The impact of Uber and Lyft on vehicle ownership, fuel economy, and transit across U.S. cities.

We estimate the effects of transportation network companies (TNCs) Uber and Lyft on vehicle ownership, fleet average fuel economy, and transit use in U.S. urban areas using a set of difference-in-difference propensity score-weighted regression models that exploit staggered market entry across the U.S. from 2011 to 2017. We find evidence that TNC entry into urban areas causes an average 0.7% increase in vehicle registrations with significant heterogeneity in these effects across urban areas: TNC entry produces larger vehicle ownership increases in urban areas with higher initial ownership (car-dependent cities) and in urban areas with lower population growth (where TNC-induced vehicle adoption outpaces population growth). We also find no statistically significant average effect of TNC entry on fuel economy or transit use but find evidence of heterogeneity in these effects across urban areas, including larger transit ridership reductions after TNC entry in areas with higher income and more childless households.

Introduction

The past decade saw the advent and growth of ridesourcing, a travel mode in which a passenger uses a mobile device to request a ride with a nearby driver at a transaction price determined in real time by a transportation network company (TNC), such as Uber or Lyft. During this period, the daily number of Americans traveling in for-hire vehicles more than doubled (Federal Highway Administration, 2019; Conway, 2018). Uber, which launched its basic level of service (UberX), in 2012, had entered 224 U.S. urban areas (approximately half of those identified by the U.S. Census) by the end of 2017, as Figure 1 and Table S2 show. In absolute terms, the 0.5% of total passenger trips now served by TNCs is still relatively small, but in urban areas, the effects can be substantial—by 2016, TNCs accounted for 15% of all intra-San Francisco vehicle trips on an average weekday (San Francisco County Transportation Authority, 2017).

Figure 1

Uber and Lyft entry over time by urban area

Uber entry date (x axis) is compared to Lyft entry date (y axis) for each of the 224 urban areas with TNC access by the end of 2017. Urban areas are depicted by bubbles proportional to population size.

While a growing body of research finds that TNCs lead to significant changes in transportation-related outcomes such as new vehicle purchases in China (Gong et al., 2017), traffic congestion (Li et al., 2016), motor vehicle homicide (Greenwood and Wattal, 2015; Lagos et al., 2019; Barrios et al., 2019), and passenger safety (Chaudhry et al., 2018), as well as other outcomes as varied as AirBnB demand (Zhang et al., 2018), entrepreneurial activity (Burtch et al., 2018), and urban crime (Weber, 2019), few studies have yet reached conclusions about effects of TNCs on energy and climate change-relevant outcomes like petroleum consumption and greenhouse gas (GHG) emissions. Both outcomes are particularly important for the transportation sector, which has long consumed the bulk of petroleum products produced and recently became a greater GHG emitter than any other sector in the U.S (Davis et al., 2019). These outcomes are a function of total vehicle travel and fuel efficiency and are related to the size of the vehicle fleet and availability of transportation alternatives. Higher vehicle ownership rates have historically been associated with both higher transportation energy consumption per capita (Newman and Kenworthy, 2006) and greater emissions from vehicle production: there were nearly 17 million new vehicles sold in the U.S. in 2018, each generating nearly 8 metric tons of GHGs from the manufacturing process (Davis et al., 2019) (Figure S1).

The introduction of TNC services could disrupt the relationship between vehicle ownership and transportation energy consumption and emissions not only in the U.S. but also globally, including in countries like China and India, where vehicle registration numbers are climbing more than 10% each year (Davis et al., 2019). There are reasons to believe, a priori, that TNC market entry could increase vehicle ownership, as aspiring TNC drivers purchase new vehicles, or decrease vehicle ownership, as passengers with access to TNC services shed or delay the purchase of a private vehicle. Similarly, TNC market entry could potentially increase average vehicle fuel economy, as TNC drivers anticipating high-mileage invest in relatively efficient vehicles, or decrease vehicle fuel economy, if TNC drivers are attracted to larger—and less fuel-efficient—vehicles to accommodate more riders and associated price premiums (through “Uber XL” and “Lyft XL” services). TNC entry could also increase transit ridership, by improving access, or decrease transit ridership, if acting as a replacement. And the relative impact of each of these (and other) effects may vary across urban areas. As TNCs continue to diffuse across global cities, understanding the impact on total vehicles, energy, and emissions becomes increasingly important.

