Do rising rents lead to longer commutes? A gravity model of commuting flows in Ireland

Abstract

Chinese

The classical monocentric city model suggests that property prices decrease and transport costs rise with distance to the urban centre, implying that employees face a trade-off between long commutes and high housing costs when making location decisions. Accordingly, some commuters might be forced to take on longer commutes due to rising rents in central locations. In this study, we investigate empirically whether the rental differential between employment centres and residential areas predicts changes in average commuting times. To this end, we consider a gravity model of commuting flows for Ireland over 2011–2016. We present results for Ireland and the metropolitan area of Dublin, which constitutes the largest commuting region in Ireland. The results imply that a 10% rise in rents in employment centres is associated with an up to 0.6 minute rise in one-way daily average commuting times nationally (about 2.2% of the average commute duration).

Keywords

commuting gravity model Ireland rental market

Introduction

The monocentric city model predicts that unit housing costs decrease with distance from the city centre where most of the employment opportunities are located, whereas the costs of transport to the centre are higher for residents in the periphery (Alonso, 1964; Mills, 1967; Muth, 1969). This suggests that workers face a trade-off between low housing costs and long com-mutes when making location decisions. If property prices increase in employment centres relative to suburban areas, it might drive residents away from the centre and force them to accept longer commutes. To get a better understanding of the existence and magnitude of this displacement effect, the present study investigates the relationship between relative rental prices and commuting behaviour using a gravity model of bilateral commuting flows in Ireland.

Ireland is a particularly suitable place to examine this relationship. The period of recovery in the Irish housing market following the property price crash of 2009 allows us to study commuting behaviour and the housing market during a period of dramatic changes, with rental price increases exceeding 30% over 2011–2016 in some areas (RTB, 2019). The tightness of the Irish housing market has been highlighted by a recent survey which found that Irish renters are willing to extend the daily commute by 23 minutes on average in order to achieve home-ownership, pointing to a trade-off between desirable housing and commuting (Corrigan et al., 2019). In addition, the five-yearly Irish census contains spatial data on commuting journeys for the whole population, and it has proved possible to link this information to administrative data on rents at the small area level.

A vast literature relates property prices to spatial structures. The basic prediction that property prices fall with distance to an exogenously defined city centre has been widely tested and generally confirmed, even though the relationship is often distorted by local amenities, infrastructure and geographical factors (Albouy and Lue, 2015). More recent empirical studies account for complex and polycentric urban structures by directly relating property prices to employment opportunities or accessibility measures (Ahlfeldt and Wendland, 2016; Osland and Thorsen, 2008). A distinct field of research analyses commuting behaviour, for example, with a focus on excess commuting (Ma and Banister, 2006), mode of transport choices (Johansson et al., 2006) or the gender gap in commuting (Crane, 2007).

In comparison, less attention has been paid to the relationship between housing costs and commuting behaviour, even though existing evidence suggests that housing costs are an important determinant for commuting behaviour. Cervero and Wu (1997) relate commuting flows to house prices at the location of residence and workplace using a cross-sectional gravity model. They find that residential house prices are negatively associated with the number of commuters between two locations, implying that low house prices in peripheral locations act as a pull factor. So et al. (2001) distinguish between four location choices: living and working in a metropolitan area, living and working in a non-metropolitan area, and cross-commuting either from or to a metropolitan area. Their multinomial logit model suggests that a 10% increase in metropolitan housing costs reduces metropolitan residence by 3.4% and increases demand for non-metropolitan housing by 1.9%. Cheshire et al. (2018) offer a public policy perspective on the relationship between housing markets and commuting. Using election outcomes as instrumental variables, they show that planning restrictions lead to an increase in average commuting distances. O’Kelly et al. (2012) develop a calibrated spatial interaction model for Ireland based on 2006 commuting data from the census-based source also used in the present article. Though data on actual rents was not available to them, the authors infer from the spatial structure of commuting trips that there were higher rents in the ‘job-rich core’ (Dublin) than in peripheral locations. However, we are not aware of an empirical analysis of the commuting–housing nexus in the context of dramatic housing market changes. This study attempts to fill this gap.

This research is motivated by the adverse effects of commuting. Aside from the monetary costs for commuters, passive commuting, which excludes walking and cycling, has been associated with negative effects on well-being and mental health (Künn-Nelen, 2016; Sandow et al., 2014; Stutzer and Frey, 2008). While the review of Saunders et al. (2013) concludes that there is some evidence that active commuting (i.e. walking or cycling) can be beneficial, the monetary costs and time spent commuting put an additional burden on workers. Commuters have also been shown to be a risk group for exposure to air pollution (Zuurbier et al., 2010). On the other hand, commuting is an essential part of modern labour markets and restrictions to commuting impede labour market efficiency (Monte et al., 2018). Commuting is also related to urban form. Low density and dispersed urban structures are associated with higher commuting times (Antipova et al., 2011; Travisi et al., 2010). In that sense, demand for long-distance commuting combined with demand for housing in non-metropolitan areas may lead to low-density and dispersed urban structures, commonly referred to as sprawl.

