Wednesday, 29 January 2014

Not Fit for Work: Statistical analysis of bias in the Work Capability Assessment.


I am repeating an analysis I have conducted a few times before, using the most up-to-date information available. The purpose of this analysis is to understand the effects of socioeconomic deprivation and health on the outcome of the Work Capability Assessment (WCA), the gatekeeper test for Employment & Support Allowance (ESA).  As educational differences were floated by @dr_wood_wca as a potential explanation of the observed effects, I’m including an analysis of education also.

Previous analyses found a complex relationship between the sets of variables, with socioeconomic deprivation and health both having unexpected relationships with the difference WCA outcomes: the proportion of people going into the WRAG, the Support Group, and those being found fit for work.

The analysis focuses on Local Authorities and similar areas in England (which I will call ‘local areas’), and not individual people. However, data from the 827,200 WCAs completed as part of the migration from Incapacity Benefit have been included, separated into the 324 administrative regions of England. This data does not include the ~60,000 claims either currently awaiting WCA or withdrawn prior to WCA.

A cursory glance at the available WCA statistics suggests problems with the test. The proportion of claimants being found fit for work varies between 13% and 37%, and the proportion of claimants going into the Support Group varies from 26% to 48%. That some areas are finding more than double the proportion of claimants fit for work than others is concerning, and warrants further investigation.

A brief statistics lesson

A correlation is a measure of the strength of relationship between two variables, and is measured in ‘r’. When r = 1.00, the two variables increase at the same rate. When r = -1.00, one variable increases as the other decreases, but the rates remain the same. r = 0.00 means there is no linear relationship between the variables. An r value that is significantly different from 0 but between -1.00 and 1.00 means there is a linear relationship but it is not perfectly proportional,  This is illustrated in the table below, and examples of how different correlations are graphed is shown here.

r = 1.00
r = -1.00
r = 0.00
Variable 1
Variable 2
Variable 1
Variable 2
Variable 1
Variable 2

Statistical tests are significant when p is equal to or less than .05. This means that the probability of getting a result like this, assuming that the relationship is not real, is 5% or less. p can have other values, such as .001. This means that the result of getting a result like this, assuming there is no true relationship, is less than 1%.


For my measure of socioeconomic deprivation, I am using a composite measure of both the proportion of a local area’s population living in one of the 20% most deprived LSOAs (Lower layer Super Output Area; a small administrative region of approximately 2,000 people) and the proportion of the local area’s population living in one of the 20% least deprived LSOAs. This composite variable will be known as “deprivation”.

Deprivation has a long established relationship with health. The relationship between life expectancy and deprivation in this dataset is r = .806, a very strong relationship. Therefore it is expected that there will be a significant effect of deprivation on WCA outcomes.

Deprivation showed a significant correlation with both the proportion of people going into the Support Group (r = -.304, p < .001) and the proportion of people being found fit for work (r = .443, p <.001). This shows that as deprivation increases, fewer people go into the Support Group of ESA, and more people are found fit for work.

In the 30 areas with the highest proportion of fit for work judgements, 29.50% of the population live in one of England’s 20% most deprived areas. In the 30 areas with the lowest proportion of fit for work judgements, this number is 1.10%.

Health and Disability

Considering ESA’s role as a benefit for those whose health inhibits their ability to work, it seems obvious that there will be a relationship between an area’s general health and local WCA outcomes. However, it is necessary to statistically control for the effects of deprivation on health and the other variables to truly see the effect that health has on WCA outcome.

There were two significant partial correlations when controlling for deprivation; life expectancy was positive related to the proportion of Support Group judgements (r = .300, p <.001) and negatively related to the proportion of fit for work judgements (r = -.287, p <.001). The same results were found with two other general health variables; local rate of early cancer deaths and local rate of early cardiovascular disease deaths. This shows that, paradoxically, a higher proportion of WCAs place claimants into the Support Group in healthier areas, while more people are found fit for work in less healthy areas.
In the 30 areas with the highest proportion of fit for work judgements there are an average of 23 (24%) more early cancer deaths per 100,000 population, 24 (51%) more early cardiovascular disease deaths and a 2.72 (3.4%) year reduction in average life expectancy.