The few published articles that do examine TNC impacts on energy and the environment are constrained to survey methods or limited geographies, and, when taken together, reach inferences and conclusions that initially may appear inconsistent with one another. Previous analyses of TNC effects on vehicle ownership, for example, have found either a decrease in vehicle ownership (Ward et al., 2019) or that “ridesourcing probably did not influence car ownership behavior” (Rayle et al., 2016). Considered with appropriate context and constraints, these findings are not necessarily inconsistent (the decline in vehicle ownership was modeled at the U.S. state level, while the no-effect finding was based on a survey specific to San Francisco, CA), but their differences point to the need for a coherent framework within which to interpret such findings and their associated implications.

As described above, plausible underlying narratives can explain either an increase or a decrease in vehicle ownership after TNC market entry (or a net-zero effect, as a result of the simultaneous presence of pressures in both directions): new economic opportunity afforded by TNC entry could motivate would-be TNC vehicle drivers to increase vehicle registrations; whereas, private vehicle-free mobility newly afforded to would-be passengers could prompt vehicle shedding and/or new vehicle purchase delays. Several working papers use surveys to understand these vehicle ownership relationships: a positive correlation for TNC drivers (Berliner and Tal, n.d.) and a negative correlation for TNC passengers (Clewlow and Mishra, 2017; Alemi et al., 2018a, 2018b; Feigon and Murphy, 2016). Other working papers use empirical panel data to observe or model a net effect, finding both increases (Gong et al., 2017; Schaller Consulting, 2018) or decreases (Hampshire et al., 2017) in vehicle ownership.

Effects of TNCs on fleet fuel economy are also ambiguous a priori, as drivers motivated by lower operating costs could plausibly migrate to newer, more efficient vehicles (Uber and Lyft require newer vehicles in some cities (Uber, 2020)), just as drivers motivated by the potential for large-party-trip price premiums may shift to larger, less efficient vehicles. Only one peer-reviewed article to date reports a TNC-fuel economy relationship, finding that ridesourcing vehicles are more efficient compared to non-ridesourcing vehicles (Wenzel et al., 2019). Two working papers agree (Gong et al., 2017; Kitchel and Wise, 2017); while another suggests that TNC vehicle fuel economies are lower than those of taxis (Wagner, 2017).

The effects of TNCs on public transit ridership are similarly equivocal, with several survey-based studies finding ridesourcing can either replace transit (Rayle et al., 2016) or complement it (Yan et al., 2018; Zhang and Zhang, 2018; Hall et al., 2018) as an effective extension of an otherwise fixed network (several working papers also suggest replacement (Alemi et al., 2018b; Norris and Xiong, 2019; Manville et al., 2018) or complementarity (Feigon and Murphy, 2016; Schweiterman and Livingston, 2018)). Importantly, two of these previous studies report heterogeneous effects—a more positive transit correlation is found either (1) in bigger cities as well as cities with smaller transit agencies (Hall et al., 2018) or (2) as a function of high population density and households with fewer vehicles (Zhang and Zhang, 2018)—and a working paper finds commuter rail usage increases while city bus usage declines, on average, with variation in effects across cities (Babar and Burtch, 2017).

We propose that heterogeneity of the effects of TNC entry in different types of cities may be responsible for some of the apparent discrepancies in reported effects, since average effects can look different depending on what subset of locations are included in the study scope and how they are aggregated. To quantify and systematically organize the potentially fundamental changes to personal travel across U.S. urban areas and associated energy and environmental outcomes, we estimate effects of Uber and Lyft entry on vehicle ownership, fleet average fuel economy, and transit ridership outcomes using a set of difference-in-difference propensity score-weighted regression models that exploit staggered TNC market entry into urban areas across the U.S. from 2010 to 2017. We combine annual individual vehicle registration data from Polk/IHS Markit with annual ZIP code-level sociodemographic data from the U.S. Census Bureau and aggregate to the urban area to estimate effects. We find that estimated vehicle ownership, fuel economy, and transit ridership effects of TNC entry vary across urban areas. Accordingly, we characterize this heterogeneity across urban areas and estimate that (1) TNC entry increases vehicle ownership on average, and this increase is larger in urban areas with higher initial vehicle ownership (“car-dependent cities”) and in urban areas with lower growth rates; (2) the effect of TNC entry on fleet fuel economy is not statistically significant on average, but it is more positive in urban areas with lower rates of childless households than in urban areas with higher rates of childless households; and (3) the effect of TNC entry on transit ridership is not statistically significant on average, but it is more negative (larger reduction) in urban areas with higher rates of childless households or higher income than in urban areas with lower rates of childless households or lower income. These results offer a novel framework for informing future energy, transportation, and urban planning decision-making.

