You are on page 1of 34

LONDON SCHOOL OF ECONOMICS

Valuing Crime from Housing Prices: An Application of


Spatial Econometric Approaches

GY460: Techniques of Spatial Economic Analysis

Candidate Number: 71459


Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

1. Introduction
Urban crime is commonly associated with other social issues such as poverty,
deprivation and unemployment. In the case of the United Kingdom and particularly
London, the government has introduced various regeneration programmes and area-
specific initiatives to reduce and prevent crimes (Imrie, et al., 2009; Home Office, 2013).
With no formal markets to price crimes, non-market methods such as hedonic pricing
have been used to elicit individuals’ willingness to pay for public safety. However, such
empirical applications, which do not give due consideration to spatial effects, could lead
to endogeneity problems (Gibbons & Machin, 2008; Ihlandeldt & Mayock, 2010). This
paper seeks to serve two objectives: i) to apply spatial econometric methods including
standard and experimentalist approaches to the valuation of crimes in London and ii) to
evaluate the advantages and limitations of the approaches used.

2. Valuation of Crimes in Previous Studies


Scholars have employed the hedonic model in valuing crimes (Thaler, 1978; Lynch &
Rasmussen, 2001; Ihlandeldt and Mayock, 2010). The model works on the notion that
homebuyers purchase a dwelling not just for its physical built but also for a range of
social attributes and amenities. As such, the price of the dwelling will reveal individuals’
willingness to pay for non-market and intangible goods such as safety in a low-crime
neighbourhood. These studies have found negative relationships between housing
prices and crimes. For instance, Thaler (1978) reported that property crime is related to
a 3% decrease in housing values. Similarly, Lynch and Ramussen (2001) found that a
one standard deviation increase in violent crime is associated with a 4% decrease in
housing price.

However, given that crime and housing are likely to exhibit spatial variations, the use of
traditional regression methods has raised concerns over the presence of endogeneity
bias. Estimations could be biased and inconsistent due to the violations of key OLS
assumptions, i.e. independence between regressors and the error term ( Cov( x,  )  0 )
and no autocorrelation ( Cov(1,  2 )  0 ) (Stock & Watson, 2012). In the case of crime-
housing relationship, Ihlandeldt and Mayock (2010) highlighted five mechanisms that
could lead to endogeneity problems. Three of the mechanisms could result in upward
biases. These include systematic targeting of neighbourhoods with more expensive
homes, the tendency for higher reporting rates in better off neighbourhoods (Skogan,
1999) and the presence of unobserved characteristics such as secluded backyards and
large windows that could increase the attractiveness of homes to criminals (Gibbons,
2004). Downward biases could also occur due to the tendency for affluent
neighbourhoods to be more protected than lower-income neighbourhoods and the
inclination for criminals to commit crimes within their home neighbourhoods that tend to
be low-income in general.

1
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

In fact, the use of spatial econometric methods in valuing crimes is unlikely to


automatically rectify the problem if no due consideration is given to methods of isolating
sources of endogeneity. Indeed, the use of Maximum Likelihood (ML) in spatial
econometrics has been criticised to be a paradox (McMillen, 2010). Specifications are
assumed to be despite the fact that ML was used in the first place due to the presence
of spatial autocorrelated errors. Gibbons & Overman (2012) thus advocated the need to
adopt an “experimental paradigm” (p. 172) that put issues of identification and causality
at the centre stage of spatial econometrices. The current study seeks to consider these
issues and approaches through the valuation of crimes from housing prices in London.

3. Research Design

3.1 Methodology
The current study will adopt three types of models in valuing crimes from housing prices.
These include the OLS model, the standard spatial lagged models and the
experimentalist models. The basic OLS model is as follows:

Y  X   (1)

whereby Y represents the log of housing price and X represents a set of explanatory
variables including crimes, structural, neighbourhood and locational attributes. 
denotes the coefficients of variables including the intercept and  refers to the error term.
Spatial effects are not considered in the basic model.

As crimes and housing are likely to be spatially correlated, spatial lag models are
employed to deal with this issue. These include the Spatial Lagged “X” (SLX) Model and
the Spatial Autoregressive (SAR) Model. The SLX model is as follows:

Y  X  WX   (2)

Apart from the inclusion of Y and X that is similar to (1), model (2) contains a set of
lagged explanatory variables denoted by WX. The coefficient  represents indirect
spillover effect from neighboring characteristics. In addition, the SAR is included to test
for nuances between standard lagged models in dealing with endogeneity issues. The
SAR model is presented as follows:

Y  WY  X   (3)

Similar to (1) and (2), Y denotes log of housing prices and X denotes a set of
explanatory variables with no spatial dimensions. WY represents outcome in
neighbouring areas and the coefficient  represents direct spillover effect from
neighbours’ actions and behaviour. In the case of crime-housing relationship, this meant

2
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

that housing prices in neighbouring region could affect housing prices in the current
region. Unlike the SLX model which uses OLS estimation, the SAR will be estimated
using ML method.

