Reader's guide to critical appraisal of cohort studies: 2. Assessing potential for confounding

Logistic regression for creation of propensity score

Creating the Study Cohorts

Published as supplied by the authors

Creating the Study Cohorts

To examine the association between atypical antipsychotic drug therapy and hip fracture, we used linked administrative healthcare databases covering over 1.3 million individuals 66 years of age and older in Ontario, Canada from April 1^st, 1997 through March 31^st, 2002. These patients have universal access to prescription drug coverage, hospital care, and physician services. Specifically, the linked databases included computerized pharmacy records of the Ontario Drug Benefit Program (ODB), which records prescription drugs dispensed to all Ontario residents 65 years of age and older. An overall error rate of less than 1% in this drug database has been reported. These data provided information on drug exposures. Hospitalization records from the Canadian Institute for Heath Information (CIHI) Discharge Abstract Database contain a detailed record of all hospital admissions. These data provided information on past medical history and were the source of information on cases of hip fracture. The Ontario Health Insurance Plan (OHIP) database provided physician billing information for inpatient and outpatient services and contains diagnostic information. This database provided information on past medical history. The Ontario Registered Persons Database (RPDB) contained basic demographic and vital statistics information, including gender and death date, for each Ontario resident. These databases were linked anonymously using encrypted individual health card numbers. Information on the identification of specific drugs, the measurement of hip fracture as the outcome, the measurement of potential confounders and other details of the methods can be obtained from GMA.

Posted as supplied by the authors

Creating a Propensity Score

The creation of a propensity score is based on a logistic regression with the dependent variable taking on the value of "1" if the individual is in the intervention group and "0" if the individual is in the comparison group, and with the potential confounders as the independent variables. The logistic regression is used to estimate the probability that, based on that individual’s potential confounders, that individual would be in the intervention or comparison group. Using the estimated coefficients from the logistic regression model and the measured potential confounders for an individual it is possible to calculate an estimated probability of being in the intervention group for each individual in the study. This estimated probability is called the propensity score for that individual.

The logistic regression determines the extent to which the potential confounders can predict whether an individual is in one group or the other, and provides an indication of the degree to which the groups are unbalanced on these confounders. The C-statistic from logistic regression provides a summary of the accuracy of the overall prediction based on the model and the coefficients for each of the independent variables provides an estimate of the effect of their independent effects.

Tables 1 and 2 contain the results of logistic regressions for Comparisons 1 and 2. For Comparison 1, a logistic regression model that contains no interaction results in a C-statistic of 0.875 indicating that the model can readily discriminate whether an individual is in the intervention or comparison group based on the potential confounders. This provides clear evidence that the groups are quite different in terms of these characteristics. The estimated odds ratio for the diagnosis of dementia is almost 18 indicating that this characteristic, even when one controls for other factors such as age and sex, is much more common in the intervention group than the comparison group. For Comparison 2, even a logistic regression with extensive interaction terms has a C-Statistic of 0.565 indicating that in this study the confounders are not very useful at discriminating between individuals with dementia who received atypical antipsyhcotics and those who received typical antipsyhoctics.

Propensity scores can be included as covariates in regression analyses, be used to define matches, or can be used to create strata. For the purposes of an example, the propensity scores were used to create strata. The results of the analysis of the association between hip fracture and atypical antipsychotic use within each of these strata are provided in a separate document.