As a measure of health, I use sex-specific age-standardised period mortality rates (ASMR) for each year between 1981 and 2019. I obtain annual deaths from all causes by district and 5-year age band from the UK Office for National Statistics (ONS). I calculate age specific mortality rates and apply the European Standard Population 2013, to give male and female annual ASMR for all-cause mortality for each district 1981–2019.

To identify coal mining areas, I quantify the concentration of the industry prior to its decline during the 1960s. I use the recently digitalised 1961 Census and identify mining areas as districts where the mining employment rate (per 1000 people) in 1961 is in the upper quartile of the distribution. To differentiate between coal and other types of mining, I use detailed industry employment data from the Census of Employment business survey from 1981 (earliest year data is freely available), identifying relevant Standard Industry Classification codes for coal mining (see Appendix A.2.1). I identify coal mining areas as having mining employment rate in 1961 in the upper quartile and coal mining employment rate greater than other mining employment rate in 1981, thereby identifying areas where coal was the primary mining activity. I take this as my primary area identification method. I retrieve all data from the ONS nomis platform unless stated otherwise (https://www.nomisweb.co.uk/).

Previous studies use the 1981 Census in their identification of coal mining areas. This captures the distribution of the industry prior to the final phase out in the 1980s but does not consider the decline observed during the 1960s. In the literature, coalfield areas have been defined as continuous groups of wards (small geographical areas) where at least 10% of the resident males in employment were employed in the energy and water sector in 1981 (Beatty & Fothergill, 1996). Using this definition, the Coalfield Regeneration Trust (CRT) identifies 55 English coalfield districts (48 using current district boundaries) targeted for regeneration programmes that Riva et al. (2011) use as their identification of coalfields (Welsh districts are not included in the CRT definition). I use two alternative identification methods to align with previous studies and to test the robustness of my results. Firstly, I identify coal mining areas as having the percentage of employed males working in the energy and water sector (hereafter referred to as mining) in 1981 in the upper quartile of the distribution, and coal mining employment rate greater than other mining employment rate in 1981. I label this method “1981”. Secondly, I define coal mining areas as coalfield districts identified by the CRT.

These area identification methods give similar results (see Table 1 and Fig. 1). The primary and 1981 methods identify the same districts as coal mining areas, with the primary method identifying an additional three districts. The CRT method largely overlaps with these methods. However, it does not identify 13 coal mining districts in Wales (as the CRT's definition only covers England) and misses 4 in England. It also identifies 8 districts as coal mining areas whose primary mining activity is not coal or have employment below one of the upper quartile employment cut-offs. Key statistics are similar with mean mining employment well above population means.

The geographical distribution of the coal mining industry shows a central belt running from Yorkshire to the Midlands, along with areas in the North East, North West and Wales (see Appendix A.2.2). This geographical distribution is very similar to those found in literature (Beatty et al., 2019; Beatty & Fothergill, 1996; Riva et al., 2011).

Socioeconomic factors that affect mortality can be both contextual, where you live, and compositional, who you are (Bambra, 2016). I include both types of factors, specifically household income, deprivation, social cohesion, air quality, population density and past manufacturing employment. For income, I use average gross disposable household income (GDHI) for each district available from 1997 to 2020 (Office for National Statistics, 2023), imputing values prior to 1997 (see Appendix A.3).

As a measure of deprivation, I use the Carstairs and Morris (CM) index (Carstairs & Morris, 1989). This index is composed of four equally weighted standardised variables; proportion of the residents without access to a car, proportion living in overcrowded accommodation, male unemployment rate, and proportion of residents classified as low social class. I use Census data from 1981 to 2021, imputing values between census years. I use this index as it can be calculated for any geography and year where constituent variables are available. As alternative measures, I use normalised male unemployment rate and male employment rate using the same methodology.

To account for the community breakdown and social capital loss that occurred in coal mining areas following pit closures, I use Congdon's (1996) measure of social fragmentation, following Riva et al. (2011). This has been shown to affect health independently from economic and deprivation variables (Fagg et al., 2008). This index is similarly composed of four standardised variables; proportion of the population living alone, proportion of unmarried adults, proportion of residents in private rented accommodation, and proportion of residents moving into the district in the previous 12 months. I obtain all data from the Census 1981–2021 and impute values between census years.

