«Does the Market Value Value-Added? Evidence from Housing Prices After a Public Release of School and Teacher Value-Added Scott A. Imberman∗ Michigan ...»
graphics. In contrast, as shown in column (2) the value-added estimates are much more weakly correlated with school demographics. Only two of the estimates are statistically signiﬁcant at the 5% level, and the R2 is only 0.22. In Column (3), we add overall API and each student subgroup’s API scores as regressors to the model. The R2 remains low at only 0.29. Thus, more than 70% of the value-added variation is unpredictable from the observable characteristics of the school, including test score levels. Further, in Columns (5) and (6) we see that these observables do an even poorer job of predicting the LAUSD value-added rank, with R2 s of 0.03 and 0.25 with and without including overall and by-category API as regressors, respectively.
Table 2 and Figures 3-4 show that the value-added data released to the public by the LA Times and LAUSD contained new and unique information about school quality that was not simply a measure of the school composition or prior test scores. Our identiﬁcation exploits this new information by identifying the impact of value-added on housing prices conditional on API along with many other observable characteristics of schools and neighborhoods. Since these characteristics are observable to homeowners as well, we are able to identify the impact of this new information given the information set that already exists.
4 Empirical Strategy Our main empirical strategy is to estimate diﬀerence-in-diﬀerence models that compare changes in property values surrounding the information releases as a function of value-added rank conditional on observable school and neighborhood characteristics, including API. Since value-added only was released for elementary schools, we ignore middle and high school zones. Our main
empirical model is of the following form:
value-added measures. This variable is set equal to zero prior to the ﬁrst release in September, 2010, is set equal to the value-added percentile in the ﬁrst release from September 2010 to April 2011, and thereafter equals the value-added percentile provided in the May 2011 LA Times’ updated release. Pooling the two releases together makes the estimates easier to interpret.
Nonetheless, it is possible that the impact of value-added may diﬀer across both the ﬁrst and second releases. Thus, we also estimated models with a separate variable that equals zero prior to each release and to that particular release’s value-added percentile thereafter. As these results were very similar to those found in the pooled model, we do not show them here. They are available by request.
While our focus is on the LA Times value-added, it is nonetheless important to account for the LAUSD value-added scores and the fact that the LA Times published API scores on the
zero prior to April 2011 and the LAUSD calculated value-added percentile thereafter. We also allow for the eﬀect of API to vary post-September 2010. Our inclusion of school ﬁxed-eﬀects in the model (γs ) implies that the coeﬃcients on the V A and AP I ∗ P ost variables represent the diﬀerence-in-diﬀerence estimate of the eﬀect of the value-added or API information release on property values.19 In order to account for the fact that there are multiple sales per school zone, all estimates are accompanied by standard errors that are clustered at the school level.
Equation (1) also includes an extensive set of controls to account for any confounding eﬀects driven by the correlation between value-added and school demographic or housing characteristics. The vector X contains the set of school observables discussed above, including two lags of API, within-LAUSD API percentile rank in the given year, and the decile of the school’s API in comparison to other “similar” schools, as deﬁned by the California Department of Education.20 The vector H is the set of house-speciﬁc characteristics and Census tract characteristics discussed above that further control for local demographic diﬀerences that are correlated with value-added and for any changes in the types of houses being sold as a function of value-added Note that unlike API, which changes each year, each value-added release provides a single value for each school, and thus main eﬀects are removed by the school ﬁxed eﬀects. In models that do not include school ﬁxed eﬀects, main eﬀects are included as controls as well.
Models that also control for student subgroup speciﬁc API provide similar results to baseline.
when each release occurs. Equation (1) contains both month-by-year ﬁxed eﬀects(λt ) and school ﬁxed eﬀects (γs ) as well, so all parameters are identiﬁed oﬀ of within-school changes in home prices over time and control for any general shocks to home prices in LAUSD in each month, including seasonal changes.
There are two main assumptions underlying identiﬁcation of the value-added parameters.
