gms | German Medical Science

Introduction: The German DIVE (Diabetes Versorgung-Evaluation) registry is a prospective, observational, multi-centre database established in 2011. It contains information on antidiabetic treatment prescription, concomitant therapies, and other covariate values of patients diagnosed with type 2 diabetes over the course of time [1]. Motivated by the identification of patient-related prognostic characteristics regarding the initiation and failure of a basal supported oral antidiabetic treatment strategy (BOT), this overview demonstrates in which way extended time-to-event techniques can adequately describe the data and how they improve effect detection. We also discuss common problems accompanied by the analysis of register-based data [2].

Methods: The temporal pattern of diabetes medication requires an application of time-to-event methodology. To obtain a precise formulation of the situation at hand, the primary endpoints (initiation-/ end-of-treatment) are seen to be subject to competing risks [3] (for instance, 'death'). Further, the specific medical issue suggested a time scale using 'onset-of-therapy' as a natural time origin; however, data on treated patients only became available after their appearance within the registry. This phenomenon is known as delayed entry or left-truncated data; failure to account for it causes length-bias. Individuals without any event until study closure and drop-outs (e.g., due to missing treatment records) are treated as right-censored observations. Risks are expressed in terms of cause-specific hazard ratios provided by univariate and multivariate Cox proportional hazards models. Missing covariate information at study entry is attacked by a multiple imputation approach using chained equations [4].

Results: Results and medical impacts regarding the DIVE database are presented. Besides (direct) effects on the hazard ratio of the event of interest, a Cox model regarding the competing endpoint additionally enables the detection of 'indirect' covariate impacts. We also compare the results to the 'ordinary' approaches like Welch's t-test or Fisher's exact test.

Discussion: Flexible time-to-event methodology using competing risks is rarely applied in register-based diabetes epidemiology. In particular, the time-dynamic documentation within the DIVE registry avoids bias when patient outcome and covariate effects after a time-dependent event is investigated. Accounting for left-truncation, the consideration of alternative time scales and the analysis of length-biased sampling is enabled. By means of multistate models, the methodology can even be extended to more complex disease histories. From a medical point of view, the identification of prognostic factors regarding BOT- initiation/-failure could be highly useful in a clinical setting when assessing the most appropriate treatment strategy for type 2 diabetes patients.

As limitations, the current setting only allows an investigation of covariates measured at (delayed) study entry and (personalized) predictions are not possible due to incomplete mortality information [3], [5].