How ride-hailing services influenced vehicle use and ownership across the Boston metropolitan region

Research about how ride-hailing influences private vehicle use has found apparently conflicting results. There is some evidence that ride-hailing may allow individuals to own or use private automobiles at a lower rate (Coogan et al. 2018; Ward et al. 2019; Bansal et al. 2020; Dong et al. 2021; Bilgin et al. 2023; Martin et al. 2024). But other studies find that ride-hailing operations have led to increased VMT within their service areas (Erhardt et al. 2019; Henao and Marshall 2019; Wu and MacKenzie 2021; Choi et al. 2022).

Research examining vehicle registrations typically identifies reductions in vehicle ownership following introduction of ride-hailing. Using a difference-in-difference model and state level aggregates of registration data, Ward et al. (2019) estimated that vehicle registrations across U.S. states were on average 3.1% lower than expected after introduction of Uber and Lyft. Bilgin et al. (2023) generated similar findings in Great Britain when looking at data aggregated to London boroughs (2.2% reduction) and rural local authority districts (1.1% reduction).

Survey-based research also often concludes that ride-hailing may enable lower personal vehicle ownership and use for some populations depending on individual characteristics. Bansal et al. (2020), using a 2017 survey of residents living in U.S. ride-hailing markets, found that ride-hailing customers are largely, but not solely, constituted of private vehicle users. When asked how they would complete their most regularly recurring trip if the mode typically utilized was not available, 66% of frequent ride-hailing users reported that they would have driven a car and only 14% said they would have taken transit. These results suggest that frequent ride-hailing users are most often replacing private vehicle trips with ride-hailing, while a minority was likely increasing their personal VMT by switching away from transit. Furthermore, about 10% of ride-hailing customers indicated that ride-hailing services allowed them to delay acquiring a personal vehicle. This last finding is echoed in the results of a more recent survey focusing on Philadelphia and Boston (Dong et al. 2021).

In another survey-based study, Coogan et al. (2018) found that 26% of respondents reported less need for a private car because of new mobility services including ride-hailing. However, this trend was not uniform across all demographics; urban commuters and single millennials were most likely to report less need for a private car. This difference in responses across groups suggests that home location and age may affect how ride-hailing shapes private vehicle use and ownership (Coogan et al. 2018).

Other studies come to the conclusion that ride-hailing operations are associated with increased VMT and congestion, but there is disagreement regarding the size of VMT growth. Some authors find large increases. For example, Erhardt et al. (2019) used traffic time data alongside TNC pick-up and drop-off data from San Francisco to model changes in congestion with and without ride-hailing services. Comparing a 2010 scenario without ride-hailing to the observed conditions in 2016 with ride-hailing available, they found a 13% increase in VMT, compared to an estimated increase of 7% if ride-hailing had not been available. They concluded that ride-hailing was the main driver of new congestion in the city. These findings are mirrored by an estimate from a study of Denver using ethnographic methods built upon rider surveys, and a researcher driving for Uber and Lyft. In this study, Henao and Marshall (2019) estimated that ride-hailing led to 83.5% more VMT than would have been generated without ride-hailing services by replacing non-automotive trips, inducing new trips, and traveling without users aboard.

But not all studies find such large associations. Choi et al. (2022) estimated that ride-hailing contributes to 0.6% average VMT growth per year using an empirical Bayes approach comparing Georgia Department of Transportation estimates for VMT in the Atlanta region post TNC introduction to a simulated counterfactual VMT estimate for the region where ride-hailing is not present. Other studies find no significant relationship between TNC operations and VMT at the state level (Ward et al. 2019).

Moreover, some survey-based studies find that VMT changes may also vary across individuals or regions. Highlighting the heterogeneous influence of ride-hailing on private vehicle use is research using data from the 2017 U.S. National Household Travel Survey. Using a propensity score matching method to analyze vehicle use and ownership cross-sectionally across levels of ridesourcing use (which included ride-hailing and taxi users), Wu and MacKenzie (2021) concluded that frequent ride-hailing users who had both driver’s licenses and access to a vehicle exhibited reduced VMT compared to occasional ride-hailing users, and that vehicle ownership declined among frequent ride-hailing customers more than occasional users. However, for individuals without a driver’s license or vehicle access, VMT increased across all levels of use. The authors also estimated that despite potential VMT reductions through frequent ride-hailing use among licensed drivers with vehicle access, at the time of the survey, ride-hailing services added 2.8 billion miles of annual VMT nationwide, or an additional 0.1%. As a whole, the study suggests that despite ride-hailing’s ability to lower vehicle use for some users and in some contexts, it inflates VMT in most situations.

