Does commuting mode choice impact health?

Governments around the world are encouraging people to switch away from sedentary modes of travel towards more active modes, including walking and cycling. The aim of these schemes is to improve population health and to reduce emissions. There is considerable evidence on the latter, but relatively little on the former. This paper investigates the impact of mode choice on physical and mental health. Using data from the UK Household Longitudinal Study, we exploit changes in mode of commute to identify health outcome responses. Individuals who change modes are matched with those whose mode remains constant. Overall we ind that mode switches affect both physical and mental health. When switching from car to active travel we see an increase in physical health for women and in mental health for both genders. In contrast, both men and women who switch from active travel to car are shown to experience a signiicant reduction in their physical health and health satisfaction, and a decline in their mental health when they change from active to public transport. health status. Using regression methods we then compare health outcomes between treated individuals and their matched controls. Conditional on the validity of selection on observables 3 , the approach identiies a causal effect of a change in commuting mode on health. Our main outcomes of interest are summary measures of mental and physical health derived from the Short Form 12 (SF12), and self‐reported satisfaction with health. Our indings show that adopting active means of travel improves health, for both men and women. Changing from an active mode to either public transport or car travel has a negative impact on health. Further analyses, comparing outcomes in the short and intermediate term, conirm our main results. beneits to physical health are more pronounced. Mode changes between car and public transport do not lead to notable affects on physical or mental health outcomes or satisfaction with health. Overall, our results lend support to UK policy initiatives designed to encourage people to move away from car commuting towards more active forms of travel. As well as the health effects estimated here, this will also help the UK government to meet its targets for reducing emissions.

information they often suffer from endogeneity bias, due to the joint choices individuals may make over commuting mode and health. 2 In order to estimate meaningful effects of the impact of mode of transport, we need to address issues of unobserved preferences and changes in mode choices occurring due to health related reasons. We tackle this by providing evidence of the effects of changes in commuting mode on health for adults in employment in the United Kingdom. Commuting is the most frequent reason for travel for working age individuals; the average UK commuter spends nearly an hour a day traveling and this is increasing over time (Department for Transport, 2017).
Taking advantage of the UK Household Longitudinal Study (UKHLS) which has a large sample size, and rich information on health and labor market experiences, we analyze the effect of changes to commuting mode on physical and mental health. A key feature of the data is that there are a suficient number of individuals who are observed to change their mode of commute (the "treated" individuals) and we have an extensive pool of potential controls for whom commuting mode remains constant. We derive estimates of average treatment effects on the treated (ATTs) by exploiting matching methods (via entropy balancing [EB]; Hainmueller, 2012;Hainmueller & Xu, 2013). We are able to obtain a close balance on confounding covariates, that in part determine both health outcomes and commuting mode choice, across treated and control individuals. Following Ho, Imai, King, and Stuart (2007), we do this by preprocessing the data via matching prior to undertaking parametric modeling. This "doubly robust" approach has the advantage of being robust to either misspeciication in the parametric model but complete covariate balance via matching, or incomplete balance through matching but correct speciication of the regression model. This can be viewed as a way to achieve balance in covariates with the objective of reducing model dependence in the subsequent regressions to extract the ATTs (Abadie & Imbens, 2011).
We follow individuals over time until they change their mode of commute, and compare their health responses to those of a matched control group. We match on socio-demographic characteristics observed pre-treatment, including initial mode, duration of commute and health status. Using regression methods we then compare health outcomes between treated individuals and their matched controls. Conditional on the validity of selection on observables 3 , the approach identiies a causal effect of a change in commuting mode on health. Our main outcomes of interest are summary measures of mental and physical health derived from the Short Form 12 (SF12), and self-reported satisfaction with health. Our indings show that adopting active means of travel improves health, for both men and women. Changing from an active mode to either public transport or car travel has a negative impact on health. Further analyses, comparing outcomes in the short and intermediate term, conirm our main results.

| RELATED LITERATURE
A number of studies have looked at the relationship between mode of commuting and health/well-being. The general consensus is that active commuting has positive effects on physical, mental and overall general health. Evidence from the United Kingdom has consistently suggested that levels of physical activity involved in active modes of commuting, such as walking or cycling, translate into health beneits for individuals; including lower BMI and body fat, enhanced mood and increases in mental and physical health. 4 For example, Laverty, Mindell, Webb, and Millett (2013), using data from the irst wave of the UKHLS, show that, in comparison to the use of private means of transport, the use of public transport, as well as walking or cycling to work, was associated with a lower likelihood of being overweight. Individuals who walked or cycled to work had a lower likelihood of having diabetes, and individuals who walked had a lower likelihood of having hypertension.
