The impact of urban form on vehicle ownership

The impact of urban form on vehicle ownership

Journal Pre-proof The impact of urban form on vehicle ownership Matthew J. Holian PII: DOI: Reference: S0165-1765(19)30383-0 https://doi.org/10.1016...

450KB Sizes 0 Downloads 29 Views

Journal Pre-proof The impact of urban form on vehicle ownership Matthew J. Holian

PII: DOI: Reference:

S0165-1765(19)30383-0 https://doi.org/10.1016/j.econlet.2019.108763 ECOLET 108763

To appear in:

Economics Letters

Received date : 1 October 2019 Revised date : 11 October 2019 Accepted date : 13 October 2019 Please cite this article as: M.J. Holian, The impact of urban form on vehicle ownership. Economics Letters (2019), doi: https://doi.org/10.1016/j.econlet.2019.108763. This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

© 2019 Published by Elsevier B.V.

Journal Pre-proof *Highlights (for review)

Highlights The Impact of Urban Form on Vehicle Ownership

relP



urn a



Implements an instrumental variables vehicle demand model using the largest US household survey An indicator variable for the presence of same gender children in the household instruments for population density I find a 10 percent increase in density causes a 0.012 decrease in the size of a household's vehicle fleet, a reduction of about half a percentage point.

Jo



pro of

Matt Holian

Journal Pre-proof *Title Page

pro of

The Impact of Urban Form on Vehicle Ownership Matthew J. Holian∗ October 11, 2019

Abstract

re-

Driving is the single biggest source of household carbon emissions, and land-use policies that encourage higher density are motivated in part by findings of lower vehicle ownership rates in compact areas. However, many previous estimates suffer from selfselection bias. Utilizing an indicator variable for the presence of same gender children in the household as an instrument for population density, I find a 10% increase in

lP

density causes a 0.012 decrease in the size of a household’s vehicle fleet, a reduction of about half a percentage point.

Keywords: vehicle demand; land-use regulation; carbon emissions.

Jo

urn a

JEL Codes: R0, Q5



Corresponding author: [email protected]. One Washington Square, San Jose, CA 95192-0114. USA. I thank Paul Lombardi and an anonymous referee for providing helpful comments. All errors are my own. Declarations of interest: none.

Journal Pre-proof *Manuscript Click here to view linked References

1

Introduction

Driving is the source of approximately 15% of the average US household’s carbon emissions

pro of

(Nordhaus, 2013, p. 161). Local regulations such as zoning are starting to be designed to reduce emissions, but enacting effective policies requires good estimates of their impacts. This study leverages micro data from the nation’s largest household survey, the American Community Survey (ACS), to present new evidence to inform policy development in this important area.

The main contribution of this paper is to present a compelling instrumental variable

re-

(IV) estimate of the effect of residential density on vehicle ownership. The most vexing problem in any study that estimates the effect of urban form on transportation behavior is that household location is not randomly assigned—households with unobserved travel preferences self-select where they live.

lP

A large literature spanning several fields has sought to understand the connection between urban form variables (like city or neighborhood population density, and distance to downtown) and transportation behaviors (such as vehicle miles travelled, transport mode,

urn a

and vehicle ownership.) Some past attempts to overcome the self-selection problem have used control variables (Brownstone and Golob, 2009, Holian and Kahn, 2015). Cao et al. (2009) survey 38 land use and transportation studies that attempt to account for self-selection and report that nearly all found a statistically significant influence. Some studies have done a better job than others, and recently, in assessing the state of this literature, Duranton

Jo

and Turner (2019, p. 171) conclude, ”...the literature has yet to identify a good source of random or quasi-random variation in neighborhood choice.” But it is possible these authors have missed one clever study in this field. Using data from Europe, Grazi et al. (2008) employ one of the more compelling instruments used in this literature. They instrument for density using an indicator for households that have same-gender children. The possibility of more easily having children of the same gender share rooms means these households have more options in denser neighborhoods where homes are typically smaller. 1

Journal Pre-proof

I find strong evidence that this gender-based room sharing and home selection story plays out in the US data. The same-gender indicator is a relevant instrument as it is a fairly strong predictor of density (F=12.8), even when using the geographically unrefined measures

pro of

of population density in the public-use data. The instrument is plausibly exogenous because child gender is, in nearly all cases, as good as randomly assigned and thus not correlated with travel preferences. I unpack this assumption further in the next section. The present study is the first that I am aware of to use this instrumental variable technique using US data, and

2

Data and Methods

re-

the ACS constitutes a significantly larger sample than the original European study.

All of the models presented below use 2012 to 2017 ACS samples, obtained from IPUMS-USA

lP

(Ruggles et al., 2018). The key model I estimate is1

V ehiclesi = Xi γ + β × logDensityi + εi ,

urn a

where V ehiclesi equals the number of vehicles household i has access to, logDensityi is the population density (people per square mile) in the Public-Use Microdata Area (PUMA) in which household i resides, and Xi is a vector of control variables. As suggested in the introduction, the coefficient β on logDensityi will be biased if there are factors in the error term εi that are correlated with density and also determine household vehicle choice. Thus I adopt an instrumental variables strategy following Grazi et al. (2008).

