- Split View
-
Views
-
Cite
Cite
J. Iparraguirre, Have winter fuel payments reduced excess winter mortality in England and Wales?, Journal of Public Health, Volume 37, Issue 1, March 2015, Pages 26–33, https://doi.org/10.1093/pubmed/fdu063
- Share Icon Share
Abstract
The historical series of excess winter mortality (EWM) in England and Wales presents a negative trend. Winter fuel payments (WFPs) are the most important benefits for people aged 65 or over directly related to Winter Mortality in the UK.
This study presents a time series analysis of the direct effect of WFPs on EWM in England and Wales.
We find a significant structural break in trend and volatility in the EWM series in England and Wales in 1999–2000. After controlling for a number of covariates, an ARIMA-X model finds that WFPs can account for almost half of the reduction in EWM in England and Wales since 1999/2000.
Almost half of the reduction in EWM since 1999/2000 is attributable to WFPs.
Introduction
Ever since 1858 excess winter mortality (EWM) has been a pressing public health issue and an active research topic in the UK.1 EWM is the difference between the number of deaths that occurred between December and March and the average number of deaths occurring the preceding August–November and the following April–July.2 This definition follows the recommendation dating from 1881 to group together the 4 months starting in December to analyse the effects of seasonality on mortality.3
Comparative studies within and across countries have shown that EWM is not higher in the colder regions,4–7 although this association varies by cause of death.8 From this comparative perspective, the UK has fared quite poorly.
Despite the wealth of research on the association between weather and mortality, and the numerous policy interventions, little has been done with regard to evaluating the policies directly aimed at reducing EWM. This article presents a time series study of the impact of the introduction of winter fuel payments (WFPs) on EWM in England and Wales.
The structure of the article is as follows: the next section reviews some attempts at measuring the effectiveness of policy interventions targeted at reducing EWM. The following section describes the data. Next, the statistical methods are presented. The penultimate section gives the results and the final section concludes.
Policy interventions to reduce EWM
In several countries, there are various interventions in place to reduce EWM. The disparity and multi-pronged nature of these policy interventions may reflect different ways of framing EWM as a public policy issue: an income poverty problem, a fuel poverty problem, a home insulation problem, a public health problem, a behavioural problem, an energy market competition problem and/or a consumer behaviour problem.9
Interventions to tackle EWM can be grouped into five domains: health, housing, behavioural, energy and income policy interventions. Recently, attempts have been made to integrate these related but disparate efforts into strategic plans. For example, in England, a number of programmes and measures are grouped under the annual Cold Weather Plans which include: the annual seasonal influenza vaccination programme, the pneumococcal vaccination programme, the ‘Keep Warm Keep Well’ advice and information programme, the local and national winter pressures reporting and winter resilience programmes, the ‘Winterwatch’ information programme for service professionals and users, the Excess Seasonal Deaths Toolkit, the Excess Winter Deaths Atlas and the Warm Homes Healthy People Fund.2
The effectiveness of these integrated plans has not been measured yet, although an evaluation study of the Cold Weather Plan for England 2012 is currently under way.2 However, on a smaller scale, there is some evidence that a combination of the severe winter weather ‘Early Warning System’ mechanism developed by the UK Meteorological Office and information and advice tools targeted at at-risk groups produced behavioural change consistent with health risk reduction of episodes of cold weather.10
Possibly the most researched health policy intervention to combat EWM is the influenza vaccination. One of the first studies found that vaccination reduced hospitalizations and deaths but the effect depended on the use of a vaccine formulated for the prevalent strain.11–13
Concerning housing interventions, beneficial health effects of the installation of central heating in households living in fuel poverty have been found in Northern Ireland,14 whereas no evidence of a significant effect of the introduction of a fuel poverty reduction scheme (i.e. the ‘Warm Front’)—aimed at improving the energy efficiency of homes of people living in fuel poverty—was found on reducing EWM.15 In contrast, using indirect, modelling techniques a study found that the Warm Front Scheme reduced excess winter deaths among older people due to cerebrovascular disease but by a very small margin: 0.4 deaths against a baseline of 27 deaths, i.e. <1.5%.16
A systematic review of the relation between EWM and excess winter hospitalization and housing quality or socioeconomic status found several methodological problems which prevented any generalization.17 Some work has been done in evaluating interventions to reduce asthma—a condition with higher prevalence in winter—in children. For example, installing central heating18,19 and ventilation systems20 in houses with children suffering from asthma significantly reduced all respiratory symptoms and the same was found of home-based, multi-trigger and multi-component interventions with an environmental component.21,22
However, a systematic review of 18 housing intervention studies to improve health outcomes in the UK concluded that small population sizes and lack of controlling for confounders limit any generalization as to the effectiveness of these interventions.23
A significant positive impact on EWM of an intervention in Australia aimed at reducing air pollution from biomass smoke which consisted of replacing domestic wood heaters with electric heaters was recently reported.24
In the UK, two benefits directly related to EWM are the WFP and the Cold Weather Payments (CWP). WFP is a non-means-tested, tax-free annual cash payment made usually in December to households with someone over Pension Credit age (currently, 65 years) to help with heating costs. It is worth about £169 per caseload on average.25 CWP is a means-tested one-off payment made to receipts of certain benefits, contingent on temperature records over 7 consecutive days. Given this contingency, the fiscal impact of this benefit varies substantially from year to year. (For example, the estimated CWP expenditure amounted to almost £620 million in 2010, £8.5 million in 2011 and £131 million in 2012 (until 31 December).24) It is worth £25 for each 7-day cold weather period.