The mental health beneits of active travel arise from the fact that it is perceived to be both more relaxing and exciting than other modes of transport (Scheepers et al., 2014). It also promotes higher life satisfaction (Morris, 2015) and is associated with a lower rate of mental distress. MacDonald, Stokes, Cohen, Kofner, and Ridgeway (2010) and Frank, Andresen, and Schmid (2004) suggest that spending more time in cars is associated with increases in obesity and blood pressure, perhaps due to the frustrations of trafic congestion (Stokols, Novaco, Stokols, & Campbell, 1978). Other studies have also concluded that car commuting is stressful and leads to negative mood among drivers 5 Contrasting evidence, however, by Anable and Gatersleben (2005) and Eriksson, Friman, and Gärling (2013) has shown that driving to work provides individuals a positive feeling through greater control and lexibility over their commute. Active travel also has positive effects on the environment since it reduces air pollution (Rabl & deNazelle, 2012), which in turn reduces the risk of cancer (Litman, 2010) and cardiovascular diseases (Genter, Donovan, Petrenas, & Badland, 2008;Hamer & Chida, 2008;Litman, 2010;Scheepers et al., 2014). Studies outside the UK report similar evidence. Turcotte (2005Turcotte ( , 2011, and Páez and Whalen (2010) using Canadian data and Friman, Fujii, Ettema, Grling, and Olsson (2013) using Swedish data, ind that active travel commuters tend to report higher satisfaction than users of other modes; public transport users were least satisied. 6 However, in terms of the effects on health, several studies have concluded that public transport users tend to be physically healthier than car commuters since they meet the recommended level of physical activity more often, as they tend to walk to reach bus or train terminals (MacDonald et al., 2010;Wener & Evans, 2007). Other studies suggest that using public transport causes travelers to experience lower levels of stress (Wener & Evans, 2011).
Little research has explored the effects of changes in travel mode on health. Martin, Goryakin, and Suhrcke (2014) explore the relationship between active travel and psychological well-being using British Household Panel Survey data from 1991-2009. The study relies on ixed effects models to investigate how choice of travel mode, commuting time and switching to active travel impact psychological well-being. They found evidence to suggest that switching to active travel was associated with an improvement in well-being compared to individuals who always commuted by car or public transport. Extending their study using the same dataset, Martin, Panter, Suhrcke, and Ogilvie (2015) examined the effect of switching from private motor transport to active travel or public transport (in the next period) on changes in BMI. They found that those who switched saw a reduction in BMI within a 2 year period.
We advance this literature by taking into account the potential for selection bias and exploiting methods of matching together with parametric regression, to improve identiication of the health impacts of commuting mode choice. We only consider individuals for whom household location is ixed but allow job locations to vary; which may be employer or employee induced. 7 A change in job location may lead to a change in commuting mode through either a change in commuting route and/or distance, or a change in job remuneration allowing, via an income effect, greater choice of travel mode.

| CONCEPTUAL FRAMEWORK
We assume that individuals derive utility (or disutility) from commuting, such that U ¼ Uðm; hðm; tÞ; zÞ, where m represents mode choice and z represents other consumption from which individuals derive utility. Individuals are also assumed to value any health impacts of their commuting mode choice, which will also be a function of time spent commuting represented by hðm; tÞ. Hence individuals derive utility, both directly and indirectly, through their choice of commuting mode. Direct utility may be positive, for example, the enjoyment of driving, the ability to relax or work on public transport, the enjoyment of exercise from walking or cycling to work, or negative, for example, frustration of sitting in heavy trafic, crowded public transport, inclement weather during active commuting. Indirect utility is derived from mode choice through the impact this has on health (Frank et al., 2004;Lancee, Veenhoven, & Burger, 2017;MacDonald et al., 2010;Wener & Evans, 2007). For example, exposure to exhaust fumes or being seated for long periods of time might impact physical health; the uncertainty of disruption during car travel may affect mental wellbeing. Accordingly, commuting mode can be seen as being valued for both a consumption property-the direct impact on utility, and an investment property-the indirect health effects (Grossman, 1972). In making choices over mode, individuals are assumed to maximize utility subject to constraints over income and time. Different forms of travel attract different prices and hence cost to the commuter and are therefore inluenced by an individual's income constraint. Individuals also face a time constraint, which, during the working day, consists of choices over time spent on leisure ðt l Þ, work hours ðt w Þ, and commuting ðt c Þ, such that (t l þ t w þ t c ¼ 24 hrs). The greater time spent commuting, the less time available for other pursuits, assumed mainly to be leisure for individuals with ixed hours of work. In this way, commuting entails an opportunity time cost to the individual and choices over mode will be inluenced by this constraint. Individuals are assumed to choose the commuting mode that maximizes their utility subject to the constraints they face at a particular point in time. Should the value individuals place on the investment and/or consumption properties of mode choice change, or should individuals face changes to their constraints (e.g., through a change in job location or road infrastructure), this may lead to a change in commuting mode.
