Using statutory health insurance data to evaluate non-response in a cross-sectional study on depression among patients with diabetes in Germany

Background: Low response rates do not indicate poor representativeness of study populations if non-response occurs completely at random. A non-response analysis can help to investigate whether non-response is a potential source for bias within a study.

Methods: A cross-sectional survey among a random sample of a health insurance population with diabetes (n=3642, 58.9% male, mean age 65.7years), assessing depression in diabetes, was conducted in 2013 in Germany. Health insurance data were available for responders and non-responders to assess non-response bias. The response rate was 51.1%. Odds ratios (ORs) for responses to the survey were calculated using logistic regression taking into consideration the depression diagnosis as well as age, sex, antihyperglycaemic medication, medication utilization, hospital admission and other comorbidities (from health insurance data).

Results: Responders and non-responders did not differ in the depression diagnosis [OR 0.99, confidence interval (CI) 0.82-1.2]. Regardless of age and sex, treatment with insulin only (OR 1.73, CI 1.36-2.21), treatment with oral antihyperglycaemic drugs (OAD) only (OR 1.77, CI 1.49-2.09), treatment with both insulin and OAD (OR 1.91, CI 1.51-2.43) and higher general medication utilization (1.29, 1.10-1.51) were associated with responding to the survey.

Conclusion: We found differences in age, sex, diabetes treatment and medication utilization between responders and non-responders, which might bias the results. However, responders and non-responders did not differ in their depression status, which is the focus of the DiaDec study. Our analysis may serve as an example for conducting non-response analyses using health insurance data.