In the first step of our empirical analysis, we develop a gravity model of commuting flows for Ireland in order to explore the link between commuting and rental prices, which we interpret as a general measure of housing user costs.¹ We link commuting data on small areas from the Irish Census, which includes the location of residence and workplace for each Irish citizen, with microdata on tenancy agreements from Ireland’s Residential Tenancy Board. With regard to the estimation strategy, we build on existing gravity models of commuting that explore how the commuting intensity between two locations depends on transport costs, distance and other geographical factors (Ahlfeldt and Wendland, 2016; Ahlfeldt et al., 2015; De Vries et al., 2009; Duran-Fernandez and Santos, 2014; McArthur et al., 2011a, 2011b; Persyn and Torfs, 2016). The gravity models allow us to predict the effect of changes in the relative rental prices of urban centre to periphery on average journey times in the second step of the analysis.

We find a significant association between commuting behaviour and rental prices over the study period 2011–2016. The main result of this study is that a 10% rise in rents in employment centres is associated with up to 0.6 minute higher one-way average journey times nationally, which corresponds to 2.2% of the average commute duration. Our results indicate that the effect is generally larger for Dublin, the largest metropolitan region in Ireland. However, the effects appear moderate in scale, which seems to reflect a relatively slow response by commuters to changes in the spatial structure of housing costs. Nevertheless, the results provide some empirical support for the existence of a trade-off between commuting and housing costs. This has implications for urban planning, as the process of sprawl and dispersion is likely to continue if rents rise faster in the urban centre relative to the commuting belt.

Data sources and data exploration

Commuting data

We obtain commuting data from the Place of Work, School or College (POWSCAR) microdata file.² POWSCAR data is collected as part of the Irish Census, which is conducted by the Central Statistical Office (CSO). We utilise data for the census years 2011 and 2016. Each data point codes the location of residence and place of work for each resident in Ireland. For the analysis of bilateral commuting flows, we aggregate the data to the level of Electoral Divisions (EDs), a geographical unit of which there are 3409 in Ireland.³ The aggregation allows us to compute the number of commuters between each pair of EDs. Intra-ED commuters, home workers and mobile workers are excluded from the analysis. Figure 1 illustrates ED-to-ED commuting flows in Ireland by plotting direct lines between the centroids of origin and destination. Darker lines indicate a larger number of commuters. The metropolitan area of Dublin clearly stands out as the largest commuting region.

Figure 1.

Map of bilateral commuting flows in Ireland in 2016.

The POWSCAR dataset includes information on self-reported one-way journey minutes and primary means of transportation for each commuter, which are summarised in Table 1 for Ireland and the metropolitan area of Dublin.⁴ The average national journey time increased from 25.9 minutes per journey to 27.3 over the course of five years, implying a rise of up to 14 minutes on a weekly basis (i.e. when assuming 10 journeys per week). With 30.5 and 32.2 minutes in 2011 and 2016, commuting times are above the national average in Dublin, which is consistent with congestion effects due to higher population and employment density.

Table 1.

Self-reported one-way journey time and means of travel in 2011 and 2016 by region.

Means of transport in %	National		Dublin
	2011	2016	2011	2016
Active commuting	13.3	13.55	15.78	16.58
Foot	10.70	10.15	11.76	11.15
Bicycle	2.60	3.40	4.02	5.43
Public transport	9.44	10.36	17.45	18.77
Bus	5.87	6.48	10.28	11.06
Train	3.57	3.88	7.17	7.71
Private motorised transport	77.22	76.08	66.75	64.64
Motorbike	0.55	0.47	0.82	0.67
Drive car	67.70	66.86	59.22	57.57
Passenger in car	4.45	4.44	3.67	3.50
Van	4.52	4.31	3.04	2.90
Working from home	0.02	0.01	0.02	0.01
One-way journey to work in minutes	25.91	27.32	30.46	32.23

Private motorised vehicles are the dominant form of transportation. In 2011, 77.2% of commuters either drove or were a passenger in a car, van or motorbike, although the share of motorised commuting dropped by more than 1 percentage point between 2011 and 2016. Public transport, on the other hand, increased from 9.4% to 10.4% over the same period. Compared with the national average, Dublin exhibits a higher level of public and non-motorised commuting. The data for Dublin also indicates a shift towards active commuting and public transport, which both increased from 2011 to 2016.⁵

A decomposition of travel times

Two factors may help to explain the rise in average journey times: first, changes in the time it takes to travel from location i to j due to congestion or infrastructure improvements, and secondly, changes in the origin–destination structure (which in turn is determined by commuters’ residence–workplace location decisions). For exploratory purposes, we decompose the rise in commuting times from 25.9 to 27.3 minutes. To this end, we first consider a simple linear regression of journey times against a constant and a binary 2016 indicator:

\begin{matrix} τ_{c, t} = \underset{(0.018)}{25.913} + \underset{(0.024)}{1.410} \times year 2016_{c, t} \\ + {\hat{u}}_{c, t}, R^{2} = 0.1 % \end{matrix}

