Investigating the impact of the English health inequalities strategy: time trend analysis

Objective To investigate whether the English health inequalities strategy was associated with a decline in geographical health inequalities, compared with trends before and after the strategy. Design Time trend analysis. Setting Two groups of lower tier local authorities in England. The most deprived, bottom fifth and the rest of England. Intervention The English health inequalities strategy—a cross government strategy implemented between 1997 and 2010 to reduce health inequalities in England. Trends in geographical health inequalities were assessed before (1983-2003), during (2004-12), and after (2013-15) the strategy using segmented linear regression. Main outcome measure Geographical health inequalities measured as the relative and absolute differences in male and female life expectancy at birth between the most deprived local authorities in England and the rest of the country. Results Before the strategy the gap in male and female life expectancy between the most deprived local authorities in England and the rest of the country increased at a rate of 0.57 months each year (95% confidence interval 0.40 to 0.74 months) and 0.30 months each year (0.12 to 0.48 months). During the strategy period this trend reversed and the gap in life expectancy for men reduced by 0.91 months each year (0.54 to 1.27 months) and for women by 0.50 months each year (0.15 to 0.86 months). Since the end of the strategy period the inequality gap has increased again at a rate of 0.68 months each year (−0.20 to 1.56 months) for men and 0.31 months each year (−0.26 to 0.88) for women. By 2012 the gap in male life expectancy was 1.2 years smaller (95% confidence interval 0.8 to 1.5 years smaller) and the gap in female life expectancy was 0.6 years smaller (0.3 to 1.0 years smaller) than it would have been if the trends in inequalities before the strategy had continued. Conclusion The English health inequalities strategy was associated with a decline in geographical inequalities in life expectancy, reversing a previously increasing trend. Since the strategy ended, inequalities have started to increase again. The strategy may have reduced geographical health inequalities in life expectancy, and future approaches should learn from this experience. The concerns are that current policies are reversing the achievements of the strategy.


Appendix 2. Model formula
We estimated whether the strategy period was associated with a greater decline in absolute inequalities between the most deprived local authorities and the rest of England compared to the before and after periods using segmented linear regression, controlling for the trend in unemployment. Specifically, we estimated the following model: : LE i,t = β 1 t 1 +β 2 Deprived + β 3 Deprived*t 1 + β 4 t 2 + β 5 Deprived*t 2 + β 6 t 3 + β 7 Deprived*t 3

+ β 8 Unemp i,t +u i +ε it
Where LE i,t is the life expectancy in LA i in year t, t 1 is annual trend term for the before period t 2 is annual trend term for the strategy period t 3 is annual trend term for the after period Deprived is a dummy variable indicating whether an LA is within the most deprived quintile. Unemp is the annual unemployment rate in each LA as measured as the proportion of 16-64 year olds claiming of unemployment benefits. u i is a fixed effect for each local authority. The breakpoints defining the beginning and end of the strategy period were defined based on the process outlined in Appendix 7.
Appendix 3. Group specific trend estimates derived from the regression models. Table A1. Trend in life expectancy in the most deprived LAs and the rest of the country -before, during and after the health inequalities strategy. Trend is shown as the annual change in life expectancy measured in months.

Men
Annual change (in months) in life expectancy 95%CI P-value for trend P-value for change in trend from previous time period

Deprived areas
Before ( Appendix 4. Alternative models Table A2. Trend in relative inequalities in life expectancy, between the most deprived LAs and the rest of the country before, during and after the health inequalities strategy. Trend is shown as the annual increase (+) or decrease (-) in the percentage difference in life expectancy between deprived LAs and the rest of the country.

Men
Annual change in the relative percentage gap in life expectancy between the most deprived 20% of local authorities and the rest of England [95%CI] Before   Note: Estimates based on fixed effects regression model using a local authority panel dataset of life expectancy from 1983 to 2015, also adjusting for local unemployment rates.
Using a continuous measure of deprivation.
To estimate the extent there was a narrowing of inequalities across all levels of deprivation we estimated our models using a continuous measure rather than a binary split between deprived areas and the rest of the country. The IMD 2004 income score was converted to a weighted rank across all local authorities (LA), from the least deprived (0) to the most deprived (1), we then fitted the model including this measure interacted with our linear spline time trend terms. The coefficients of this model can then be interpreted as the absolute annual change in the Slope Index of Inequality 1 (i.e the change in the gap between the most deprived and least deprived LAsassuming a linear relationship between change in inequalities and deprivation). Note: Estimates using a local authority panel dataset of life expectancy from 1983 to 2015, also adjusting for local unemployment rates.

