Nasal Polyps

Selection bias was assessed by comparing respondents and non-respondents to the CRISP Questionnaire on EHR data that included age, race/ethnicity, sex, history of receiving Medical Assistance (surrogate for socioeconomic status (SES))[14 ] the Charlson co-morbidity index [15 ] and EHR sampling strata. Next, prevalence of CRS and CRS symptoms was estimated using SAS Proc SURVEYFREQ [16 ] to account for sampling fraction and adjust for differential selection using previously published probabilities and participation weights, a standard methodology for referring back to the source population [17 –18 ].
We conducted unadjusted and adjusted analyses to identify sociodemographic, smoking, and health factors associated with CRS and CRS symptom subgroups. Sociodemographic factors evaluated included sex, race/ethnicity, age, and lifetime Medical Assistance status. Health outcomes included self-reported physician diagnosis of CRS, nasal polyps, asthma, aspirin-intolerant asthma, and hay fever, as well as migraine and fatigue determined by previously validated scales [19 (link)]. In unadjusted analysis we used the chi-square test to compare these factors between current CRS and never CRS and across CRS symptom subgroups. In adjusted analysis, we evaluated sociodemographic and health factors separately. For sociodemographic factors we created separate logistic regression models for current CRS (vs. never CRS) and each of the CRS symptom subgroups (vs. never CRS). Each model included sex (male, female); race/ethnicity (white, other), age (linear and quadratic terms to allow for non-linearity), smoking status (current, former, never), and Medical Assistance status (yes, no). To identify health outcomes associated with CRS, we used separate logistic models for each health outcome. Each model included a categorical variable for CRS status (never (reference), OBS/DIS, PP, SL, PPSL) as the primary independent variable and the sociodemographic and smoking variables as covariates.
All modeling was done using SAS Proc SURVEYLOGISTIC procedure, which accounted for the complex sampling design, and utilized sampling and participation weights [17 –18 ]. Weighted analysis balances bias with precision in association estimates; bias is reduced with weighted analysis but precision is also reduced (larger standard errors leading to larger confidence intervals) [17 –18 ]. One of the sampling weights was an order of magnitude larger than any other (weight = 150), because of the under-sampling of white subjects in the “No sinus codes” group. Using this weight severely inflated standard errors, a known consequence of large sampling weights [17 –19 (link)], so it was truncated to the next largest value (weight = 32), a standard method for dealing with large weights.