Likewise, Martin et al. (2024) found evidence that ride-hailing’s relationship with VMT may differ between metropolitan areas. Through survey methods focusing on 3 large TNC markets, the authors estimate that TNC users in San Francisco and Los Angeles slightly increased their annual VMT on average (~ 240 additional miles), but in Washington D.C., the authors estimate a small reduction (~ 83 miles).

Another important phenomenon to understand ride-hailing’s influence on VMT is “deadheading,” a term referring to the portion of travel by ride-hailing vehicles without a paying passenger aboard, in between trips. Many studies conclude that deadheading is a significant contributor to ride-hailing VMT (Komanduri et al. 2018; Henao and Marshall 2019; Wu and MacKenzie 2021; Schaller 2021). Deadheading may help explain how research has identified evidence for both lower vehicle ownership and higher VMT related to ride-hailing services. Even if a ride-hailing user forwent vehicle ownership, or traveled fewer miles compared to driving a personal vehicle, ride-hailing vehicles themselves may produce more net VMT after accounting for deadheading.

Some observers have suggested that ride-hailing and transit could operate alongside each other as complements resulting in a more sustainable transport system if a sizeable number of urban travelers utilized ride-hailing to connect to or from transit, and that in turn meant they were more likely to use transit for part of their travel instead of using a personal vehicle for most of it. A simulation model using survey data from the San Francisco Bay Area found that nearly a third of morning commuters who drove alone could have reduced the generalized costs of their trips by using ride-hailing to connect to the Bay Area Rapid Transit network (Alemi and Rodier 2017). The authors estimated that if all of these commuters switched, then over 600,000 miles of total daily VMT could be avoided.

However, much of the existing research finds that ride-hailing substitutes for transit except in areas with poorly developed transit options (Tirachini 2020). And there is strong evidence of an association between ride-hailing operations and declines in transit ridership. Using longitudinal agency-level panel data from the National Transit Database, Graehler et al. (2019) found that across 22 large U.S. cities, heavy rail ridership declined 1.3% and bus ridership declined 1.7% each year after Uber launched in the area. A later study using data from San Francisco similarly found that bus ridership declined roughly 10% over the first five years of Uber operations (Erhardt et al. 2022). These findings also show variation depending on transit mode; for example, in the latter study, Uber’s introduction was not found to have a significant relationship with light rail ridership.

Second, numerous studies have found that ride-hailing trips rarely connect to transit, while nevertheless overlapping with transit services. Henao and Marshall (2019), using passenger surveys from the Denver area, found that just 5.5% of Uber and Lyft riders incorporated an additional mode into their trip. Likewise, survey research from Santiago, Chile found that among ride-hailing users, only 3.8% reported use of other forms of transportation to complement ride-hailing trips (Tirachini and del Río 2019). Research looking at Philadelphia and Boston found that 5 to 6 percent of surveyed ride-hailing users reported using their most recent ride-hailing trip to access transit (Dong et al. 2021). Taking into account that many ride-hailing trips originate or end near transit stops (Rayle et al. 2016), these findings imply that ride-hailing often substitutes for transit services.

Despite ride-hailing appearing to act as a substitute for transit, there is also evidence indicating complementary effects in some situations. A study using Uber market penetration data in conjunction with the National Transit Database found evidence that ride-hailing may be a complement to transit in U.S. metropolitan areas with weak transit ridership. Analyzing data from every metropolitan statistical area with public transit, the authors estimated that Uber’s introduction led to a 6% jump in ridership for transit agencies with ridership below the national median, and a 2.1% reduction in ridership for agencies above the median. The authors argue this result is likely because ride-hailing provides added service flexibility for riders within the context of infrequent transit services that access relatively few destinations. Altogether, the authors estimated a 5% increase in ridership for the average U.S. transit agency two years after the local launch of Uber (Hall et al. 2018).