Results

Analyses are conducted using a series of difference-in-difference models with inverse probability of treatment weighting (IPTW). The difference-in-difference method is a quasi-experimental technique that compares trends before and after a treatment—i.e., TNC market entry in this study—against counterfactual trends in an untreated control group, and IPTW weights help ensure the control group is appropriately comparable to the treatment group (see Transparent Methods in the SI for details and Table S10 for a comparison with unweighted results). We specify models to estimate average TNC entry effects on vehicle ownership, fuel economy, and transit ridership and then characterize heterogeneity in this effect by (1) leveraging two exploratory methods (heterogeneous treatment effect [HTE] and cluster analysis) to identify explanatory variables associated with differences in treatment effects and (2) estimating interaction effects of these identified variables in our primary specification (testing the null hypothesis that these variables do not influence treatment effects). Specifically, in our first exploratory method, we conduct HTE analysis to estimate urban area-specific TNC effects and identify factors that differ between urban areas with positive versus negative estimated effects. In a second, complementary exploratory method, we cluster similar urban areas, calculate cluster-specific TNC effects, and identify factors that differ between clusters with statistically significant versus non-significant estimated effects. Finally, in our primary specification, we add targeted treatment interaction terms identified by the HTE and cluster analysis to our regression model to confirm whether these urban area characteristics explain the heterogeneity in TNC entry effects across urban areas. We present average effect results first (Table 1), followed by HTE (Figure 2, Table 2), cluster analysis (Figure 3, Table 3), and results of our primary specification (Table 4, Figure 4).

Table 1

Results for average effects

	Dependent variable, log:
	Vehicle registration, per capita	Average fuel economy	Transit trips, per capita
Treatment	0.007∗∗	3.00 × 10-4	5.19 × 10-4
	(0.004)	(8.76 × 10-4)	(0.0116)
Covariate controls	Y	Y	Y
Time fixed effects	Y	Y	Y
Group fixed effects	Y	Y	Y
Group time trends	Y	Y	Y
Observations	3395	3395	1848
Deg. freedom	2894	2894	1569
Adjusted R-Sq.	0.948	0.979	0.998

Covariate, time fixed effects, group fixed effects, and group time trend coefficient estimates not shown. ∗p<0.1; ∗∗p<0.05; ∗∗∗p< 0.01.

Average treatment effects of TNC entry on urban areas in the U.S. from three regression models estimating (1) vehicle registrations per capita, (2) average fuel economy, and (3) transit ridership (coefficients for control variables, fixed effects, and linear time trends are excluded for brevity). Expanded results and a comparison with OLS results are presented in Table S10.

Figure 2

Heterogeneous treatment effect results

Effects on TNC entry on per capita vehicle registrations (left) and fleet average fuel economy (right), ranked by urban area from lowest to highest; only statistically significant effects are shown. The center blue line illustrates treatment effects, and the gray bands indicate 95% confidence intervals. Detailed results are presented in Table S5.

Table 2

Urban-area attributes influencing treatment effects: HTE results

	Dependent variable:
	Indicator: TNC entry has a statistically significant [1 = positive, 0 = negative] HTE-estimated effect for per capita registrations	Indicator: TNC entry has a statistically significant [1 = positive, 0 = negative] HTE-estimated effect for fleet fuel economy
Vehicle registrations per capita	0.284∗∗	0.628∗∗
	(0.141)	(0.141)
Population, log	−0.040∗∗	−0.033∗∗
	(0.017)	(0.017)
Δ population, log	−0.448	0.533
	(0.765)	(1.051)
Income	−0.282∗∗∗	−0.161∗
	(0.095)	(0.095)
Transit commuters, log +1	1.396∗∗	2.528∗∗∗
	(0.662)	(0.637)
Unemployment rate	−1.133	0.936
	(0.822)	(0.834)
Childless household rate	0.818∗∗∗	−2.238∗∗∗
	(0.294)	(0.305)
Gasoline price	−0.003	−0.002
	(0.003)	(0.003)
Observations	3395	3395
Degrees of freedom	2895	2895
Adjusted R-squared	0.042	0.057

∗p < 0.1; ∗∗p < 0.05; ∗∗∗p < 0.01.