However, all three approaches highlighted above are likely to be prone to estimation
errors due to the violation of OLS assumptions 1 . In particular, even though ML
estimation is used in (3) to achieve identification, it is unlikely to be inaccurate given that
it works under the basic assumption that the true model is already identified.
Experimentalist approaches such as spatial differencing, i.e. fixed effects (FE), and
instrumental variables are thus included to give a greater focus on identification and
causal issues. The experimentalist model is as follows:

[Y  m(Y | W )]  [ X  m( X | W )]  [  m( | W )] (4)

whereby all variables are deviated from the local neighbor averages to remove
unobservables that are constant across neighbouring wards as a source of endogeneity.
Three IV are also included to instrument the two crime variables including density of
burglary, criminal damage and theft in non-residential buildings. This approach is
adopted from Gibbons (2004)’s study on the cost of property crimes in London. The IVs
will be further validated to ascertain their relevance (Cov (Z, X) ≠0) and exogeneity
(Cov(Z ,ɛ)=0) in the context of this paper.

3.2 Data Availability and Weight Construction

The current study uses ward-level data on housing prices, crimes and socio-economic
indicators for the year 2001. Given that there is no crime data for the City of London, a
total of 624 wards (excluding the city) are used for this analysis. Queen contiguity-
based weight matrices are constructed from the polygon shapefile provided. Figure 1
shows that all wards share borders with other wards, ranging from 1 neighbour to 13
neighbours.

1
The four key assumptions are: expected mean value of residual to be zero (E(ɛ) = 0), ii) constant variance, iii) independence
between regressors and the error term (Cov (X, ɛ) =0 ) and iv) no autocorrelation (Cov (ɛ1, ɛ2) =0).

3
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Figure 1 Neighbours per ward

4. Descriptive Statistics

4.1 Summary of Data


Table 1 summarises the variables selected. These variables are further explored in a
spatial context in the following sections.

Table 1 Summary of Data (^Central London refers to Holborn and Covent Garden)

4.2 Spatial Patterns for Housing and Crime


As of 2001, the mean housing price in London is £160,940 with a standard deviation of
£53,763.46. As shown in Figure 2, properties with the highest values seem to be
concentrated in the central western region close to the Thames River as well as the
northwestern and southeastern suburbs.

4
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Figure 2 Distribution of Housing Prices

In contrast with the distribution of housing prices, the density of burglary and criminal
damage in dwellings are the highest in inner London as indicated in Figure 3a and b.
The means of burglary and criminal damage in dwellings are 80.66 cases per square
kilometre and 40.69 cases per square kilometre respectively.

Figure 3a Distribution of burglary in dwellings Figure 3b Distribution of criminal damage in dwelling

4.3 Global and Local Moran’s I Analyses


To ascertain if significant spatial clustering exists, both the Global and Local Moran’s I
Analyses are conducted. Table 2 shows the Moran’s I value and significance for all
variables. All Moran’s I values are positive, indicating that high-value regions tend to be
surrounded by high-value neighbours. Similarly, low-value areas are clustered with low-
value neighbours. Given that the p-values for all global Moran’s Is are less than 5%, we

5
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

can reject the null hypothesis that spatial distributions are due to complete spatial
randomness. See Appendix A for the Moran I’s scatter plots.

Table 2 Global Moran's I Analysis

In addition, the localised Moran’s I analysis is conducted to identify spatial differences


across wards. Table 3 shows the distribution of wards by clusters (high-high and low-
low), outliers (low-high and high-low) and spatial randomness (not significant). Among
the nine variables, ln_Price, Burglary in Dwelling, Criminal Damage in Dwelling,
Population Density and Distance to Central London show the highest level of clustering
(40% and above). In comparison, spatial randomness is most common for proportion of
vacant housing and floor size (70% and above).

6
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Table 3 Local Moran's I Analysis

As shown in Figure 4a and b, clusters of high property values in the central western
region are found to be statistically significant. This is similar to findings in Figure 2. Low
property values are also found to cluster in the northeast region.

