travel time), we might expect total employment to be positively related to SOV share. The mode} has a correlation coefficient adjusted for degrees of freedom (the so- called R2) of 0.53, indicating that 53% of the variation in SOV mode shares is "explained by" these four parameters. Table 13 presents the 1990 SOV share mode! for the central city to central city commuter market. This mode! contains a similar set of variables: vehicle availability, transit service level, and parking cost. Again the transit service level variable is negative and significant. This is certainly to be expected, particularly in the central city market. l:nterestingly, the parking cost variable is significant at only the 14% level. while the vehicle availability variable is significant at the 4% level, the reverse of the relative importance of these two vanables in the suburb to central city mode} in Table 12. For commuting within a central city where vehicle ownership is typically lower, it is logical that vehicle availability would be a more important driver of SOV share. Likewise, we know that SOV shares are generally lower for the within central city commute compared to the suburb to central city commute; those using SOV for the within city market may represent a residue of travelers that are less likely to pay for parking. Alternatively, those who live and work in the central city and commute by private vehicle may have higher incomes, and are thus less sensitive to parking prices. Table 13. 1990 SOV Share Model for Central City to Central City Commute Veritable ~ ~ ~= ~ ~ ~ ~ ~ ~ Parameter ~ t-Stotistic Private vehicles per worker not working at home 0.3215 2.195 Transit vehicle hours per worker not working at -0.0586 -2.965 home Downtown parking cost -0.0053 -1.510 Percent central city housing built prewar -0.0026 -4.103 N Adjusted R2 . El-asticitr +0.68 -0.16 -0.05 ~ 11 -0.14 32 0.77 Source: Charles River Associates, 1997. instead of the central city employment variable in the suburban model, this mode! has a variable meant to capture the density or deve;topment pattern of the central city area. This variable is the percentage of housing units in the central city built prewar, a proxy for the general typology of the city (clense, old, eastern, vs. sparse, 39
new southwestern, for example). This is the most significant variable in the model; its negative sign indicates that density is a significant restraint on SOV shares within a city. This is to be expected, since this type of development pattern will mean more congestion, less available parking, and higher parking costs, as well as easier use of non-private vehicle modes such as transit and walking. It should be pointed out that this variable is not highly co-linear with either the transit service variable or the parking cost variable. However, it may explain the relatively weak coefficient on parking cost, since it does measure some of the same effects. The mode! has an adjusted R2 of 0.77, indicating that 77% of the variation in SOV mode shares are explained by these four parameters, a higher result than that for the suburban model. Private Vehicle Occupancy Models Table 14 presents the 1990 mean occupancy mode! for the suburb to central city commute market. This mode! is somewhat different from the share models as it attempts to explain the average number of persons per vehicle rather than the relative number of total trips by private vehicles in this commute market. The table shows that the variables that best explain mean occupancy include measures of household size and HOV lane availability. The workers not working at home per household in suburbs variable measures the average number of people in each household who need to commute to work, and is specific to the area where workers in this market segment live. We should expect that with more workers in each household, there will be more opportunities for ridesharing, causing a positive impact on vehicle occupancies. This is confirmed by the positive and significant coefficient found for this vanable. Data for this variable were not available for the city of Providence, and there are therefore only 32 observations used in this model. 40
