Professional Documents
Culture Documents
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
The MIT Press is collaborating with JSTOR to digitize, preserve and extend access to The Review of
Economics and Statistics
This content downloaded from 190.116.36.68 on Thu, 01 Jun 2017 19:29:49 UTC
All use subject to http://about.jstor.org/terms
LAND VALUES IN A NEWLY ZONED CITY
Abstract-The introduction of a new zoning ordinance to Chicago in 1923 Zoning's effects would be far easier to detect with a
offers a natural experiment that allows us to determine the effects of
natural experiment. Suppose we were to observe a sample of
zoning on relative land-value growth rates. Policymakers claimed at the
otherwise identical parcels of land, all of which are initially
time that land-use zoning would raise aggregate land values by reducing
negative externalities associated with mixed land use. After controlling for
unzoned. A new ordinance is adopted under which some
initial land use and the endogeneity of zoning decisions, we find that
parcels are zoned residential and others are zoned commer-
residential zoning led to higher land value growth rates than commercial
zoning. cial. If we know land values before and after the new
ordinance, then the difference in the land-value growth rate
I. Introduction between residential and commercial parcels provides a
clean measure of residential zoning's effect relative to
potentially conflicting land uses often are close to one zoning. Of course, if no parcels remain un-
commercial
another in densely populated urban areas. Stores, gas
zoned, we cannot determine whether growth rates are higher
stations, and even factories are sometimes in the same block
overall after the zoning ordinance than would have pur-
as houses. It is possible that everyone is better off by the
veyed without the new law. But if the zoning ordinance is
arrangement: businesses get a ready supply of workers and
necessary to eliminate unwanted intrusions of commercial
customers, who in turn have their transport costs reduced.
land into residential areas, we should find that land-value
But it also is possible that businesses create nuisances-
growth rates are relatively high in areas that are zoned
such as noise and traffic-to households, and that the
exclusively residential.
negative externalities are ignored when firms make their
This difference-in-differences approach cannot be imple-
location decisions. The potential market failure may be
mented using recent data because nearly all American cities
averted through land-use zoning: if negative externalities
were zoned for the first time in the 1920s, and large-scale
are limited to a small area around businesses, then separat-
revisions of ordinances are rare.2 Moreover, the necessary
ing business and residential areas may improve on the
panel of land-value data is rarely available. In this study, we
market allocation of land. Zoning then will increase resi-
use a remarkable data set to determine the effects of zoning
dential land values because homeowners are willing to pay
on relative land-value growth rates in Chicago. Chicago
a premium for exclusively residential areas.
adopted its first zoning ordinance in 1923. Land use was
Verifying the effect of zoning on land values has proved
completely unregulated before this date, but every block of
extraordinarily difficult.1 The reason is twofold. First, the
the city was zoned as of April 16, 1923. Chicago has a
competitive market may already have separated land uses.
unique data source, Olcott's Land Values Blue Book of
Large tracts of land are often developed by firms that would
Chicago, which provides annual estimates of land value for
forego profits if they ignored homeowners' preferences for
every block in the city. We obtained maps showing the use
exclusivity. Siegan (1972) is the standard reference. He
of every building in 1921 and the new zoning category for
argues that Houston, which is the only major American city
every block. For a sample of 1,013 blocks, we recorded the
without a zoning ordinance, has similar land-use patterns to
zoning category, the 1921 and 1924 land values, and the
other cities although there all land-use decisions are made
initial use, along with standard control variables such as
privately. Similarly, McMillen and McDonald (1999)
distance from the city center. All the pieces are in place to
present evidence that land use was highly segregated in
determine the effect of zoning on relative land-value growth
Chicago even before the city adopted its first zoning ordi- rates.
nance. The second reason empirical studies have failed to
The potential endogeneity of the zoning decisions poses
establish conclusively that zoning affects land values is that
econometric challenges. Zoning officials may have been
zoning may be flexible, mimicking the competitive market
aware in 1923 that land values were growing slowly in some
(although probably with a lag). McMillen and McDonald
areas. If blocks with high land-value growth rates were
(1991) present evidence for this view using data from
more likely to be zoned residential, then residential zoning
suburban Chicago. They find that changes in relative land
would appear to cause higher growth rates even if zoning
values lead to changes in zoning categories. But no study
had no independent effect on land values. The availability of
has documented the frequency of variances or simple toler-
land value data before and after zoning permits us to use
ance of zoning code violations that many researchers sus-
recent extensions (Heckman, Smith, & Clements, 1997;
pect make zoning far from binding.
