Reviewing deaths in British and US hospitals: a study of two scales for assessing preventability

Abstract

Background: Standardised mortality ratios do not provide accurate measures of preventable mortality. This has generated interest in using case notes to assess the preventable component of mortality. But, different methods of measurement have not been compared. We compared the reliability of two scales for assessing preventability and the correspondence between them.

Methods: Medical specialists reviewed case notes of patients who had died in hospital, using two instruments: a five-point Likert scale and a continuous (0-100) scale of preventability. To enhance generalisability, we used two different hospital datasets with different types of acute medical patients across different epochs, and in two jurisdictions (UK and USA). We investigated the reliability of measurement and correspondence of preventability estimates across the two scales. Ordinal mixed effects regression methods were used to analyse the Likert scale and to calibrate it against the continuous scale. We report the estimates of the probability a death could have been prevented, accounting for reviewer inconsistency.

Results: Correspondence between the two scales was strong; the Likert categories explained most of the variation (76% UK, 73% USA) in the continuous scale. Measurement reliability was low, but similar across the two instruments in each dataset (intraclass correlation: 0.27, UK; 0.23, USA). Adjusting for the inconsistency of reviewer judgements reduced the proportion of cases with high preventability, such that the proportion of all deaths judged probably or definitely preventable on the balance of probability was less than 1%.

Conclusions: The correspondence is high between a Likert and a continuous scale, although the low reliability of both would suggest careful measurement design would be needed to use either scale. Few to no cases are above the threshold when using a balance of probability approach to determining a preventable death, and in any case, there is little evidence supporting anything more than an ordinal correspondence between these reviewer estimates of probability and the true probability. Thus, it would be more defensible to use them as an ordinal measure of the quality of care received by patients who died in the hospital.

Keywords: Adverse events, epidemiology and detection; Medical error, measurement/epidemiology; Mortality (standardized mortality ratios); Quality improvement methodologies.

Figure 1

Association between the continuous (0–100) preventability scale and Likert categories. (A) UK study. (B) US study. The line connects the median preventability scores within each Likert category. Three points in the UK study have been circled as logically inconsistent and were removed from subsequent analyses.

Figure 2

Estimated calibration plots for reviewers (22 UK; 13 USA) from a random slopes ordinal regression model. (A) UK study. (B) US study. A point on a curve represents the average latent score assessment given by an individual reviewer to a case note with a given percent preventability score. The shaded grey area defines a 95% prediction region over the population of reviewers for the mean latent score assigned to a given preventability percentage. The horizontal lines represent divisions on the latent scale corresponding to the Likert categories L1–L5.

Figure 3

Likert category distributions: (i) raw data and (ii) predictive case note distribution after model-fitting. (A) UK study. (B) US study. For both distributions, the median lies in Likert category 4, that is, ‘probably not preventable’.

Figure 4

Distributions of percentage preventability: (i) raw data and (ii) predictive case note distribution after fitting a categorical model. (A) UK study. (B) US study.