Table 14. 1990 Private Vehicle Occupancy Model for Suburb to Central City Commute Variable |~Parameter | t-§tatistic | Elasticity Workers not working at home per household in 0.0735 2.564 suburbs l Private vehicles available per household in -0.0849 -2.077 suburbs Total metropolitan area HOVlane miles | 0.0006 | 3.605 N Adjusted R2 +0.07 -0.14 +0.01 33 0.34 Source: Charles River Associates, 1997. Like the SOV share models, this mode] also contains a measure of vehicle availability. The fewer vehicles available in each household, the more there is a need for ridesharing. This is confirmed by the negative and significant coefficient (fewer vehicles means higher occupancy, or conversely, more vehicles means a lower average occupancy). The mode] also contains a vanable measuring the total HOV lane miles for each city in 1990. Given that HOV lanes often provide less congestion and its consequent faster travel times (particularly for the suburb to central city flow of interest here), we should expect (other things equal) that more HOV capacity would encourage higher occupancies. The mode] indeed shows a positive and significant relationship in this case. Table 15 presents the 1990 average occupancy mode! for the within central city commute market. This mode} also includes a vehicle availability variable, but one that instead measures vehicles available to central city residents. This is the most significant variable in the model, and as expected we again observe a negative relationship. The household size variable is again positive, but not significant at a high level of confidence. While more household members provide more opportunities for ridesharing, the greater number of travel alternatives in the central city (relative to the suburbs) make this measure less important. In addition, the closer proximity of residences and jobs makes the opportunities to carpoo} with people outside of the household greater relative to the suburbs. This notion is further confirmed by the positive and significant coefficient on the variable characterizing employment density. This variable measures the percentage of the total CMSA/MSA labor force resident in the central city, where jobs are more closely concentrated, and carpooling is consequently easier. 41
Table 15. 1990 Private Vehicle Occupancy Model for Central City to Central City Commute _ _ .. me. . . . . . . : Parameter ~ t-Statistic ~ ~ . ~ ~ ~ ~ ~ aft. ~ ~ ~ ~ ~ ~ ~ ~ if-. ,- .~. ~ ~ ~ ~ ~ ~ ~ ~ If ~ ~ ~.~ ~ ~ ~ ~-~ ~ ~ .~. ~ ~ . ~. ~ ~ If. ~ .~ i. ~ - ~. If. ~ ~ ~ .~ ~ .~ ~ ~ Central city percent of MSA labor force 0.0403 1.898 Private vehicles available per household in -0.0823 -7.169 central city Total metropolitan area HOV lane miles 0.0002 2.273 Average household size for central city 0.0183 1.201 households . N Adjusted R E~ ~ . ~ . ~ ~..~..~ ~ ~ ~ ~ ~ .... ~lastlclty ~ ... ~ ~ ~ ~.~ ~..~ .~ i.. . ~ ~ . . ~ +0.01 -0.09 +0.00 +0.04 33 0.65 Source: Charles River Associates, 1997. The mode! also includes the HOV lane miles variable which is again positive and significant. It is important to remember that many HOV facilities are at least in part within the limits of the central city area. The mode! has a relatively high adjusted R2 at 0.65. Public Transit Share Models Table 16 presents the results of the 1990 public transit share mode! for the suburb to central city commute. This mode! shows that about sixty percent of the variation in transit shares across metro areas in this market is now explained by vehicle availability, parking costs, and the density or development pattern of the central city area. The availability of vehicles for commuters has a strong negative influence on transit share. An alternative interpretation would be that higher transit shares are evidenced in places with relatively few vehicles available. Parking cost is positive and significant, indicating a positive relationship between parking prices and transit market shares. 42
Table 16. Public Transit Share Model for Suburb to Central City Commute ; Variable - ~Parameter t-Stat~stic : ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ I. . ~ ~ . ~ Private vehicles available per worker not working -0.2738 -3.143 at home Downtown parking cost 0.0053 1.959 Percentage of central city housing units built 0.0015 3.1 40 prewar N Adjusted R2 ~ . ~ ~ 55 . ~! ~ ~ ~ 55~ . ~ I: Elasticity I. ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ -4.91 +0.40 +0.66 32 0.59 Source: Charles River Associates, 1997. Finally, the mode] includes a variable representing the percentage of central city housing units built prewar, the same variable used in the SOV share mode] (Table 12), and also meant to act as a proxy for the general density and development pattern of the city.T The variable is positive and significant, indicating as expected that the older, denser, cities tend to have higher transit market shares. This may reflect their historically more developed transit infrastructure, but also most certainly reflects the fact that these kinds of development patterns are more conducive to transit use. Table 17 presents the transit share mode! for the within central city market. This mode! is quite similar, containing variables for vehicle availability, parking cost, and employment density. Vehicle availability is again the most important influence, with more vehicles causing lower transit shares. Parking cost is also significant at better than the i% level. The percent of the MSA labor force resident in the central city is also positive and significant at a high level of confidence, associating higher employment densities with higher transit shares. This is a logical result in that as jobs are more concentrated, fixed-route transit services can reach them more effectively. in this model, these three vanables explain 83% of the variation in transit shares. Data for this variable were not available for the city of Providence, and there are therefore only 32 observations used in this model. 43
