Many epidemiological studies assess the effects of time-dependent exposures, where both the exposure status and its intensity vary over time. One example that attracts public attention concerns pharmacoepidemiological studies of the adverse effects of medications. The analysis of such studies poses challenges for modeling the impact of complex time-dependent drug exposure, especially given the uncertainty about the way effects cumulate over time and about the etiological relevance of doses taken in different time periods. We present a flexible method for modeling cumulative effects of time-varying exposures, weighted by recency, represented by time-dependent covariates in the Cox proportional hazards model. The function that assigns weights to doses taken in the past is estimated using cubic regression splines. We validated the method in simulations and applied it to re-assess the association between exposure to a psychotropic drug and fall-related injuries in the elderly.

Pharmacoepidemiology investigates associations between time-varying medication use/dose and risk of adverse events. Applied research typically relies on a priori chosen simple conventional models, such as current dose or any use in the past 3 months. However, different models imply different risk predictions, and only one model can be etiologically correct in any specific applications. We first formally defined several candidate models mapping the time vector of past drug doses (X (t), t  =  1, … ,u) into the value of a time-varying exposure metric M(u) at current time u. In addition to conventional one-parameter models, we considered two-parameter models accounting for recent dose increase or withdrawal and a flexible spline-based weighted cumulative exposure (WCE) model that defines M(u) as the weighted sum of past doses. In simulations, we generated event times assuming one of the models was correct and then analyzed the data with all candidate models. We demonstrated that the minimum AIC criterion is able to identify the correct model as the best-fitting model or one of the equivalent (within 4 AIC points of the minimum) models in a vast majority of simulated samples, especially with 500 or more events. We also showed how relying on an incorrect a priori chosen model may largely reduce the power to test for an association. Finally, we demonstrated how the flexible WCE estimates may help with model diagnostics even if the correct model is not WCE. We illustrated the practical advantages of AIC-based a posteriori model selection and WCE modeling in a real-life pharmacoepidemiology example.

OBJECTIVEGlycemia is a major risk factor for the development of long-term complications in type 1 diabetes; however, no specific genetic loci have been identified for glycemic control in individuals with type 1 diabetes. To identify such loci in type 1 diabetes, we analyzed longitudinal repeated measures of A1C from the Diabetes Control and Complications Trial.RESEARCH DESIGN AND METHODSWe performed a genome-wide association study using the mean of quarterly A1C values measured over 6.5 years, separately in the conventional (n = 667) and intensive (n = 637) treatment groups of the DCCT. At loci of interest, linear mixed models were used to take advantage of all the repeated measures. We then assessed the association of these loci with capillary glucose and repeated measures of multiple complications of diabetes.RESULTSWe identified a major locus for A1C levels in the conventional treatment group near SORCS1 (10q25.1, P = 7 × 10−10), which was also associated with mean glucose (P = 2 × 10−5). This was confirmed using A1C in the intensive treatment group (P = 0.01). Other loci achieved evidence close to genome-wide significance: 14q32.13 (GSC) and 9p22 (BNC2) in the combined treatment groups and 15q21.3 (WDR72) in the intensive group. Further, these loci gave evidence for association with diabetic complications, specifically SORCS1 with hypoglycemia and BNC2 with renal and retinal complications. We replicated the SORCS1 association in Genetics of Diabetes in Kidneys (GoKinD) study control subjects (P = 0.01) and the BNC2 association with A1C in nondiabetic individuals.CONCLUSIONSA major locus for A1C and glucose in individuals with diabetes is near SORCS1. This may influence the design and analysis of genetic studies attempting to identify risk factors for long-term diabetic complications.

In an article published in Annals of Internal Medicine in 2001, Redelmeier and Singh reported that Academy Award-winning actors and actresses lived almost 4 years longer than their less successful peers. However, the statistical method used to derive this statistically significant difference gave winners an unfair advantage because it credited an Oscar winner's years of life before winning toward survival subsequent to winning. When the authors of the current article reanalyzed the data using methods that avoided this "immortal time" bias, the survival advantage was closer to 1 year and was not statistically significant. The type of bias in Redelmeier and Singh's study is not limited to longevity comparisons of persons who reach different ranks within their profession; it can, and often does, occur in nonexperimental studies of life- or time-extending benefits of medical interventions. The current authors suggest ways in which researchers and readers may avoid and recognize this bias.