Another study used general transit feed specification (GTFS) data and ride-hailing passenger surveys from Toronto to assess which ride-hailing trips had an alternative transit trip available. They found that about 30% of ride-hailing trips could have been completed by transit with a 15 min or less time penalty, and interpreted this finding to mean that these trips were substituting for transit. However, they also found that nearly 27% of the examined trips had a transit alternative that would have increased travel time by at least 30 min. In these cases, the authors state ride-hailing was likely complementing transit, and taken as a whole, they believe that the results show that ride-hailing acts as a substitute to transit in areas with robust transit services, and a complement where transit options are sparce (Young et al. 2020). These results reflect similar findings from New York City (Jin et al 2019) and Chengdu (Kong et al. 2020).

To calculate the average daily VMT of individual vehicles, we used the disaggregate Researcher Files of the MAVC as compiled by the Metropolitan Area Planning Council (MAPC). This is a comprehensive dataset of vehicle inspections merged with registration information for individual vehicles in Massachusetts from 2009 to 2014 (Metropolitan Area Planning Council 2016). The dataset contains vehicle identification numbers (VINs), inspection dates, and Massachusetts property tax parcel numbers (Parcel IDs) for each inspection record, and we used these identifiers to track vehicle mileage and registered location over time. For each inspection record, the dataset provides odometer readings (mileage), a calculated average mileage per day, emissions data, an adjusted miles-per-gallon rating assuming growing inefficiency over time and use, and other vehicle characteristics such as make, model, model-year, and manufacturer's suggested retail price (MSRP). Table 1 shows the summary statistics for the mileage per day records when excluding inspections missing odometer readings, and Fig. 1 shows a density distribution histogram of mileage per day.

Bild vergrößern

Note that the distribution of the mileage-per-day variable is positively skewed, with a long, right tail, and a vertical asymptote at 0 miles per day (Fig. 1). None of these features are surprising; some vehicles are heavily used while others are not used at all. As noted later, the distribution of the dependent variable required an appropriate functional form in regressions.

The MAVC includes 34 million individual records from across the state of Massachusetts, and 24.8 million include odometer readings from vehicle inspections. Of those, 22 million also include parcel information. We found no systematic pattern regarding records missing mileage or parcel information. Using the MAVC records including odometer and parcel data, we identified mileage records for each unique VIN-Parcel ID pair for vehicles located within the study area as defined by the extent of the Boston Region MPO corresponding to each year from 2009 to 2014. We then calculated the average daily VMT for a given year based on the average daily VMT between inspections and the number of days between inspections in the given calendar year, for each unique VIN-Parcel ID pair. 1,668,215 vehicles with at least 1 record for the five years of the study period were included in this study as were 1,873,996 unique VIN-Parcel ID pairs. In total, the dataset used in the individual vehicle level analysis had 6,309,506 observations due to multiple observations of the same vehicles in the same locations over the observation period. Note that if a vehicle changed registered residence, then it may have multiple observations for a single year.

We used parcel information to match vehicles to Census tracts, which allowed the inclusion of spatial variables including Uber availability, transit access density, and population demographics. For the second part of our analysis using a panel of Census tracts, we aggregated vehicle data to calculate VMT per vehicle-day, vehicle turnover, and vehicles per population at the Census tract level.

While there is no variable in the MAVC indicating whether a vehicle is used to provide ride-hailing services, all private vehicles in Massachusetts, including ride-hailing vehicles, must undergo annual inspections by law. Thus, Uber vehicles registered in Massachusetts are included in our data, but it is unknown which are Uber vehicles. Furthermore, some unknown fraction of Uber vehicles operating in Massachusetts over this period were likely registered outside the state.

Boston is one of Uber’s oldest markets, beginning operations within the city on October 27th, 2011 (Uber 2016). In their earliest years of operation, Uber maintained a relatively small service area limited to the cities of Boston and Cambridge. A few trips in these early years extended out of this zone, but Uber avoided expanding their service area until July 2013 when they began operating in suburbs south of the urban core (Uber 2013). In January 2014, they expanded again to the north (Uber 2014a). Finally, by October 2014, Uber operations had expanded west of Boston’s inner suburbs towards central Massachusetts (Uber 2014b).