We are interested in identifying the health effects of commuting mode choice. Our approach considers those individuals who change mode at time t as treated and those who do not change mode as potential controls. By matching controls to treated individuals at time tÀ1 we assume that the average utility of the two groups, prior to treatment, is equivalent. Matching is undertaken on a set of potential confounding characteristics thought, a priori, to inluence both mode choice and health; this includes initial mode and commuting time, health status, and household income among other factors. 8 Adopting a potential outcomes framework, the above procedure assumes that conditional on the set of confounding covariates, x, selection into treatment, d, is independent of potential outcomes, such that ðh 0 ; h 1 Þ ⊥ d|x: where h 0 and h 1 are potential health outcomes for treated individuals without treatment, h 0 , and with treatment, h 1 , respectively. This is often termed the conditional independence assumption (Heckman and Robb (1985)). Where this holds, the can be estimated by replacing the unobserved component Eðh 0 | x; d ¼ 1Þ with its observed counterfactual Eðh 0 | x; d ¼ 0Þ. The conditional independence assumption is required to hold for us to identify the causal effect of change in mode on health outcomes. This requirement is challenging in any empirical application that relies on techniques based on selection on observables (this is also true of methods that are based on selection on unobservables). However, we emphasize that we match on a wealth of pretreatment characteristics, and importantly include pre-treatment mode choice and health. These characteristics are also used in the regression analysis which follows EB. By conditioning on prior health we are effectively considering changes to health due to a change in mode, mimicking a ixed effects approach. We further balance on variables that represent the sequence of observations observed for individuals in the panel dataset. This is intended to balance for potential attrition bias, which, in part, is likely to be driven by unobservables. Following matching, we estimate the treatment effect using a regression framework. The latter helps to mitigate bias resulting from less than perfect matching.

| EMPIRICAL APPROACH
Our empirical strategy exploits changes to mode of commute observed in the data at time t, but occurring somewhere between tÀ1 and t, to identify the responses on health outcomes at time t þ 1. We compare outcomes for 'treated' individuals who experience a change to their mode of commute with outcomes for observationally identical (as of tÀ1) controls, who do not experience a change to their commuting mode. 9 Prior to the occurrence of the change, observational equivalence is deined by a wide set of potential confounding variables, including demographic and individual factors such as age and sex, baseline health and educational achievement; household characteristics such as cohabiting status, number of kids, and household income; and labor market characteristics such as job hours and baseline mode of commuting. These variables are expected to determine, in part, both household health and choices over mode of transport. Conditioning on baseline health and commuting mode is important in deining suitable controls for individuals who are observed to change commuting mode. Including baseline health in both the matching and subsequent regression has the further advantage of identifying health effects from a change to commuting conditional on individual-speciic underlying level of health. This helps to remove unobserved effects speciic to individuals and their health.
Our approach follows the principles set out in Ho et al. (2007) to use matching methods to preprocess the data prior to parametric modeling of outcomes. 10 The aim is to reduce model dependency by using matching to create balance in covariate distributions across treated and control groups. Successful (perfect) matching renders treatment independent of control variables. Subsequent parametric regression modeling of the preprocessed data is therefore less dependent on speciication assumptions and hence more likely to identify consistent causal effects. Where matching proves to be less than perfect, the application of regression techniques conditional on the same set of confounding variables controls for the lack of perfect balance. The approach can be viewed as an extension of the usual matching techniques, which rely on comparisons of means of the matched data.