Jo

Considering proto-typical American families with two kids, families with same gender children will find room sharing easier, and thus have more housing options in higher-density neighborhoods where homes are typically smaller but in more walkable and transit friendly neighborhoods. If the gender of the children is random, it would not be correlated with the 1

The data for this study is archived at OpenICPSR: https://www.openicpsr.org/openicpsr/ project/112134.

2

Journal Pre-proof

household’s travel preferences, and thus be an exogenous instrument.2 To estimate this model I use the subsample of white, married-couple households with precisely two children. The gender of children in these households is as good as randomly

pro of

assigned, but many of these households will share characteristics, including importantly cultural norms regarding room-sharing.3 I control for household income, education, and other observables. This choice of subsamples provides for a clear interpretation of the results, though it does come at the cost of a loss in generality.

Results

re-

3

The regression results appear in Table 3. Table 1 and Table 2 present variable descriptions and summary statistics, respectively. In column one of Table 3, I estimate the OLS model,

lP

and we see that the coefficient on logDENSTY is -0.0906. This means a 10% increase in density is associated with about a 0.01 decrease in a household’s vehicle fleet. The average household in the sample has 2.28 vehicles, so this is a reduction of less than a half a percentage point.

urn a

Columns 2 and 3 of Table 3 present a motivating regression, and the first stage of the two-stage least squares estimation, respectively. The motivating regression in (2) provides evidence for the causal story about how child gender affects density, which works through a room-sharing argument. Indeed, in column 2 we see the coefficient on SAMEGENDER (an indicator for households with same gender child pairs) have fewer bedrooms; this result is

Jo

small in magnitude but highly significant (t=15.59). In column 3 we see households with same gender children also live at higher density and the effect is significant (t=3.58). Given SAMEGENDER is the only instrument in the model, the F-statistic on the instrument is 2

One possibility which would violate exogeneity is if households with same gender children economize on trips; for example households with only girls may just go to ballet practice, while households with a boy and a girl may have to go to ballet and football. I have not evaluated the extent to which this is true. 3 Child gender is not always random. Some families practice sex-selection, and if this practice is correlated with travel preferences, the instrument would not be exogenous. Luckily, recent research in family economics finds, ”...no evidence that sex selection has spread beyond the race groups (Chinese, Asian Indian, and Korean) identified in previous work.” (Blau et al. 2016, p. 3).

3

Journal Pre-proof

the square of 3.58, or 12.8. This is greater than the value of 10 that Stock et al. (2002) suggest rules out a weak instrument. The last column of Table 6 presents the second stage results. The coefficient on logDEN-

pro of

SITY, at -0.1184 is similar in magnitude though more negative than the OLS estimates. This means a 10% increase in density causes a reduction in household vehicle fleet by just over a half a percent. This may seem small, but the average US households lives in a neighborhoods with low population density, which is to say a 10% increase in density is small given many neighborhoods in the US have room to drastically increase their density. While we might have expected that IV estimate would have been smaller in magnitude (less negative) than

re-

the OLS estimates presented in column 1, it is important to remember that the IV estimate is a Local Average Treatment Effect (LATE), and households that are most responsive to the same gender treatment may also have more elastic vehicle demand. While this makes a

lP

direct comparison across models (1) and (4) complicated, the IV results do not suggest the OLS estimates are severely biased.

I performed four robustness checks. I added state fixed effects, which I excluded in

urn a

my preferred specification because family composition can plausibly influence cross-state as well as intra-urban sorting patterns. I also used an estimation subsample of all races in models with and without state fixed effects. Finally, I implemented sample weighting in the preferred specification. The OLS results are similar across all models. In some of the alternative specifications, the IV estimate of density is about three times larger than in the preferred specification. Finally, the coefficient on SAMEGENDER was significant in the first

4

Jo

stage across all robustness checks, but the F-statistic ranged from 3 to 9.4 Although the lower first-stage t-statistics raises weak instrument concerns, in some ways it is surprising that the SAMEGENDER instrument does so well given the crude geographical measure we have for households in the public-use version of the data. There are only 2,378 PUMAs in the US, and population density varies significantly within many PUMAs, but there are 73,058 tracts in the US. The restricted use version of the data includes tract and even block group and block identifiers, strongly suggesting the IV strategy I develop here would be highly relevant with more geographically refined neighborhood land use measures.