Using the publicly available history on Uber’s expansion in the Boston area, we created a dummy exposure variable indicating whether individual vehicles reported in the MAVC were registered to an address both within the Boston Region MPO and within a Census tract with Uber availability for each year of the study. We used April 1st of each year as the critical date determining Uber availability because April 1st is the reference day for U.S. Census data, and we used Census tract data in our model as control variables. If a Census tract fell within Uber’s service area on April 1st of a specified study year, then it was designated as such using the dummy variable. Thus, for 2010 and 2011, none of the vehicles were designated as having Uber availability. For 2012 and 2013, only vehicles registered in Boston or Cambridge were designated as exposed to Uber availability because Uber did not expand out of the core urban area until later in 2013. In 2014, all Census tracts in the MPO (and all vehicles registered therein) were designated as having Uber availability since Uber services had expanded to areas outside the MPO by that time. While we are not certain that every area in the MPO had Uber availability by April 1st, 2014, most if not all did. We therefore treat Uber availability as a binary exposure variable. If spatially fine-grained ridership or wait-time data were available, it might be possible to specify low- and high-exposure Census tracts, but such data are not available for this period of time.

Table 2 summarizes operational definitions of Uber availability over time, and a map highlighting Boston, Cambridge, and MPO boundaries is shown in Fig. 2 (City of Boston 2020; City of Cambridge 2020; Massachusetts Department of Transportation 2019). Note that other ride-hailing TNCs besides Uber, namely Lyft, began operating in the Boston area near the end of the study period.

Bild vergrößern

Because of the timeframe of the vehicle census data from 2009 to 2014, the current study’s focus is constrained to the early years of ride-hailing when ridership was less developed. Data on annual Uber trip totals is unavailable for this time period, as is data on individual Uber trips, but according to Uber, it completed 28 million trips in the Boston area between October 2011 and October 2015 (DeCosta-Klipa 2015). In the coming years, Uber ridership would expand. In 2019, over 53 million ride-hailing trips started in the cities of Boston and Cambridge alone (Massachusetts Department of Public Works 2019).

To measure transit accessibility, we calculated the density of transit stops and stations for each Census tract in the Boston Region MPO for each year of the study. We used date-specific spatial data on MBTA facilities from the agency’s GTFS archive (Massachusetts Bay Transportation Authority 2021), using the record closest to April 1st of each year to create shapefiles representing the location of MBTA facilities for each study year. We examined the data, and removed non-access nodes (such as maintenance facilities), ferry docks, airport shuttle bus stops, and duplicate entries. We reviewed the accuracy of the data by comparing it to a written history of MBTA service changes (Belcher 2021), and then updated the data to reflect which of three types of services operated at each stop or station: traditional bus, rapid transit (including heavy rail, light rail, and bus rapid transit), and commuter rail. Finally, we spatially joined the remaining transit node points onto a polygon shapefile of Census tracts (U.S. Census Bureau 2010), and calculated the density of transit points in each Census tract. While most of the variation in the transit density scores can be seen between Census tracts, there is some variation over time within Census tracts as well since some bus stop locations changed, and other stops or stations closed or opened. The transit density scores were then merged onto the MAVC data by Census tract so that each vehicle received a transit density score for the Census tract in which it was located for each study year. Transit density scores for 2014 are shown in Fig. 3 at the regional scale; Fig. 4 shows detail for Boston and Cambridge on commuter and metro station density.

Bild vergrößern

Bild vergrößern

We do not observe the demographic characteristics of vehicle owners (these are not available in the MAVC data), so for demographic controls we used year-specific, U.S. Census and ACS 5-Year demographic characteristics of the Census tract in which the vehicles were registered as a proxy for the average population characteristics of the vehicle owners. We created a population density variable based on the square mileage of each Census tract, as well as 5-year estimates of median household income, the proportion of owner-occupied housing units, the percent of residents younger than 25, the percent of residents older than 65, and the percent of residents reported in the Census-defined race and ethnic categories as white alone for every year of the study (U.S. Census Bureau 2010, 2019a, 2019b, 2019c). These data were then merged with the MAVC vehicle level data according to the Census tract in which the vehicle was registered.

