Recommended peers is a list of 10 trusts that are similar to yours according to a number of factors that affect productivity. These are factors the trust cannot change in the short to medium term, and that have an impact on cost per weighted activity unit (WAU).
The recommended peer set has been used to calculate many of the productivity opportunities featured in the Model Hospital.
Updates December 2018
Following feedback we've updated the recommended peers list in December 2018: we’ve improved the algorithm, now including factors such as casemix, PFI and Major Trauma status, and validated the results by cross-checking the peers with those selected by trusts themselves. Future developments will include recommended peers for individual specialties and compartments.
Factors used to select recommended peers
The list of variables used is extensive but includes:
- demographic data such as age ranges, deprivation indices and urbanity
- economic data such as operating expenditure and market forces factors (MFF) payment index
- geographic data such as eastings, northings, number of attendances and number of sites
- performance data such as delayed transfer of care (DTOC) rate and staff absence rate
- logistic data such as percentage of emergency admissions and student staff rates
- clinical data such as casemix at HRG level, degree of specialism and a disease prevalence metric
- estates data such as PFI and Major Trauma sites
How the list was developed
The Model Hospital’s analytics team developed this list building on work with NHS Digital.
In doing this work we spoke to a lot of trusts about what peers they like to compare themselves against. We found they used a range of different criteria and there was no single right way to do this. The method we used applies the same criteria objectively to each trust and prioritises similarity on specific factors that the data tell us are most important in affecting productivity.
The Recommended peers list uses a range of mathematical techniques to approximate the best possible set of peers. It considers as many useful variables as possible and weights these variables on their relevance to a trust’s cost per weighted activity unit (WAU).
We took a 2-stage approach to development:
- First we built a supervised model (a sophisticated algorithm) to try to capture the relationship between all the available variables and each trust’s productivity, as measured by their overall cost per WAU. This provided a set of coefficients (multipliers) that show the ‘importance’ of each variable.
Linear regression, looking for relatively simple relationships between variables and productivity, yields a straightforward set of coefficients, but this method poorly predicts cost per WAU. So instead we used a more sophisticated method called gradient boosting, which uses a decision tree approach to take into account relationships between variables. Gradient boosting gives a much better approximation of cost per WAU, and it is easy to extract the importance of each metric from the resulting rule tree.
- We then used the resulting multipliers in a weighted Euclidian distance ranking tool, inputting data for every variable for every trust, to see which trusts are closest to one another in terms of the factors that predict productivity.
Using this approach we are able to rank every trust with regard to its suitability as a peer for every other trust. We can then choose the top 10 ranked peers as a suggested set of peers.