Abstract
Background/Aims Propensity scores provide an alternative to traditional regression modeling for assessing the effect of exposure on an outcome in the presence of confounders. We describe methods to establish propensity scores for a cohort of Kaiser Permanente Northwest (KPNW) members receiving surgical procedures of the cervix.
Methods The HMORN Virtual Data Warehouse (VDW) was used to obtain health plan enrollment, diagnosis, procedure, and demographic data for a cohort of 14- to 53-year-old KPNW women during the study years, 1998–2009. Unexposed women were matched to exposed women on age and calendar year. We performed logistic regression to model the propensity of receiving surgical procedures of the cervix. Potential covariates included smoking status, race and ethnicity, sexually transmitted infections (STI), immunocompromised status, length of health plan enrollment, and number of encounters per enrollment year. Propensity score balance was assessed using histograms and boxplots. The common support condition was imposed. Matching on propensity scores was performed using a greedy 5 to 1 digit match.
Results There were 86,898 women in our cohort. Women receiving a surgical procedure were more likely to be current tobacco users, immunocompromised, have a history of STI, have longer prior enrolment in the health plan, and have more healthcare encounters per enrolment year than unexposed women. In our cohort, propensity scores for receiving invasive cervical procedures for exposed women ranged from 0.02 to 0.56 (n=4138; mean=0.08; median=0.06), whereas propensity scores for unexposed women ranged from 0.01 to 0.99 (n=82760; mean=0.05; median=0.04). We excluded 1,819 (2%) unexposed women based on the common support condition. A 1:3 match ratio produced the best matched population. No matches were found for 43 (1%) of our exposed population.
Discussion Use of the HMORN VDW enabled us to create propensity scores for estimating the probability of receiving cervical procedures. Matching and imposing the common support condition decreased our study population minimally. However, women excluded from analysis may be intrinsically different from those who remained. One limitation to our study is that some potential variables we wished to include in our model, such as marital status, were not available in HMORN VDW data which may introduce bias.