To account for urban-rural mortality differences I include population density. I use population estimates for each year and district land area (land area data from the 2011 Census (Office for National Statistics, 2013)). To measure air quality, I use a proxy variable of carbon dioxide emissions which exhibit co-pollutant elasticities with various health damaging air pollutants (Zwickl et al., 2021). I obtain total carbon dioxide emission concentrations for each district 2005–2019 (Department for Business, Energy & Industrial Strategy, 2020), imputing missing values. Heavy industry and manufacturing workers tend to suffer from higher mortality and higher incidence of various health issues (Katikireddi et al., 2017; McDonald et al., 2005). To account for health effects from these industries, such as asbestos exposure in shipbuilding (Goswami et al., 2013), I include the past concentration of the manufacturing industry as a covariate control using the 1961 Census and calculating the “production” employment rate for each district. For models that use the alternative coal mining area identification methods I use the percentage of employed males working in manufacturing from the 1981 Census.

Coal mining areas have different profiles from other areas (Table 2). These areas suffer from higher male and female mortality rates, both in the past (1981) and more recently (2019). This gap has narrowed for male mortality, showing some convergence in the mean area mortality rate. This is not the case for female mortality where the difference in mean area mortality rate is roughly the same in 1981 and 2019. The socioeconomic profiles of these two groups are also different. Coal mining areas are characterised by lower average income, higher deprivation and higher unemployment rates. These areas also tend to be more rural with corresponding lower carbon dioxide emissions.

The aim of this paper is to separate the legacy health effect of coal mining from unemployment, deprivation and social capital effects. Estimation via standard regression controlling for relevant covariates may be problematic, as the socioeconomic profiles of coal mining and other areas differ substantially (Table 2). There is an issue of common support with 64 (24%) other areas having incomes higher than the coal mining area with the highest income. This means that high-income districts, together with low-income districts, jointly set model parameters that estimate the effect of coal mining on mortality. Essentially, high-income districts in London and the South East are used as potential counterfactuals for coal mining areas in Yorkshire and the North East that on average have lower income and higher deprivation. Results will be model dependent and unbiased estimation may be problematic, particularly with the presence of control units far outside the covariate range of treated units (Ho et al., 2007; Sekhon, 2009).

I use matching to reduce model dependence and mitigate this common support issue. Matching aims to mimic a randomised experiment by creating treatment and control groups that are similar on observable characteristics, where treatment assignment is as-if-random. I use propensity score matching to reduce the covariate imbalance between coal mining and non-mining areas, matching socioeconomically similar areas to create a series of district pairs. The matching algorithm prioritises area similarity by choosing pairs with the minimum propensity score difference first, within a caliper. The treatment group is composed of coal mining areas identified using the methods detailed in section 2.2. For the control group, I restrict other areas to non-mining areas that have negligible mining employment (coal or otherwise). For the primary area identification method, I define non-mining areas as areas with the mining employment rate in 1961 below the median value. This definition ensures that non-mining areas have negligible mining employment while also giving a sizeable control pool of districts (see Appendix A.4.1).

For the alternative area identification methods, I follow a similar procedure. For the 1981 alternative I define non-mining areas as districts with percentage of employed males working in mining in 1981 below the median value and exclude districts with mining employment rates in 1961 in the upper quartile (removing areas with significant mining employment in 1961). For the CRT alternative, I include all districts not identified as CRT Coalfield districts as non-mining areas. This gives three sets of treatment and controls pools, where treatment is defined as the presence of the coal mining industry followed by its subsequent disappearance, and control is defined as never having coal mines in the area. From these pools, I obtain matched pairs of districts using the propensity score obtained via a logit model, regressing a coal mining area binary indicator on the socioeconomic covariates outlined above (income, carbon dioxide emissions, population density, deprivation, social fragmentation and manufacturing employment rate) using average values for 1981–2019.

This procedure results in 27 matched pairs of districts using the primary area identification method. Importantly, covariate balance is substantially improved and results in the selection of coal mining and non-mining areas with similar socioeconomic conditions over time (see Appendices A.4.2 and A.4.3). There is now a lack of districts in London and the surrounding area in the South East included (Fig. 2). The alternative area identification methods show similar matching results.

I use a feasible generalised least squares (FGLS) random effects estimator as the model contains both time-variant and invariant regressors. This estimator accounts for AR(1) type serial correlation in the errors (Wooldridge, 2012). Serial correlation results from the fact that coal mining areas are time invariant and that mortality rates are likely to be correlated over time due the cumulative nature of health impacts.