This is the first article evaluating the effects of an income policy intervention on EWM by means of econometric methods.
Data
We studied annual data for the all-age EWM for England and Wales.26 This time series is available since 1950/51, but for the ARIMA-X models we studied the period 1970–2012 due to the available data on the covariates.
With regard to WFPs, we used the series of payments per beneficiary in real terms (2012/3 prices).27
When applicable, depending on the model specification, we have included one or more of the following variables to control for their confounding effects on EWM:
Mean or minimum recorded outside temperature in winter months.28
Mean internal temperature in household dwellings. Source: For 1970–71 and 2011–12.29 Extrapolated for 2012–13.
Annual average of the proportion of household expenditure spent on domiciliary gas (in the UK as a whole).30
Statistical methods
This article presents two time series analyses. The starting point is the detection of a structural change in the series in 1999–2000. In fact, the long series starting in 1950/51 presents two other breaks but in the series since 1970, upon which the rest of the article is based, only one significant structural break is detected in 1999–2000. The other two significant breaks correspond to 1962–63 and 1969–70. The winter of 1962–63 was the coldest for more than 200 years.31–33 The break in 1969–70 coincides with a virulent influenza A/Hong Kong (H3N2) virus pandemic.34,35 It is worth noting, however, that none of the statistical techniques have detected a structural break in the winter of 1989–90, when England and Wales also suffered from an influenza epidemic.36
Secondly, we run ARIMA-X models, which from an econometric perspective can be understood as ARIMA models with additional covariates, and from a policy-focused perspective, as intervention models with other plausible explanatory variables apart from the intervention dummy.
Structural breaks
Structural breaks are one type of non-stationarity in a time series which arises from changes in the population regression parameters at one or more dates and whose overlook may result in wrong inference and forecasting.
Several statistical tests for the detection of structural changes in time series have been proposed (see refs 37 and 38 for an overview). Some unit root tests allow for the presence of one or more structural breaks in univariate time series.39,40
We have run the modified F tests,41,42 the algorithm for simultaneous estimation of multiple breakpoints43 and the empirical fluctuation tests on the cumulative sums of the scaled residuals of an ordinary least squared regression—OLS-based CUSUM test44–46 (these tests are implemented in the strucchange package47 in R48). The null hypothesis is that EWM follows a standard Brownian motion; the alternative hypothesis is that EWM shifted either in levels, trend or both at some point over the time period.
With regard to unit root tests with structural breaks, we ran the Zivot and Andrews test49 which detects one structural break endogenously and the Narayan–Popp (NP) M2 test which can accommodate two breaks simultaneously in the intercept and the trend.50 (The Zivot and Andrews test was run with urca51 the R package. We thank Deepa Bannigidadmath from Deakin University who kindly ran the NP test for us.) The NP test exhibits superior performance than other tests that allow for two structural breaks, such as the Lumsdaine and Papell and the Lee and Strazicich tests,52,53 especially on short series.47
Furthermore, we ran the modified robust Brown–Forsythe version of the Levene test of equality of variances54,55 to test whether the variability of the time series of EWD presented a significant change before and after 1999–2000 (we used the lawstat56 R package).