Table 17. Public Transit Share Model for Central City to Central City Commute _ ~ ~ ~ ~ ~ \/drialile ~ ~ ~ ~ -Parameter ~ ~t-Statistic ~ ~.-.~ ~ ~ ~ ~.~ ~ ~ . .. ~ ~-~ ~ ~. ~ ~.~ ~ . ~.~ ~.~ it.. ~.~ ~ alp. ~ ~ ,... ~ ~ ~ ~ ~ ~:~ ~ ~ it,, if.. ~ ~ ~ ~ ~ ~ ~ .. i. ~ ~.~.~ .~.. .~ i. Private vehicles available per household in -0.3594 -9.722 central city Downtown parking cost 0.0097 4.123 Central city percent of MSA labor force 0.1576 2.527 _ Adjusted R Elasticity -3.13 +0.37 +0.31 33 0.83 Source: Charles River Associates, 1997. 980-1 990 Trend Models From explanatory models based solely on the differences between metropolitan areas in 1990, we next turn to examine the levels of change in those cities between 1980 and 1990. Because of the problems with changes in geography described in Chapter 2, these models have been estimated only for the total commute market, rather than separately for the suburb to central city and within central city markets. Table IS presents the cross-sectional regression mode] for the change in single- occupant vehicle share between 1980 and 1990. The results show that over three- quarters of the change in SOV share can be explained by four variables. The most important variable is the change in private vehicles per capita. This variable is positive and significant at less than the ~ % level, indicating that continued increases in vehicle ownership during the 19SOs have contributed to the growth in SOV share. 44
Table 18. 1980-1990 Change in SOV Share Model . ~ ~ ~ ~ a: ~ ~ ~.~.~.~. ~ ~ ~.~ ~ ~ .- ~ ~ ~ ~ ~ I- ~ ~ I. . ~ ~.~ I. ~ ~ ~ ~ ~ ~ ~ ~ ~:~ ~ .~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~.~.~ - ~.~ .~ ~ ~.~ . ~ ~ ~ ~ ~.~ ~ ~ ~.~ ~.~ .. Id--- ~ I: ~ ~ Variable- ~ ~ ~ ~ ~ ~ ~ ~ I-: ~ ~ ~ :~Pa:rameter~ ~ ~ ~ ~ ~ ~ ~ ~ ~ .~ ~ ~: . ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ . ~ ~ ~.~ ~.~ ~ ~ ~ ~ ~ ~. ~.~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ . ~ ~ At- -I ~ ~ ~ ~ ~ .~ ~ ~ ~ ~ ~ ~ . ~ ..~ ~ ~ ~ ~ :~ ..~ Change in private vehicles per capita 0.5157 Change in workers not working at home per household -0.3916 Coefficient of variation in departure time to work (1990) 0.1208 . Change in percentage of MSA employment in suburbs 0.0067 . N Adjusted R2 . . .. .. . . .. . . . .. . . . . . .. 9.051 -4.200 1.923 1 .835 33 0.76 Source: Charles River Associates, 1997. The second variable is the change in workers not working at home per household. This variable is significant at better than the 1% level and has a negative sign. Household sizes continued to decline between 1980-1990, and this result shows that increases in SOV shares are strongly related to shrinking numbers of workers per household. Given that much of the increase in SOV shares in the 19SOs came at the expense of carpooling rather than transit or other modes, this is a logical result: fewer workers per household means fewer opportunities to carpool. This result is especially important given that, as described in Chapter :l, the NETS indicates that people are most likely to carpoo] with other household members. The chance in the percentage of MSA emnIovment in the suburbs also is positive ~ ~ r . . . i. . . . . and significant. As jobs have been dispersed into the suburbs, the opportunities for carpooling and the feasibility of using transit to reach these jobs have correspondingly declined. SOV use has consequently increased as a result, and this is confirmed strongly by the model. Finally, the mode] contains a variable representing the variation in departure times among the metro areas for 1990. Data on departure time to work has only begun to be collected with the 1990 Census, so it was not available for 1980. The mode] does indicate, however, that those cities with more variation in departure times have seen larger increases in SOV share. This is consistent with the fact that transit services are generally geared toward serving peak period travel, and are thus at a disadvantage in serving workers leaving at off-peak hours. Table 19 presents the results of the mode! examining changes in private vehicle occupancy between 1980 and 1990. This mode] is quite similar to the static occupancy models for 1990, in that it explains changes in occupancies as a function of changes in vehicle availability, household size, and residenceJworkplace dispersion. in this case, change in percent suburban population is used to measure the dispersion aspect. The results are very 45
