Hospital-level associations with 30-day patient mortality after cardiac surgery: a tutorial on the application and interpretation of marginal and multilevel logistic regression

Abstract

Background Marginal and multilevel logistic regression methods can estimate associations between hospital-level factors and patient-level 30-day mortality outcomes after cardiac surgery. However, it is not widely understood how the interpretation of hospital-level effects differs between these methods. Methods The Australasian Society of Cardiac and Thoracic Surgeons (ASCTS) registry provided data on 32,354 patients undergoing cardiac surgery in 18 hospitals from 2001 to 2009. The logistic regression methods related 30-day mortality after surgery to hospital characteristics with concurrent adjustment for patient characteristics. Results Hospital-level mortality rates varied from 1.0% to 4.1% of patients. Ordinary, marginal and multilevel regression methods differed with regard to point estimates and conclusions on statistical significance for hospital-level risk factors; ordinary logistic regression giving inappropriately narrow confidence intervals. The median odds ratio, MOR, from the multilevel model was 1.2 whereas ORs for most patient-level characteristics were of greater magnitude suggesting that unexplained between-hospital variation was not as relevant as patient-level characteristics for understanding mortality rates. For hospital-level characteristics in the multilevel model, 80% interval ORs, IOR-80%, supplemented the usual ORs from the logistic regression. The IOR-80% was (0.8 to 1.8) for academic affiliation and (0.6 to 1.3) for the median annual number of cardiac surgery procedures. The width of these intervals reflected the unexplained variation between hospitals in mortality rates; the inclusion of one in each interval suggested an inability to add meaningfully to explaining variation in mortality rates. Conclusions Marginal and multilevel models take different approaches to account for correlation between patients within hospitals and they lead to different interpretations for hospital-level odds ratios.

Variations in Use of Optimal Medical Therapy in Patients With Nonobstructive Coronary Artery Disease: A Population‐Based Study

“…The MOR is defined as the median of the set of ORs that could be obtained by comparing 2 patients, 1 at higher risk and 1 at lower risk of the outcome, with identical characteristics from 2 different randomly chosen hospitals 15. Specifically, the MOR is a measure of variation between different hospitals that is not associated with the modeled risk factors 15.…”

Using a Cardiac Arrest Registry to Measure the Quality of Emergency Medical Service Care

“…Multilevel models were used because of 17,18 We included patient-level factors and fiscal year as fixed-effects while allowing for hospital site as a random intercept. Fixed-effects are presented as ORs and 95% CIs.…”

Current practices in feeding tube placement for US acute ischemic stroke inpatients

“…17,18 We present the proportional change in variance as the percent change of between-hospital variance in comparison to a null model. 19 The MOR is defined as the median of a set of odds ratios (ORs) that could be obtained when comparing 2 randomly chosen hospitals with different random effects. The MOR can be interpreted as the median increased risk by moving from a hospital with lower risk to a hospital with higher risk, assuming the same patient covariates.…”

“…The MOR can be interpreted as the median increased risk by moving from a hospital with lower risk to a hospital with higher risk, assuming the same patient covariates. [18][19][20] The ICC, proportional change in variance, and MOR are calculated from the multilevel multivariable regression models. 21 We constructed 2 mixed-effects multilevel multivariable logistic regression models, adjusting for patient characteristics (model 1) and patient and hospital characteristics (model 2).…”