Quantifying selection bias in national institute of health stroke scale data documented in an acute stroke registry

Abstract

Background - As a measure of stroke severity, the National Institutes of Health Stroke Scale (NIHSS) is an important predictor of patient- and hospital-level outcomes, yet is often undocumented. The purpose of this study is to quantify and correct for potential selection bias in observed NIHSS data. Methods and Results - Data were obtained from the Michigan Stroke Registry and included 10 262 patients with ischemic stroke aged ≥65 years discharged from 23 hospitals from 2009 to 2012, of which 74.6% of patients had documented NIHSS. We estimated models predicting NIHSS documentation and NIHSS score and used the Heckman selection model to estimate a correlation coefficient (ρ) between the 2 model error terms, which quantifies the degree of selection bias in the documentation of NIHSS. The Heckman model found modest, but significant, selection bias (ρ=0.19; 95% confidence interval: 0.09, 0.29; P<0.001), indicating that because NIHSS score increased (ie, strokes were more severe), the probability of documentation also increased. We also estimated a selection bias-corrected population mean NIHSS score of 4.8, which was substantially lower than the observed mean NIHSS score of 7.4. Evidence of selection bias was also identified using hospital-level analysis, where increased NIHSS documentation was correlated with lower mean NIHSS scores (r=-0.39; P<0.001). Conclusions - We demonstrate modest, but important, selection bias in documented NIHSS data, which are missing more often in patients with less severe stroke. The population mean NIHSS score was overestimated by >2 points, which could significantly alter the risk profile of hospitals treating patients with ischemic stroke and subsequent hospital risk-adjusted outcomes.