Heckman, Ichimura, & Todd, 1997, 1998; Heckman et al.,
1998) of Rosenbaum and Rubin's (1983) propensity score
Received for publication April 2, 1998. Revision accepted for publica-
tion February 13, 2001.
* University of Illinois-Chicago. 2 The first comprehensive zoning ordinance was passed in New York
City in 1916. Mills (1979) reports that by 1925 nearly 500 American cities
We thank Dan Black, Chris Bollinger, John Engberg, Robert Moffitt, and
the referees for helpful suggestions. had zoning ordinances, and by 1930 three-fifths of the urban population
1 See Fischel (1990) and Pogodzinski and Sass (1991) for reviews. lived in zoned communities.
This content downloaded from 190.116.36.68 on Thu, 01 Jun 2017 19:29:49 UTC
All use subject to http://about.jstor.org/terms
LAND VALUES IN A NEWLY ZONED CITY 63
This content downloaded from 190.116.36.68 on Thu, 01 Jun 2017 19:29:49 UTC
All use subject to http://about.jstor.org/terms
64 THE REVIEW OF ECONOMICS AND STATISTICS
Proportion Residential
The simple partial equilibrium model is entirely static.
Suppose that landowners are forward looking. Grandfather
clauses permitted existing uses to continue indefinitely. The
zoning ordinance created local monopolies for commercial
residential than in commercial use for any level of p. A enterprises in the middle of areas zoned residential (Bab-
competitive market allocates the land to residential use cock, 1957). Because these establishments could be sold,
(point A 1). The opposite situation occurs when the residen-
land values might rise in mixed land-use areas that were
tial land-value function falls to R3. Commercial firms
zoned residential. In contrast, landowners might have ex-
always pay more for the land than do homeowners, and pected
the that the grandfather clauses would eventually fade
competitive market assigns all land to commercial useleading the new zoning provisions to be binding.5 Land
out,
(point A3). The most interesting case occurs when the then would fall in mixed land-use areas that were
values
residential land-value function is R2. Here, businesses out-
zoned residential if commercial firms expected eventually to
bid homeowners for parcels on exclusively residential
leave. In a growing city, it is reasonable to expect that
blocks, but homeowners outbid businesses for parcels in
commercial firms may eventually seek out sites in areas that
exclusively commercial areas. Point A2 then is a stable
currently are predominantly residential. Land values might
equilibrium, and the competitive market's tendencyfalltoin residential areas that are zoned commercial if current
equalize land values leads to mixed land use.4 landowners are wary of the possibility of invasion by
In the 1923 Chicago zoning ordinance, new businesses
conflicting land uses in the future, and residential landown-
were not permitted in areas zoned residential, but either use
ers may pay a premium for the insurance of neighborhood
was permitted in areas zoned commercial. Ignore grandfa-
stability provided by residential zoning.
ther clauses for now, and assume that a block zoned resi-
We must also allow for the potential endogeneity of
dential would subsequently have p = 1. When the residen-
zoning decisions. Zoning authorities were probably aware
tial land-value function is as high as R1, residential zoning
of the effects of their decisions on land values. They
is redundant: the market has already led to exclusively
claimed at the time that eliminating nuisances would raise
residential land use. Commercial zoning also is meaningless
land values, which suggests that they did not want to take
in a hierarchical zoning system because houses are permit-
actions that would lower land-value growth. An area with
ted and the land is worth more in residential use. In contrast,
commercial land would be zoned residential only if its land
suppose that the residential land-value function is as low as
R3 and that the area is currently all commercial. Land
5 Indeed, the city reneged on the grandfather clause in 1957 when the
values remain at A3 if the area is zoned commercial, butordinance
fall was revised. Conflicting businesses had ten years to sell, after
to A3* under residential zoning. The best case for zoning
which all land had to conform with the new zoning. The new zoning
occurs under mixed land use. Here, land values remainordinance
at was not hierarchical, and residential uses also were required to
leave when an area was zoned commercial. The 1957 ordinance would
have been viewed as Draconian in 1923, and would not have passed court
review.
4 The mixed-use equilibrium is stable if both ,r and P, are positive, but Still, there may have been some early suspicion that the grandfa-
3, > P,. See McDonald and McMillen (1998). ther clauses would eventually be phased out.
