# RECORD b8c37b8ea0a12e9cbed953c9880d16a2673b2159
# TITLE: Measures of labour market accessibility. What can we learn from observed commuting patterns?
# YEAR: 2020.0  VENUE: nan  DOI: 10.18335/region.v7i1.261
# AUTHORS: Liv Osland; Arnstein Gjestland; I. Thorsen
# PAGES: 22  CANDIDATE TABLE PAGES (1-based): [13, 17, 22]
# ABSTRACT: It is well known that measures of labour market accessibility explains spatial variation in housing prices even in markets with polycentric labour market structures. This paper examines whether data on observed commuting patterns can replace or supplement gravity-based measures representing the commuting potential at specific locations. We use data from a region in Western Norway,and we find that measures based on observed commuting flows and commuting time cannot replace a gravity-based measure of labour market accessibility. Based on, inter alia, the spatial Durbin estimator we find that measures of observed commuting flows increase the explanatory power of a hedonic house price model.


===== PAGE 1 (score 7, numeric-paren 0, table-words 0, var-words 7) =====

Volume 7, Number 1, 2020, 49–70
journal homepage: region.ersa.org
DOI: 10.18335/region.v7i1.261
Measures of labour market accessibility. What can we
learn from observed commuting patterns?∗
Arnstein Gjestland1, Liv Osland1 and Inge Thorsen1
1 Western Norway University of Applied Sciences
Received: 13 February 2019/Accepted: 28 February 2020
Abstract. It is well known that measures of labour market accessibility explain spatial
variation in housing prices, even in markets with polycentric labour market structures.
This paper examines whether data on observed commuting patterns can replace or
supplement gravity-based measures representing the commuting potential at speciﬁc
locations. We use data from a region in Western Norway, and we ﬁnd that measures
based on observed commuting ﬂows and commuting time cannot replace a gravity-based
measure of labour market accessibility. Based on, inter alia, the spatial Durbin estimator
we ﬁnd that measures of observed commuting ﬂows increase the explanatory power of a
hedonic house price model.
1
Introduction
The relationship between house prices and access to workplaces is a central theme in
both theoretical and empirical housing market research. There are many reasons why this
relationship is important. Travelling to work is a regular and bounded trip. According to
V˚agane et al. (2011), travelling to work constitutes 18 per cent of all travels in Norway.
During workdays, approximately 25 per cent of the travels are journey to work. Most
commuters travel at the same point in time every day, which also creates congestion. Road
transportation infrastructure is often given a capacity to deal with such traﬃc peaks. At
the same time, investments in the road network aﬀect labour market accessibility, which
is in turn capitalized into house prices. In order to reduce many of the transportation
problems related to commuting, planners may seek to locate houses in areas where job
accessibility is assumed to be high.
According to Handy, Niemeier (1997), there is no consensus in the literature on a
good measure of accessibility. In explaining housing prices, gravity-based accessibility
measures have been suggested as a generalization of modern polycentric labour market
structures. Although more recent research has shown that gravity-based accessibility
measures explain signiﬁcant spatial variation in housing prices, Handy, Niemeier (1997)
show that the gravity-based measures are not the only weighted measures that can be
used to capture the job opportunity density of a given area.
Gravity-based measures of labour market accessibility reﬂect the potential for commut-
ing from a speciﬁc residential area. In this paper we introduce three other measures, based
on actual commuting patterns. One measure is origin speciﬁc: the percentage of the total
working population living in the zone and working in a diﬀerent zone. The second measure
∗We are grateful to two anonymous reviewers for very helpful comments.
49