consistent with those described earlier: that as population has moved out into the suburbs and households have become smaller but with more available vehicles for each worker, the opportunities and incentives for riciesharing have diminished. As a result, these factors have caused occupancies to decline. Table 19.1980-1990 Change in Private Vehicle Occupancy Model 1 ~:~: ~ i': ~ ~ ~ ~ ~ ~ :~' ~ ~ ~ ~ In' ~ ~ ~ ~.~ ''aim i: ~:~ ~ - ~ ~ ~ ~ ~ ~ ~- :~ ~ ~ ~ -I ~ ~.~-~ ~ hi.: ~ hi: ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~:.~:. ~ ~ ~ ~ ~ ~ ~ ~ ~.~ ~ ~ : i: ~ :~ ~ ~ ~ : :: ~ I- ~ ~ - Variable ~: :: ~ Parameter ~ -0.1378 ~ ~ ~ ~ ~0.3263 . Chat ~ ~ ~ ~ ~^ -0.1060 N 1,, Adjusted R2 Source: Charles River Associates, 1997. ~ ~ ~ ~ ~ ~ .~ ~ . .. ~ . ~ ~ ~t-iStatistic~ ~ ~ ~ i. .~ ~ ~ ~ ~ ~ ~ ~ i- ~ .~ ~ ~ ~ -3.447 11 5.068 -2.305 33 0.52 Table 20 presents the mode] of the change in public transit share between 1980 ant! 1990. The results are very consistent with those obtained for the SOV share model. About 70% of the change in transit shares is explained by five variables. The most important variable is the change in households without any private vehicles. The positive sign implies that increases in the number of zero vehicle households may have a positive effect on transit share, or alternatively, where these households have declined, so has the market share of public transit. Table 20.1980-1990 Change in Public Transit Share Model i ~ ~ ~ ~ I~ of; ~ ~ ~ ~ ~ ~ . ~ ~ ~ ~ ~ . _ ~ .. ~ ~ ~ ~ . ~ I ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ A ~ ~ ~ ~ ~ ~ ~ ~ . ~ ~.~ . ~ I . . . ~ . ~ ~ ~ ~ ~ ~ - ~ : :: :: :: ~ : :: ~\I^ - em A: ~ :~:: : : :~ : ::: ::: i::: : ::: :: ~ : : ~ : i:: : ~ ~ ::.: : ~:~::~ ~ i: : :: ::: -::: ~^e.^e~t.^e~ : .~ ~: ~ ~V~-~ ~ J ~ ~.~ ~ ~. ~ ~ ~ ~ ~ ~ . -ball Change in zero vehicle households 0.7455 Change in percent suburb-to-suburb commute flow -0.3647 Change in household size 2.4649 Percentage of commutes over 45 minutes (1990) 1.2368 Change in HOV lane miles -0.0090 N Adjusted R2 Statistic 5.726 -2.689 4.285 3.660 -1.211 33 0.71 Source: Charles River Associates, 1997. On the national level, Pisarski shows that both the percentage and absolute number of households without vehicles changed little between 1980 and 1990, but this result masks some important variation among metropolitan areas. In places that experienced a large influx of new immigrants (who are unlikely to own vehicles) such as Phoenix, Sacramento, Los Angeles, and Houston, there were corresponding increases in vehicle-less households. 46
The mode! also implies that changes in household size are positively related to changes in transit share. Again, since household sizes have been declining, it is most appropn ate to interpret this to mean that falling household sizes are an important factor in the fall in transit shares we have observed during the penod. The percentage of commutes over 45 minutes for 1990 has a positive sign and is also significant at better than the too level. Of this result. First, higher commute times are an indicator of congestion, and we expect higher congestion to increase transit shares, other things being equal. An alternative interpretation is that commuters may be more likely to use transit for long trips (if available), because of the stress and time associated with driving, particularly under congested conditions. There are two possible interpretations The change in percent suburb to suburb commute flow measures the general dispersion of jobs and housing that has continued to occur during the 1980s. This dispersion has made it more difficult for transit to serve many commute trips, and the mode! shows that it has had a strongly negative influence on transit markets shares. Finally, the increase in HOV lane miles has had a small but negative impact on transit share, according to the model. The coefficient is not statistically significant at the 20% level, but its negative sign suggests that the addition of highway capacity in the form of HOV lanes has provided an alternative to transit for SOV commuters. Table 21 shows the results of the estimated mode! of changes in vehicle miles of travel (VMT) in the 1980-1990 period. This mode} explains increases in VMT as a function of the increase in population and employment, the dispersion of residences and workplaces, and increases in income. As we might expect given that this is a mode} of total VMT (for all purposes and at all times of day), population and employment are the most important variables. Total VMT was used as this was the only statistic available, but the mode! nevertheless shows that rising total employment has been an important factor in the increase in VMT. 47