This content downloaded from 190.116.36.68 on Thu, 01 Jun 2017 19:29:49 UTC
All use subject to http://about.jstor.org/terms
LAND VALUES IN A NEWLY ZONED CITY 65
I = P3'X,
Yri1 Xi ++Uri
ul, (2)
(2) Equation (5) measures the difference in land-value growth
rates for sites that have the same initial land use but were
and subsequently zoned for different uses. In contrast to the
traditional selection bias estimator, which assumes that
, = Pc'Xi + uL. (3) conditioning on Z eliminates selection bias in the post-
zoning time period, equation (5) assumes that conditional
The superscript indicateson the
Z selection bias is constant
time over time soafter
period that it cancelsthe e
of the zoning ordinance.out in differencing.
This content downloaded from 190.116.36.68 on Thu, 01 Jun 2017 19:29:49 UTC
All use subject to http://about.jstor.org/terms
66 THE REVIEW OF ECONOMICS AND STATISTICS
Equation (5) can be estimated using a variantestimate is the standard deviation of the 10,000 new values
on match-
ing procedures (Rosenbaum & Rubin, 1983, 1984). TheWe
of MA. first
also present 90% and 95% confidence intervals
step is to estimate the probability of residential zoning
using by
the percentile method: the values of A are sorted, and
probit: P(Z) = (Dy'Z). The next step is to compare
entries
land-value growth rates across zoning categories, ()B and (1 - 2)B form the confidence interval,
placing
more weight on observations that have similar estimated
where a = 0.05 and 0.10.
probabilities of residential zoning:
Heckman et al. (1998) demonstrate that the matching
estimator is effective at reducing selection bias in the
A = 2o(i)[(y y) - yO) _ W(i,j)(y, - y)], (6)
iEII jEjIo evaluation of a job training program. They decompos
selection bias into three parts. The first two arise because of
where I1 is an index for observations zoned residential and
different densities and different weighting of E( yX, D) for
Io is an index for observations zoned commercial.6 The first
the I1 and Io observations. These two components ar
set of weights, ((i), can be used to account for heteroske-
"selection on observables." The other component arises
dasticity and scale. In our application, o(i) equals 0 when
when differences in outcomes continue to exist after con-
an I1 observation has no counterpart in Io with a similartrolling for Z (which includes X and other exogenous
probability of residential zoning, and w(i) = llnl other-variables), and is the form of selection bias that is handled
wise, where nl is the number of observations with o(i) by: standard estimators. Heckman et al. (1998) show that
0. The second set of weights, W(i, j), sum to 1 and depend
matching estimators are effective at reducing the first two
on the estimated probability of residential zoning.
components of selection bias, and find that empirically these
Following Heckman, Ichimura et al. (1997), we use Fan's
components comprise a much greater proportion of overall
(1992) local linear estimator to construct W(i, j). We use
15
bias than traditional selection on unobservables. Thus, in
their application the matching procedure is a better method
the biweight kernel, K(si) = 1-6(Si - 1)2I(sij < 1),
than traditional IV procedures for controlling for exogenous
where sij = P(Zi) - P(Zj) and h is the bandwidth. The variables.
weights are normalized to sum to 1. The local linear esti-
In our context, the first two components of selection bias
mator is a variant of the nearest neighbor estimator and is
arise when there are few sites with commercial zoning that
less sensitive to boundary effects than standard kernel
have similar values of P(Z) to sites that are zoned residen-
estimators. The estimates are easy to construct by weighted
tial. In the extreme, all sites zoned commercial have P(Z) =
least squares (Pagan & Ullah, 1999). Let Q = ri - y? and
0 and all sites that are zoned residential have P(Z) = 1. A
Qcj = yj - yj. Then, jElo W(i, j)(yj - y?) is equivalent
to the predicted value from a weighted least squares regres- comparison of land values across the two regimes then has
little meaning because the sites are inherently different. In
sion of Qcj on P(Zj) with weights of K(sij)1/2. Thus, the
match for Qri is the predicted value from a weighted least contrast, two sites that both have P(Z) = 0.4 but which are
squares regression of Qc on the estimated probabilities, zoned differently are similar in all measurable ways. If their
using the estimated probability for observation i as the land values differ, the difference is attributable to zoning.
