Natural Sciences
Life Sciences
Scientific Computing
Scientific Computing

Regina Stegherr, Institut für Statistik, Universität Ulm

DKFZ Communication Center, K2, Im Neuenheimer Feld 280

Heidelberger Kolloquium Medizinische Biometrie, Informatik und Epidemiologie

In observational studies left-truncation, i.e. delayed entry after time zero, is a common problem. For example, in pregnancy studies women usually enter the study with a doctoral visit and not at conception (time-origin). Another example are diabetes registers where one relevant timescale is ‘time-since-first-antidiabetic-medication’. Due to left-truncation some challenges in data analyses arise. One challenge are possibly unmeasured baseline covariates for individuals who enter a study after the time-origin. For example, a baseline covariate in diabetes patients is the glycated haemoglobin (HbA1c) level. Since data are collected in calendar time, some individuals enter the study at their time-origin, e.g., start of treatment, but others have a known data of treatment start before data collection started. Covariates for the latter individuals are measured the first time at the delayed study entry and, hence, not at the time-origin (baseline), but, e.g., HbA1c will have changed in the random and patient-specific time interval between start of medication and study entry. As a possible solution to this longstanding open problem, we propose to investigate the impact of possibly unmeasured baseline covariates on the risk of having an event using joint models. This contrasts with the standard use of joint models where the aim is to analyze the effect of the current value of the covariate on the hazard of an event. Our approach shows proper performance in a simulation study and was applied to data from a German diabetes register to evaluate the effect of HbA1c at therapy initiation on the risk of treatment failure. Another challenge in the analyses of left-truncated time-to-event data is possible dependence between left-truncation time and time-to-event. For instance, in studies on adverse drug reactions during pregnancy, study entry times and the hazard of elective terminations may be related. We propose a new semiparametric estimator of the marginal cumulative incidence function that does not require independence. To this end, the dependence between entry times and time-to-event is modeled using Cox proportional hazards models and the marginal estimates are derived via inverse probability weighting arguments. Simulations, as well as a data example, show that the new estimator pro-vides at least for a sensitivity analysis when compared to the standard Aalen-Johansen estimator. Finally, we briefly comment on stabilizing statistical inference that may be compromised by early overly small risk sets because of delayed study entry.

Event data:
Import event data into Outlook Calendar