Do people have suggestions or criticisms of how to handle risk-adjustment at hospital/center level when there are too few observations? NSQIP and CMS use shrinkage estimators but I worry this is a poor, unreliable assumption when there are too few observation. I would prefer to just drop centers with too few observations from being “risk-adjusted.” Are there better statistical solutions. Appreciate any formal criticisms or citations against shrinkage estimation.
My understanding is that it is worse to remove small centers than to use the (huge) shrinkage they will receive. Some general documents that may touch on the subject are here. Even a center with a few observations will contribute information to the whole analysis.
I would still include all centers to generate coefficients in the prediction model. My issue is how to report on the risk-adjusted rates for centers with too few observations. The shrinkage methods described will pull risk-adjusted estimates towards the grand mean, but at certain point that may be an unreasonable approach that ignores case-mix. Risk-adjustment itself seems like a very arbitrary concept when model fit may be poor to start with, yet this is done routinely for both mortality and hospitalization risk.