target point. Io observations with estimated probabilities If our concern were only the distributional or functional
that are close to P(Zi) get the heaviest weight in construct- form assumption behind the standard selection model, we
ing the match for observation i. could use a semiparametric estimator such as those of Ahn
The weighted least squares regression has no observa- and Powell (1993), Cosslett (1991), or the approaches
tions, where no is the number of observations with Io = 1. discussed by Pagan and Ullah (1999). We are also con-
Following Silverman (1986), we set the bandwidth using cerned about the plausibility of our exclusion restrictions,
the rule h = 1.066nol/5, where 62 = var(P(Z)). We show which leads us to prefer the matching approach.7 Matching
that the results are not sensitive to moderate variation in the
does not correct for classical selection bias, which focuses
bandwidth. The regression cannot be estimated for obser-on the error terms. It does correct for the influence of
vation i unless there are at least two Io observations withobservables without making strong functional form assump-
K(sij) > 0. In practice, we require three observations with tions as in Barnow, Cain, and Goldberger (1981). As with
K(sij) > 0, and set w(i) to 0 otherwise. other estimators for selection models, the matching proce-
We use a simple bootstrap procedure to construct stan-dure hinges on the specification of Z. Our data set provides
dard errors for RA. We make B = 10,000 independentenough strong predictors of zoning that we are confident
draws from the I, and Io observations, and form new
that we have accounted for the key variables used to draw
estimates of M for each draw. The bootstrap standard error
Chicago's zoning map, and that remaining unobserved in-
fluences are not closely correlated with land-value changes.
6 We do not use Heckman, Ichimura et al.'s (1997) regression-adjusted
local linear matching estimator because our values of X do not differ over
time, which causes the regression adjustment to cancel out of the two 7 Heckman et al. (1999) provide an excellent description of alternative
land-value growth rate terms in equation (6). methods for evaluating the effects of labor market programs.
This content downloaded from 190.116.36.68 on Thu, 01 Jun 2017 19:29:49 UTC
All use subject to http://about.jstor.org/terms
LAND VALUES IN A NEWLY ZONED CITY 67
The matching estimator is better suited to our data set than TABLE 1.-DESCRIPTIVE STATISTICS
This content downloaded from 190.116.36.68 on Thu, 01 Jun 2017 19:29:49 UTC
All use subject to http://about.jstor.org/terms
68 THE REVIEW OF ECONOMICS AND STATISTICS
This content downloaded from 190.116.36.68 on Thu, 01 Jun 2017 19:29:49 UTC
All use subject to http://about.jstor.org/terms
LAND VALUES IN A NEWLY ZONED CITY 69
the most information on relative land values for these probability of resident
observations.
For mixed land use, we have ample numbers of blocks in
use blocks with residential zoning makes a comparison of
each zoning category with comparable estimated probabil- land-value growth rates less reliable. The difference in
ities of residential zoning. The numbers are smaller, butreliability is reflected in larger bootstrap standard errors for
residential-use blocks also have comparable observations in
the commercial-use blocks. It is also reflected in the number
each category. However, the small number of commercial- of observations given no weight (io = 0 in equation (6)) in
constructing the difference-in-difference estimates. Matches
are available for all residential and mixed-use blocks that
9 Following Silverman (1986), we use a Gaussian kernel with bandwidth
h = 1 .066n-5, where 6 is the standard deviation of the probit
were zoned residential, but only six of 26 blocks have o,i
probability estimates and n is the number of observations for the land use
0 for commercial blocks. These six blocks have the lowest
and zoning category under consideration.
FIGURE 3.-KERNEL DENSITY ESTIMATES: MIXED LAND USE FIGURE 4.-KERNEL DENSITY ESTIMATES: COMMERCIAL LAND USE
3.5 35
3.0 30
2.5 25
2.0 20
1.5 15
1.0 10
0.5 5
0.0
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
probability of residential
probability
zoningof resi
This content downloaded from 190.116.36.68 on Thu, 01 Jun 2017 19:29:49 UTC
All use subject to http://about.jstor.org/terms
70 THE REVIEW OF ECONOMICS AND STATISTICS
Estimate of residential
residential land values, all blocks faced a risk that an
minus commercial incompatible nonresidential use could move in prior to the
land-value
zoning ordinance. As we have shown, post-zoning land
growth 0.195 0.184 0.477
Bootstrap standard error 0.110 0.095 0.307 values reflect this assurance provided by residential zoning.
Percentile method 90%
confidence interval 0.035, 0.379 0.109,0.246 0.065, 1.002 VI. Spatial Autocorrelation
Percentile method 95%
confidence interval -0.010,0.428 0.063, 0.261 0.015, 1.113
Bandwidth 0.129 0.120 0.107
Autocorrelation is a potential problem in spatial data sets.