Removing outliers.
Initially we estimated the trend in life expectancy for each local authority between 2004-2012. We then removed all local authorities that had a trend during this period that was +/-2 standard deviations greater or lesser than the mean (5 deprived LAs and 17 LAs from the rest of England), and re-estimated our model.  deprived LAs and the rest of the country before, during and after the health inequalities strategy. Trend is shown as the annual increase (+) or decrease (-) in the absolute gap in life expectancy measured in months.
As there was evidence of autocorrelation in the time series, we initially estimated the maximum lags required to take into account the autocorrelation structure using Newey and West's (1994) automatic bandwidth selection procedure. 2 This indicated a maximum lag of 16 was appropriate.

Men
Annual change (in months) in absolute gap in life expectancy between the most deprived 20% of local authorities and the rest of England Before ( Table A10. Controlling for trends in migration. Trend in absolute inequalities in life expectancy, between the most deprived LAs and the rest of the country before, during and after the health inequalities strategy. Trend is shown as the annual increase (+) or decrease (-) in the absolute gap in life expectancy measured in months.
We only investigated change in inequalities at the area level. It is possible that the observed trends in health inequalities are due to a change in the composition of the populations in those areas, rather than a reduction in inequalities in individual mortality risks. To investigate this we estimate whether the declining trend in health inequalities during the strategy period changed when we adjusted for differential trends in migration.
Data were only available on migration at the local authority level from the ONS for the years 2004 to 2014. To investigate whether migration patterns were likely to influence our results we estimated further models limited to this period with time trend terms for the strategy period (2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012) and the period following the strategy (2013-2014). We calculated migration inflow and outflow rates for international and internal migration for each local authority in each year using migration flows data from the ONS. 3 We then estimated models without (A) and with (B) controls for migration to investigate whether controlling for migration changed the estimate of the trend in health inequalities during the strategy period.
These show that adjusting for migration patterns did not affect the trend in inequalities during the strategy period. Appendix 5. Simulation study investigating likely errors that would result from using a comparison between Spearhead and non-Spearhead areas to identify a change in the trend in spatial socioeconomic inequalities.

