by Bao Tran-Phu
I. Research question
In “Identifying the Impacts of Rail Transit Stations on Residential Property Values,” Bowes and Ihlanfeldt (2001) study the relationship between transit station proximity and local property values. Specifically, they use least-square regressions to study the property value effects of having a Metropolitan Atlanta Rapid Transit Authority (MARTA) rail station nearby. To understand the factors driving this impact, they disaggregate it into four distinct effects: a direct positive effect (from heightened access to transit), a direct negative effect (from physical aspects of the station itself, such as noise and pollution), an indirect positive effect (as stations attract retail development), and an indirect negative effect (as stations serve as a gateway for crime). Previous studies have investigated this transit-property value impact, but by not quantifying the effects of these individual drivers, they have failed to build a comprehensive picture of what exactly causes the impact.
Given a Tiebout amenities-based approach to evaluating the economic benefits of public transit, the benefits (as well as costs) of having a transit station nearby would be fully capitalized in local property values. Thus, by analyzing the relationship between transit station proximity and property values, in theory they are conducting a comprehensive accounting of all of the rail’s social benefits. More importantly, however, they employ regression analysis to better understand the drivers behind this transit-property values link. This study is important for transit policy decisions for two reasons: first, for proposed transit networks, it can help transit authorities choose optimal locations for their stations and to understand what their effects will be on the surrounding neighborhoods. Second, for existing transit networks, it can help local governments tailor policy in order to promote the benefits and mitigate the costs caused by the transit stations.
II. Empirical analysis
This paper employs a series of multivariate least-squares linear regressions. To briefly illustrate the mechanisms behind ordinary least-squares (OLS) regression, assume a response variable Y, where Y is determined by the following relationship:
Y= α0 + α1X1 + α2X2 + α12X1* X2
X1 and X2 are explanatory variables that are hypothesized to collectively explain Y. X1*X2 is an “interaction” of the two explanatory variables, capturing any synergistic effects above and beyond the sum of the two effects individually. Finally, α1, α2, and α12 are coefficients that quantify the linear effects of their respective explanatory variables, while a0 gives the average value for Y when all explanatory values equal zero. Explanatory variables can either be continuous, discrete (for ex, binary), or mixed. If explanatory variables are continuous, then their coefficients give the slope of the variable’s effect, holding fixed all other variables. If they are binary “dummy variables,” then the coefficient quantifies the difference-in-means effect of X=1 (relative to X=0), holding fixed all other variables. From there, we set up a regression, given i observations:
In the estimated regression, ui is the residual, signifying the variation in Ythat is not explained by the regression model. Statistical software can then find the coefficient values that minimize the sum of square residuals. In their paper, Bowes and Ihlandfeldt use three regression models that are estimated separately.
Hedonic price model (OLS regression)
The first model is used is to quantify station distance’s effect on property values, and to disaggregate that effect into the four main effects listed above.
P = α0 + αZZ + βLL + βCC + βRR + γ SS
P is house sale prices, C is neighborhood crime density, R is neighborhood retail employment, S is rail station proximity, and Z and L are sets of other house and neighborhood characteristics respectively that are independent of rail station proximity (for example, the number of rooms in the house). They opt for a semilog regression, (whereby the response and explanatory variables relate exponentially) over a linear regression, and they find that doing so achieves higher statistical significance. Also, station proximity (S) is captured by a set of dummy variables for being less than one-quarter mile, between one-quarter and one-half mile, between one-half and one mile, between one and three miles, and greater than three miles away. They opt for this technique over using a continuous proximity variable based on the intuition that the effects of station proximity will differ depending on how far away the neighborhood is from it to begin with.
Crime and retail employment auxiliary models (random effects regression)
The two auxiliary models are then used to provide a closer look at how exactly crime and retail development indirectly affect property values.
C = ψ0 + θNN + ψSS
R = φ0 + πQQ + φSS
As before, C is crime density, R is retail employment, and S is a set of station proximity dummy variables. N and Q are respective sets of crime- and retail employment-correlated neighborhood factors that are independent of station proximity. In both cases, since they use panel data on 206 census tracts from 1991-1994, random effects models are used instead of the standard OLS (“fixed effects”) regression model. This is a variation of OLS, whereby the census tracts are treated as randomly-selected samples from a large pool, as doing so avoids having to employ 206 dummy variables to control for the tract-level effects.