(1)

Here, $τ_{c, t}$ is the travel time of commuter $c = c (i, j)$ who commutes from i to j in year t. The year dummy equals 1 in year 2016, 0 otherwise. The resulting point estimate on the year dummy corresponds to the percentage-point increase in average commuting times. When adding origin–destination fixed effects,

\begin{matrix} τ_{c, t} = \underset{(0.015)}{0.705} \times year 2016_{c, t} + {\hat{μ}}_{ij} \\ + {\hat{u}}_{e, t}, R^{2} = 70 %, \end{matrix}

(2)

the point estimate falls to 0.71 percentage points and can be interpreted as the change in travel times due to congestion effects and changes in infrastructure, that is, when holding the spatial structure of origin–destination pairs constant. The increase in the overall R² from 0.1% to 70.0% provides evidence that commuting times are predominantly driven by location choices rather than by changes in travel time for given origin–destination pairs. The empirical analysis of this study thus focuses on using a gravity model to explain how commuters choose the location of residence and work.

Rent price data

Rental prices are taken from Ireland’s Residential Tenancy Board (RTB), which provides information and dispute resolution services to tenants and landlords.⁶ All tenancy agreements in Ireland are required to be registered with the RTB in accordance with the Residential Tenancies Act 2004. Landlords may face fines of up to €4000 and/or six months’ imprisonment if they fail to register. In addition to the rent level, the data records property-level characteristics (e.g. number of rooms, type of house). More than 78% of observations have Eircodes (Irish postal codes), which permit the rental data to be matched with EDs.⁷

We have rental data available going back to 2007, Quarter 3. Since the commuting data is only available for two points in time (2011 and 2016), we calculate average rents for the years 2007–2011 and 2012–2016, respectively, which we relate to commuting in 2011 and 2016. This approach also accounts for persistence in commuting behaviour, as commuting flows recorded in 2011 and 2016 are likely to be determined by conditions that manifest throughout the previous years. Thus, taking only current or lagged rents seems inappropriate. Furthermore, we discard EDs for which there are fewer than 30 rent data points. Sufficient rental price data is available for 1970 out of 3409 Irish Electoral Divisions. Summary statistics are provided in the Supplemental Material.

Figure 2 shows a time-series graph of average rents where EDs are split into quartiles based on the empirical employment density distribution. The rental price level is strictly ordered by density groups, with high-density EDs exhibiting higher rent levels across all years. The drop in rents after 2007 is striking. Rents in the second group (between first and second quartile) fell by 28.7% between the second half of 2007 and 2011. A prolonged recovery only started in 2013, but the speed of recovery varied across groups. Rental price growth has been more pronounced for above-median EDs and EDs above the 3rd quartile, with growth rates of 19.4% and 26.3% over 2011–2016. For comparison, the growth figures for the EDs below the 1st and 2nd quartile are 9.5% and 9.7%, respectively, suggesting a divergence in rental prices between areas with high and low employment density.

Figure 2.

Time series graph of average rents grouped by employment density quartile.

A cross-section of Dublin

Given its relative size and importance, we discuss the Dublin metropolitan area separately. Figure 3 provides insights into the urban structure of Dublin, the largest commuting area in Ireland. Sub-figures (a) to (d) show average journey time (in minutes), the job density, monthly rent (in €) and population density on the vertical axis. The horizontal axis is the distance to the centre of Dublin city, which we take to be the Spire of Dublin (a 120 metre-high monument on O’Connell Street). While this is a somewhat arbitrary location, it is only used here for illustrative purposes. The points in Figure 3 are averages per ED and the lines are fitted values from linear regressions with higher-order polynomials. Note that sub-figures (b) to (d) use logarithmic scales for the vertical axis.

Figure 3.

Dublin cross-section. Journey time (minutes), job density, monthly rent (in €) and population density as a function of distance to the centre of Dublin.

The relationship between journey time and distance resembles an inverted u-shaped function. Journey times are between 20 and 30 minutes in direct proximity to the centre and increase up to a peak of 35 minutes on average at a distance of around 35 to 40 km. Above 40 km from the centre, travel times decrease; this is presumably because many residents in these areas do not commute to the centre of Dublin, but rather travel to alternative employment centres around Dublin. Job density, monthly rents and population density decrease with distance, but level off above 40 km. Rents exhibit a much wider spread around the centre, in the range of €750 to €2000. By comparing the job and population density, it can be seen that, as typical for urban structures, the centre of Dublin shows a higher density of jobs than population, while the two regression curves seem to align as we move away from the centre.