I regress sex-specific male and female mortality rates on a coal mining area binary indicator. I test for convergence in mortality rates by interacting this indicator with a linear time variable. General improvements in mortality rates over time common to all areas are accounted for by adding a time trend, including a quadratic term to account for non-linearities. These general improvements include, for example, improvements in cardiovascular mortality and in the National Health Service. This regression is given by equation (1).

For district i, year t ϵ {0, 1, …, 38} (equivalent to 1981 to 2019), where ASMRit is the sex-specific age-standardised mortality rate, CoalMiningi is the coal mining area binary indicator and Cov_ijt are the j socioeconomic covariates outlined above.

The quantities of interest in this estimation are β1 and β2. β1 indicates the effect of coal mining on mortality rates in 1981 (at t = 0), positive coefficients will show an initial effect of higher mortality in coal mining areas. β2 is the parameter for the interaction term between coal mining and time. A β2 coefficient of the opposite sign to that of the coal mining indicator (β1) will indicate a diminishing gap, i.e. convergence, in mortality rates over time. At any given time, the effect of coal mining on mortality and its duration will depend on a combination of these two parameters (β1 + β2*t).

I display regression results in Table 3 using the primary area identification method and the two alternatives, running the models for samples with matched districts only (models 1–3 and 7–9) and comparing results to samples using all districts (models 4–6 and 10–12). I also plot the marginal effect coefficients and 95% confidence intervals for the coal mining effect in 1981 and 2019 (Fig. 3).

For the matched sample models an initial mortality gap is observed, with higher mortality rates in coal mining areas, followed by convergence in mortality rates over time. Estimated coefficients for the coal mining effect are 122.6 (103.7 and 146.1 for alternative specifications) deaths per 100,000 for male mortality rates and 66.5 (50.6 and 66.1 for alternative specifications) for female mortality rates. All coefficients are statistically significant at the 5% level. For the matched sample models, the interaction term is negative and statistically significant for both male and female rates, indicating an additional decrease in mortality rates for coal mining areas over time. This combination indicates initially higher mortality rates in coal mining areas followed by a reduction of this mortality gap over time, i.e. convergence. Male mortality rates show elimination of 91% (92% and 108% for alternative specifications) of the mortality rate gap in 1981, indicating almost complete convergence for 1981 to 2019. Female mortality rates show partial convergence, with 70% (64% and 75% for alternative specifications) of the initial mortality rate difference eliminated, though the 95% confidence interval of the coal mining marginal effect in 2019 includes zero, meaning complete convergence cannot be ruled out.

Comparing estimation results for the matched versus all districts samples shows some variation in results. Coal mining areas continue to exhibit initially higher mortality rates. However, the amount of convergence decreases when including all districts. For male mortality rates the interaction term coefficient decreases in magnitude, indicating convergence but at a lower level (67–75%). For female mortality rates, this interaction term decreases in size and becomes statistically insignificant (at the 5% level), indicating a lack of convergence for models that include all districts.

These models estimate convergence in mortality rates as a linear function of time. However, the rate of convergence need not be constant over time. To allow for variation in the rate of convergence, I estimate a flexible model interacting the coal mining indicator with a series of year dummies, shown by equation (2). This estimates a coal mining effect parameter for each time period. I plot the three-year rolling average coal mining marginal effects for male and female mortality rate for the matched samples (Fig. 4).

For male mortality rates, the coal mining effect decreases throughout this period. Initially the effect decreases slowly, indicating a low convergence rate, until the early 1990s. The effect subsequently decreases at a higher rate, indicating a higher rate of convergence, until the early 2000s, after which the curve flattens and there is little further convergence. This leaves a small coal mining effect remaining at the end of the period, of approximately 30 deaths per 1000.

The pattern for female mortality rates shows some similarities. However, the effect of coal mining initially increases, indicating divergence in mortality, peaking in the early 1990s. The effect then decreases at a slightly declining rate over time, indicating convergence at a rate decreasing over time.

I further test the robustness of these results using several robustness tests (see Appendix A.5 for regression tables and marginal effects plots). Firstly, I vary the covariate values used in the propensity score matching model. I use 1981 and 2019 values for socioeconomic variables rather than average values. Results remain largely consistent, particularly for male mortality rates, though there is more variation in the coefficients for female mortality rates. Examining the socioeconomic profiles over time for matched coal mining versus non-mining areas, all three matching procedures result in samples with similar profiles albeit via different matching pairs (see Appendix A.5.3).