Coefficients of a linear model estimating whether the HTE-estimated urban area effect for vehicle registrations or fuel economy is statistically significant (p < 0.05) and positive (dependent variable indicator = 1) versus negative (dependent variable indicator = 0) as a function of other covariates used in the primary regression. Results are used to identify candidate urban area characteristics that may be important in determining response to TNC entry.

Figure 3

Cluster analysis results

TNC treatment effect on the change in per capita vehicle registrations (top) and average fuel economy (bottom) varies by urban area typology and is consistently significant and positive (indicated in blue) for one-to-two clusters of urban areas across a sweep of exogenously specified cluster numbers (excluding the single-urban-area cluster containing New York City). Statistically significant effects are highlighted in blue, and estimates that are not significant are gray. The size of each circle reflects the number of urban areas in each cluster; note that the weighted average (by number of urban areas per cluster) of cluster effects is consistent across the number of clusters and with the average estimates in Table 1. Detailed results are presented in Table S6.

Table 3

Urban-area attributes influencing treatment effects: cluster analysis results

	Dependent variable:
	Indicator: TNC entry [1 = has, 0 = does not have] a statistically significant positive cluster-estimated effect for per capita registrations	Indicator: TNC entry [1 = has, 0 = does not have] a statistically significant positive cluster-estimated effect for fleet fuel economy
Vehicle registrations per capita	1.023∗∗∗	1.023∗∗∗
	(0.083)	(0.083)
Population, log	−0.032∗∗∗	−0.032∗∗∗
	(0.009)	(0.009)
Δ population, log	−5.663∗∗∗	−5.663∗∗∗
	(0.439)	(0.439)
Income	−0.243∗∗∗	−0.243∗∗∗
	(0.052)	(0.052)
Transit commuters, log +1	0.695	0.695
	(0.516)	(0.516)
Unemployment rate	−0.080	−0.080
	(0.386)	(0.386)
Childless household rate	1.155∗∗∗	1.155∗∗∗
	(0.146)	(0.146)
Gasoline price	0.004∗	0.004∗
	(0.002)	(0.002)
Observations	3395	3395
Degrees of freedom	2895	2895
Adjusted R-squared	0.1892	0.1892

∗p < 0.1; ∗∗p < 0.05; ∗∗∗p < 0.01.

Coefficients of a linear model estimating whether the cluster containing each urban area has or does not have a significant (p < 0.05) positive estimated TNC effect on vehicle registrations or fuel economy as a function of other covariates used in the primary regression. Results are identical for both dependent variables because the set of clusters with positive effects are identical for the 3-cluster case. Results for the 4-cluster case are presented in Figure S11. Results are used to identify candidate urban area characteristics that may be important in determining response to TNC entry.

Table 4

Results for primary model specification

	Dependent variable, log:
	Vehicle registration, per capita	Average fuel economy	Transit trips, per capita
Treatment	0.006	0.002∗∗	0.003
	(0.008)	(0.001)	(0.027)
Treatment×1Top50%_VehicleRegistrations	0.009∗∗	−2.2 × 10−5	−0.002
	(0.004)	(4.2 × 10−4)	(0.013)
Treatment ×1Top50%_Population	−0.006	8.5 × 10−5	0.024
	(0.008)	(6.9 × 10−4)	(0.021)
Treatment×1Top50%_PopulationGrowth	−0.009∗∗	−5.2 × 10−4	−0.002
	(0.004)	(3.8 × 10−4)	(0.013)
Treatment ×1Top50%_ChildlessHHRate	0.003	−0.001∗∗	−0.026∗∗
	(0.004)	(0.000)	(0.013)
Treatment×1Top50%_Income	0.000	−4.6 × 10−4	−0.051∗∗∗
	(0.005)	(4.4 × 10−4)	(0.017)
Treatment ×1Top50%_TransitRidership	0.004	−5.5 × 10-4	0.027
	(0.005)	(4.2 × 10−4)	(0.020)
Covariate controls	Y	Y	Y
Time fixed effects	Y	Y	Y
Group fixed effects	Y	Y	Y
Group time trends	Y	Y	Y
Observations	3395	3395	1584
Deg. freedom	2407	2407	1034
Adjusted R-sq.	0.972	0.996	0.998

Covariate, time fixed effects, group fixed effects, and group time trend coefficient estimates not shown. †computed post-hoc and not directly estimated; ∗p < 0.1; ∗∗p < 0.05; ∗∗∗p < 0.01.

