Inside England and Wales's prisons crisis: Annex and regression methodology

Annex 

This table shows all prisons that are in the worst 20% of prisons for at least half of the metrics we have data for (the worst performing) and all prisons that are in the best 20% for at least half the metrics we have data for (the best performing). Lowdham Grange has also been included in the worst-performing category, despite being in the worst 20% for only 4 out of 9 metrics where we have data; this is because of significant data gaps and performance in the prison being so bad during 2023 and 2024 that it was taken back into public ownership. The numbers in the table show the decile each prison falls into for each category, with 1 meaning the best 10%, 2 meaning between 10-20%, and so on. 

We selected these performance metrics from the Ministry of Justice’s published performance statistics to cover a variety of areas of performance. We did not include any scores from the prisons inspectorate or other sources, only directly measured variables. Some metrics, such as the rate of escapes/absconds, were excluded as levels are so low they do not show significant variation across prisons or over time.

Regression methodology

Data

The dataset for this analysis includes all public prisons in England and Wales. [Data is taken from several Ministry of Justice datasets: 

• ‘HMPPS Annual Digest, April 2023 to March 2024’
• ‘HM Prison and Probation Service workforce quarterly: March 2024’
• ‘Annual Prison Performance Ratings 2023 to 2024 Supplementary Tables’. 

Private prisons (14 out of 119 total prisons) are excluded because they are not required to report several key metrics, including staffing and the share of prisoners in purposeful activity. 

For some prisons in some years, data is not available in the MoJ data because publishing the data would be disclosive and so values are suppressed. We exclude these prisons from analysis when using these variables. In total there are 101 public prisons in 2023/24, but seven have unavailable data for some key metrics like incidents at height. 

The results presented in the body of the report are all based on data from 2023/24, as this is when we have the most complete data to let us control for as many factors as we can. Where possible, we have tested the results with data from other years to confirm that the relationship still held, which it did in all cases. 

To understand the effect of different types of staff on outcome variables of interest, we group staff into two groups: non-operational staff (those not involved in day-today interactions with prisoners) and operational staff. We also include variables for the share of operational staff in different categories: Band 2 (junior operational staff), Band 4 supervisors and Band 5s (senior officers) and Band 6+ (senior managers). The excluded category are prison officers in Bands 3 and 4. Coefficients on these shares should be interpreted as the effect of a 1ppt of total operational staff change from Band 3 and 4 officers to the category. 

Where relevant, all variables are expressed as either a rate per 1,000 prisoners or a share of all prisoners. This is specified in regression tables and in the list below.

Explanation of approach taken 

Our main results are based on cross-sectional linear regressions in 2023/24. This is the year in which data is available for all variables of interest. In addition, we conduct robustness checks including all years 2020/21 to 2023/24 where data is available (for some variables, specified in the tables below, data is not available in other years). For regressions covering multiple years, we cluster standard errors at the prison level. 

Interpretation and presentation 

The analysis controls for as many prison characteristics as the data allow, but it is possible that there are other features of prisons that we cannot control for that drive the results we observe. As a result, the regressions should be interpreted as a description of the observed relationship between prison characteristics, rather than necessarily causal relationships. 

To explain the findings of the regressions more clearly, in some places in the report we demonstrate the implied effect of one variable on another at the median. To do this, multiply the coefficient on the independent variable by the difference between the 50th and 75th percentile of that variable. We then compare this to the median of the dependent variable to demonstrate the estimated size of the effect. 

Full results 

Drivers of violence and protest in prison 

In our analysis, we looked at five measures of violence and protest in prisons (each measured as a rate per 1000 officers): prisoner-on-prisoner assaults, assaults on staff, self-harm incidents, incidents at height (an established prison protest) and barricades. As there are fewer barricade incidents and incidents at height, the value is suppressed for some small prisons. Results for 2023/24 are presented in Table 1. 

Results for 2020/21 to 2023/24 are presented in Table 2.

Relationship between different measures of violence and protest, 2021/22-2023/24 

We also explored whether the different measures of violence and protest in prisons were correlated with one another. The results are shown in table 3.

Drivers of purposeful activity in prison 

We also conducted regressions looking at the best predictors of purposeful activity in prisons. We looked at two variables: the share of prisoners undertaking purposeful activity, and the share receiving vocational qualifications. 

Results for 2023/24, and combined results for 2022/23 and 2023/24, are presented in Table 4. As this data is only available from 2022/23, we cannot use earlier years of data.

Drivers of staff resignation in prison 

To understand the predictors of staff sentiment, we conducted regressions with staff resignation rates and staff sickness rates as the dependent variables. Data on staff resignation rates is not available in 2022/23, and data on staff sickness is not available in 2020/21, 2021/22 or 2022/23.

Drivers of cost in prison 

We also explored the drivers of total prison costs per prisoner. In one specification, we do not control for staffing (so staff costs are included). In another specification, we do control for staffing, so results are conditional on staffing levels. This data is only available in 2022/23

Drivers of post-prison outcomes 

The final set of regressions explores the relationship between prison characteristics and post-prison outcomes (whether an individual is housed on the first night after prison, and whether they are employed six weeks later).