The matching method we propose, together with subsequent regression modeling assumes that selection into a change of commuting mode can be captured by the set of conditioning variables used. If this assumption does not hold and selection is also a function of unobservable characteristics, then techniques such as instrumental variables would be required. 11 As with many empirical applications, it is profoundly dificult to ind appropriate instruments to identify causal effects; and even where these can be found identiication often leads to very localized treatment effects which suffer from a lack of generalizability. An important consideration when selecting variables to match controls to treated individuals is that these are based on observed characteristics measured prior to treatment. By doing so, we eliminate the possibility of treatment inluencing the set of conditioning variables. As well as the set of characteristics used in previous studies, we further condition on initial commuting mode and health. In this way, the model mimics a ixed effects approach to dealing with individual-speciic unobservable characteristics thought to inluence outcomes and mode choice.
We match individuals based on their characteristics measured at tÀ1, which can lie anywhere between the irst observed wave and the antepenultimate wave (for our main results outcomes are measured at time t þ 1). Sample attrition bias in panel data might arise due to healthier individuals remaining in the panel longer than less healthy counterparts. Health, at least in part, is assumed to determine and be determined by transport mode choice. To mitigate concerns that attrition bias may arise differentially across treated and control individuals we further match on the wave of mode change together with two variables constructed to better approximate the panel proile of treated and control individuals across waves of the data. These are informed by the literature on testing for attrition bias (Jones, Koolman, & Rice, 2006;Verbeek & Nijman, 1992). For each individual we construct variables to represent the total number of waves and the number of consecutive waves they are observed in the dataset. 12 These are included in the matching step.
Matching is undertaken for each of the observed treatments deined by changes in commuting mode: car-public, car-active, public-active and their converse. We then regress outcomes on the set of controls and a treatment effect separately for each of the six matched samples as follows: where β d identiies the treatment effect of interest; the change in mode at time t on health outcomes, h i at time t þ 1. The set of variables used to match controls to treated individuals prior to treatment are represented by X i;tÀ1 (see Table 5 for the variables) and their corresponding relationship with outcomes, β x . 13 λ i;tÀ1 are wave indicators to recognize that mode changes may occur in different calendar years; ϵ i;tþ1 is the usual idiosyncratic error term. Regression weights derived from EB are applied to Model (1). Models for cardinal outcomes are estimated using ordinary least squares; ordered categorical outcomes are estimated with ordered probits. All regressions contain robust standard errors. We use matching techniques to adjust the covariate distribution of the control group data by reweighting and/or discarding units such that it becomes more similar to the covariate distribution in the treatment group. We apply EB (Hainmueller, 2012), which involves a reweighting scheme that directly incorporates covariate balance into the weight function that is applied to the sample units. This is done by selecting a set of weights for each observation in the control group that minimize an entropy distance metric subject to balance and normalizing constraints. This ensures that the weights are nonnegative and sum to unity. These weights satisfy a set of balancing constraints that involve specifying exact balance on moments of the covariate distributions (in our case the mean and variance) in the treatment and the reweighted control group.
All individuals are considered untreated in the irst wave. An individual is assigned only once to the treatment group, when they irst change their mode of commute, any subsequent changes in commuting mode are excluded from analysis. 14 Treated individuals never act as potential controls at any other point in time. Potential control individuals are those who never change their mode of commute while they are observed.
We are concerned with three different commuting modes; car, public transport and active travel; and consider the following changes: car to active travel, public transport to active travel, active travel to car, and active travel to public transport. We have additionally considered switches between car and public transport, but as these do not involve a switch into or out of more active modes, which are often the policy goal, these are not the main focus of our analysis. For each change in mode we match control individuals to treated individuals and then perform regression analysis on the balanced data. Matching is undertaken at tÀ1, mode change is observed at time t and outcomes at t þ 1. We further repeat the analyses (including matching and regression on outcomes) to compare short-run outcomes at time t, and longer term outcomes at t þ 2. 15 An important feature of the literature on commuting is the difference in travel behavior between men and women, with men, on average, undertaking longer commutes. Further, Roberts, Hodgson, and Dolan (2011), ind that the wellbeing of women, but not men, is adversely affected by increased commuting times, while Jacob, Munford, Rice, and Roberts (2019) provide evidence that this is due to the different labor markets in which women and men operate. Accordingly, we undertake heterogeneity analysis by gender and apply EB and regression analysis within gender for each of the mode changes.