In the first stage of our modeling, we performed negative binomial fixed-effects panel regressions, with vehicles as units of analysis. The negative binomial model form was chosen over Poisson or tobit specifications to account for the distribution of mileage data, with significant overdispersion (see Fig. 2). We combined the vehicle identification number (VIN) with the parcel identification code (Parcel ID) in order to treat vehicles as the same units of analysis over time only when owned in the same location, partly as a control for changes in ownership and partly as a control for changes in spatial variation. The panel was structured on a year-by-year basis. Average daily VMT was the dependent variable, calculated as a weighted average of the average daily VMT between odometer readings based on the share of days falling in a given calendar year. In the negative binomial fixed-effects panel regression, the conditional mean, μ_it, of mileage y_it for an individual vehicle i in year t, was modeled as follows (Cameron and Trivedi 2013; Hausman et al. 1984):where y_it is the mileage variable for vehicle i in year t; x_it is a vector of explanatory variables for vehicle i in year t; c_i is the constant term for vehicle i; β is a vector of coefficients to be estimated (including a constant term); and α_i represents the time-invariant, vehicle-specific “fixed effect.”

Descriptive statistics for the dependent and independent variables, and the number of observations (VIN-Parcel ID Pairs) for each year of the study period are shown in Table 3 below.

The main independent variable in the regression models was the Uber availability or “exposure” variable, defined as equal to one for all Census tracts in years when Uber services were made available. We included a dummy variable designating whether the vehicle was registered in the Boston and Cambridge urban core, along with vehicle-specific variables (adjusted miles per gallon and vehicle age), and demographic control variables measured at the Census tract level. To further understand the mediating effects of transit service availability, we also included rapid transit stop density, commuter rail stop density, and bus stop density measured at the Census tract level. We also interacted these transit variables with the Uber availability variable. All variables were measured on a year to year basis to enable a fixed effects panel method (see Table 3 for means by year).

Using the panel regression method allows us to consider each vehicle in the MAVC data as an individual observation while treating Uber availability as a categorical exposure variable to analyze how the introduction of ride-hailing was associated with vehicle use and ownership by people living in areas with and without ride-hailing service. In the final year (2014), all vehicles and tracts in the study area had Uber availability, so each vehicle in earlier years acts as its own pre-exposure control for later years, akin to a before-and-after approach; while the vehicles that lacked Uber availability in early years act as non-exposure controls for those who gained Uber availability early on, akin to a cross-sectional approach within observation years.

This method treats Uber exposure as uniform across Census tracts once included in Uber’s service area. In fact, some Census tracts undoubtedly had higher-quality exposure to Uber than others. Unfortunately, additional data to enable a more finely detailed exposure measure, such as wait times for Uber rides by Census tract, are not available.

The MAVC has a limited number of time-variant vehicle characteristics including adjusted MPG, which assumes degradation of fuel efficiency over time based on a vehicle’s model year, and vehicle age. We included the adjusted MPG variable, given that fuel economy relates to vehicle use, with owners purchasing more fuel-efficient vehicles when they intended to drive longer distances. We also included vehicle age as a proxy measure for the income of vehicle owners, with the expectation that newer vehicles may be driven more as household income is positively associated with driving distance.

Table 4 shows the preferred model, chosen based on the log-likelihood ratio along with Akaike information criterion (AIC) and Bayesian information criterion (BIC) test scores. Because of the large number of observations, we defined statistical significance at the 99% level. We tested our fixed effects model specification verses a random effects alternative using the Hausman specification test (Hausman 1978), which confirmed endogeneity of vehicle specific effects. In Table 4, we report exponentiated coefficients (incidence risk ratios, or IRRs) to ease interpretation. IRRs can be interpreted as a multiple of average daily VMT (from its mean observed value) associated with a unit change in the corresponding independent variable. For example, an IRR of 1.1 can be interpreted as a 10% increase in average daily VMT for a unit change in the independent variable, while an IRR of 0.9 would denote a 10% decrease.

To examine neighborhood level changes, we also estimated three models at the Census tract level. We first aggregated the vehicle level MAVC data by Census tract, locating individual vehicles within tracts based on the parcel location. Three dependent variables were calculated at the Census tract aggregate level: [1] vehicle turnover, [2] VMT per vehicle-day, and [3] vehicles per population.

Vehicle turnover (also called churn) is intended as a proxy for neighborhood change. It is a measure of the total number of vehicles that moved into or out of the Census tract relative to the total number of vehicles in the Census tract. We identified vehicles moving in or out by the continuity of the unique VIN-Parcel IDs in each Census tract. For example, if a unique VIN-Parcel ID record began in 2012, then that vehicle was classified as moving into the Census tract in 2012. If a unique VIN-Parcel ID record ended in 2013, then that vehicle was classified as having moved out of the Census tract in 2014. A limitation of this methodology is that vehicles moving within a Census tract are overcounted.