I also included a brief analysis of disability, using data from the 2011 census. The data consists of self-reports of disability, and the proportion of respondents reporting severe or minor disability based upon how much it affected their day-to-day lives. Given that I have already established a relationship between general health, deprivation and WCA outcome, I will control for these effects when examining the relationship between disability severity and WCA outcome.

Population reporting no disability
Population reporting moderate disability
Population reporting severe disability

As the controlled analysis and graphs show, there is a significant variation in WCA outcomes across areas reporting different degrees of severe, moderate and little disability. As levels of moderate and severe disability rise, the proportion of claimants going into the Support Group – the group of ESA intended for the most severely disabled claimants – goes down.


Finally, it was posited that education would play a role in WCA outcomes. This is feasible, as a good education could confer an ability to research and use information relevant to a WCA, such as understanding the criteria or an ability to explain symptoms and conditions sufficiently.

To this end, I conducted a partial correlation between the proportion of GCSE students achieving 5 A*-Cs including Maths and English and WCA outcome, controlling for deprivation and health for their known role in educational attainment. There were two significant findings; a positive correlation with Support Group judgements (r = .244, p < .001) and a negative correlation with WRAG judgements (r = -.226, p <.001).

To get a better idea of how all these factors interact to affect WCA outcomes, I built a path model to analyse the effects on both Support Group and Fit for Work outcomes, as there were not enough data to legitimately model all three WCA outcome variables. This model was a reasonable fit for the data (GFI = .910, RMSEA = .239), and illustrates the complex relationship between the variables.

The model shows that, for example, as deprivation increases by 1 standard deviation, life expectancy decreases by 0.81 standard deviations, and as life expectancy increases by 1 standard deviation, the proportion of people found fit for work decreases by 0.43 standard deviations.

From this model, it appears that deprivation itself has a small but real effect upon WCA outcome; higher deprivation increases the proportion of fit for work judgements slightly, and increases the proportion of Support Group judgements. However, the consequences of deprivation have a much more powerful impact upon WCA results. Lowered life expectancy, for example, results in more claimants judged fit for work, but lowered life expectancy carries some of that effect forward by making people more disabled, which in turn reduces the number of people entered into the Support Group.

Given this model’s relatively poor statistical fit, it is intended mostly as an illustration of how the effects I’ve discussed can interact, and is not intended to be a watertight portrayal of the ‘truth’ of what is really happening with Work Capability Assessments.


Given the rather intricate relationship between the variables I studied, it is difficult to draw firm conclusions about their exact relationship. However, one thing is clear: the WCA is unfair.

People in poorer areas are at a disadvantage when claiming ESA and going through the Work Capability Assessment. It is clear that this is in part due to indirect effects (e.g. lowering educational attainment), but it also seems likely that there is also a direct effect. For example, in the 30 areas where people are found fit for work the most often, 30% of the population lives in one of the 20% poorest LSOAs. In the areas where people are found fit for work the least often, this is 1%.

Secondly, people living in the unhealthiest areas are at a severe disadvantage in the WCA, being found fit for work more frequently than people in healthier parts of the country. For example, in the areas that most frequently place claimants into the Support Group (average 45%), there are 100 early cancer deaths per 100,000 people per year. In the areas that place the least people in the Support Group (average 29%), there are 115 early cancer deaths, an increase of 15%.

It’s worth bearing in mind that this is a repeat of an analysis I conducted previously. The results are broadly the same, but this analysis was more in-depth. These are largely similar to the results that I sent my MP, who in turn forwarded them to Mark Hoban (then Minister of Employment, forwarded by Esther McVey, Minister for Disabled People).  His response talked about delays, tribunals, the Harrington review, Atos training, the supposed benefits of employment, and the fact that ESA and the WCA were Labour inventions. Basically, everything EXCEPT the issues I raised. For fairness, his letter is available at this Imgur link. The dataset I used, and my sources, are available at this Google Docs link - every source is Government data, easily accessible to the general public.