Background — Data in health research are frequently structured hierarchically. For example, data may consist of patients treated by physicians who in turn practice in hospitals. Traditional statistical techniques ignore the possible correlation of outcomes within a given practice or hospital. Furthermore, imputing characteristics measured at higher levels of the hierarchy to the patient-level artificially inflates the amount of available information on the effect of higher-level characteristics on outcomes.
Methods — Conventional logistic regression models and multilevel logistic regression models were fit to a cross-sectional cohort of patients hospitalized with a diagnosis of acute myocardial infarction. The statistical significance of the effect of patient, physician, and hospital characteristics on patient outcomes was compared between the 2 modeling strategies.
Results — The 2 analytic strategies agreed well on the effect of patient characteristics on outcomes. According to the traditional analysis, teaching status was statistically significantly associated with 5 of the 9 outcomes, whereas the multilevel models did not find a statistically significant association between teaching status and any patient outcomes. Similarly, the traditional and multilevel models disagreed on the statistical significance of the effect of being treated at a revascularization hospital and 3 patient outcomes.
Conclusions — In comparing the resultant models, we see that false inferences can be drawn by ignoring the structure of the data. Conventional logistic regression tended to increase the statistical significance for the effects of variables measured at the hospital-level compared to the level of significance indicated by the multilevel model.
View full text
Coronary disease/Myocardial infarction