===== PAGE 13 (score 100, numeric-paren 67, table-words 4, var-words 2) =====

A. Gjestland, L. Osland, I. Thorsen
61
Table 2: Estimated results from alternative hedonic house price models
Variable name
M0
M1
M2
M3
M4
M5
Constant
12.970
12.417
12.942
12.085
12.543
11.430
(71.89)
(63.52)
(71.75)
(59.50)
(49.53)
(38.87)
LotSize
-0.173
-0.204
-0.167
-0.203
-0.177
-0.221
(-3.00)
(-3.62)
(-2.91)
(-3.55)
(-3.1)
(-3.92)
LotSize2
0.022
0.025
0.022
0.025
0.023
0.026
(4.62)
(5.27)
(4.55)
(5.22)
( 4.72)
(5.59)
RurLotSize
-0.025
-0.021
-0.027
-0.023
-0.025
-0.019
(-10.28)
(-8.83)
(-10.15)
(-9.64)
(-10.11)
( -8.36)
Age
-0.2017
-0.192
-0.205
-0.199
-0.204
-0.196
(-11.75)
(-11.38)
(-11.84)
(-11.72)
(-11.75)
( -11.47)
Age2
0.017
0.016
0.018
0.017
0.018
0.016
(5.75)
(5.23)
(5.92)
(5.65)
(5.85)
( 5.38)
Garage
0.042
0.043
0.040
0.042
0.042
0.040
(7.21)
(7.44)
(6.93)
(7.16)
(7.19)
(7.06)
LivingArea
0.510
0.505
0.510
0.503
0.511
0.504
(42.77)
(42.69)
(42.80)
(42.66)
( 42.99)
(42.88)
YearDum04
0.097
0.098
0.097
0.097
0.097
0.097
(8.89)
(9.03)
(8.90)
(9.05)
(8.89)
(8.98)
YearDum05
0.206
0.207
0.205
0.205
0.205
0.205
(18.75)
(18.92)
(18.68)
(18.90)
(18.68)
(18.84)
YearDum06
0.370
0.371
0.370
0.370
0.370
0.370
(33.55)
(33.67)
(33.49)
(33.82)
(33.51)
(33.82)
YearDum07
0.558
0.558
0.558
0.557
0.558
0.555
(53.95)
(54.17)
(53.85)
(54.44
(53.93)
(54.39)
DistCBD
-0.056
-0.104
-0.043
-0.114
-0.076
-0.171
(-3.14)
(-5.36)
(-2.28)
(-5.57)
(-3.90)
(-7.40)
DistCBD2
-0.046
-0.027
-0.051
-0.024
-0.043
-0.0150
(-13.18)
(-5.84)
(-12.22)
(-4.53)
(-10.86)
(-2.85)
ACCESS
0.064
0.096
0.114
(6.48)
(7.79)
(7.06)
IN-COM
-0.030
-0.083
(-2.06)
(-4.90)
OUT-COM
-0.043
0.001
(0.03)
(2.31)
MCT
0.393
0.399
(2.45)
(2.29)
MCT2
-0.081
-0.054
(-2.41)
(-1.45)
n
4392
4392
4392
4392
4392
4392
R2
0.822
0.824
0.823
0.826
0.823
0.826
R2(adj)
0.822
0.824
0.822
0.825
0.822
0.825
Log-likelihood
1184.02
1207.05
1187.48
1226.90
1187.60
1228.14
VIF
16.36
17.56
16.06
17.60
45.87
46.20
Ramsey reset
0.767
0.760
0.464
0.316
0.8003
0.5364
APE
510405
508469
510366
507978
510486
508223
SRMSE
0.289
0.288
0.290
0.288
0.289
0.288
RLM-lag
1.70
0.86
1.38
0.17
1.70
0.39
RLM-error
323.59
313.09
324.15
307.78
320.36
313.01
REGION : Volume 7, Number 1, 2020


===== PAGE 14 (score 7, numeric-paren 0, table-words 0, var-words 7) =====

62
A. Gjestland, L. Osland, I. Thorsen
Note: The dashed lines refer to M3, while the solid line refers to a corresponding model speciﬁcation
without a gravity-based accessibility measure (M2). Predicted housing prices for variation in OUT-COM
based on M2 is not shown in the Figure to the right, given that this variable does not contribute to
explain the variation in housing prices in these two models.
Figure 3: Predicted house prices of a standard house
Note: The solid line refers to M1 where the commuting variables are excluded. The dotted line refers to
M3, and the dashed line refers to M5. The values of the accessibility variables is mean-normalized.
Figure 4: Predicted house prices of a standard house
Figure 6, however, there is a substantial diﬀerence between the model M5, and the
more parsimonious model M4. Technically, this is due to the diﬀerent parameter values
estimated for the quadratic term, but it can also be argued that the results reﬂect changes
in the characteristics of the urban structure.
As pointed out in Section 3, a high MCT may reﬂect either a peripheral location, or
a centrally located area with heterogenous agents and considerable excess commuting.
For the model M5, the eﬀects of labour market heterogeneity and a high level of spatial
interaction in the central parts of the urban area seem to dominate, since MCT is
estimated to have a positive impact on housing prices. Over time, the job growth in the
Stavanger urban area has come in areas that used to be the outskirt of the city, mainly in
suburban industrial parks. As a consequence, workers living in residential areas close to
the city centre, see Figure 1, no longer have on average shorter commuting times than
workers living in some of the suburban areas. Still, some of the residential areas close
to the city centre have very high housing prices. These areas have traditionally been
fashionable residential locations, with neighbourhoods that are popular, beyond the pure
urban attraction eﬀect. This is one possible explanation why housing prices are predicted
to be an increasing function of MCT.
The ﬁrst, rising part of the curve resulting from model M4 in Figure 6 can be explained
from the same line of reasoning as above. In this model, however, the labour market
accessibility measure is not incorporated. Since MCT is closely and negatively correlated
to labour market accessibility (Pearson’s correlation coeﬃcient is -0.8677), the labour
REGION : Volume 7, Number 1, 2020


