Applying Hodges-Lehmann scale parameter estimates to hospital discharge times
BACKGROUND: Clinical trials aimed at shortening the time-to-discharge need to have rational and easily understood effect size estimates for health care management organizations. A natural choice is a scale model, where the distribution of time to discharge on one treatment is assumed to be the same as rho times that on the other. If, for example rho=0.6, then one treatment is associated with a 40% reduction in discharge time. In designing and analyzing these studies, we need to have the capability to accommodate right censored data, as it is plausible that some patients may never meet the discharge criteria, even though they do get discharged.
PURPOSE: Utilizing the ideas of Hodges and Lehmann, to provide methods for analysis of trials aimed at shortening hospital discharge times, using point and interval estimates of scale parameters based on the Gehan generalization of the Mann-Whitney-Wilcoxon test, which accommodates right censoring for situations where patients never meet discharge criteria (+infinity).
METHODS: For every value of rho > 0, we shall test the null hypothesis that the distribution of discharge times on one treatment is the same rho times that on the other. The values of rho that we fail to reject form the confidence interval for the true rho.
RESULTS: The methods were developed and applied to a real clinical trial for times to meet the three objective discharge criteria in knee replacement surgery for two post-operative pain control strategies (usual care plus a perineural infusion of either placebo or 0.2% ropivocaine, until the morning following surgery). Based on 48 randomized patients, the point estimate (95% confidence limits) for rho was 0.47 (0.32-0.67), favoring ropivocaine.
LIMITATIONS: The methods cannot as yet be applied to group sequential designs or studies with more than two treatments.
CONCLUSION: This methodology is an effective way to analyze two-arm trials involving continuous hospital discharge time data.