Methodology

The empirical analysis is based on gravity models. An approach using journey times directly as the dependent variable would require us to take either the workplace or the place of residence as given, with the distance to the workplace or distance to the residence as the commuters’ choice variable. However, residence–job decisions are often taken simultaneously. A gravity model on the other hand allows us to explain bilateral commuting flows through both origin and destination factors. Another approach allowing for simultaneous location choice is to use multinomial models (So et al., 2001; Vega and Reynolds-Feighan, 2008). These models, however, rely on ad hoc categorisation (e.g. urban versus non-urban) and impose a limit on the number of spatial units, whereas a gravity model allows us to explore commuting at a spatially disaggregate level.

A cross-sectional model of commuting

Ahlfeldt and Wendland (2016) consider the cross-section commuting model

π_{ij} = τ t_{ij} + o_{i} + d_{j} + u_{ij} .

(3)

The dependent variable π_ij is the logarithm of the probability of commuting from residence i to place-of-work j, that is, π_ij = lnC_ij/(P_i), where C_ij is the number of commuters from i to j and P_i is the number of residents; t_ij is a measure of geographic distance or travel time between i and j. The parameters o_i and d_j are push and pull factors that capture the attractiveness of i and j as locations of residence and workplace. For example, o_i may include proximity to natural sights and d_j could depend on the accessibility of location j.

The model (1) provides insights into the spatial decay of commuting probabilities, and has been used to study urban structures. There are, however, a few drawbacks to the model: first, the additive fixed effects o_i and d_j capture global push and pull factors. In other words, model (1) assumes that, for example, ED j is equally attractive as a location for work for all i. However, the attractiveness of ED j as a workplace is likely to vary across i, even when observable measures of distance and travel time are accounted for. Secondly, this cross-section model does not allow one to disentangle the effect of specific location factors, such as rental prices, from other local characteristics captured by the push and pull factors. Both issues can be addressed in a two-period panel framework, which can exploit changes in commuting behaviour over time.

A linear first-difference gravity model of commuting

We consider an extension of the above base specification:

π_{ij, t} = f (r_{i, t}, r_{j, t}) + x'_{i, t} θ + x'_{j, t} δ + μ_{ij} + ε_{ij, t}

(4)

Here, the dependent variable is defined as above, but we add the time index t. The vectors $x_{i, t}$ and $x_{j, t}$ are observable, time-varying characteristics of locations i and j. These include the number of jobs, the number of residents and socio-economic characteristics. The control variables include property characteristics from the RTB database as well as the number of residents, the number of jobs, other demographic factors and socio-economic variables which are taken from the CSO Census.⁸

The pair-wise fixed effects µ_ij captures the time-invariant attractiveness of commuting from residence location i to place-of-work location j. µ_ij depends on distance which is in principle observable (as Euclidean or road distance) and other factors that are not easily captured with readily available data. For example, the demand for commuting from i and j may be low if, for example, i has a high share of luxury apartments and j is dominated by low-skilled manual labour, implying that residents of i have a low propensity of commuting to j. To account for these unobserved factors, we estimate (4) in first differences:

Δ π_{ij, t} = Δ f (r_{i, t}, r_{j, t}) + Δ x'_{i, t} θ + Δ x'_{j, t} δ + Δ ε_{ij, t}

(5)

The main interest lies in the role of r_i and r_j, which denote the rental price at location i and j in logarithmic terms. We consider two specifications for the link function f(r_it, r_jt):

f (r_{i t}, r_{j t}) = α (r_{j t} - r_{i t}) (rental differential)

(6)

f (r_{i t}, r_{j t}) = β_{1} r_{i t} + β_{2} r_{j t} (additive effect)

(7)

The aim is to estimate the effects α, β₁ and β₂. We refer to r_jt− r_it as the rental differential. The rationale for considering the rental differential specification is that location decisions may be determined by relative prices between centre and periphery rather than the absolute levels only. The testable hypothesis is that, ceteris paribus, a larger rental price gap between destination and origin location makes commuting more worthwhile, thus increasing the commuting probability. For example, a positive coefficient on the rental differential, that is, α > 0, implies that the commuting probability is increasing in the rental differential. The additive model provides insights into whether location decisions are primarily driven by pull factors (low rents in the periphery) or by push factors (high rents in the centre). With regard to the additive model, we expect β₁< 0 and β₂ > 0.

Post-double-selection Lasso

Since gravity models are concerned with the relation between two units, that is, origin and destination location, there is a need to adjust for confounding factors both at the origin and the destination. Given the large set of putative controls, an empirical concern is to select the right set of control variables. The total number of variables included in x _i,t (residence-specific controls) and x _j,t (place-of-work controls) is 192, many of which are highly collinear. Omitting relevant control variables induces an omitted variable bias, whereas including too many leads to over-fitting and inefficient causal inference.