First, the model assumes that home prices were not trending diﬀerentially by value-added prior to each of the data releases. Using the panel nature of our data, we can test for such diﬀerential trends directly in an event-study framework. In Figure 5, we present estimates of the eﬀect of each value-added and API release, where V ALAT P ooled and V ALAU SD are interacted with a st st series of indicator variables for time relative to the August 2010 LA Times release. Note that all of the estimates shown in this ﬁgure are from a single regression and in this particular model V ALAT P ooled is set equal to the initial value-added percentile from the August 2010 release for st all periods prior to September 2010. The top two panels of Figure 5 show no evidence of a pre-release trend in home prices as a function of LAT or LAUSD value-added. The estimates exhibit a fair amount of noise, but home prices are relatively ﬂat as a function of future valueadded rank in the pre-treatment period. Thus, there are no clear pre-treatment trends for either information release that would bias our estimates. For API, while there appears to be a slight downward trend in earlier months, it is not statistically diﬀerent from zero and by 7 months prior to the release, property values ﬂatten as a function of API.
Figure 5 also previews the main empirical ﬁnding of this analysis: home prices do not change as a function of value-added nor API post-release. However, these estimates are relatively imprecise, as the estimates in Figure 5 are demanding of the data. We thus favor the more parametric model given by equation (1). Nonetheless, Figure 5 demonstrates that pooling the estimates over each post-release period does not mask any time-varying treatment eﬀects.
The second main identiﬁcation assumption required by equation (1) is that the value-added percentile, conditional on school characteristics, is not correlated with unobserved characteristics of households that could aﬀect prices. While this assumption is diﬃcult to test, given the rich set of observable information we have about the homes sold, examining how these observables shift as a function of value-added will provide some insight into the veracity of this assumption. Thus, in Table 3, we show estimates in which we use neighborhood characteristics, school demographics and housing characteristics as dependent variables in regressions akin to equation (1) but only including API percentiles, time ﬁxed eﬀects and school ﬁxed eﬀects as controls.21 Each column in the table comes from a separate regression, and each estimate shows how the observable characteristic changes as a function of value-added percentile after the LA Times data releases. Overall, the results in Table 3 provide little support for any demographic or housing type changes that could seriously impact our estimates. There are 58 estimates of housing and neighborhood characteristics in the table, one that is signiﬁcant at the 5% level or higher and four more that are signiﬁcant at the 10% level. While clearly these variables are not independent, if they were we would expect to falsely reject the null at the 10% level six times.22 Furthermore, the estimates, even when signiﬁcant, are small. For example, a 10 percentage point increase in LA Times value-added post-release is associated with a decrease in the percent of household with children of 0.04 percentage points, which is a very small effect relative to the mean of 32%. Additionally, the signs of the estimates do not suggest any particular patterns that could cause a systematic bias in either direction.
Another concern is that the release of a value-added score may induce changes in the number of homes sold in a school catchment area. Since we only observe prices of homes that are sold, we may understate the magnitude of the eﬀect if having a lower value-added reduces the number of homes sold and this reduction comes from the bottom of the price distribution. To test this hypothesis, we estimate a version of equation (1) in which we aggregate the data to the schoolmonth level and use the total number of sales or the total number of sales with a valid sales price in each school-month as the dependent variable.23 We ﬁnd little evidence of a change in the number of sales. The estimate of the eﬀect of LA Times value-added on total sales24 The school characteristics estimates in Panel C use data aggregated to the school-year level. Since we do not have a year of data after the second LA Times release, we examine only the ﬁrst LA Times value-added estimates in Panel C.
While not shown in the table, the estimates for the LAUSD value-added measure are signiﬁcant at the 10% or lower level seven times and the estimates for API*post are signiﬁcant at the 10% or lower level for only two variables. The former is slightly more than would be expected by random chance and thus provides some additional justiﬁcation for focusing on the LA Times release. These results are available by request.
We include mean Census Tract characteristics of properties sold in a school zone and school characteristics but do not control for aggregate individual property characteristics as these may be endogenous in this regression.
Our data only cover the three most recent sales of a property. Thus, our measure of total sales will be slightly underestimated.
is -0.0066 with a standard error of 0.0051. Taken at face value, this would suggest that a 10 percentile increase in value added only reduces monthly sales by 0.07 oﬀ of a mean of 8.4. For sales with price data, the estimate is -0.0043 with a standard error of 0.0025. Estimates for LAUSD value added are smaller and close to zero.
