Purpose: Patients need to be properly informed to choose between health care providers. We aim to compare different methods for analyzing outcome and ranking providers according to their performance in stroke care.
Method: The Netherlands Stroke Survey was conducted in 10 centers in the Netherlands, with prospective and consecutive enrolment of 579 patients with acute stroke. Poor outcome was defined as dead or disabled at 1 year (modified Rankin scale ≥3). We calculated unadjusted and adjusted between center differences in poor outcome using fixed and random effects logistic regression models. We ranked the centers using a new method, which calculates the expected rank (ER). The ER incorporates both the magnitude of the differences between centers and the uncertainty in these differences.
Result: At 1 year, 268 patients (53%) had a poor outcome. The fraction of patients with poor outcome varied between 29% to 78% between centers (unadjusted χ2 = 48, 9 df, p<0.0001). Adjusting for 12 patient characteristics led to halving of the differences seen in unadjusted analysis (χ2 = 23, 9 df, p=0.006). Further changes were seen with random effect estimation that accounted for uncertainty, with shrinkage of differences to zero, especially for smaller centers. Ranking the centers by using the ER led to substantial shrinkage of the original ranks (1-10) towards the median rank of 5.5.
Conclusion: To provide patients with reliable information to choose between providers, relatively advanced statistical methods should be used that account both for case-mix adjustment and uncertainty. If ranking is attempted, it should be made with expected rank.
Candidate for the Lee B. Lusted Student Prize Competition