===== PAGE 17 (score 84, numeric-paren 62, table-words 1, var-words 2) =====

A. Gjestland, L. Osland, I. Thorsen
65
Table 3: Estimated direct, indirect, and total impact from variants of M3 and M5 using
the spatial Durbin Estimator
M3
M5
Variable Name
Direct
Indirect
Total
Direct
Indirect
Total
Lotsize
-0.201
0.150
-0.051
-0.173
0.012
-0.161
(-6.47)
(2.21)
(-0.70)
(-4.92)
(0.15)
(-1.84)
LotSize2
0.025
-0.013
0.013
0.024
-0.005
0.020
(9.69)
(-2.20)
(2.01)
(8.73)
(-0.67)
(2.64)
RurLotsize
-0.028
0.007
-0.021
-0.041
0.024
-0.017
(-8.89)
(1.76)
(-8.10)
(-3.92)
(2.21)
(-5.24)
Age
-0.196
0.038
-0.158
-0.161
-0.097
-0.285
(-11.93)
(1.05)
(-4.13)
(-9.39)
(-2.51)
(-6.22)
Age2
0.016
-0.004
0.012
0.009
0.020
0.028
(5.89)
(-0.71)
(1.84)
(2.95)
(3.09)
(4.13)
Garage
0.040
0.026
0.066
0.039
0.009
0.048
(6.48)
(2.06)
(4.70)
(6.78)
(0.63)
(2.86)
LivingArea
0.503
-0.037
0.466
0.487
0.094
0.581
(52.91)
(-1.84)
(21.11)
(53.64)
(3.72)
(20.94)
YearDum04
0.099
-0.029
0.070
0.097
-0.001
0.096
(9.58)
(-1.24)
(2.86)
(9.05)
(-0.04)
(2.22)
YearDum05
0.209
-0.027
0.182
0.207
0.038
0.245
(19.93)
(-1.16)
(7.48)
(19.40)
(1.03)
(5.72)
YearDum06
0.373
-0.014
0.359
0.371
-0.019
0.353
(35.41)
(-0.57)
(14.40)
(35.31)
(-0.53)
(8.50)
YearDum07
0.560
-0.022
0.538
0.557
0.001
0.558
(56.92)
(-0.92)
(21.51)
(54.24)
(0.04)
(14.02)
DistCBD
-0.081
-0.073
-0.153
-0.036
-0.103
-0.139
(-3.45)
(-2.09)
(-4.86)
(0.40)
(-0.97)
(-4.23)
DistCBD2
-0.024
0.008
-0.016
0.014
-0.033
-0.019
(-4.20)
(0.95)
(-1.94)
(0.37)
(-0.87)
(-2.66)
ACCESS
0.084
0.035
0.120
0.161
-0.051
0.109
(7.23)
(1.87)
(6.99)
(2.00)
(-0.63)
(5.22)
IN-COM
-0.073
-0.064
-0.136
(-4.67)
(-2.62)
(-6.18)
OUT-COM
0.040
0.033
0.073
(2.36)
(1.87)
(2.72)
MCT
0.247
-0.122
0.125
(1.80)
(-0.88)
(3.43)
Note: Z-values in parentheses.
The spatial Durbin model yields the same sign on all the estimated parameters, as
for the ordinary least squares estimator. The direct impacts yield results that are within
the 95% conﬁdence region of the ordinary least squares regression results for the model
variant based on M3. For the speciﬁcation based on M5, this is not always the case. For
most of these variables the total impact is within the 95% conﬁdence region. The impacts
related to age variables is not within the mentioned 95% conﬁdence region. For M1, not
reported, the total impact of ACCESS is not within the 95% conﬁdence region. The
indirect impact is negative, but not signiﬁcant.
Indirect impacts are signiﬁcant at the 5% level for the variables LivingArea and Age2
for the model based on M5. The same is the case for the spatial Durbin variant of
M1. In the model based on M3, only IN-COM has a signiﬁcant indirect impact at the
5% signiﬁcance level. The indirect impact is negative. This means that an increase in
IN-COM of the houses neighbouring a house i, will on average have a negative impact
on the price of house i. Hence, there are negative spillover eﬀects of having a house
located in an area with a high relative level of in-commuting. This result is in line with
our interpretations of regression results found in previous sections. The average indirect
REGION : Volume 7, Number 1, 2020


