Case-mix methodology for the NHS outcomes framework GP patient survey questionnaire data

Objective

The objective of the research described in the current report was to explore alternative methodologies which could be used to determine whether the health status of people living with long-term conditions in England is changing over time, all other factors being equal.

Method

Data from the Health Survey for England (HSE) were used in the analyses and EQ-5D was used to represent health related quality of life (HRQoL). The proposed case-mix ratio approach which utilised ordinary least square regressions (with the EQ-5D preference-based score as the dependent variable) was replicated, and alternatives using logistic regressions and two-part models (both using the responses to the EQ-5D health dimensions as the dependent variables) were explored. An alternative method using the HSE year as a performance indicator (PI) was explored and results presented for the four most prevalent health conditions. Results were compared in terms of errors in predicted scores and the ability to capture changes in the distributions of the preference-based scores. Both expected and simulated values were compared.

Results

The EQ-5D data were not normally distributed irrespective of survey or health condition. The annual fluctuations in mean EQ-5D scores, and the proportions in full health, were relatively small overall but differed substantially by health condition. The annual fluctuations in mean EQ-5D scores did not necessarily describe the shifts in the EQ-5D distributions.

Comparing the predicted results from the ordinary least squares (OLS) regressions and the health dimensions models, magnitude and statistical significance of the coefficients in the models differed by health condition. While the linear model was more accurate in terms of errors in the mean of predicted values for the base year (2003), it was less accurate than the logistic models for two of the remaining four surveys. The approaches were not particularly accurate at predicting EQ-5D scores across the full range of the EQ-5D index. However, the dimension models replicated the observed distributions well, unlike the linear models which produced a normally distributed sample with a proportion of scores outside the bounds of the index. The substantial errors in the predicted scores had implications with regard to the face validity of using a case-mix adjustment factor, which was based on a ratio of individual observed and predicted scores.

4

The results for the performance indicator models were promising and again the logistic dimension models out-performed the linear models. The magnitude and statistical significance of the coefficients in the models were both condition and health dimension specific. The linear models again predicted mean EQ-5D scores more accurately than the dimension models, but the latter performed better across the range of the EQ-5D index in terms of mean errors and mean absolute errors. This was reflected in distributions of predicted scores as the linear models predicted scores outside the EQ-5D range, covered a truncated range and did not capture the characteristics of the actual data.

Conclusion

While linear models obtained using OLS regressions performed well on the aggregate level, they did not capture the underlying distributions of the EQ-5D scores and were not able to detect shifts in these. The bias in the errors of predicted values raised questions relating to confidence in any case-mix adjustment derived from a ratio based in individual predicted scores. The results from the logistic models appeared to capture the underlying distributions far better than the linear models but additional research is required to develop this approach further.