The value-added releases we study come at a time of high volatility in the housing market, as home prices declined during this period throughout most of the United States. In the Los Angeles MSA prices declined by 4.5%.25 This was also a period with a large number of foreclosures in Los Angeles. If foreclosure rates are correlated with the value-added releases, it could bias our home price estimates because foreclosures tend to be sold at below market value. In order to provide some evidence on this potential source of bias, we use the number of foreclosures in each month and zip code in LAUSD that were collected by the RAND Corporation.26 We aggregate prices to the school-month level and use the zipcode-level data to approximate the number of foreclosures in the school catchment area in each month. The resulting estimates show little evidence of a correlation between value-added post-release and the number of foreclosures. None of the estimates for any of the three data releases is statistically signiﬁcant.
The estimate on the pooled LA Times release is only 0.003 (0.009), which indicates that a 10 percentile value-added increase post-release increases the number of monthly foreclosures in a school zone by 0.03, oﬀ of a mean of 5.7. Overall, the estimates described above along with those provided in Table 3 and Figure 5 provide support for our identiﬁcation strategy.
5.1 Pre-Release School Quality Valuation Estimates Before presenting the main diﬀerence-in-diﬀerences estimates, it is important to establish that some measures of school quality are indeed valued by LA residents. Whether public school quality, or public school characteristics more generally, are capitalized into home prices in Los This calculation comes from the Federal Housing Finance Agency’s seasonally adjusted home price index.
Note that this decline was smaller than the 7% rate for the US as a whole during this period.
These data are available at http://ca.rand.org/stats/economics/foreclose.html.
Angeles is not obvious, as LAUSD has an active school choice system in which students can enroll in their non-neighborhood school. There also is a large charter school and private school presence in the District. Thus, any ﬁnding that property values do not respond to value-added information could be driven by a general lack of association between local school characteristics and property values. Nonetheless, there are a few reasons to believe that this is not a major concern in the Los Angeles context. First of all, the open-enrollment program is small relative to the size of the district. In 2010-2011, only 9,500 seats were available district-wide, accounting for at most 1.5% of the district’s 671,000 students. Second, while LAUSD has a number of magnet programs, they are highly sought after and oversubscribed, hence admission to a desired magnet is far from guaranteed. Third, Los Angeles is a very large city with notorious traﬃc problems and poor public transportation, making it diﬃcult for parents to send their children to schools any substantial distance from home.
To further address this issue, we estimate boundary ﬁxed eﬀects models using pre-release data from April 2009 to August 2010 in which API percentile is the dependent variable. This model is similar to the one used in Black (1999) as well as in the subsequent other boundary ﬁxed eﬀects analyses in the literature (Black and Machin, 2011) and allows us to establish whether average test scores are valued in LA as they have been shown to be in other areas.
Table 4 contains boundary ﬁxed eﬀects results, comparing home prices within 0.1 mile of elementary attendance zone boundaries. In column (1), which includes no other controls, properties just over the border from a school with a higher API rank are worth substantially more. For ease of exposition, all estimates are multiplied by 100, so a 10 percentage point increase in API rank is associated with a 4.5% increase in home values. In column (2), we control for housing characteristics, which have little impact on the estimates. However, controlling for school demographics in column (3) signiﬁcantly reduces this association. This result is not surprising given the ﬁndings in Bayer, Ferreira and McMillan (2007) and Kane, Riegg and Staiger (2006). Nonetheless, in column (3), we ﬁnd a 10 percentage point increase in API rank leads is correlated with a statistically signiﬁcant 1.3% increase in property values. This estimate is roughly equivalent in magnitude to those in Black (1999) and Bayer, Ferreira and McMillan (2007). Thus, this school characteristic appears to be similarly valued in Los Angeles as in the areas studied in these previous analyses (Massachusetts and San Francisco, respectively). It remains unclear, however, whether the capitalization of API scores is driven by valuation of schools’ contribution to learning or by valuation of neighborhood or school composition that is correlated with API levels.27 Our analysis of capitalization of value-added information is designed to provide insight into resolving this question, which is very diﬃcult to do without a school quality measure that is not highly correlated with student demographics.