ARIMA-X models
Discussions of intervention models in an otherwise univariate ARIMA time series setting can be found in some time series econometrics textbooks.57,58 If the date of a policy intervention is known, these models can determine whether it has caused any significant immediate and long-term impact on the levels and/or trend of a variable under study. Interventions can be incorporated in different ways, which can be reduced to a pulse variable (the intervention only lasts one period) or a step variable (the effects are felt since a given period). In each of our models, the intervention dummy is introduced as a step response function with no lags—that is, we assume that if the winter allowances had an impact, it was felt in the same winter they were paid.
Box and Tiao introduced the analysis of policy interventions within a univariate ARIMA model with covariates59—that is, in the presence of a series which depends on its past values and on other variables and whose noise structure can be modelled as a moving average of past shocks.
An ARIMA-X model is similar to the general specification above with additional covariates following an ARIMA process. It is common to express these models as transfer function models:
Results
Structural breaks
All the structural break tests detect a significant change in 1999/2000 (the NP test also detects a significant break in 1995). Table 1 presents the statistics and the P-values, critical values or Bayesian Information Criterion statistics for each test on the EWD series between 1970 and 2012:
Test . | Break points . | Stat . | P-value/critical value . |
---|---|---|---|
supF | 1999 | 14.3 | 0.0035 |
aveF | 1999 | 7.0 | 0.0005 |
expF | 1999 | 4.6 | 0.0011 |
efp-OLS | 1985 and 1999 | 1.5 | 0.0227 |
ZA | 1999 | −5.8 | −4.93 (1% c.v.) |
Test . | Break points . | Stat . | P-value/critical value . |
---|---|---|---|
supF | 1999 | 14.3 | 0.0035 |
aveF | 1999 | 7.0 | 0.0005 |
expF | 1999 | 4.6 | 0.0011 |
efp-OLS | 1985 and 1999 | 1.5 | 0.0227 |
ZA | 1999 | −5.8 | −4.93 (1% c.v.) |
supF, Quandt test; aveF, average Lagrange multiplier test; expF, average exponential Lagrange multiplier test; efp-OLS, OLS-CUSUM empirical fluctuation process test; ZA, Zivot-Andrews test with trend and 1 lag.
Test . | Break points . | Stat . | P-value/critical value . |
---|---|---|---|
supF | 1999 | 14.3 | 0.0035 |
aveF | 1999 | 7.0 | 0.0005 |
expF | 1999 | 4.6 | 0.0011 |
efp-OLS | 1985 and 1999 | 1.5 | 0.0227 |
ZA | 1999 | −5.8 | −4.93 (1% c.v.) |
Test . | Break points . | Stat . | P-value/critical value . |
---|---|---|---|
supF | 1999 | 14.3 | 0.0035 |
aveF | 1999 | 7.0 | 0.0005 |
expF | 1999 | 4.6 | 0.0011 |
efp-OLS | 1985 and 1999 | 1.5 | 0.0227 |
ZA | 1999 | −5.8 | −4.93 (1% c.v.) |
supF, Quandt test; aveF, average Lagrange multiplier test; expF, average exponential Lagrange multiplier test; efp-OLS, OLS-CUSUM empirical fluctuation process test; ZA, Zivot-Andrews test with trend and 1 lag.
The results from the modified robust Brown–Forsythe Levene test reject the null of homogeneity of variances (tests statistic based on the absolute deviations from the median = 4.4014; P-value based on 1000 bootstrapped replications = 0.048), so we can conclude that the variability in the EWD series exhibits a significant change before and after 1999/2000.
ARIMA-X models
We ran ARIMA-X models which incorporate WFP as a dummy variable equal to 0 until 1998/1999 and equal to 1 thereafter, following the results above, as well as one or more of these additional covariates: We ran five models on (the logs of) EWM (for the period 1970–71 and 2012–13). In all these models, we based the time series specification of the dependent variable on an ARIMA(2,1,2) following results from the auto-arima function in the forecast R package60 extended with covariates (results available from the author).
Average recorded outside temperature in winter months
Mean internal temperature in household dwellings
Average household expenditure on gas as a proportion of total household expenditure
Models 1 and 2 include the mean outside winter temperature, whereas Models 3 and 4 use the mean internal temperature. Moreover, Models 2 and 4 include gas expenditure, while Models 1 and 3 omit this variable. Finally, Model 5 includes the four covariates.