VMT per vehicle-day is a measure of vehicle use at the Census tract level. This variable represents the total VMT by the vehicles registered within a Census tract normalized by the number of vehicles in that Census tract, and the number of days a vehicle resided in the Census tract (vehicle-days). This is an aggregate level average daily VMT measure that accounts for vehicles moving in and out of the Census tract and weights them accordingly.

Finally, vehicles per population is defined as the total number of vehicles registered in a Census tract for the year divided by the population of the Census tract. We found that vehicle counts in some areas of the Boston metropolitan area are disproportionately low in the processed MAVC dataset. This error is likely due to missing parcel records or geocoding issues that are systematic geospatially but likely not temporally. This issue is mitigated to some degree by the use of fixed-effects panel regression methods which account for geospatial errors (errors over the panel variable) better than random-effects models because fixed-effects regressions consider each geospatial observation (Census tract ID) as a unique categorical variable.

For each of these three dependent variables, we performed ordinary least squares (OLS), fixed-effects panel regressions with the Census tract identification code as the panel variable and the observation year as the time variable. The general form of the three types of OLS regression was:where DV is the dependent variable (vehicle turnover, VMT per vehicle-day, or vehicles per population), \({\alpha }_{i}\) is the fixed effect term for Census tract \(i\), \({\beta }_{0}\) is the constant term, \({\beta }_{1}, {\beta }_{2}, \dots {\beta }_{k}\) are the regression coefficients for the independent variables, \({X}_{it1}, {X}_{it2},\dots {X}_{itk}\) are the independent variables for Census tract i for year t, and \({u}_{it}\) is the unobserved error term for Census tract i for year t.                                                                                                                                        

Descriptive statistics for all the variables utilized in model selection for each year of the study period are shown in Table 5. The three models are shown in Table 6. We specified the models in a similar way as the individual vehicle level analysis, which interacted Uber availability, Boston and Cambridge, and transit stop densities variables. The same set of Census tract characteristics were used.

These models complement our individual vehicle level analyses in three ways. First, they allow us to examine whether neighborhood change, as proxied by vehicle turnover, is correlated with VMT changes. Second, they provide a supplemental analysis of how ride-hailing relates to VMT when aggregated to the Census tract level. Third, and perhaps most usefully, they allow us to examine how ride-hailing relates to the rate of vehicle ownership, which cannot be examined at the individual vehicle level because we observe vehicles, and not individuals or households, in the disaggregate dataset.

Perhaps the most notable result of this study is a much weaker-in-magnitude relationship between ride-hailing and VMT than many previous studies. One study relying on city-wide trend data, surveys, and TNC service patterns estimated that ride-hailing nearly doubled the growth of congestion and VMT in San Francisco (Erhardt et al. 2019). Another study concluded that ride-hailing contributed to 83.5% more VMT in Denver than would have been generated without ride-hailing services (Henao and Marshall 2019). In contrast, we found only a 0.6% increase from the average daily VMT when looking at changes occurring in individual vehicles in the Boston metropolitan area. Our findings are in line with research showing a much smaller relationship between ride-hailing and VMT, such as Choi et al. (2022), who similarly approximated a 0.6% annual growth in VMT related to ride-hailing in Atlanta, and Wu and MacKenzie (2021) who estimated a 1% increase in national VMT related to ride-hailing.

These differences may be for a number of reasons. First, the methods being used in studies showing large VMT increases rely on simulations and estimates based on assumptions regarding unobserved travel including deadheading, aggregate city-level time series data, and/or survey data, rather than observed individual vehicle use. These methodological shortcomings mean that they can only imperfectly control for factors such as secular changes in auto use over time unrelated to Uber, and they cannot directly control for other factors that may be correlated with changes in auto use. Our data and methods account directly for deadheading and other difficult to estimate TNC related travel because our dependent variable measures all vehicles, including Uber and any other TNC vehicles, and all mileage within the study area, at least to the extent that vehicles registered within the study area are those used within the area.

Second, our data are from Boston between 2011 and 2014, during the early years of Uber operations. The situations in other cities in different time periods might, of course, be somewhat different. Boston is one of the oldest cities in the United States, and its development patterns are typically denser and less auto-oriented than most American cities. Perhaps, therefore, our findings might be thought to be particularly relevant in similar metropolitan areas such as New York, Philadelphia, Baltimore, and Chicago. Our use of Boston in this study was due to the quality and availability of the vehicle census data. As far as we are aware, no other US metropolitan area has similar data.

