Abstract
Background/Aims Non-response is an important potential source of bias in survey research. It is usually difficult to obtain meaningful information on non-respondents to adequately assess bias on estimates of population statistics, especially when the survey involves sensitive or intrusive questions. We combined Electronic Health Record (EHR) and survey data to assess the impact of non-response on estimates of urinary incontinence (UI) prevalence. We used data from the baseline survey of the General Longitudinal Overactive Bladder Evaluation – Urinary Incontinence (GLOBE-UI), a population-based study of the natural history of UI in women = 40 years of age. We conducted the survey on a random sample of 7,125 Geisinger Clinic Primary Care patients. The response to the baseline survey was 57%. We also used electronic health records (EHR) data to obtain demographic, care utilization, health behavior, social history, and clinical information on all eligible respondents of the GLOBE-UI survey.
Methods We complete a two-stage model. First, we used logistic regression to assess baseline characteristics associated with response status. The predicted probabilities were used as a weight in the second stage model. UI status was defined for each response based on survey data. An Inverse Probability weighting (IPW) logistic regression model was applied to adjust non-response bias and to determine prevalence of UI based on baseline risk factors associated with UI status, including baseline characteristics, health utilizations, social and educational history, and comorbidities. The non-response adjusted estimate of prevalence was compared with prevalence obtained from respondents.
Results Age at survey, smoking, marital status, number of outpatient encounters per year, CHF, dementia, and severe liver disease were significantly associated with response status and UI status. Adjusting for non-response bias resulted in a reduction of the UI prevalence estimate from an observed value of 31% to an adjusted value of 28% for those 60 and older. The observed and adjusted estimates were not different for younger respondents.
Conclusions Non-response bias in women with UI is limited. It may lead to a marginal overestimate of UI prevalence for older patients. This could be explained by the non-response group being relatively younger.