Table 2 presents the results (we ran the package TSA in R61; see also ref. 62):
. | Model 1 . | Model 2 . | Model 3 . | Model 4 . | Model 5 . |
---|---|---|---|---|---|
Ar1 | 0.985 | −0.879 | −0.250 | 0.759 | −0.923 |
(5.884) | −(2.264) | −(0.299) | (4.791) | −(2.025) | |
ar2 | −0.110 | −0.043 | −0.204 | −0.162 | −0.115 |
−(0.679) | −(0.244) | −(1.329) | −1.041) | −(0.629) | |
ma1 | −1.996 | −0.100 | −0.784 | −1.993 | −0.102 |
−(13.995) | −(0.263) | −(0.907) | −(20.035) | −(0.217) | |
ma2 | 1.000 | −0.900 | −0.216 | 1.000 | −0.898 |
(7.008) | −(2.375) | −(0.251) | (10.059) | −(1.906) | |
Mean outside winter temperature | −0.007 | −0.010 | −0.008 | ||
−(1.406) | −(1.732) | −(1.422) | |||
Mean internal temperature | −0.057 | −0.051 | −0.045 | ||
−(2.034) | −(4.022) | −(1.468) | |||
Gas expenditure (% total household expenditure) | −18.902 | −16.751 | −17.981 | ||
−(1.171) | −(1.943) | −(1.183) | |||
WFP (=1 since 1999/2000) | −0.247 | −0.272 | −0.045 | −0.145 | −0.112 |
−(3.979) | −(3.022) | −(0.374) | −(2.073) | −(0.816) | |
Training set error measures | |||||
ME | −0.036 | −0.024 | −0.003 | 0.012 | 0.003 |
RMSE | 0.228 | 0.230 | 0.229 | 0.208 | 0.224 |
MAE | 0.184 | 0.188 | 0.190 | 0.170 | 0.190 |
MPE | −0.386 | −0.279 | −0.080 | 0.076 | −0.020 |
MAPE | 1.765 | 1.799 | 1.824 | 1.626 | 1.817 |
MASE | 0.686 | 0.701 | 0.711 | 0.635 | 0.709 |
. | Model 1 . | Model 2 . | Model 3 . | Model 4 . | Model 5 . |
---|---|---|---|---|---|
Ar1 | 0.985 | −0.879 | −0.250 | 0.759 | −0.923 |
(5.884) | −(2.264) | −(0.299) | (4.791) | −(2.025) | |
ar2 | −0.110 | −0.043 | −0.204 | −0.162 | −0.115 |
−(0.679) | −(0.244) | −(1.329) | −1.041) | −(0.629) | |
ma1 | −1.996 | −0.100 | −0.784 | −1.993 | −0.102 |
−(13.995) | −(0.263) | −(0.907) | −(20.035) | −(0.217) | |
ma2 | 1.000 | −0.900 | −0.216 | 1.000 | −0.898 |
(7.008) | −(2.375) | −(0.251) | (10.059) | −(1.906) | |
Mean outside winter temperature | −0.007 | −0.010 | −0.008 | ||
−(1.406) | −(1.732) | −(1.422) | |||
Mean internal temperature | −0.057 | −0.051 | −0.045 | ||
−(2.034) | −(4.022) | −(1.468) | |||
Gas expenditure (% total household expenditure) | −18.902 | −16.751 | −17.981 | ||
−(1.171) | −(1.943) | −(1.183) | |||
WFP (=1 since 1999/2000) | −0.247 | −0.272 | −0.045 | −0.145 | −0.112 |
−(3.979) | −(3.022) | −(0.374) | −(2.073) | −(0.816) | |
Training set error measures | |||||
ME | −0.036 | −0.024 | −0.003 | 0.012 | 0.003 |
RMSE | 0.228 | 0.230 | 0.229 | 0.208 | 0.224 |
MAE | 0.184 | 0.188 | 0.190 | 0.170 | 0.190 |
MPE | −0.386 | −0.279 | −0.080 | 0.076 | −0.020 |
MAPE | 1.765 | 1.799 | 1.824 | 1.626 | 1.817 |
MASE | 0.686 | 0.701 | 0.711 | 0.635 | 0.709 |
Note 1: All models follow an ARIMA(2,1,2) specification for the dependent variable (i.e. the log of Excess Winter Deaths).
Note 2: t-statistics between brackets.
