Objective:
Complete documentation in large scale datasets such as administrative data or disease registries is often difficult. Given that the subset of patients with complete data documentation are most likely not a random sample of patients, selection bias threatens the validity of results if a complete case analysis is used. To demonstrate, we will assess the presence and magnitude of selection bias in ischemic stroke patients with documented National Institute of Health Stroke Scale (NIHSS) [[Unable to Display Character: –]] which is often incomplete [[Unable to Display Character: –]] using the Heckman Selection Model.
Methods:
Patient level variables including demographics, comorbidities, clinical EMS and admission variables, and medical history/comorbidities were obtained from 10,717 ischemic stroke patients aged 65 and older in the Michigan Stroke Registry in 2009-2012. The Heckman Selection Model assesses the presence and magnitude of selection bias by estimating a correlation coefficient between error components of a linear regression model predicting patient NIHSS score [[Unable to Display Character: –]] the outcome model [[Unable to Display Character: –]] and a binary probit model predicting NIHSS documentation [[Unable to Display Character: –]] the selection model [[Unable to Display Character: –]] conditional on patient and hospital predictors. The outcome model predicting NIHSS score was specified using a backward selection process with stepwise deletion of non-significant predictors. The selection model included all variables in the outcome model, plus additional significant predictors of NIHHS documentation. Quasi-maximum likelihood estimation was used to produce robust standard errors. All analyses were done using PROC QLIM procedure in SAS.
Results:
7,956 cases (74.2%) of cases had NIHSS documented. Significant predictors in the outcome and selection models are shown in the Table. The Heckman Selection Model found a statistically significant but modest correlation coefficient of
ρ
=0.1089 (SE=0.0119, p<0.0001). The positive correlation indicates that NIHSS was more likely to be documented in patients with higher NIHSS scores, i.e., more severe strokes.
Conclusions:
We found statistically significant albeit weak selection bias in the documentation of NIHSS in stroke patients. The Heckman Selection Model is a novel method that can be used to assess the presence and magnitude of selection bias when missing data is common.
How are ecological approaches justified in European rural development policy? Evidence from a content analysis of CAP and rural development discourses