Table 21.1980-1990 Change in Total VMT Model ~ ~ if. ~ . ~ ~ ~ . ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ i. ~ ~ ~ ... ~ ~ ~ . .... ~ .. ~ ~. ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~.~ ~ ~ ~ . ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ .~ ~ ~. ~ ~ ~ ~ . ~ ~ ~ .. ~ . ~ ~ ~ ~ If- ~ ~ ~ ~ ~ ~ -: ~ ~ ~-~ ~ ~ ~ : ~ - ~ ~ i: ~ ~\/ar~iab~le~-~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ -- ~ ~ ~ ~ - Parameter ~ ~ ~ i. ~ ~ ~ ~ ~ ~ ~ ~ ~ . . . ~ ~ .~. ~ ~ ~ ~ ~ ~ ~ . ~ ~ ~ ~ ~ ~ ~ ~ . . ~ ~ ~ ~ ~ ~ ~ . Change in total employment 1.1476 l Change in Federal Aid Urbanized Area population 0.1147 ~ Change in percent suburb-to-suburb commute flow 0.8067 Change in median household income | 0.6873 | Housing density (1990) -0.0007 N Adjusted R2 : t-Statistic ~ ~ . 3.084 2.485 1.957 2.268 -1.837 33 0.57 Source: Charles River Associates, 1997. The mode] also contains a variable for the change in "Federal Aid Urbanized area population," the geographic unit for which the FHWA computes VMT data. This variable serves two functions: it measures the effect of increasing population on VMT, and also controls for the changes in the geographic definition of the areas that occurred between 1980 and 1990. This variable is positive and significant, as expected, but it is not highly correlated with the employment variable, and therefore does not introduce bias into the model. The dispersion of workplace and residential locations is measured through a variable defined as the change in percent suburb-to-suburb commute flow. This is the percentage of total work trips that are classified as suburb-to-suburb by the Census, and it has increased markedly between 1980 and 1990. The mode] shows that this reorientation of commuting to a less concentrated patterns has had a strong positive effect on private vehicle travel. Finally, rising household incomes have also had a positive effect on VMT; analysts have long observed that the propensity to travel increases with income, and this result is confirmed by the model. TIME SERIES REGRESSION MODELS FOR NEW YORK CITY This section contains the results of the time-series analysis for New York City. These models were developed using a similar procedure to that described already for the cross-sectional models. We first developed a theoretical model, and tested various combinations of the explanatory variables to achieve the best results. However, in this case, the process was as much limited by the availability of time series data as by theoretical constraints. 48
As described in Chapter 2, the models were estimated on cordon counts of trips entering the core area during the AM peak period. The core area represents the New York City central business district (CBD) which is the portion of Manhattan south of 60th St. Trips entering the core by all modes are counted at a total of forty-two locations or "cordons", including all of streets intersecting 60th street and all of the bridge or tunnel crossings (whether for common camer modes or private vehicle) connecting to the CBD. Counts are usually on one representative fall business day in each year. Trips are reported separately by mode, allowing us to analyze data specifically for private vehicle trips, and the counts are made hourly. Models were estimated for travel in the AM peak period of 7:00 to 10:00 AM for both total private vehicle trips and the private vehicle share of all trips. Table 22 presents the time-series mode} for total private vehicle trips entering the core dunng the AM peak period. The mode} is estimated for the period 1978- 1994. It has been estimated as a log-Iog model, and as such the coefficients can be interpreted directly as elasticities. The mode} contains three parameters: CBD employment, average bridge/tunnel tolls, and average transit fare. The table shows that all of the variables are statistically significant at the l% level. The employment elasticity is about +0.S, indicating that a 10% increase in employment will produce a 8% increase in private vehicle trips. This is a reasonable magnitude given that for travel during the peak period, when the great majority of trips are work trips, we would expect the elasticity to be close to 1.0. That is, since most AM peak period trips to the core area are work trips, ant! most people who work in the core area travel to work during the AM peak period, changes in core area employment should be nearly directly proportional to changes in core area peak period trips. Table 22. Time Series Model of Private Vehicle Trips Entering Core Area in AM Peak Period for New York City . . ~ ~: .