. | Model 1 . | Model 2 . | Model 3 . | Model 4 . | Model 5 . |
---|---|---|---|---|---|
Ar1 | 0.985 | −0.879 | −0.250 | 0.759 | −0.923 |
(5.884) | −(2.264) | −(0.299) | (4.791) | −(2.025) | |
ar2 | −0.110 | −0.043 | −0.204 | −0.162 | −0.115 |
−(0.679) | −(0.244) | −(1.329) | −1.041) | −(0.629) | |
ma1 | −1.996 | −0.100 | −0.784 | −1.993 | −0.102 |
−(13.995) | −(0.263) | −(0.907) | −(20.035) | −(0.217) | |
ma2 | 1.000 | −0.900 | −0.216 | 1.000 | −0.898 |
(7.008) | −(2.375) | −(0.251) | (10.059) | −(1.906) | |
Mean outside winter temperature | −0.007 | −0.010 | −0.008 | ||
−(1.406) | −(1.732) | −(1.422) | |||
Mean internal temperature | −0.057 | −0.051 | −0.045 | ||
−(2.034) | −(4.022) | −(1.468) | |||
Gas expenditure (% total household expenditure) | −18.902 | −16.751 | −17.981 | ||
−(1.171) | −(1.943) | −(1.183) | |||
WFP (=1 since 1999/2000) | −0.247 | −0.272 | −0.045 | −0.145 | −0.112 |
−(3.979) | −(3.022) | −(0.374) | −(2.073) | −(0.816) | |
Training set error measures | |||||
ME | −0.036 | −0.024 | −0.003 | 0.012 | 0.003 |
RMSE | 0.228 | 0.230 | 0.229 | 0.208 | 0.224 |
MAE | 0.184 | 0.188 | 0.190 | 0.170 | 0.190 |
MPE | −0.386 | −0.279 | −0.080 | 0.076 | −0.020 |
MAPE | 1.765 | 1.799 | 1.824 | 1.626 | 1.817 |
MASE | 0.686 | 0.701 | 0.711 | 0.635 | 0.709 |
. | Model 1 . | Model 2 . | Model 3 . | Model 4 . | Model 5 . |
---|---|---|---|---|---|
Ar1 | 0.985 | −0.879 | −0.250 | 0.759 | −0.923 |
(5.884) | −(2.264) | −(0.299) | (4.791) | −(2.025) | |
ar2 | −0.110 | −0.043 | −0.204 | −0.162 | −0.115 |
−(0.679) | −(0.244) | −(1.329) | −1.041) | −(0.629) | |
ma1 | −1.996 | −0.100 | −0.784 | −1.993 | −0.102 |
−(13.995) | −(0.263) | −(0.907) | −(20.035) | −(0.217) | |
ma2 | 1.000 | −0.900 | −0.216 | 1.000 | −0.898 |
(7.008) | −(2.375) | −(0.251) | (10.059) | −(1.906) | |
Mean outside winter temperature | −0.007 | −0.010 | −0.008 | ||
−(1.406) | −(1.732) | −(1.422) | |||
Mean internal temperature | −0.057 | −0.051 | −0.045 | ||
−(2.034) | −(4.022) | −(1.468) | |||
Gas expenditure (% total household expenditure) | −18.902 | −16.751 | −17.981 | ||
−(1.171) | −(1.943) | −(1.183) | |||
WFP (=1 since 1999/2000) | −0.247 | −0.272 | −0.045 | −0.145 | −0.112 |
−(3.979) | −(3.022) | −(0.374) | −(2.073) | −(0.816) | |
Training set error measures | |||||
ME | −0.036 | −0.024 | −0.003 | 0.012 | 0.003 |
RMSE | 0.228 | 0.230 | 0.229 | 0.208 | 0.224 |
MAE | 0.184 | 0.188 | 0.190 | 0.170 | 0.190 |
MPE | −0.386 | −0.279 | −0.080 | 0.076 | −0.020 |
MAPE | 1.765 | 1.799 | 1.824 | 1.626 | 1.817 |
MASE | 0.686 | 0.701 | 0.711 | 0.635 | 0.709 |
Note 1: All models follow an ARIMA(2,1,2) specification for the dependent variable (i.e. the log of Excess Winter Deaths).
Note 2: t-statistics between brackets.
The MASE statistic favours Model 4, which includes both mean internal temperature and gas expenditure as covariates apart from the WFP dummy (=1 since 1999/2000). The coefficient for the mean internal temperature is statistically significant and negative: as it would be expected, higher internal temperatures would reduce EWM. In turn, the proportion of household income spent on gas is significant at a 5.2% level, and also has the correct sign (the correlation coefficient between both variables is −0.53, which rules out any collinearity problems). The coefficient for the WFP dummy (−0.145) results in a predicted reduction in EWM of 13.5% since 1999–2000 compared with before this year.