| UK Household Longitudinal Study
The UKHLS is a nationally representative sample of UK households, containing panel information on around 100,000 individuals in 40,000 households. We use seven waves of data from 2009 to 2016, containing rich information on socio-economic, health, and labor market characteristics. Health is measured using component scores derived from the SF12 questionnaire. The SF12 uses twelve questions to measure functional health and wellbeing; the responses are aggegated to form the Physical (SF12-PCS) and Mental (SF12-MCS) Component Scores. These are cardinal representations of underlying health status, designed to lie between 0 (lowest level of health) and 100 (highest), and have a mean of 50 and a standard deviation of 10 for the general population (Ware, Keller, & Kosinski, 2002). As an additional outcome we also use responses to questions on satisfaction with health, recorded on a ive point ordered categorical scale, where 1 is least satisied and 5 is most satisied. 16 Our measure of commuting mode is taken from the question "How do you usually get to your place of work?" which is asked only to people who state they are in employment. Responses are categorized as Car (drivers and passengers), Public transport (bus, train, undergound) and Active travel (cycle, walk) with Other (taxi, moped, other mode) as an alternative group that we do not consider due to small sample sizes. To control for individual preferences we condition on characteristics typically used in the literature, including age, educational attainment, the number of children in a household, a married/cohabiting identiier, and log equivalised monthly household income (delated to 2005 prices, and equivalised using the OECD modiied scale, detailed in Foster, 2009). Table 1 presents information on the basic inclusion criteria for the estimation sample. The seven waves of the UKHLS contain information on N ¼ 83; 287 individuals who are observed across waves to provide NT ¼ 333; 773 observations. We remove individuals who are observed in only a single wave; individuals not employed and individuals who change place of residence. The criterion of being observed in at least two consecutive waves allows us to consider short-run outcomes at time t following balancing on covariates at time tÀ1. Our working age (16-65 years) sample consists of 31,736 individuals for whom there are a total of 106,195 observations. 17 Descriptive statistics for this sample are provided in Table 2. The mean scores on SF12 PCS (physical health) and SF12 MCS (mental health) are 52.9 and 49.9, respectively, while the mean for health satisfaction 3.5. There are slightly more observations on females than males; mean age is 42 years; 45% have a university level qualiication, average usual hours of work is 33; and average log equivalized monthly household income is £7.55 (equivalent to £1900/month.) These igures are in line with average values of the UK workforce obtained from the 2011 Census and estimates from the UK Labour Force Survey.
First, the data are stratiied into treated and control groups, where the treated are observed to change mode, for example, from car to active travel and the control group never change. Secondly, for this sub-sample, matching controls to treated individuals through EB is undertaken followed by weighted regression of outcomes (here at time t). Exact sample sizes will vary across the four possible mode changes observed. Our main outcome of interest is observed at time t þ 1. Similarly, when considering long-run effects (t þ 2), the initial basic sample is further reined to exclude individuals with less than four waves of data before matching and regression analysis.  Table 3 breaks down the descriptive statistics of commuting time by gender and mode of transport. Males, in general, experience longer commutes (27.8 min one-way compared to 23.6 for women), with the differential between men and women remaining irrespective of the mode of transport. Public transport is associated with the longest commuting times (an average of 48 min) and cycling the shortest (16 min). The distribution of commuting times for  active travel and non-active (users of public transport or car) is provided in the Appendix as Figure A1. As expected there is a greater concentration of short commute durations for active commuters compared to non-active. Figure 1 shows the percentage of individuals who use each of the three modes over time. The percentage of people using a car is relatively stable at around 70% in each wave. The percentage using public transport drops between waves 1 and 2, but then steadily increases. There has been a slight decline in the number of people walking or cycling. Figure 2 shows the associated commuting times. All three modes have experienced a gradual increase in commuting time, but this is largest for walking and cycling. Table 4 reports the transition probabilities between waves t and t þ 1. Among car users at time t, 95% will remain so in the following wave, with 2% switching to public transport and 3% to walking or cycling. Amongst initial public transport users, 81% remain whereas 13% switch to car and 6% switch to active modes. Finally, among initial active commuters, 78% remain so, whereas 16% and 5% switch to car and public transport, respectively. So in summary, there is much more resilience to switching away from car than the other two modes.