===== PAGE 18 (score 12, numeric-paren 0, table-words 1, var-words 10) =====

66
A. Gjestland, L. Osland, I. Thorsen
impacts of OUT-COM and ACCESS are positive, but not signiﬁcant at the 5% level.
The results of this robustness check also show that all the direct impacts related to the
studied variables have the same sign as the ones documented in Table 2. Given the fact
that the correlation between the variables are relatively high, it is to be expected that
the inclusion and exclusion of variables will have an impact on the value of the estimated
coeﬃcients. The direction of any existing bias is not possible to determine based on any
of the robustness checks.
7.2
Reverse Causality
The second relevant type of endogeneity problem could be reverse causality. By way of
example, neighbourhoods with higher housing prices could attract high-income earners,
perhaps also with higher educational attainments. Locations close to these places could be
relevant for speciﬁc ﬁrms in order to get access to an attractive pool of labour. This type
of reasoning is based on the hypothesis that “jobs follow people”, which subsequently may
improve labour market accessibility in places with higher housing prices. Hoogstra et al.
(2017) provide a useful meta-analysis of the related literature regarding this question.
Their main ﬁnding is that the evidence is inconclusive. The causality between jobs and
people could run in diﬀerent directions (pages 371-372). If this holds true, for our data,
the resulting bias could be minor, and unpredictable.
If there exists reverse causality, identiﬁcation of the impact of accessibility would
necessitate an instrument.
One possible instrument could be an exogenous change
in measures of accessibility. However, we do not have access to any such instrument.
Moreover, we focus on several variables, which are potentially endogenous. Testing for
exogeneity is, hence, not straightforward according to e.g. Baum et al. (2007). We have
to use several instruments, and in these cases, the traditional IV-estimators could be
biased and inconsistent (see also Nordvik et al. 2019, for further discussions of this issue).
Finally, it is also possible to argue that the need for an instrument is less important
in our case, given the results from the spatial Durbin model. This model accounts for
indirect spill over impacts and the spatial Durbin model, in general, is robust to omitted
systematic spatial variation of characteristics (LeSage, Pace 2009).
8
Conclusions
In explaining spatial variation in housing prices, gravity-based accessibility measures have
been suggested as a generalization of modern polycentric labour market structures. From
a computational and data collection perspective, it would be convenient if easily available
information on actual commuting patterns could replace a more complex measure of
commuting potential. According to our results, two of the hypotheses formulated in
Section 3 have to be rejected, however. The labour market accessibility eﬀect is not
adequately represented by the proposed characteristics of observed commuting patterns.
We used a wide range of diﬀerent methods to obtain robust conclusions. Labour market
accessibility in relation to housing prices is best captured by the gravity based potential
variable.
Our results, on the other hand, provide support for the hypotheses HB
0 and HD
0 .
Observed measures of commuting patterns are found to contribute with information that
adds to the eﬀect of a potential measure in explaining spatial variation in housing prices.
In particular, the results from the ordinary least squares and the spatial Durbin estimator
support the hypothesis that a relatively high level of commuting into a zone corresponds
to negative externalities, such as noise, pollution, or other negative eﬀects of heavy
traﬃc and/or industrial activities. There is only weak support for the hypothesis that a
relatively high level of out-commuting from a zone corresponds to attractive neighbourhood
characteristics, which are positively related to housing prices. Average commuting time,
a priori, reﬂect something about the actual commuting costs for households. However,
our interpretation of the positive impact of this variable is that a high level of spatial
interaction in the central parts of the urban area seem to dominate. Over time, the
job growth in the Stavanger urban area has come in areas that used to be the outskirt
of the city. Still, some of the residential areas close to the city centre have very high
REGION : Volume 7, Number 1, 2020