To appropriately account for confounding factors, we consider the post-double-selection strategy of Belloni et al. (2014), which provides a data-driven method for selecting controls that is rooted in the Machine Learning literature. The framework has been extended to the panel data setting with fixed effects in Belloni et al. (2016). The methodology relies on employing the Lasso estimator from Tibshirani (1996). The Lasso is a linear regularisation method that, like OLS, minimises the squared error, but at the same time penalises the absolute size of co-efficient estimates. Due to this penalisation approach, the Lasso sets some coefficient estimates to exactly zero, and thus removes some variables from the regression model.

The double-selection technique allows exploitation of the strength of the Lasso as a prediction technique for causal inference. Unlike standard approaches for variable selection, such as the general-to-specific approach, the method does not suffer from pre-testing bias issues. The double-selection algorithm proceeds as follows: (a) the dependent variable (here, π) is regressed against the set of controls (i.e. x _i, x _j); (b) the variables of interest, that is, either r_i and r_j or(r_j− r_i), are regressed against the controls; and (c) the final estimate is the OLS estimate of π against the variables of interest and the controls selected by the Lasso in steps (a) and (b). The advantage of the post-double-selection approach is improved and robust causal inference on the parameters of interest (β₁, β₂ and α in our case) without the need for manual and error-prone model selection. A drawback is that the approach cannot provide valid inference for coefficients on individual controls.

Over-dispersion of the dependent variable

Another concern for the estimation is due to the nature of the dependent variable. As is common with count data, the number of commuters is highly over-dispersed: 51.4% have only one commuter and 90.2% have 10 or fewer commuters.

We consider two different approaches to address the issue. First, as in Ahlfeldt and Wendland (2016), we employ a linear model where the dependent variable is the logarithm of the probability of commuting ED-to-ED, rather than the number of commuters. This seems to address the issue of over-dispersion to some extent (see Figure S1 in the Supplemental Material). In addition, we also consider a fixed effects Poisson model with the count of commuters as the dependent variable, which is shown to be robust even in the presence of over-dispersion (Wooldridge, 1999). While the fixed effects negative binomial estimator is another popular choice in this setting, Guimarães (2008) emphasises that it only controls for fixed effects under a narrow set of assumptions.

Results

Rental prices and commuting flows

We first present results for the linearised gravity model in (5) with rental differential and additive rental effects as defined by equations (6) and (7). The model is estimated in first differences to eliminate the pairwise fixed effect. The estimation results are shown in Table 2. The first two models include bilateral commuting flows with at least one commuter, while the third and fourth models include flows with at least five commuters to verify the robustness of results. In each table, we present OLS results without additional controls and post-double-selection results, where the set of control variables is chosen by the Lasso estimator as proposed by Belloni et al. (2014). We confine the discussion and interpretation of the results to the rental variables.⁹

Table 2.

Commuting flows in Ireland and Dublin. Linear specification in first-differences.

	(1)	(2)	(3)	(4)
Dependent variable: $Δ π_{ij, t}$
Panel A: Rental differential
OLS
$Δ (r_{jt} - r_{it})$	0.204***	0.224***	0.0879*	0.0921
	(0.0257)	(0.0345)	(0.0381)	(0.0524)
Post-double-selection
$Δ (r_{jt} - r_{it})$	0.0317	0.00373	0.234***	0.525***
	(0.0359)	(0.0601)	(0.0515)	(0.0924)
Panel B: Additive effect
OLS
$Δ r_{jt}$	0.159***	0.176***	0.0598	−0.0835
	(0.0308)	(0.0509)	0.0419)	(0.0667)
$Δ r_{it}$	−0.245***	−0.257***	−0.119**	−0.257***
	(0.0310)	(0.0431)	(0.0459)	(0.0684)
Post-double-selection
$Δ r_{jt}$	−0.0456	−0.0450	0.227***	0.891***
	(0.0483)	(0.101)	(0.0645)	(0.161)
$Δ r_{it}$	−0.127**	−0.0102	−0.211**	−0.286**
	(0.0481)	(0.0733)	(0.0674)	(0.111)
Obs.	78,620	58,270	32,471	23,526
Region	National	Dublin	National	Dublin
Threshold	1	1	5	5

Notes: Standard errors in parentheses are robust to arbitrary forms of heteroskedasticity. *p < 0.05, **p < 0.01, ***p < 0.001.

The OLS regression results in Panel A in Table 2 indicate that commuting probabilities are increasing in the rental differential when adjusting for origin–destination fixed effects. The post-double-selection methodology, which adds Lasso-selected controls to the estimation, confirms this insight. For example, the point estimates in Panel A, column (3) suggest that a 1% rise in the rental differential at the origin is associated with a 0.2% rise in commuting probability. We find the association to be significant at the 5% level in five out of eight specifications. Panel B in the same table separates the rental differential into prices at the location of residence and workplace. The results reveal that the statistical association appears to be more robust for rental prices at the location of residence.