| RESULTS
The success of any matching strategy is achieved through obtaining close covariate balance and common support between treated and controls. This relies on the availability of an adequate number of potential control individuals. From our sample, 82% (26,177) of individuals report no change in their commuting mode (the controls), while 12% (3654) report having changed mode once across the sample period. The remaining observations are observed to change mode twice (5%) or more. A full breakdown is provided in Table A1. Table 5 illustrates EB, on the irst and second moments (mean and variance), for a mode change from car to public transport. Matching takes place on covariates measured at time tÀ1. Treated individuals undergo change in mode, controls remain as car users. EB equates the moments of the covariate distribution across treated and control groups. As can be seen, following EB the mean and variance of the set of covariates are very similar across treated and control individuals. This is reassuring as it provides support that the conditional independence assumption, ðh 0 ; h 1 Þ ⊥ d|x:, set out in Section 3 holds. EB for other mode changes and for men and women separately (not reported here), follow a similar pattern.
The results in Table 6 exploit changes to commuting mode occurring between tÀ1 and t to identify health outcomes observed at t þ 1. EB is used to preprocess the data using information on the set of controls prior to parametric modeling. The results suggest that mode changes from car to public transport and vice versa, do not impact health outcomes. Estimated effects are generally small and do not attain statistical signiicance. In contrast, when considering a mode change from car to active travel, we observe a large positive effect on mental health (SF12-MCS). The effect is observed, in similar magnitude, for both men and women. There is also an indication that physical health (SF12-PCS) improves for women, signiicant at the 10% level. Interestingly, individuals who switch mode from active to car report a signiicant decrease in physical health. Again these effects are observed overall and for men and women separately. We also observe a decrease in satisfaction with health for the overall sample (at 10% signiicance). It would appear, therefore, that the effect of a change from car to active travel is felt more strongly through improvements to mental health, whereas the effect of a change to car from active travel is felt through decreases to physical health. We speculate that the asymmetry might be due to the immediacy of feeling the effects. Improvements in mental health brought about through exercise are likely to be felt more quickly (e.g., through increased adrenaline and release of endorphins). One might expect this to be apparent when switching from car to regular active travel. A converse change in mode, however, might not produce such immediate effects. When we stop exercising while we might initially miss the high that this produced, over time we are more likely to pay greater attention to a feeling of lethargy and weight gain, and assign this to a decrease in physical, rather than mental, health. However, we do not observe the same effects when considering changes from public transport to active travel and vice-versa. Individuals who switch from public to active forms of travel report increased health satisfaction, predominantly men, but we do not observe signiicant effects for mental or physical health; however, this may be due to small sample sizes. The reverse mode change from active travel to public transport is associated with a reduction in mental health, particularly for men. 18 The contrast in results from a switch from active travel to car (decrease in physical health) compared to the switch from active travel to public transport (decrease in mental health) suggests that the experience of car and public transport confer different effects on the commuter. A user of public transport is passive, has no control over the journey and often is subjected to overcrowding. In contrast car users have some control over the journey and are actively engaged in the commuting process. The lack of control and passive nature of public transport may lead to a more noticeable effect on mental health. As a consequence, a switch from active travel to car is more likely to be experienced as a decrease in physical health. A graphical illustration of these results, for switches to or from active travel across both men and women is shown in Figure A2.
Overall, we do not observe effects on health from changes in mode between public transport and car use, or viceversa, but do observe effects when moving between active and other forms of travel. However, effects appear generally small, typically less than a tenth of a standard deviation. In comparison to other studies that use the SF12 health measure, Ziebarth (2010) shows that the difference in means for the mental health score of the SF12 is 6.2 and physical health is 3.6 (when rescaled between 0 and 1), when comparing health for the lowest income percentile group to the highest percentile group. While the study does not explicitly consider changes in income and instead compares means across groups, the results do provide context to the size of effects found in this paper for observed changes in commuting mode. In general, our indings indicate that changing commuting mode has a notable impact on health. A change of mode from car to active travel for women has an approximate equivalent effect on physical health of one sixth of the effect of moving between the lowest and highest income percentile groups. The corresponding effect on mental health for both men and women is approximately equivalent to one eighth of the effect of changing income percentile groups.

| Immediate and longer run effects
Here we investigate the possible immediate effects (at time t), as well as longer-run effects (at time t þ 2), of a change in commuting mode (occurring at time t). Full results are reported in Tables A2 and A3. Results are broadly Note: Individuals present for at least 3 waves. Dependent variables measured at t þ 1, and are increasing in good health. Controls matched to treated using entropy balancing at tÀ1, prior to regression of outcomes on treatment (at t), conditioning on covariates and wave dummies (at tÀ1). We also balance on attrition variables, consecutive waves and number of waves. Covariates include age, number of kids, job hours, marital status, household income, commuting time and initial health. Estimates for Health Satisfaction are coeficients from an ordered probit model. Standard errors in parentheses.