4

Journal Pre-proof

4

Conclusion

Although the ACS is the nation’s largest household survey, few researchers have used it to

pro of

study the causal effect of land use on transportation behavior. This study has taken some promising first steps by adapting the model for the US context and data. It set the stage for future research using different outcome variables in the ACS, such as commute time, distance and mode. Finally, the results here are encouraging for future research that uses the geographically more refined density measures available in the restricted-use version of the data.

re-

References

lP

Blau, F. D., Kahn, L. M., Brummund, P., Cook, J., and Larson-Koester, M. (2017). Is there still son preference in the united states? Technical report, National Bureau of Economic Research. Brownstone, D. and Golob, T. F. (2009). The impact of residential density on vehicle usage and energy consumption. Journal of urban Economics, 65(1):91–98.

urn a

Cao, X., Mokhtarian, P. L., and Handy, S. L. (2009). Examining the impacts of residential self-selection on travel behaviour: a focus on empirical findings. Transport reviews, 29(3):359–395. Duranton, G. and Turner, M. A. (2018). Urban form and driving: Evidence from us cities. Journal of Urban Economics, 108:170–191. Grazi, F., van den Bergh, J. C., and van Ommeren, J. N. (2008). An empirical analysis of urban form, transport, and global warming. Energy Journal, 29(4):97.

Jo

Holian, M. J. and Kahn, M. E. (2015). Household carbon emissions from driving and center city quality of life. Ecological Economics, 116:362–368. Nordhaus, W. D. (2013). The climate casino: Risk, uncertainty, and economics for a warming world. Yale University Press. Ruggles, S., Flood, S., Goeken, R., Grover, J., Meyer, E., Pacas, J., and Sobek, M. (2018). Ipums usa: Version 8.0 [dataset]. minneapolis, mn: Ipums, 2018. Stock, J. H., Wright, J. H., and Yogo, M. (2002). A survey of weak instruments and weak identification in generalized method of moments. Journal of Business & Economic Statistics, 20(4):518–529. 5

Journal Pre-proof

Jo

urn a

lP

re-

pro of

Blau et al. (2017); Brownstone and Golob (2009); Cao et al. (2009); Duranton and Turner (2018); Grazi et al. (2008); Holian and Kahn (2015); Nordhaus (2013); Ruggles et al. (2018); Stock et al. (2002)

6

Journal Pre-proof

Tables Table 1: Variable Descriptions Description

DENSITY logDENSITY BEDROOMS VEHICLES WORKERS SAMEGENDER COLLEGE HHINCOME logHHINCOME

Population in PUMA / PUMA land area (sq. mi.) The natural log of DENSITY Number of bedrooms in home Number of vehicles available to household Number of workers in the household Indicator equal to one if both children have same gender Indicator equal to one if head of household has college degree Household income Natural log of HHINCOME

re-

pro of

Variable

Table 2: Summary Statistics Mean

1,910.34 6.08 4.39 2.28 1.83 0.48 0.52 125,632.40 11.45

urn a

DENSITY logDENSITY BEDROOMS VEHICLES WORKERS SAMEGENDER COLLEGE HHINCOME logHHINCOME

Median

St. Dev.

Min

Max

434.30 6.07 4 2 2 0 1 97,900 11.49

5,094.19 1.89 0.91 0.81 0.57 0.50 0.50 111,136.80 0.79

0.25 −1.39 1 0 0 0 0 1 0.00

108,979.20 11.60 13 6 4 1 1 1,892,000 14.45

lP

Statistic

Jo

Note: N=379,117 for all variables, which are described in Table 1. The caption to Table 3 describes the sample.

7

Journal Pre-proof

Table 3: Regression Results Dependent variable: logDENSITY

VEHICLES

OLS

OLS

OLS 1st stage

IV 2nd stage

(1)

(2)

(3)

(4)

−0.0906∗∗∗ (0.0007)

SAMEGENDER

logHHINCOME

pro of

BEDROOMS

0.1803∗∗∗ (0.0021)

−0.1184∗∗∗ (0.0015)

−0.0412∗∗∗ (0.0027)

0.0211∗∗∗ (0.0059)

0.3627∗∗∗ (0.0027)

0.4174∗∗∗ (0.0048)

0.1919∗∗∗ (0.0022)

0.1858∗∗∗ (0.0032)

0.5504∗∗∗ (0.0068)

−0.0775∗∗∗ (0.0029)

re-

logDENSITY

VEHICLES

−0.0928∗∗∗ (0.0028)

WORKERS

0.2883∗∗∗ (0.0025)

−0.0111∗∗∗ (0.0026)

−0.3910∗∗∗ (0.0053)

0.2774∗∗∗ (0.0026)

Constant

0.2913∗∗∗ (0.0229)

0.2076∗∗∗ (0.0294)

1.7454∗∗∗ (0.0531)

0.3400∗∗∗ (0.0232)

379,117 0.1360 0.8423

379,117 0.0754 1.8152

379,117 0.1197 0.7577

urn a

lP

COLLEGE

Observations R2 Residual Std. Error

379,117 0.1236 0.7560

Jo

Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. The sample consists of white, married-couple households with exactly two children and no others living in the home, where the head of household’s age is between 25 and 55. ACS samples 2012-2017. All models include survey year fixed effects. Model (1) is a baseline model which may suffer from omitted variables bias, (2) is an auxiliary regression highlighting the effect of SAMEGENDER on household bedroom choice, and (3) is the 1st stage, where the coefficient on SAMEGENDER is significant (t=3.58 and F=12.8). (4) is the second stage of the IV model.

8