Background Reducing time-to-care is crucial in many acute and chronic diseases. Quality indicators based on target delays derived from guidelines are used to compare hospital performance but there is no accepted methodology for comparing performance when no target delay has been established.

Aim To explore by different statistical methods the uncertainty in hospital comparisons that are based on delay indicators, when no target delay is available.

Methods Data for hospital door-to-needle time were extracted from a 2010 study of 1699 patients in 57 hospitals with ST-elevated myocardial infarction. We determined whether the times in each hospital were statistically different from the overall mean time or the median time for all hospitals by (i) one-way analysis of variance (ANOVA), (ii) non-parametric ANOVA with Nelson–Hsu adjustment (ANOVA R) and (iii) the proportional hazard model (PHM). We also tested for the assumptions underlying the methods: normal distribution for ANOVA, homogeneity of variances (homoscedasticity) for ANOVA and ANOVA R, and proportionality for PHM.

Results Door-to-needle times were available for 889 patients in 44 hospitals. Data distribution was not Gaussian. Test assumptions were verified for ANOVA R (homoscedasticity) for one data subset (>48-h times (48H) excluded) and for PHM (proportionality) for two data subsets (48H or >95th percentile (P95) times excluded). The same five significantly better performers were identified in each case (although ANOVA R missed one). ANOVA R (48H) identified two significantly poorer performers, PHM (48H) identified three and PHM (P95) just one. Poorer performers differed according to method.

Conclusions The tested statistical methods yielded broadly similar results but no method was truly satisfactory. A transparency statement should therefore always specify the ranking method used to compare hospital performance.