***p < 0.01, **p < 0.05, *p < 0.1. JACOB ET AL. similar to those reported in Table 6. Mode changes from car to public transport or vice-versa, do not lead to changes in reported health or health satisfaction (an exception is that women report increased health satisfaction from a change from public to car at time t þ 1). A shorter run effect of a change from car to active travel is observed for women's mental health and for health satisfaction overall. We also observe a decrease in reported physical health for men in the short-run when switching from active travel to car. These results echo those observed for the main results at time t þ 1. We also observe shorter-run effects of a reduction in health satisfaction (at p < 10% level). Similarly, mode change from public transport to active travel results in a short-run increase in reported health satisfaction, driven predominantly by men. However, we do not observe signiicant shorter-run effects from switches from active travel to public transport.
In the longer-run at time t þ 2 we ind positive and signiicant effects on health satisfaction from a mode change from public to active travel driven mainly by women, and a corresponding decrease in health satisfaction for changes in the opposite direction (for men only). Mode changes from car to active travel at time t do not affect health at time t þ 2. It is worth recalling that the health outcomes are self-reported and may be subject to differential reporting behavior (King, Murray, J., & Tandon, 2004). That is, individuals with similar levels of underlying health may report these differently due to perceptions of what may be regarded as healthy and preferences over health compared to other attributes. As long as such reporting behavior is time-invariant it does not present an issue for the analysis. However, it is possible that individuals adapt to changes in health over time, such that we observe initial health changes due to a change in commuting mode in the short to medium term but this then dissipates as time progresses and individuals become accustomed to their changed health. This might be the reason we observe less effects at t þ 2 than at time t þ 1. Oddly, while a mode change from active to car travel leads to a lowering of physical health in the longer run, we also see an improvement in mental health, particularly for women.

| Seasonality in mode choices
It is possible that individuals may change their choice of commuting mode depending on weather conditions. Progressing into summer, individuals may increasingly opt to switch to active travel. Conversely individuals are more likely to switch to car or public transport in winter months. To control for seasonal effects we include the lag of the month of interview in our balancing and regression model. Results are reported in Table A4 and are consistent with the main results. Again, we observe an increase in mental health for both men and women when they switch from car to active travel and a decline in physical health for both groups when they switch from active to car. Similarly, the transition from public to active transport increases health satisfaction for men while the reverse transition decreases their mental health signiicantly, as previously observed.
As a further step, we divide the sample by seasons. Due to small sample sizes we combine spring/summer (typically warm and dry) and autumn/winter (typically cold and wet). These results are presented in Tables A5 and A6 and show that in summer and spring, mental health and satisfaction with own health for women increases when they switch from car to active modes of travel. Similarly, for men we observe an increase in physical health and satisfaction with own health when they change from public to active travel. Furthermore, we observe a decrease in mental health for men along with a decrease in physical health for women when they switch from active to public travel. In the winter months, we observe an increase in mental health for men when they switch from car to active modes of transport and a decline in physical health when they change from public to car travel. The only effect that we observe for women shows a decrease in their physical health when they switch from active to car travel. Broadly, in spring and summer months we observe more positive beneits of switching to active forms of travel, whilst in autumn and winter months we observe more negative health effects of switching from active to non-active forms of transport.

| Constant household location and job
So far, our estimation sample consists of individuals who do not change household address but we placed no restriction on their job characteristics. However, changes in commuting mode can also occur if individuals change jobs leading to a greater distance to travel. In further analysis, we select a subsample of individuals who report no change in household location or job characteristics. These estimates are reported in Table A7. Once again, these effects conirm our main results, although each of these effects are of a slightly higher magnitude compared to those in Table 6. Again, the main effects are observed for mens' health satisfaction which increases when they change from public to active travel and a signiicant decline in their mental health when they switch from active to public transport. We observe a decrease in physical health for women when they switch from active to car travel and an increase in physical health (at lower levels of signiicance) when they move from car to active transport.