Treatment effects of TNC entry in US urban areas from three regression models for (1) vehicle registrations per capita, (2) fleet average fuel economy, and (3) transit trips per capita, each including interactions with categorical measures of vehicle ownership, population, population growth, childless household rate, income, and transit commuters. 1Top50%_X is an indicator function that is 1 for urban areas with values of characteristic X in the top 50% of all urban areas in the data set (0 otherwise). Coefficients for controls, fixed effects, and linear time trends are omitted for brevity. Alternative specifications and robustness checks are presented in Tables S9–S16.

Figure 4

Summary of results for primary model specification

Visual summary of regression model findings estimating TNC market entry effects on vehicle ownership, fuel economy, and transit ridership. Average effects for each outcome are shown in blue (with an error bar indicating a 95% confidence interval), and heterogeneous effects are shown in gray (with arrows indicating the estimated interaction effects). Data from Tables 1 and 4.

Average effect

On average, we find TNC entry in an urban area increases per capita vehicle registrations by 0.7% (95% confidence interval: 0.1–1.3%), and we find no significant average effect on fleet average fuel economy or per capita transit trips (Table 1). Each estimate can be interpreted as a weighted average across urban areas (Table S2, Figure S4). The estimated average effect on registrations is also robust to a battery of robustness checks and sensitivity analyses (Table S9), including randomized treatment (Figure S7), leave-one-out analysis (Figures S8 and S9), leave-multiple-out analysis (Figure S10), alternative treatment encoding (Table S12), disaggregation of transit ridership data (Table S13), and an event study (Figure S6). When data are aggregated to the state level of resolution, rather than the urban area level, estimated average effects are negative (Table S11), consistent with a study by Ward et al. (2019) using state-level data. This suggests heterogeneous effects: When effects of TNC entry differ across urban areas, averaging across states, rather than urban areas, can produce different estimates – in this case changing sign. In the following sections, we use HTE and cluster analysis to identify features of urban areas that are important for determining the type of response urban areas have to TNC entry, and we use these results to inform our primary specification, testing for interaction effects with these urban area features.

Heterogeneous treatment effects

In a first approach to exploring heterogeneity, we estimate urban area-specific TNC entry effects and find estimates ranging from an 11.0% decrease in per capita vehicle registrations in Redding, CA to a 15.7% increase in Gainesville, FL, and from a 1.9% decrease in fleet average fuel economy in Greely, CO to a 2.6% increase in Thousand Oaks, CA. Transit ridership data are unavailable for nearly half of urban areas and so are not reported in this approach. Figure 2 shows the distribution of those estimated effects in urban areas that are statistically significant, and Table S5 provides the full results.

TNC entry is associated with a significant decline in vehicle registrations per capita in 38 urban areas (17% of 224 treated urban areas) and an increase in 58 (26%) urban areas. Conducting a linear regression using a binary variable equal to 1 if an urban area has a statistically significant positive estimated HTE effect and 0 if it has a statistically significant negative estimated HTE effect (other urban areas are excluded) (Table 2), we find that urban areas with positive estimated effects on vehicle registrations tend to be smaller, lower income, and have more vehicle registrations per capita, transit commuters, and childless households than urban areas with negative estimated effects. TNC entry is also associated with a significant decline in fleet average fuel economy in 30 (13%) urban areas and increase in 38 (17%) urban areas; urban areas with positive effects on fleet average fuel economy also tend to be smaller, lower income (at the 90% confidence level), and have more vehicle registrations per capita and transit commuters but fewer households without children than urban areas with negative effects. Estimated effects on per capita vehicle registrations and average fuel economy are not significant in 57% and 70% of urban areas, respectively (not shown). Because each treated urban area observes only one TNC entry event, our ability to confidently identify effects for specific urban areas is limited, and we use this analysis primarily to identify candidate hypotheses for which urban area attributes are important in determining the type of response an urban area will have to TNC entry.