Even so, Boston was one of Uber’s earliest markets, and as of 2017, use of ride-hailing in the Boston metropolitan area was one of the highest in the country (Schaller 2018). Given the early rollout, and eventual popularity of Uber in the Boston region by 2017, we would expect an analogous relationship between ride-hailing and VMT in Boston compared to many other US cities.

Finally, these differences may be due to data limitations with the current study. The MAVC only covers the years 2009–2014, when ride-hailing’s influence on VMT and vehicle ownership may have been less pronounced. It is possible that ride-hailing’s impact on VMT may have increased over a longer time period. Moreover, we looked at changes occurring in individual vehicles registered in the MPO, and Boston has commuters from outside the region. The fraction of out-of-region, in-state commuters was 19.6% in 2011 (Kuttner et al. 2017). If a large percentage of Uber drivers commuted from outside the Boston MPO, then these data may be insufficient.

Our results suggest other factors influence vehicle use more than the availability of ride-hailing. For example, in the vehicle level model, for every additional mile of Adjusted MPG, vehicles on average traveled 2.5% more than the mean VMT holding other variables constant. Similarly, we found that for every year of Vehicle Age, vehicles traveled 7.15% less. Our dummy variables for the study year also provide mixed evidence that vehicle use in general may have changed over time. In the individual level analysis, each subsequent year was significantly related to an increase in VMT across the study area with the average daily VMT in 2014 being 20.5% higher compared to the mean VMT in 2010. However, this finding was contradicted in the Census tract level analysis which indicate small decreases in VMT per vehicle day for each subsequent year. Likewise, the Census/ACS demographic controls of median income, homeownership rate, and percentage white population were also found to be highly significant in our individual level model, though coefficient magnitudes were small.

We did not find any indications of the density of transit stops and stations mediating Uber availability’s relationship with VMT. In interaction with Uber, all but one of the transit variables were insignificant in both the individual vehicle level and Census tract level models regressing over VMT. The one variable that was significant, commuter rail station density in interaction with Uber, is significant only in the individual vehicle level model, and only in Boston and Cambridge where Uber on its own was found insignificant with average daily VMT.

In the individual vehicle level model Uber availability was associated with a slightly weaker relationship between the density of commuter rail stations and reductions in VMT within Boston and Cambridge. In these cities, commuter rail density was associated with a 1% decrease in VMT from the mean for every stop per Census tract square mile, but the combination of commuter rail and Uber availability was associated with a smaller net decrease of about 0.8%. These results suggest that in the areas of Boston and Cambridge served by commuter rail, auto use declined less than it might have without Uber availability.

Some previous research studies have suggested that ride-hailing substitutes for transit in many contexts (Graehler et al. 2019; Henao and Marshall 2019; Tirachini and del Río 2019; Tirachini 2020; Erhardt et al. 2022), but may complement transit where transit accessibility is not high. While we do not find any direct evidence for a substitutional relationship, the individual vehicle level model does suggest that ride-hailing could weaken transit’s ability to reduce private vehicle use in more central areas. Given that many commuter rail users also consider auto use, it is not unreasonable to hypothesize that transit may be less appealing to potential auto users if ride-hailing is available. It may be that people living in areas with high rates of transit use or who are themselves frequent transit riders may be more likely to increase their auto use due to TNCs becoming more available, particularly if they were not auto owners prior to the introduction of TNCs, and if TNCs are competitive on price or travel time with transit. In contrast, people who did formally own a car or who were about to buy one may find that TNC availability made it possible to forestall auto ownership, use their autos less, or even shed their (expensive) autos.

We find no evidence that neighborhood change affected VMT variation after Uber’s introduction in either the individual vehicle level or Census tract level analyses. The aggregate models at the Census tract level indicate that outside of Boston and Cambridge, Uber was associated with 0.013 fewer vehicles turning over per existing vehicle, and 0.022 fewer vehicles per population. Reduced turnover implies reduced neighborhood change or displacement, and reductions in vehicles per population suggest that the individuals living in these Census tracts gave up some vehicle ownership or that any incoming population did not bring vehicles with them. Either interpretation suggests that gentrification and/or displacement did not play a role in whether and how Uber affected VMT in Boston over the study period.