Each model is analyzed both with and without interactions. Including the interactions allows the authors to study how the effects of station proximity vary depending on station characteristics (for example, whether or not the station has a parking lot), neighborhood income, and distance from the CBD.
III. Results and implications
For the hedonic price model regression, R2 = 0.475, where R2 measures the percentage of the variation in the response variable explained by the regression. Effects of both crime density and retail employment on property values were statistically significant, providing validation for the two auxiliary models. Excluding interactions, being less than one-quarter of a mile away from a station causes property values to be 20% lower than equivalent houses more than 3 miles away, but houses between 1-3 miles away are worth the most (holding all other variables constant). When the interactions are included, Bowes and Ihlanfeldt additionally find that the premium paid for being between 1-3 miles from a station is greater for high-income neighborhoods, and being less than a mile from a station is worth more in neighborhoods further away from the CBD.
For the crime auxiliary model, R2 = 0.824. Without interactions, it is found that being less than half a mile from a rail station led to higher crime density. Including the interactions then shows that rail stations that have parking lots increase crime for neighborhoods less than one-quarter mile away, but decrease crime for neighborhoods between one-quarter and one mile away. This may indicate that station parking lots serve as a magnet for crime, attracting it in the station’s immediate vicinity but pulling it away from areas a bit further away. Also, the authors find an increase in crime as neighborhood median income rises, but only for tracts less than one-quarter mile from the station. Finally, the station proximity effect on crime is smaller for neighborhoods further away from the CBD. This could suggest that in the suburbs, criminals may see themselves as more easily identifiable as an outsider, and the higher perceived probability of being caught could thus deter them.
Finally, for the retail auxiliary model, R2 = 0.670. Excluding interactions, station proximity has no significant effect on retail employment. When the interactions are added, however, it is shown that further away from the CBD (greater than 10 miles away, for example), having a station nearby has a significant positive effect on retail employment.
Overall, these three regression analyses demonstrate that station proximity affects property values both indirectly (by affecting crime and retail development) and directly (possibly by improving transit accessibility and/or by imposing negative physical externalities) (see Figure 1). Direct effects are generally found to be larger than the indirect crime or retail effects. Further, Retail effects are larger in magnitude than crime effects, except for downtown neighborhoods closest to the stations. Most importantly, though, the study demonstrates that the station proximity effect differs greatly depending on neighborhood income, distance from the CBD, and distance from the station. The biggest benefactors of rail stations are neighborhoods that are far from the CBD and very close (one-quarter to one-half mile) from the station, in which case the positive retail effect dominates. The biggest losers are higher-income downtown neighborhoods closest (less than one-quarter mile) to the station, in which case the negative crime and direct externality effects dominate.
Figure 1. The three models demonstrate station proximity’s direct and indirect effects on property values.
In planning public transit, local governments may assume that stations will have a dominantly positive proximity effect on surrounding property values, due to the heightened transit accessibility that the stations provide. As these results show, however, the reality is more complex. To accurately predict the effect of a nearby transit station, the underlying drivers as well as the neighborhood’s characteristics must be considered. Additionally, local governments should seek to maximize the benefits and minimize the costs of existing transit stations. In low-income neighborhoods, governments should focus on mitigating the stations’ negative externalities, while the focus should be on reducing transit-driven crime in high-income neighborhoods.
In this paper, Bowes and Ihlanfedlt are meticulous to separate the competing indirect effects of rail stations driven by crime-related and retail employment-related factors. Curiously, a similar effort is not made to disaggregate the direct effects of station proximity, even while two major competing effects (transit accessibility and negative physical externalities) are identified.
In addition, property values are used here as a proxy for the local demand for transit, as theorized by the amenities view of public goods. However, given that real estate buying decisions are likely temporally and conceptually displaced from household transportation decisions, property values may be only an imprecise proxy. A possible extension could thus be a similar study, replacing property value with household transportation costs as the response variable. This study would investigate the effects of nearby transit stations on actualized transportation costs for surrounding households, which would be rooted in actual transit use. These transportation cost savings could then be aggregated with the other direct and indirect effects studied in this model to create single hedonic model, providing a more complete illustration of the economic impact of public transit.
The full document can be found here: TP_Bao
Bowes, David R., and Keith R. Ihlanfeldt. “Identifying the Impacts of Rail Transit Stations on Residential Property Values.” Journal of Urban Economics 50 (2001): 1-25.