Standard probit estimates remain consistent when errors are
autocorrelated, but the standard errors are biased. Bootstrap
estimated probability procedures
of residentialhave not been developed
zoning. for spatially
Thecorrelated
small
ber of residentially zoned error structures.
commercialDriscoll and Kraay (1998)
blockspresent a con-
accou
the large standard error, and is a clear warning sistent covariance matrix for continuous dependent variable th
results for this use should be viewed with caution. models with spatially dependent panel data, but no such
estimator has been developed for cross-sectional spatial data
The results of the nonparametric matching estimator are
presented in table 4. The results are nearly the same asor infor models with discrete dependent variables. Table 6
table 2 for residential blocks: residential zoning increasespresents the results of three alternative model specifications
land values by about 19.5% compared with commercial that help control for spatial autocorrelation. The "spatial
zoning. The bootstrap standard error is high, but we can expansion" model is commonly used by geographers (Ca-
setti,
reject the null of no difference in land-value growth rates at 1972; Jones & Casetti, 1992) to control for spatial
the 10% level. The value of the matching estimator is seen effects. This model adds a low-order polynomial of the
in the last two columns. Table 4 indicates that residential geographic coordinates to the base model. The latent vari-
zoning increases relative land values by 18.4% in mixed-use able for the probit model becomes D* = y'Zi + f(Ni, Ei) +
vi, in
blocks. The effect is statistically significant, whereas where N and E represent distances north and east of the
table 2 we found no difference in land values for this city center. We use a second-order polynomial, adding N, E,
N2, E2, and N X E as explanatory variables. The motivation
land-use category. Table 4 also indicates that residential
behind this procedure is that spatial autocorrelation is usu-
zoning increases land values for blocks that were initially
commercial, but the effect is reduced as compared ally caused by missing variables that are correlated over
to table
2 to a more reasonable figure of 47.7%. The results inspace
table(but hopefully not with Z). The spatial expansion
4 are consistent across land-use category and leadhelps reduce the extent of the correlation in the errors by
to the
taking into
conclusion that residential zoning increased land values in account broad spatial trends.
the period shortly after the adoption of the zoningThe other two alternatives follow the literature on spatial
ordi-
nance. probit models (Case, 1992; McMillen, 1992) by focusing on
The results of nonparametric estimators the are
heteroskedasticity
sometimes that is generally associated with spa-
not robust to changes in the bandwidth. Table 5 presents the autocorrelation per se. Heteroske-
tial models rather than
results when the bandwidth is varied. The results are not dasticity is potentially more serious than autocorrelation
highly sensitive to the bandwidth choice for residential and
mixed-use blocks. The results are more variable for com- TABLE 5.-BANDWIDTH VARIATION
mercial blocks, but they are stable for a fairly wide range of Estimate of Residential Minus Commercial
Land Value Growth
0.05 to 0.11. Table 5 suggests that our results are not an
artifact of the choice of bandwidth. Bandwidth Residential Mixed Commercial
The results of this study are in some contrast to the 0.05 0.162 0.177 0.512
findings in our earlier study of Chicago land values prior to 0.06 0.178 0.183 0.503
0.07 0.191 0.186 0.492
zoning (McMillen & McDonald, 1993). In that study, we 0.08 0.194 0.185 0.484
found that the presence of commercial or light manufactur- 0.09 0.196 0.184 0.480
ing use on a block face did not reduce residential land 0.10 0.198 0.184 0.478
0.11 0.199 0.184 0.477
values in the (as of 1921) unregulated land market. How- 0.12 0.198 0.184 0.740
ever, as we show in table 4, blocks with residential zoning 0.13 0.195 0.183 0.758
increased in value more than did other blocks. What can 0.14 0.192 0.183 0.675
0.15 0.189 0.182 0.660
explain this seeming discrepancy in empirical results ob- 0.16 0.186 0.182 0.632
tained from the same data source (Olcott's)? A simple 0.17 0.185 0.181 0.753
answer is that the residential zoning designation provides a0.18 0.184 0.181 0.754
0.19 0.184 0.180 0.729
guarantee that there will be no additional commercial or0.20 0.184 0.179 0.697
industrial use on the block of any kind, an assurance that did
This content downloaded from 190.116.36.68 on Thu, 01 Jun 2017 19:29:49 UTC
All use subject to http://about.jstor.org/terms
LAND VALUES IN A NEWLY ZONED CITY 71
The first entry in each cell is the estimate of residential minus commercial land-value gro
This content downloaded from 190.116.36.68 on Thu, 01 Jun 2017 19:29:49 UTC
All use subject to http://about.jstor.org/terms
72 THE REVIEW OF ECONOMICS AND STATISTICS
This content downloaded from 190.116.36.68 on Thu, 01 Jun 2017 19:29:49 UTC
All use subject to http://about.jstor.org/terms