On average, between 1970–71 and 1998–99 and since 1999–2000, EWM fell by 29.5%. Therefore, the results suggest that almost half of this reduction would be attributable to the introduction of WFPs.
Figure 1 plots the time series for EWM since 1970–71 and the fitted results for the best-fitting intervention model and the ARIMA-X models.
Discussion
Main finding of this study
WFPs can account for almost half of the reduction in EWM in England and Wales since 1999–2000.
What is already known on this topic
WFPs are the most important benefits for people aged 65 or over directly related to Winter Mortality in the UK but no previous attempts had been made to measure their impact on EWM.
What this study adds
The first rigorous econometric estimation of the impact of WFPs on EWM.
Limitations of this study
The aggregate nature of the data prevented from an individually based study and unavailability of longer time series precluded any detailed regional analyses or the application of vector autoregressive econometric techniques.
Following Docherty and Smith,63 this final section presents a structured discussion of the article (we are grateful to one of the reviewers for suggesting the inclusion of a structured discussion).
Statement of principal findings
This article presents the first econometric analysis of the effects of the introduction of WFPs in the levels of EWM in England and Wales. We find that the introduction of WFP has been a policy intervention with very positive effects: EWM went down by ∼13% a year on average since the amounts of WFP paid per beneficiary stabilized in 1999–2000 (2 years after the introduction of this benefit), accounting for ∼46% of the recorded reduction in EWM since then. Therefore, the evidence presented in this article supports the efficacy of WFPs in reducing EWM in England and Wales.
Strengths and weaknesses of this study
We consider the rigorous application of statistical techniques as the main strength of this article. However, theoretically sound, applied statistical analyses need equally sound data. In this sense, a potential weakness of this study has to do with the accuracy of the energy-related data used in the models. The mean internal temperature in household dwellings is derived from a domestic energy consumption model (the Cambridge Housing Model) developed by Cambridge Architectural Research, a private consultancy.64 Given that most of the inputs to the model are based on data from housing surveys, which exhibit good agreement in global analyses but are subject to uncertainties that may introduce major differences at local levels.65
The other energy-related variable in the model—i.e. the household expenditure spent on domiciliary gas—is less subjected to sample and measurement errors. These estimates are based on the Domestic Fuel Inquiry collected by the Department of Energy and Climate Change from eight gas suppliers, which in essence is a census of ∼95% of all domestic energy customers in the UK and therefore is not based on sampling techniques.66 Furthermore, the statistical procedures to arrive at these estimates conform to the internationally agreed methodology to compile National Accounts.
Meaning of the study: possible mechanisms and implications for policy-makers
Considering the results reported in this article, we recommend that WFPs should be incorporated into the annual Cold Weather Plans as part of a better joined-up policy approach to combat the deleterious health effects of cold weather upon vulnerable groups, particularly the elderly.
As mentioned above, WFPs are framed in various ways in public policy discussions. Recently, the UK government has investigated the effectiveness of using WFP to incentivize the take-up of energy saving improvement initiatives such as the Green Deal cashback scheme.67 Furthermore, some studies looked into the labelling effects of WFP on domestic fuel consumption.68,69 Other recent policy discussions around WFP included extending the benefit to severely disabled people regardless of age, bringing forward the timing of payments for households not connected to the main gas grid, means testing the allowance re-defining the eligibility criteria, withdrawing the allowance from pensioners paying higher rate income tax and making it a taxable benefit. These as well as any other future options for reforming the WFP scheme should take into account the positive effects in reducing EWM reported in this article and therefore the likely impact on mortality of each policy alternative.
Unanswered questions and future research
We applied the same statistical techniques to detect structural breaks described above to the somewhat short (between 1991–92 and 2011–12) available regional series of EWM for the population as a whole by NUTS-1 (Nomenclature of Territorial Units for Statistics) English region. We detected one significant break, in 1999–2000, in each region except West Midlands. In addition, using regional time series of EWM for the population aged 85 or over, we found a significant break in 1999–2000 in every regional series. These results should be read as additional confirmatory findings of those presented in the structural breaks section above. However, due to data unavailability, it was neither possible to run ARIMA-X models by region nor any statistical analyses at lower geographical units (for example, local authorities). Furthermore, we could not distinguish between causes of death. These are two avenues for future research.
Alternative statistical approaches—such as vector autoregressive, structural vector autoregressive and structural vector error correction models—allow for further investigation into causality, prediction and impulse responses. However, these models would require longer series than existing ones for all the relevant variables. In the future, it would be worth revisiting the results reported in this article with these alternative econometric techniques.