~ ~ .~ ~ - . . I. ~ . ~ . I-. ~ ~ . ~ ~. ~ all: ~ ~ ~ ~ ~ ~ Variables ~ ~: ~I: ~ Parameter t-Statisti~c - ~ ~ ~ ~ ~ ~ ~ ~ .~ ~ ~ ~ ~ ~ ~ At. ~ ~ ~ . . ~. ~ ~ ~ . ~ ~ . ~ Log of core area employment | 0.8073 4; Log of average bridge/tunnel toll -0.2335 Log of average MTA transit fare 0.2903 . N Adjusted R2 3.063 -3.213 3.241 17 0.73 Source: Charles River Associates, 1997. The coefficient for the toll variable has a negative sign, as expected, but a low elasticity of around -0.2. This result is generally consistent with recent modeling results obtained by Hirschman et ait. for several New York City toll facilities. 49
The coefficient on the transit fare variable is positive as expected, indicating that higher transit fares mean more private vehicle trips (or conversely, that lower fares mean fewer trips). Because it represents a cross-elasticity, the most appropriate value of this variable is not clear. Evidence of cross-elasticities is relatively scarce, and values can vary widely with local conditions, particularly the relative shares of transit and private vehicle. The value of 0.29 is higher that the range of observed by Mayworm, et. al, in their somewhat dated survey of transit elasticities, but they do point out that cross-elasticities are much higher for trips to the core area, particularly where there is a greater propensity to use transit.2 They report results of auto cross elasticities with respect to rapid rail fares in the range of +0.06 to +0.13 for all trips, and cross-elasticities with respect to bus fares in the range of +0.03 to +0.~S for inns to the central business district. On the other - -r . , ~ hand, the recent study of New York facilities referenced above reports auto demand elasticities with respect to transit fares as high as +0.23. if we believe the value of 0.29 is indeed to high, one possible explanation might be a potential bias in the coefficients because of omitted vanable~s) that there is some important variable that has not been included in the estimation. The adjusted R2 statistic shows that over a quarter of the variation in trips remains unexplained and certainly there are potentially important variables that could not be included because of a lack of available time-genes data. Examples of such variables might include downtown parking costs and the level of congestion on facilities providing access to the core area. Table 23 presents the time-series mode! of the private vehicle market share. As stated earlier, the private vehicle share represents the share of all trips entering the urban core area (Manhattan south of 60th St.) during the AM peak period. The mode] is again estimated on data for the 1978 to 1994 penod. Table 23. Time Series Model of Private Vehicle Share Entering Core Area in AM Peak Period for New York City If: Variable ~ ~ ~; Parameter Log of average bridge/tunnel toll -0.1866 Log of gasoline price -0.1631 Log of average MTA transit fare 0.2819 . Adjusted R2 ~ - U ~ t-Statistic~ :: :: ::: : :: ::: ::::: :: -3.814 -3.605 4.143 17 0.51 Source: Charles River Associates, 1997 . 50
While there are only three parameters in the mode} due to the lack of available data, the table nevertheless shows the interesting result that over 50% of the variation in private vehicle shares can be explained entirely by pnces. Both the average toll and gasoline price vanables are negative and significant at better than the I% level. The toll elasticity is -0.19, and the gasoline price elasticity is slightly lower, as we would expect, at -0. 16. While there are only seventeen years of data in the model, gasoline prices vaned considerably dunng this penod, and the coefficient is statistically significant at better than the ~ % level. The cross elasticity with respect to transit fare is similar to that observed in the total trips mode] in Table 22. This result is higher than other observed values, and it may reflect the omission of some important variable in the estimation such as parking costs. The result is not the effect of senal correlation, however, as a test for senal correlation was performed, but was rejected at better than the 99% confidence level. ~ Hirschman, I., McKnight, C., et. al., "Bridge and Tunnel Toll Elasticities in New York: Some Recent Evidence." Transportation, Vol. 22, No. 2, May 1995. 2 Mayworm, P., Lago, A., et. al., Patronage Impacts of Changes in Transit Fares and Services. Prepared for the Urban Mass Trans Transportation Administration, Washington, DC (1980~. 51