===== PAGE 22 (score 101, numeric-paren 75, table-words 1, var-words 4) =====

70
A. Gjestland, L. Osland, I. Thorsen
A
Appendix: Results of spatial error model estimations
Table A.1: Estimated results for the hedonic house price models based on the spatial
error model formulation as speciﬁed in equation (9)
Variable Name
M0
M1
M2
M3
M4
M5
Constant
13.060
12.506
13.031
12.180
12.710
11.596
(120.88)
(75.10)
(119.33)
(66.87)
(48.96)
(37.09)
Lotsize
-0.187
-0.204
-0.183
-0.202
-0.189
-0.214
(-6.11)
(-6.61)
(-5.94))
(-6.56)
(-6.15)
(-6.93)
LotSize2
0.024
0.026
0.024
0.026
0.024
0.026
(9.55)
(9.98)
(9.39)
(9.99)
(9.58)
(10.30)
RurLotsize
-0.025
-0.021
-0.026
-0.023
-0.025
-0.019
(-8.76)
(-7.26)
(-8.86)
(-7.86)
(-8.73)
(6.47)
Age
-0.1809
-0.178
-0.183
-0.182
-0.182
-0.180
(-10.71)
(-10.51)
(-10.78)
(-10.82)
(-10.78)
(-10.67)
Age2
0.013
0.012
0.013
0.013
0.013
0.013
(4.64)
(4.43)
(4.75)
(4.77)
(4.75)
(4.60)
Garage
0.039
0.040
0.038
0.039
0.039
0.039
(6.71)
(6.82)
(6.59)
(6.70)
( 6.68)
(6.67)
Size of house
0.493
0.491
0.493
0.490
0.493
0.491
(52.31)
(52.19)
(52.31)
(52.27)
(52.47)
(52.36)
YearDum04
0.0955
0.096
0.096
0.096
0.096
0.096
(9.87)
(9.91)
(9.87)
(9.93)
(9.88)
(9.90)
YearDum05
0.199
0.199
0.199
0.199
0.199
0.199
(20.41)
(20.45)
(20.39)
(20.46)
(20.41)
(20.43)
YearDum06
0.370
0.370
0.370
0.370
0.370
0.370
(38.33)
(38.34)
(38.34)
(38.36)
(38.32)
(38.35)
YearDum07
0.5573
0.557
0.557
0.557
0.557
0.556
(58.91)
(58.88)
(58.92)
(58.93)
(58.93)
(58.82)
DistCBD
-0.0574
-0.101
-0.040
-0.105
-0.074
-0.163
(-2.58)
(-4.14)
(-1.58)
(-3.93)
(-2.96)
(-5.71)
DistCBD2
-0.047
-0.029
-0.053
-0.028
-0.045
-0.018
(-11.06)
(-5.07)
(-9.17)
(-3.96)
(-9.48)
(-2.82)
ACCESS
0.059
0.088
0.109
(4.36)
(5.81)
(6.07)
IN-COM
-0.033
-0.082
(-1.78)
(-4.39)
OUT-COM
-0.006
0.034
(-0.22)
(1.65)
MCT
0.315
0.306
(1.48)
(1.47)
MCT2
-0.063
-0.035
(-1.44)
(-0.80)
λ
0.374
0.364
0.373
0.359
0.374
0.358
(23.67)
(22.85)
(23.60)
(22.30)
(23.60)
(22.24)
Log-likelihood
1420.26
1429.58
1422.56
1440.23
1422.69
1440.57
Note: Z-values in parentheses.
c⃝2020 by the authors. Licensee: REGION – The Journal of ERSA, European
Regional Science Association, Louvain-la-Neuve, Belgium. This article is distri-
buted under the terms and conditions of the Creative Commons Attribution, Non-Commercial
(CC BY NC) license (http://creativecommons.org/licenses/by-nc/4.0/).
REGION : Volume 7, Number 1, 2020