A. Estimates for time periods not controlling for migration
We compared life expectancy in the most deprived quintile of local authorities to the rest of the country, between 1983 and 2015, to investigate trends in geographical health inequalities before, during and after the health inequalities strategy. We outline below the reasons why we used this grouping based on income deprivation rather than comparing the Spearhead areas to the rest of the country and provide a simulation analysis to test our assumptions.
The Spearhead areas were selected as local authority areas that were in the bottom fifth nationally for three or more of the following five indicators, between 1995-1997:  Male life expectancy at birth  Female life expectancy at birth  Cancer mortality rate in under 75s  Cardio Vascular Disease mortality rate in under 75s  Index of Multiple Deprivation 2004 (Local Authority Summary), average score These LAs were therefore not just identified because they were socioeconomically deprived but also because they were outliers for low life expectancy and high premature mortality in 1995-1997. There are a number of reasons why the change in the gap in life expectancy between Spearhead areas and the rest of the country may not reflect trends in spatial socioeconomic inequalities (i.e the difference in health between areas defined purely by their socioeconomic conditions).
Firstly as Spearhead areas were selected in part because they had low life expectancy and high under 75 year old mortality in 1995-1997, falls in life expectancy in Spearhead areas during the strategy period may be due to 'regression to the mean' rather than the effect of the strategy. In other words there could be a significant narrowing of the gap between Spearhead areas and the rest of the county when there was no narrowing of socioeconomic inequalities. Secondly , because the selection criteria also mean that Spearhead LAs had a relatively wide range of levels of socioeconomic deprivation there could be no significant narrowing of the gap between Spearhead areas and the rest of the county whilst there was a significant narrowing of socioeconomic inequalities.
To test these two potential sources of bias we conducted two simulation analyses. Firstly we simulated 100 datasets which were the same as the study data except that there was no difference in the trend in male and female life expectancy between LAs between 1983 and 2015i.e there are parallel trends with random variation simulated based on the variance in life expectancy within LAs found in the data. We then simulated cancer and circulatory mortality rates based on the correlations found in the study data. For each iteration, we applied the Spearhead selection criteriaidentifying the LAs that are in the bottom fifth nationally for three or more of the five indicators between 1995-1997, and then conducted the analysis using this 'Spearhead' grouping and separately using our preferred grouping based solely on the income deprivation score of the IMD2004. By design in these simulations the only difference in the trends between local authorities is due to random noise there is no narrowing of socioeconomic inequalities.
In these simulations 73% of the Spearhead models reported a significant narrowing of the gap during the strategy period, although there was actually no difference in trends in life expectancy, by design. This reflects the effect of regression to the mean. Only 6% of the deprivation models reported a significant narrowing of inequalities in these simulations. This is approximately what would be expected since we are using a 5% threshold to define statistical significance (see Table A10).
Secondly we the simulated a further 100 datasets where life expectancy for each LA in each year was drawn from a random normal distribution such that the gap in mean life expectancies between LAs was set to narrow between the most deprived areas and the rest of the country, during the strategy period (2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012). We then simulated premature cardiovascular and cancer mortality for 1995-1997 as above, applied the Spearhead selection criteria in each iteration and conducted the two analyses (1) comparing 'Spearhead' and 'non-Spearhead areas' and (2) comparing income deprived and nonincome deprived areas.
In these simulations 26% of the Spearhead models failed to detect a narrowing of inequalities (when there was one by design), whilst none of the deprivation models failed to detect a significant narrowing of inequalities (see Table A10).
Differences in life expectancy between the most income deprived quintile of local authorities and the rest of the country provides a measure of geographical health inequalities, that would be sensitive to the impact of the strategy, whilst not being affected by the biases outlined above. Thirty-five of this group of 40 deprived local authorities (88%) were Spearhead areas. They will have therefore been affected by actions targeted at the Spearhead areas as well as other broader policies that were targeted more generally at more deprived groups and areas (e.g. The allocation of additional NHS and local government resources to deprived areas, Sure Start, Health Action Zones, New Deal for Communities, introduction of minimum wage and tax and benefit changes). In addition, assessing progress on health inequalities, based on differences in health between groups defined by their socioeconomic status (e.g income), rather than their baseline health status, is more consistent with current definitions of health inequalities used in Europe and expert guidance on their measurement. 30-32 Table A11. Results of 100 simulations for each scenario, showing the % of models indicating a significant reduction in the gap during the strategy period (when there was no difference in trends between LAs by design) and % of models indicating NO significant reduction in the gap during the strategy period when inequalities were simulated to narrow between the most deprived areas and the rest of the country. (2) Inequalities narrow between the most deprived areas and the rest of the country, during the strategy period (2003-2012).

Appendix 6. Trend in life expectancy in Spearhead and non-Spearhead areas
When investigating the trends in life expectancy between Spearhead and non-Spearhead areas it is worth noting that because of the way they were selected, several Spearhead LAs were not particularly income deprived and the less income deprived Spearhead LAs tended to have lower life expectancy in 1995-1997 for their level of deprivation (see Figure A1).

Appendix 7. Identifying 'natural' breakpoints in the trend in inequalities.
As the strategy developed incrementally and it is likely that there was a lag between implementation and any impacts on life expectancy, it was not possible to determine apriori precisely at which time points we might expect the trend in inequalities to change. We therefore investigated empirically whether there was a significant change in the trend in health inequalities around the time of the beginning of the strategy period (between 1997 and 2006) and around the time of the end of the strategy (between 2008 and 2015). We use an iterative search procedure to identify which combination of two breakpointsone at the beginning and one at the end of the strategy provided the best fit for the data by comparing all models with these alternative breakpoints, as well as models with just one of these, or no breakpoints. We then plotted the R-squared values from each of these models to identify the combination of breakpoints that provided the best fit with the data. In other words we fitted 88 separate models each with a different initial and final break points. Figure A4 shows the R-squared from each of these modelsindicating that an initial breakpoint at 2003 and a final breakpoint at 2012 provides the best fitting model compared to all the other alternative break points.

Figure A 4 R-squared from 88 regression models with different breakpoints indicating the best fitting model has an initial breakpoint at 2003 and a final breakpoint at 2012.
i.e the model allowing the trend to change at these points was a better fit than the alternative models.