Figure 4a Cluster Map for ln_Prices Figure 4b Significance Map for ln_Prices

The cluster and significance maps for the two crime variables also show high-high
clusters in inner London and low-low clusters in the outer regions as indicated in Figure
5 and 6. Comparison of the housing and crime spatial patterns seem to indicate that

7
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

crime rates tend to be higher in areas with high housing prices, albeit that the clustered
areas do not overlap completely.

Figure 5a Cluster Map for Burglary in Dwellings Figure 5b Significance Map for Burglary in Dwelling

Figure 6a Cluster Map for Criminal Damage in Dwelling Figure 6b Significance Map for Criminal Damage in Dwelling

Apart from housing price and the crime variables, cluster and significance maps are
also generated for the other explanatory variables. Distinctions between inner and outer
London can be observed. In general, inner London seems to be associated with low-low
clusters of bedrooms and floor size as well as high-high clusters of aged dwellings,
population density and vacancy (only pertaining to the central western region). Refer to
Appendix B for the maps. Given the consistency in spatial concentration in inner
London, one would expect that these variables would have some influences (either
direct or spatial) on housing prices as well. Further testing will be required to ascertain
the direction and significance of relationship between housing prices and the set of
explanatory variables.

8
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

5. Regression Results
Table 4 presents regression results from the five models derived from equations (1) to
(4). Refer to Appendices C-G for the regression outputs.

Table 4 Regression Results (Standard errors are given in parentheses under the coefficients. The individual
coefficient is statistically significant at the 10%*, 5%** or 1%*** significance level.)

In general, a one unit increase in criminal damage per square kilometre is associated
with a decrease in housing prices for all models. The coefficients range from a decrease

9
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

of 0.3% for the OLS model to 0.1% for the FE with IV. These are statistically significant
at the 1% level for all models. In contrast, burglary in dwelling does no yield significant
results (p-value > 0.05).

For the OLS, a semi-log specification is used with the ad hoc inclusion of age square to
improve the functional form. The SLX model excludes distance to Central London and
adds on two lagged variables for burglary in dwelling and criminal damage in dwelling.
This is under the assumption that the occurrence of crime in neighbouring areas may
also be factored into the purchase decisions of homebuyers in the area of interest. The
SAR model includes a lagged ln_Price instead of lagged explanatory variables. The FE
models are regressed on variables (excluding lagged variables) that deviate from local
averages. The three IVs used namely burglary, criminal damage and theft in non-
residential dwellings are also found to be strong (F-statistic = 40.39 > 10, p-value< 0.01)
and exogenous instruments (J-statistic = 1.02 < 3.84 for df = 1 and α = 0.05).

Coefficients vary as lagged variables are introduced into the model. These values
stablise with the use of the two FE models. Standard errors also improve greatly with
the introduction of lagged and FE models. It is interesting to note that regression
diagnostics suggest that the use of FE models remove problems of multicollinearity as
the multicollinearity condition number improves from 69.284 for the OLS model to 3.946
for the FE model. Spatial autocorrelation in the error term is also reduced as indicated in
Figure 7. Appendices H-K show the cluster and significance maps for the residuals.

Figure 7a Scatter Plot for OLS Residuals Figure 7b Scatter Plot for SLX Residuals

10
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Figure 8c Scatter Plot for SAR Residuals Figure 7d Scatter Plot for FE Residuals

6. Discussion and Conclusion

The above study compares the robustness of various approaches in valuing crime from
housing prices after taking identifications issues into consideration. The standard lagged
models, i.e. SLX and SAR, are improvements over the OLS given the inclusion of
lagged variables and the use of identification methods, i.e. ML in the case of this study.
However, these methods are likely to be inadequate in dealing with endogeneity issues.
In the case of the SLX, OLS assumption of independence of error terms from X and WX
is likely to be violated due to mobility of individuals across space and the presence of
unobserved variables unaccounted for in the model. Figure 7b shows the tendency for
spatial autocorrelation for the residual across space. The use of ML estimates in the
SAR model as an identification strategy is also problematic, given its assumption that
the model is already correctly specified.

Given the limitations of standard spatial econometric methods, the use of


experimentalist approaches could potentially improve the robustness of the study.
Indeed, the use of the FE models provides more consistent coefficients. In addition,
spatial autocorrelation is greatly reduced (as shown in Figure 7d). The three
instruments are also tested for their relevance and exogeneity, contrasting with the use
of spatial IVs that are rather arbitrary in nature.

11
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Still, there are limitations to the use of experimentalist approach here. For instance, the
variables are deviated from local averages under the assumption that these averages
are representative of the region. Outliers could potentially skew these values. In addition,
the use of contiguity-based weights suggests that spatial effects are defined according
to the administrative boundaries of the wards. This may not be the case for issues such
as crimes. Distance-based weighting could be used as an alternative method for weight
construction provided that coordinate data is available. However, this too would require
careful judgment on the most appropriate distance.