One concern is that the linear specification does not appropriately capture the relationship between rents and commuting. Table 3 considers the fixed effects Poisson estimator, both with rental differential and additive effects, while we include the same set of controls as in the linear specifications above. The effects are generally larger in magnitude compared with the linearised first-difference gravity model in Table 2. In particular, the association with workplace rents is statistically significant, with point estimates of 0.22 and 0.79 for the national and Dublin sample. The discrepancy in magnitude between the linearised specification and the Poisson model might be because the former is less robust to over-dispersion and also known to be biased in the presence of heteroskedasticity (Silva and Tenreyro, 2006).

Table 3.

Commuting flows in Ireland and Dublin. Fixed effects Poisson.

	(1)	(2)
Dependent variable: C_ij
Panel A: Rental differential
$(r_{jt} - r_{it})$	0.203***	0.354***
	(0.0359)	(0.0596)
Panel B: Additive effect
$r_{jt}$	0.217***	0.787***
	(0.0436)	(0.108)
$r_{it}$	−0.0863	−0.0706
	(0.0472)	(0.0722)
Obs.	157,240	116,540
Region	National	Dublin

Notes: Standard errors in parentheses are robust to both arbitrary heteroskedasticity and within-ED correlation. All models include time effects and ED-level fixed effects. *p < 0.05, **p < 0.01, ***p < 0.001.

We refrain from drawing causal inferences based upon the estimation results in the previous sections. Rents themselves are determined by the interaction of property demand and supply, which in turn is affected by past rents. However, the gravity model allows us to establish a statistical association that we can exploit for the second part of the analysis. The ultimate aim of the analysis is to predict the partial equilibrium effect of changes in the spatial structure of rents on commuting times.

Commuting times

The previous section has provided evidence for a statistical association between rental prices and commuting flows. We now predict the effect of changes in rental prices on commuting times. For this purpose, we induce a 10% shock in employment centres which we define as EDs for which the job density is above the 50th, 75th and 90th percentiles. The predicted changes in one-way journey times are shown in Table 4 and are based on the fixed effects Poisson models. For the calculation, we assume that bilateral journey times are fixed, that is, we rule out changes in the journey duration from i to j due to, for example, changes in infrastructure. We obtain the expected change in total journey times by first predicting the change in the number of commuters between each origin and destination following the 10% shock in rents. We then multiply the predicted change in the number of commuters by the average journey time in 2016, which we hold constant.

Table 4.

Predicted change in one-way journey minutes.

	National			Dublin
	Mean	Low	High	Mean	Low	High
Difference
50th percentile	0.244	0.115	0.374	0.431	0.214	0.649
75th percentile	0.176	0.087	0.265	0.310	0.161	0.460
90th percentile	0.083	0.043	0.123	0.146	0.079	0.213
Additive
50th percentile	0.387	0.159	0.614	1.667	1.118	2.216
75th percentile	0.227	0.109	0.346	0.915	0.626	1.204
90th percentile	0.101	0.051	0.150	0.395	0.273	0.517

Notes: ‘Mean’ is the average rise in journey times. ‘Low’ and ‘high’ are the averages of the 95% confidence interval.

Average predicted changes in journey minutes along with confidence intervals are presented in Table 4. Point estimates range between 0.08 and 0.39 on the national level and between 0.15 and 1.67 for Dublin. For example, the predicted change in journey minutes based on the differential model is 0.24 and 0.43 for Ireland and the Dublin region when relative rents in EDs with above-median employment density increase by 10%. Naturally, the effect size is smaller when only the top 25% or 10% of EDs are exposed to the shock. This finding is in line with the expectation that high rental costs in urban centres provide an incentive for commuting, while high costs in suburban areas discourage commuting.

Conclusion

A common narrative holds that high housing costs force people to commute over longer distances. In this study, we investigate the statistical association between commuting patterns and housing costs, which we approximate using rental prices. We relate bilateral commuting flows to spatially disaggregated rent data utilising a panel gravity model that accounts for pairwise, that is, origin–destination, fixed effects. The results for Ireland are consistent with the notion that the demand for commuting increases with the rental differential between two locations. In other words, we find evidence that workers face a trade-off between high commuting costs and low housing costs when making location decisions. For example, a 10% rise in rents within the top quartile of employment centres is associated with daily one-way commutes being longer by around 0.1 to 0.3 minutes nationally and by around 0.2 to 1.2 minutes for the Dublin metropolitan area. While our method accounts for time-invariant factors affecting commuting and we apply data-driven methods to adjust for a range of observable confounding factors, future work is needed to establish if the associations we have reported are causal.