Investigating natural breakpoints within each of the two groups of local authorities.
To investigate whether there was a change in trend in either or both of the two groups of local authorities (1-deprived areas and 2-the rest of the country) before and after the strategy, we applied the same iterative search procedure separately for these two groups. This confirms what is shown in the full regression results in Appendix 4, that there was a significant upturn in the trend in life expectancy in both deprived and non-deprived areas around 2003. However this change in trend was greater in the more deprived areashence inequalities narrowed. Similarly there was a significant downturn in the trend in life expectancy around 2012 across the country, but this change in trend was greatest in more deprived areas widening inequalities. The decline in inequalities we observed between deprived and less deprived local authorities may not necessarily reflect a decline in inequalities at the neighbourhood or individual level. It is possible that inequalities between local authorities could be reduced if the health of more affluent groups within the deprived local authorities improved more than more deprived groups within those local authorities. In other words there could have been an increase in health inequalities within more deprived local authorities even though inequalities between local authorities reduced. To investigate whether this had occurred or not we analysed data on potential years of life lost (PYLL) in 1997-2001and 2008-2012 for lower level super output areas (LSOA) obtained from the underlying indicators of the IMD2004 and the IMD2015. LSOA are small geographical areas used in reporting small areas statistics, each including a population of around 1600 people and 650 households. We only included LSOAs whose boundaries had not changed between the 2001 and 2011 censuses giving 31671 LSOA for analysis. We then calculated the change in PYLL for each LSOA between these two periods. We then plotted the change in PYLL against the IMD2004 income deprivation score for each LSOA, for England as a whole, for the most deprived local authorities and for the rest of England.
As Figure A6 shows there tended to be a greater decline in premature mortality in the more deprived neighbourhoods, reducing inequalities. Within the most deprived local authorities there was actually a greater decline in inequalities, than was observed in the less deprived local authorities. This suggests that the decline in inequalities observed at the local authority level following the English health inequalities strategy was also observed at the neighbourhood level and that this was achieved in part through reducing inequalities within deprived local authorities as well as between these local authorities and the rest of the country. Appendix 9. Trend in inequalities, poverty measures and government expenditure. Figure A7 shows that income inequalities as measured by the Gini index increased from 1979 to 1990. Although the Gini index remained stable from then on, poverty amongst pensions and children fell substantially from the mid-1990s to 2010. These reductions in poverty were the result of specific tax and benefit measures. 4 Total government expenditure increased markedly between 1997-2010 (see Figure A 8). This was particularly due to increases in spending on health and education, spending on housing and community amenities also increased markedly during this period (see Figure A 8). Part of the strategy was that the distribution of this increase in resources was equity-focused and targeted at the most deprived areas. As can be seen from Figures A8 and A10 increases in NHS and Local Government funding were particularly targeted at the most socio-economically deprived areas rather than at the Spearhead areas specifically. Both the increases in funding and the reductions in poverty could have contributed to the reductions in health inequalities that were observed in this study. To investigate the effect of population revisions following the 2011 census on the trend in health inequalities we recalculated life expectancies, for deprived areas and the rest of the country, using the old unrevised population estimates and compared the trend in the inequality gap using these estimates with the trend using the revised and more accurate population estimates. The gap was reduced slightly, from 2006 using the new population estimates ( Figure A12). Appendix 12. Age specific trends in inequalities.
To investigate whether the trends in inequalities in life expectancy we observed were due to a change in inequalities in mortality in particular age groups we replicated our model using age adjusted mortality rates for 5 age groups 0-19 year olds, 20-44 year olds, 45-64 year olds, 65-74 year olds and over 75 year olds. We then added three way interaction terms to the model between age group, deprivation area, and time trend spline terms. We log transformed the age adjusted mortality rates in order to estimate the trend in relative inequalities in mortality rates, as relative measures are more comparable between age groups. Figure A15 shows that the reduction in inequalities during the strategy period was particularly due to reduced inequalities in mortality in people under the age of 65. The reversal in this trend has largely been in the same age groups, although inequalities in female 0-19 year old mortality continued to decline. Appendix 13. Relative change in deaths under 65. As a sensitivity analysis to check whether our results are influenced by changes in the population estimates over time rather than changes in the number of deaths, we replicated our model using the log of the number of deaths in each LA as the outcome. The model then provides an estimate of the annual change in the relative percentage gap in deaths under 65 between the most deprived 20% of local authorities and the rest of England before, during and after the health inequalities strategy. As this analysis does not use population denominators it cannot be influenced by inaccuracies in population estimates.  Before  0.545