We’re putting together a study of an intervention in the emergency department (ED). The outcome is length of stay, LOS (which we hope will be reduced). It is a step-wedge design with several clusters of two hospitals each. Each cluster will transition from control to intervention at 1-month intervals. There will be a two-month run-in phase.
The current version of the analysis model is a Generalised linear mixed model:
LOS ~ Study_arm + Study_site + Cluster + Time_of_presentation
Cluster is the month of crossover for a site.
Time of presentation is the month of presentation for an individual.
Cluster and Study site are random. Time of presentation and Study arm (Control or Intervention) are fixed.
I have three questions/uncertainties:
- LOS is also affected by business of EDs. In NZ this has been increasing over time (staffing and physical space not keeping up with demand). How best to account for this?
- Lockdowns and Covid rates affect LOS in ED. How best to account for these? [I may not have available Covid rates by site, but may have vaccination rates which could act as a surrogate].
- The LOS is a funny distribution. The figure attached is from an earlier observation study in the same cohort following a different kind of intervention. The two or three peaks/humps are real given the clinical pathway. When I look at qqplots for linear models of LOS with this data they don’t look good. Should I transform LOS first (eg by a restricted cubic spline with at least 5 knots)?