The developments in the Irish rental market described in this article reflect a period of pronounced changes. Other places and times are likely to experience fewer extreme developments in price levels and relative prices. Nevertheless, our results underline the potential conflict between trends in local planning restrictions and broader environmental and social objectives such as climate policy or efforts to improve public health. It is well established that restrictive planning can raise house prices (Hilber and Vermeulen, 2016). Local pressure for restrictions is likely to be more pronounced where prices of existing properties are higher and local interest groups are longer-established and better resourced (Taylor, 2013; Taylor and Hurley, 2016). In a monocentric city these places would likely be closer to the city centre. This mechanism could provide an impetus for the spatial housing cost gradient to steepen over time, contributing to growing commuting distances.

If journey times for commuters continue to rise, this is likely to have negative monetary, health and environmental consequences. The costs of longer journeys will fall partly on individuals but also on government budgets, as many countries or cities subsidise particular transport modes. The existing challenges of developing appropriate decarbonised transport systems will be compounded. Long commutes can harm well-being and mental health, and commuters may be exposed to additional air pollution. Our analysis also suggests that demand for housing in suburban areas rises with increasing rents in the city centre, which in turn may lead to sprawl: a dispersed low-density urban structure. Finally, there may be distributional consequences to added commuting, as some socio-economic groups bear more of the costs of adjustment than others.

Supplemental Material

USJ910698_Supplemental_Material – Supplemental material for Do rising rents lead to longer commutes? A gravity model of commuting flows in Ireland

Supplemental material, USJ910698_Supplemental_Material for Do rising rents lead to longer commutes? A gravity model of commuting flows in Ireland by Achim Ahrens and Sean Lyons in Urban Studies

Footnotes

Acknowledgements

We are grateful for funding from the ESRI Environmental Research Programme, which in turn is funded by Ireland’s Environmental Protection Agency. We would also like to thank three anonymous reviewers for their comments, the Central Statistics Office of Ireland and the Residential Tenancy Board for providing the data used in this study, as well as participants at the 2019 Irish Economic Association Conference and the 58th European Regional Science Association Congress for helpful comments. The usual disclaimer applies.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the ESRI Environmental Research Programme, which in turn is funded by Ireland’s Environmental Protection Agency.

Notes

ORCID iD

Sean Lyons

References

Ahlfeldt

Wendland

(2016) The spatial decay in commuting probabilities: employment potential vs. commuting gravity. Economics Letters 143: 125–129.

Ahlfeldt

Redding

Sturm

, et al. (2015) The economics of density: Evidence from the Berlin Wall. Econometrica 83(6): 2127–2189.

Ahrens

Hansen

Schaffer

(2018) PDSLASSO: Stata module for post-selection and post-regularization OLS or IV estimation and inference. Statistical Software Components, Boston College Department of Economics. Available at: https://ideas.repec.org/c/boc/bocode/s458459.html (accessed 4 March 2020).

Ahrens

Hansen

Schaffer

(2019) lassopack: Model selection and prediction with regularized regression in Stata. Available at: http://arxiv.org/abs/1901.05397 (accessed 4 March 2020).

Albouy

Lue

(2015) Driving to opportunity: Local rents, wages, commuting, and sub-metropolitan quality of life. Journal of Urban Economics 89: 74–92.

Alonso

(1964) Location and Land Use: Toward a General Theory of Land Rent. Cambridge, MA: Harvard University Press.

Antipova

Wang

Wilmot

(2011) Urban land uses, socio-demographic attributes and commuting: A multilevel modeling approach. Applied Geography 31(3): 1010–1018.

Belloni

Chernozhukov

Hansen

(2014) Inference on treatment effects after selection among high-dimensional controls. Review of Economic Studies 81: 608–650.

Belloni

Chernozhukov

Hansen

, et al. (2016) Inference in high dimensional panel models with an application to gun control. Journal of Business & Economic Statistics 34(4): 590–605.

10.

Borgersen

Sommervoll

(2012) Housing careers, price–rent ratios and rental equivalence. Housing, Theory and Society 29(3): 276–287.

11.

Cervero

(1997) Polycentrism, commuting, and residential location in the San Francisco Bay area. Environment and Planning A: Economy and Space 29(5): 865–886.

12.

Cheshire

Hilber

Koster

(2018) Empty homes, longer commutes: The unintended consequences of more restrictive local planning. Journal of Public Economics 158: 126–151.

13.

Commins

Nolan

(2011) The determinants of mode of transport to work in the Greater Dublin Area. Transport Policy 18(1): 259–268.

14.

Corrigan

Cotter

Hussey

(2019) The housing aspirations and preferences of renters. IGEES Research Paper. Available at: https://www.housing.gov.ie/sites/default/files/publications/files/20190306b_renters_aspirations_and_preferences.pdf (accessed 4 March 2020).

15.

Crane

(2007) Is there a quiet revolution in women’s travel? Revisiting the gender gap in commuting. Journal of the American Planning Association 73(3): 298–316.

16.

De Vries

Nijkamp

Rietveld

(2009) Exponential or power distance-decay for commuting? An alternative specification. Environment and Planning A: Economy and Space 41(2): 461–480.

17.