Cluster effects

As a complementary approach to characterizing heterogeneity, we use hierarchical clustering to identify and group similar urban areas and then estimate TNC entry effects for each cluster. In Figure 3 we show for alternative exogeneous specifications for the number of clusters the resulting effect in each cluster of TNC entry on per capita vehicle registrations (detailed results for the 3- and 4-cluster cases are presented in Table S6, and results for alternative clustering methods and clustering features are presented in Figures S12 and S13). Significant estimates range from a 1%–3% increase (excluding the single-urban-area cluster containing New York City, which contains only one treatment observation), and all estimates range from a 1.8% decrease to a 3.0% increase. Conducting a linear regression for the case of 3 clusters using a binary variable equal to 1 if an urban area is in a cluster with a statistically significant positive estimated effect and 0 otherwise, we find that the significant-and-positive-effect clusters include urban areas that are, on average, smaller with lower incomes and population growth rates, higher vehicle ownership rates, and more households without children compared to the clusters with estimated effects that are not statistically significant (Table 3). Results for 4 clusters are presented in Figure S11.

Figure 3 also shows for each cluster the effect of TNC entry on average fuel economy, significant estimates of which range from a 0.4%–0.5% increase and all estimates of which range from a 0.07% decrease to a 0.4% increase. As was the case for heterogeneity in estimated vehicle registration effects, the cluster with significant-and-positive-effects includes urban areas that are, on average, smaller with lower incomes and population growth rates and higher vehicle ownership rates and more households without children compared to the clusters with insignificant-effects, as Table 3 again shows.

Estimates are overall robust to the battery of robustness checks already described as well as to additional checks for alternative clustering methods (with further detail the SI). New York City appears as a “cluster” of just one urban area in each figure and is not explicitly explored further in this analysis because the intent is to identify trends across cities, and single-city estimates based on a single TNC entry event may not be reliable. For the effect of TNC entry on transit ridership, none of our clusters had significant effects. Like the HTE analysis, we use the results of this analysis primarily to identify candidate hypotheses for which urban area attributes are important in determining the type of response an urban area will have to TNC entry.

Primary model specification

In our primary specification for characterizing differences across urban areas, we introduce treatment interaction regressions that include the six dimensions of urban area attributes suggested by either the HTE results or the clustering analysis results as being important to determining the direction of TNC entry effects: vehicle registrations per capita (in a pre-treatment reference year, 2010, to avoid endogeneity in modeling vehicle registrations in 2011–2017), population, population growth, income, the percentage of commuters who travel by transit, and the percentage of households without children (all also measured in a pre-treatment reference year, 2011) (Table S8). As Table 4 and Figure 4 show, we find evidence of heterogeneity in TNC effects across all outcomes studied: larger increases on vehicle ownership in urban areas with higher initial vehicle ownership and lower growth rates; larger increases on average fuel economy in urban areas with lower childless households rates; and larger decreases in transit ridership in urban areas with higher incomes and higher childless households rates. Estimates are generally robust to a battery of robustness checks (Table S9), including randomized treatment (Figure S7), leave-one-out analysis (Figures S8 and S9), leave-multiple-out analysis (Figure S10), alternative quantile resolution (Table S15), continuous interactions (Table S16), and corrections for multiple hypothesis tests (Table S14).

Discussion

We apply HTE analysis, cluster analysis, and a regression analysis with interaction effects as three different ways to characterize the heterogeneity of vehicle ownership, fuel economy, and transit ridership impacts as Uber and Lyft enter into urban areas. Our results suggest that access to these TNC services has significant and heterogeneous effects on vehicle ownership (per capita vehicle registrations). All three methods agree that TNC entry tends to produce larger increases in vehicle ownership for those urban areas with higher initial vehicle ownership (Table S8). So, “car-dependent cities” (urban areas in the top 50% of vehicle registrations per capita) remain or become even more so after the introduction of TNCs, whereas car ownership is not necessarily increased when TNCs enter less car-dependent areas. And our primary specification suggests that while TNC-stimulated new vehicle acquisitions tend to outpace new residents in slower-population-growth areas, in faster-population-growth areas, more new residents afforded the flexibility of being potential TNC passengers depresses overall per capita private vehicle registrations.