Secondly, I use male unemployment and employment rate variables as alternative measures to the CM deprivation index. While this index is a robust indicator of mortality, it has been shown to be less predictive of excess mortality over time (Hanlon et al., 2005). Unemployment rates have been shown to not entirely reflect the employment situation following the coal mine closures, with substantial increases in long-term sickness rates and retirement rates of working age males in coal mining areas, as former miners withdrew from the labour market (Beatty & Fothergill, 1996). I therefore use the male unemployment rate and male employment rate as alternatives to the CM deprivation index. Results remain consistent with those from the primary specifications, again with more variation for female mortality rates.

This is the first study to my knowledge that employs a matching causal inference research design to isolate the legacy health effect of the coal mining industry in the UK and analyse its temporal dynamics. Using longitudinal data to estimate effect dynamics enables me to enrich previous findings with one of a progressive decrease in the coal mining effect on mortality. Through the use of matching, I estimate effects for coal mining and non-mining areas that are socioeconomically similar, greatly improving covariate balance and removing districts with very different socioeconomic characteristics. This differs from previous studies, which examine the effect of coal mining on health outcomes at particular points in time, either controlling for socioeconomic differences by adding covariates to regressions (Riva et al., 2011) or simply looking at differences in means (Beatty et al., 2019; Foden et al., 2014; Shucksmith et al., 2010). Matching reduces the model dependency of estimates, meaning results are less driven by particular modelling assumptions, making inferences more reliable. Thus, by restricting my sample to matched pairs of districts with similar socioeconomic profiles, I am better able to assess the impact of different causal mechanisms affecting health in coal mining areas.

While this research design aims to isolate coal mining legacy health effects from other mechanisms, remaining alternative explanations cannot be entirely ruled out and may contribute to explaining the observed coal mining effect. Other explanations must result from, and be correlated with, coal mine closures and not be sufficiently accounted for by the covariates included. Of course, the variables used (household income, deprivation and social fragmentation indices) may not fully capture the unemployment, deprivation and social capital loss that occurred following coal mine closures. If this is the case, then the estimated coal mining effect will contain the part of these factors not captured by the variables. A second possible explanation is local pollution effects of coal mines (Pless-Mulloli et al., 1998) disappearing with mine closures, that is not sufficiently captured through the proxy variable carbon dioxide emissions, for which I impute values prior to 2005. Thirdly, unemployment induced out-migration following coal mine closures, or subsequent in-migration due to reduced house prices and increased commuting, could provide an alternative explanation. However, the evidence on out-migration is mixed (Beatty & Fothergill, 1996; Green, 1994), with Hollywood (2002) finding a lack of migration among miners. Finally, it is possible that the observed convergence in mortality rates occurs due to improved NHS access and care improvements in coal mining areas. However, such improvements would have to be specific to coal mining areas, above and beyond improvements in more deprived areas.

Analysis at local authority district level is consistent with previous studies that assess health outcomes in coal mining areas (Riva et al., 2011; Shucksmith et al., 2010). Analysis at this spatial scale may bundle together coal mining and non-mining local areas in the same district, potentially weakening effects. However, identification at the district scale shows mining as an important sector in the local labour market. Coal mining areas have an average of almost 15% rising to over 45% of employed males working in coal mining in 1981. Using districts also minimises the impact of migration, as UK Census Flow data shows that most migration occurs within districts. Migration will thus become more problematic using smaller local areas, which may not effectively capture the industry's effect on health over such a long time period.

The theoretical case for higher female mortality in coal mining areas and subsequent convergence is less obvious than for male mortality. It was almost exclusively men that worked in the mines and suffered the direct health consequences from the hazardous working environment. There is some evidence for a female specific coal mining health effect, with wives and cohabitors of miners exhibiting a higher prevalence of respiratory and asbestos related health issues (Goswami et al., 2013; Higgins et al., 1959). The smaller yet similar trend in female mortality rates could also be explained by a common factor affecting both men and women, potentially in combination with sex specific mechanisms. Possible common factors that could affect both male and female mortality rates include the alternative explanations highlighted above, namely pollution, migration, improved NHS access and any residual effect of unemployment, deprivation and social capital loss not captured by the variables used. Future research could investigate female specific effects and the possibility of one of the aforementioned common factors driving the observed trends.

A small minority of districts change boundaries over time. All recent boundary changes have been to merge districts. These cases were identified (via district code changes) and data summed/weighted averages taken. This applies to all data except 1961 Census data which were mapped to 2021 district boundaries via small area (Output Area) lookups downloaded from the ONS Open Geography Portal.

Data will be made available on request.