Moreover, while income was found to be a significant predictor of VMT at the 99% confidence level in the individual vehicle level negative binomial regression analysis, the magnitude of its relationship was very small (IRR of 1.000192, which is reported in Table 4 as 1.000). The same is true in the Census-tract level OLS regression analysis using vehicles per population as the dependent variable, in which median Census tract level household income produced a negative, but very small-in-magnitude coefficient (-0.0000031, which is reported in Table 6 as -0.000). These results suggest that while income may affect vehicle use and ownership, any change in Census tract median income likely did not affect VMT or vehicle ownership to a meaningful degree over this period of time.

Previous research has produced evidence for a relationship between ride-hailing and both increased VMT (Erhardt et al. 2019; Henao and Marshall 2019; Choi et al. 2022) and reduced auto dependency (Coogan et al. 2018; Ward et al. 2019; Bansal et al. 2020; Dong et al. 2021; Bilgin et al. 2023), but evidence for either outcome typically comes from separate studies. Here we find evidence for both within the MAVC data, but only for areas outside of Boston and Cambridge. In the Census tract level models Uber availability was associated with an increase in VMT (0.344 more miles per vehicle day), as well as 0.022 fewer vehicles per population, and less vehicle turnover. The reduction in vehicles per population could signal lower dependence on private vehicles, and when paired with the small increase in VMT, suggests that ride-hailing related reductions to auto dependency could occur alongside VMT growth in certain areas such as inner suburbs and metropolitan peripheries.

One noteworthy study that identifies both trends in the same data comes from Wu and MacKenzie (2021). However, their VMT and ownership estimates were based on 2017 National Household Travel Survey data. Similarly, Martin et al. (2024) found both potentials through TNC rider and driver surveys deployed in San Francisco, Los Angeles, and Washington DC. Interestingly, Wu and MacKenzie (2021) found that access to a driver’s license and household vehicle may influence ride-hailing’s relationship with VMT, and Martin et al. (2024) found evidence that VMT changes may vary between regions. Wu and MacKenzie (2021) contend that ride-hailing propensity to lower VMT exists only among those with driver’s licenses and access to a household vehicle who frequently use ride-hailing. This conclusion suggests that certain population factors shape ride-hailing’s potential to reduce VMT, and that prospective reductions in VMT exist within groups who can meaningfully reduce the number and length of their private vehicle trips by utilizing ride-hailing. We identify a comparable potential resting within potential users living in non-central parts of a metropolitan area—areas where private vehicle trips are more likely to be the norm.

A consistent finding across our study is that ride-hailing on its own relates to VMT outside of Boston and Cambridge but not within. This distinction between whether a vehicle is registered in Boston or Cambridge and elsewhere suggests there is a geographical element to ride-hailing’s relationship with passenger vehicle use that may affect ride-hailing’s influence over VMT within metropolitan markets. We offer two possible reasons why this may be the case. First, automobile users who live in highly urban areas may not be replacing many personal vehicle trips with ride-hailing trips. In urban cores, where auto dependency is already lower than suburban and exurban areas, private vehicle trips that can be replaced by ride-hailing trips may be less prevalent. For example, a resident in a central city who owns a car may use non-automotive modes for most regularly occurring trips, but still keep a car to leave the city on weekends, or for large shopping trips. Previous research from Boston suggests that ride-hailing draws users from multiple modes including driving, transit, biking, and walking (Dong et al. 2021). It may be that for urban car owners, ride-hailing replaces some of their non-automotive trips, but not many of their personal vehicle trips. Second, ride-hailing drivers may be more likely to be outside of urban cores; if so, VMT reductions from suburban car owners replacing car trips with ride-hailing could be offset by the increased VMT produced by vehicles used by ride-hailing drivers.

More broadly, the disparity in results within Boston and Cambridge and elsewhere in the MPO indicates that ride-hailing, transit, and private vehicle use all interrelate differently within urban cores compared to inner suburban or peripheral areas. Our results are consistent with the hypothesis that the potential to reduce vehicle use through complementary ride-hailing and transit services may exist primarily outside central areas—perhaps because ride-hailing mainly replaces the last-mile travel related to park-and-ride transit services in peripheral areas, and these services are less common in urban cores.