| Panel attrition
As with all analyses that rely on panel data, results may be sensitive to attrition bias due to non-random drop out from the survey. Recalling that, for the main analysis, we include respondents observed across a minimum of three waves (tÀ1, t and t þ 1) where in wave t we observe a change to travel mode. The point at which the change to travel mode takes place can be in any particular wave (t) of the panel as long as the individual is also observed both in the previous (tÀ1) and subsequent (t þ 1) waves. When we consider a longer run effect, we observe outcomes at wave t þ 2. Accordingly, the sample of respondents are observed across four waves (tÀ1 to t þ 2). Due to attrition from the survey, the sample sizes for the sub-sets of respondents with observed outcomes at t þ 2 is smaller than for the sub-sets observed at wave t þ 1. This can be seen from a comparison of N in the results presented in Table 6 and those presented in Table A3. This change in sub-sample size allows us to investigate the likely role that survey attrition bias may play. When we compare the set of variables used in the EB and subsequent regression analysis across the two subsets of respondents, we observe the summary statistics presented in Table A8. As can be seen the means and standard deviations are very close. This provides prima-facie evidence for no signiicant attrition bias.

| CONCLUDING REMARKS
This paper evaluates the impact of a change in mode of commute on health. There is evidence on the gains to health from active modes of travel. Therefore, schemes to encourage active travel in the form of walking or cycling are being adopted by countries around the world. The majority of this evidence relies on (often dated) cross-sectional data and thus does not examine the effect of changes in travel mode on health. Of those few studies that do explore the effect of changes in mode, Martin et al. (2014) use ixed effects regressions to address the potential for selection bias. We improve on identiication by employing an empirical strategy that combines matching techniques together with regression based analysis, to provide new evidence on the effect of commuting mode change on health. The approach has the advantage of being "doubly robust" to either poor matching but correct speciication of the regression model, or complete covariate balance and misspeciication of the regression model. Identiication is, however, conditional on the validity of selection on observables. While we mitigate against failure of this assumption by matching on a wealth of pre-treatment characteristics including mode and health, some caution should be applied in interpreting the results due to the possibility of selection on unobservable characteristics inluencing outcomes.
Using rich data taken from the UKHLS covering 2009-2016, we compare health outcomes (at various time periods) for individuals in employment who never change mode throughout the survey, with those who experience a mode change. Our main results indicate a signiicant increase in physical and mental health for commuters switching from car to active forms of transport, particularly for women. We further observe a decline in physical health for individuals of both sexes who switch from active travel to car. A change in mode from active travel to public transport leads to a decrease in reported mental health, largely for men, but we do not observe signiicant decreases in physical health. Mode changes in the opposite direction from public transport to active travel are associated with increases in reported satisfaction with health. The lack of an effect on physical health when changing between active and public transport may be due to accessing public transport requiring exercise, via walking to or from a bus or train station. As this is not the case for switches to and from car travel to active travel the beneits to physical health are more pronounced. Mode changes between car and public transport do not lead to notable affects on physical or mental health outcomes or satisfaction with health. Overall, our results lend support to UK policy initiatives designed to encourage people to move away from car commuting towards more active forms of travel. As well as the health effects estimated here, this will also help the UK government to meet its targets for reducing emissions.    Note: Individuals present for at least 4 waves. Dependent variables measured at t þ 2, and are increasing in good health. Controls matched to treated using entropy balancing at tÀ1, prior to regression of outcomes on treatment (at t), conditioning on covariates and wave dummies (at tÀ1). We also balance on attrition variables, consecutive waves and number of waves. Covariates include age, number of kids, job hours, marital status, household income, commuting time and initial health. Estimates for Health Satisfaction are coeficients from an ordered probit model. Standard errors in parentheses.
***p < 0.01, **p < 0.05, *p < 0.1.    Note: Individuals present for at least three waves. Dependent variables measured at t þ 1, and are increasing in good health. Controls matched to treated using entropy balancing at tÀ1, prior to regression of outcomes on treatment (at t), conditioning on covariates and lags of wave and month of interview (at tÀ1). We also balance on attrition variables, consecutive waves and number of waves. Covariates include age, number of kids, job hours, marital status, household income, commuting time and initial health. Estimates for Health Satisfaction are coeficients from an ordered probit model. Standard errors in parentheses.

-
***p < 0.01, **p < 0.05, *p < 0.1.  Note: Individuals present for at least 3 waves. Dependent variables measured at t þ 1, and are increasing in good health. Controls matched to treated using entropy balancing at tÀ1, prior to regression of outcomes on treatment (at t), conditioning on covariates and wave dummies (at tÀ1). We also balance on attrition variables, consecutive waves and number of waves. Covariates include age, number of kids, job hours, marital status, household income, commuting time and initial health. Estimates for Health Satisfaction are coeficients from an ordered probit model. Standard errors in parentheses.