Duran-Fernandez

Santos

(2014) Gravity, distance, and traffic flows in Mexico. Research in Transportation Economics 46: 30–35.

18.

European Commission (2017) Technical manual on owner-occupied housing and house price indices. Technical report. Available at: https://ec.europa.eu/eurostat/documents/7590317/0/Technical-Manual-OOH-HPI-2017/ (accessed 4 March 2020).

19.

Guimarães

(2008) The fixed effects negative binomial model revisited. Economics Letters 99(1): 63–66.

20.

Hilber

Vermeulen

(2016) The impact of supply constraints on house prices in England. Economic Journal 126(591): 358–405.

21.

Johansson

Heldt

Johansson

(2006) The effects of attitudes and personality traits on mode choice. Transportation Research Part A: Policy and Practice 40(6): 507–525.

22.

Künn-Nelen

(2016) Does commuting affect health? Health Economics 25(8): 984–1004.

23.

Banister

(2006) Excess commuting: A critical review. Transport Reviews 26(6): 749–767.

24.

McArthur

Kleppe

Thorsen

, et al. (2011a) The impact of monetary costs on commuting flows. Papers in Regional Science 92(1): 69–86.

25.

McArthur

Kleppe

Thorsen

, et al. (2011b) The spatial transferability of parameters in a gravity model of commuting flows. Journal of Transport Geography 19(4): 596–605.

26.

Mills

(1967) An aggregative model of resource allocation in a metropolitan area. The American Economic Review 57(2): 197–210. Available at: https://www-jstor-org.web.bisu.edu.cn/stable/1821621 (accessed 4 March 2020).

27.

Monte

Redding

Rossi-Hansberg

(2018) Commuting, migration, and local employment elasticities. American Economic Review 108(12): 3855–3890.

28.

Muth

(1969) Cities and Housing: The Spatial Patterns of Urban Residential Land Use. Chicago, IL: University of Chicago Press.

29.

O’Kelly

Niedzielski

Gleeson

(2012) Spatial interaction models from Irish commuting data: Variations in trip length by occupation and gender. Journal of Geographical Systems 14: 357–387.

30.

Osland

Thorsen

(2008) Effects on housing prices of urban attraction and labor-market accessibility. Environment and Planning A: Economy and Space 40(10): 2490–2509.

31.

Persyn

Torfs

(2016) A gravity equation for commuting with an application to estimating regional border effects in Belgium. Journal of Economic Geography 16(1): 155–175.

32.

RTB (2019) Rent index 2019 quarter 1. Technical report, Dublin. Available at: https://onestopshop.rtb.ie/images/uploads/Comms%20and%20Research/RTB_Rent_Index_2019_Q1_V2.pdf (accessed 4 March 2020).

33.

Sandow

Westerlund

Lindgren

(2014) Is your commute killing you? On the mortality risks of long-distance commuting. Environment and Planning A 46(6): 1496–1516.

34.

Saunders

Green

Petticrew

, et al. (2013) What are the health benefits of active travel? A systematic review of trials and cohort studies. PLoS One 8(8): 1–13.

35.

Silva

JMC

Tenreyro

(2006) The log of gravity. The Review of Economics and Statistics 88(4): 641–658.

36.

Orazem

Otto

(2001) The effects of housing prices, wages, and commuting time on joint residential and job location choices. American Journal of Agricultural Economics 83(4): 1036–1048.

37.

Stutzer

Frey

(2008) Stress that doesn’t pay: The commuting paradox. The Scandinavian Journal of Economics 110(2): 339–366.

38.

Taylor

(2013) Do house values influence resistance to development? A spatial analysis of planning objection and appeals in Melbourne. Urban Policy and Research 31(1): 5–26.

39.

Taylor

Hurley

(2016) Do objections count? Estimating the influence of residents on housing development assessment in Melbourne. Urban Policy and Research 34(3): 269–283.

40.

Tibshirani

(1996) Regression shrinkage and selection via the Lasso. Journal of the Royal Statistical Society. Series B (Methodological) 58(1): 267–288.

41.

Travisi

Camagni

Nijkamp

(2010) Impacts of urban sprawl and commuting: A modelling study for Italy. Journal of Transport Geography 18(3): 382–392.

42.

Vega

Reynolds-Feighan

(2008) Employment sub-centres and travel-to-work mode choice in the Dublin region. Urban Studies 45(9): 1747–1768.

43.

Viegas

Martinez

Silva

(2009) Effects of the modifiable areal unit problem on the delineation of traffic analysis zones. Environment and Planning B: Planning and Design 36(4): 625–643.

44.

Wooldridge

(1999) Distribution-free estimation of some nonlinear panel data models. Journal of Econometrics 90(1): 77–97.

45.

Zuurbier

Hoek

Oldenwening

, et al. (2010) Commuters’ exposure to particulate matter air pollution is affected by mode of transport, fuel type, and route. Environmental Health Perspectives 118(6): 783–789.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.22 MB