Our interaction effects model and our HTE model also agree that TNC entry tends to produce larger increases in fleet average fuel economy in urban areas with fewer childless households, but our cluster analysis results suggest the opposite. One possible interpretation for the importance of childless households is that the purchase of efficient new vehicles by new TNC drivers results in a smaller relative efficiency increase in urban areas with a higher concentration of smaller cars than in urban areas with a higher concentration of family cars. Increases in average fuel economy can reduce fuel consumption and associated emissions, although the net effect also depends on potential changes in travel behavior, which was not studied here due to lack of annual vehicle travel distance data at the urban area level.

Finally, the interaction effects model suggests that TNC entry tends to produce larger decreases in transit ridership for urban areas with higher incomes and childless household rates, (we did not identify statistically significant results for transit in our HTE and cluster analysis models). A potential interpretation is that TNCs have a greater tendency to displace transit in cities where potential passengers possess a greater ability to pay for more expensive, more convenient travel modes and where potential passengers are childless and may have fewer family constraints, such as car seats, and different travel behavior, such as frequenting bars and restaurants. Declines in transit ridership affect transit revenues and potentially affect service availability and access. Additionally, diversion of trips from transit to TNC travel has implications for energy use, emissions, and road congestion.

Limitations of the study

Our analysis controls for all described urban area covariates using an approach “doubly robust” to model specification by including these controls in the regression specification as well as in determining propensity score weights using a flexible functional form via gradient boosting. The difference-in-difference model specification also controls for unobserved effects, to the extent that they are time-invariant in each urban area (captured by group fixed effects) or consistent across urban areas (captured by time fixed effects). The addition of linear time trends provides further flexibility to control for different trends across urban areas. Remaining factors for which we (and similar econometric models) cannot control are any unobserved nonlinear changes in trends of unobserved variables that differ across urban areas. We are not aware of additional factors that changed nonlinearly and heterogeneously across urban areas during the analysis period that are correlated with treatment and substantially influence our dependent variables enough to bias our estimates, and our robustness checks protect against some potential effects of this form (e.g.: via leave-one-out tests).

Interpreting the estimated effects presented here as causal relies on three key assumptions of the difference-in-difference approach: (1) exogeneous intervention, (2) parallel trends, and (3) no spillover. For the exogeneous intervention condition, informal discussions with TNCs about market entry strategy suggests that decisions were not made by monitoring differential nonlinear changes in trends across urban area variables other than monitoring frequency of local web searches for terms like “uber” and “lyft”. Our event study results (Figure S6) are consistent with exogeneous intervention because, for registrations as the dependent variable, we observe no statistically significant effects prior to entry and observe statistically significant and positive effects every year after entry. For the parallel trends condition, use of linear time trends provides flexibility in establishing parallel trends relative to potentially distinct linear trends in each urban area, and it is encouraging that our event study (Figure S6), which does not assume parallel trends, also produces a positive (although larger) effect estimate after entry. For the spillover condition, we view it as safe to assume the effect of residents in one city changing vehicle ownership or transit ridership patterns in response to entry of TNCs in other cities is negligible.

Our analysis identifies net overall outcomes after TNCs enter urban areas. We cannot identify more detailed changes to vehicle fleet mix or changes in specific cities with the available data, and there are potentially multiple alternative—and sometimes competing—narratives that might explain these trends. Additional study of the effect of TNC market entry on vehicle fleet composition and travel behavior across the fleet is needed for deeper insight about the mechanisms that produce these outcomes. Our analysis is also constrained to a relatively short time period and relatively simple measures of TNC availability and use; whereas, future study can build on this work with more detailed data. Given the global growth trends of TNC services, vehicle manufacturing, and travel-related GHG, and air pollutant emissions, understanding and mitigating the environmental impacts of TNC use in urban areas is critical.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Jeremy Michalek (jmichalek@cmu.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

The data set and code used in this study have been deposited to Mendeley Data [https://doi.org/10.17632/8tpstn78dh.1], with the exception of vehicle registration and fuel economy data. Vehicle registration and fuel economy data are propriety to and available from IHS Markit; restrictions apply to the availability of these data, which were used under license for the current study, and so are available for purchase from IHS Markit.

Methods

All methods can be found in the accompanying Transparent Methods supplemental file.