In sum, while there is no single approach that could entirely deal with endogeneity
issues in the valuation of crimes, attempts should be made to incorporate more
experimental approaches particularly when applying spatial econometric methods. This
could potentially lead to more robust and meaningful results from spatial studies.

12
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Bibliography
Gibbons, S. & Machin, S., 2008. Valuing school quality, better transport, and lower crime: Evidence from
house prices. Oxford Review of Economic Policy, 24(1), pp. 99-119.

Gibbons, S. & Overman, H. G., 2012. Mostly Pointless Spatial Econometrics. Journal of Regional Science,
52(2), pp. 172-191.

Home Office, 2013. Reducing and Preventing Crime. [Online]


Available at: https://www.gov.uk/government/policies/reducing-and-preventing-crime--2
[Accessed 29 April 2013].

Ihlandeldt, K. & Mayock, T., 2010. Panel data estimates of the effects of different types of crimes on
housing prices. Regional Science and Urban Economics, Volume 40, pp. 161-172.

Imrie, R., Lees, L. & Raco, M., 2009. Regenerating London. Oxon: Routledge.

Lynch, A. K. & Rasmussen, D., 2001. Measuring the impact of crime on house prices. Applied Economics,
Volume 33, pp. 1981-9.

McMillen, D. P., 2010. Issues in Spatial Data Analysis. Journal of Regional Science, 50(1), pp. 119-141.

Skogan, W., 1999. Measuring what matters: crime, disorder, and fear. In: R. H. Langworthy, ed. Measing
what matters: Proceedings from the policing research institute meetings. Washing, DC: US Department of
Justice, pp. 37-53.

Stock, J. H. & Watson, M. M., 2012. Introduction to Econometrics. 3rd ed. Essex: Pearson.

Thaler, R., 1978. A note on the value of crime control: evidence from the property market. Journal of
Urban Economics, Volume 5, pp. 137-145.

13
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Appendix A – Scatter Plots for All Variables (Global Moran’s I)

Ln_Price

Burglary in Dwelling

14
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Criminal Damage in Dwelling

Number of Bedrooms

15
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Floor Size

Age of Dwelling

16
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Population Density

Proportion of Vacant Housing

17
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Distance to Central London

18
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Appendix B – Cluster and Significance Maps for Explanatory Variables (LISA)

Number of Bedrooms

19
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Floor Size

20
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Age of Dwelling

21
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Population Density

22
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Proportion of Vacant Housing

23
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Distance to Central London

24
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Appendix C – Regression Output for OLS2

2
The variables are coded as follows: Combined_7 - ln_Price; Combined_31 – Burglary in Dwellings; Combined 33 –
Criminal Damage in Dwellings; Combined_10 – Bedrooms; Combined_11 – Floor Size; Combined_12 – Age;
Combined_35 – Population Density; Combined_41 – Vacant Housing; Combined_25 – Distance to Central London

25
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Appendix D – Regression Output for SLX3

3
The variables are coded as follows: Combined_7 - ln_Price; Combined_31 – Burglary in Dwellings; Combined 33 –
Criminal Damage in Dwellings; Combined_10 – Bedrooms; Combined_11 – Floor Size; Combined_12 – Age;
Combined_35 – Population Density; Combined_41 – Vacant Housing; LaggedBud – Burglary in Dwellings (Lagged);
LaggedCDD – Criminal Damage in Dwellings (Lagged)

26
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Appendix E – Regression Output for SAR with ME4

4
The variables are coded as follows: Combined_7 - ln_Price; W_Combined_7 – ln_Price (Lagged); Combined_31 –
Burglary in Dwellings; Combined 33 – Criminal Damage in Dwellings; Combined_10 – Bedrooms; Combined_11 –
Floor Size; Combined_12 – Age; Combined_35 – Population Density; Combined_41 – Vacant Housing

27
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Appendix F – FE Model5

5
Variables are deviations from local averages

28
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Appendix G – FE Model with IV6

6
Variables are deviations from local averages

29
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Appendix H – Moran’s I Analysis for OLS Residuals

30
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Appendix I – Moran’I Analysis for SLX Residuals

31
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Appendix J – Moran’s I Analysis for SAR Residuals

32
Valuing Crime from Housing Prices: An Application of Spatial Econometric Approaches
Candidate Number: 71459

Appendix K – Moran’s I Analysis for FE Residuals

33

You might also like