Fitting parametric random effects models in very large data sets with application to VHA national data

Abstract

BackgroundWith the current focus on personalized medicine, patient/subject level inference is often of key interest in translational research. As a result, random effects models (REM) are becoming popular for patient level inference. However, for very large data sets that are characterized by large sample size, it can be difficult to fit REM using commonly available statistical software such as SAS since they require inordinate amounts of computer time and memory allocations beyond what are available preventing model convergence. For example, in a retrospective cohort study of over 800,000 Veterans with type 2 diabetes with longitudinal data over 5 years, fitting REM via generalized linear mixed modeling using currently available standard procedures in SAS (e.g. PROC GLIMMIX) was very difficult and same problems exist in Stata’s gllamm or R’s lme packages. Thus, this study proposes and assesses the performance of a meta regression approach and makes comparison with methods based on sampling of the full data.DataWe use both simulated and real data from a national cohort of Veterans with type 2 diabetes (n=890,394) which was created by linking multiple patient and administrative files resulting in a cohort with longitudinal data collected over 5 years.Methods and resultsThe outcome of interest was mean annual HbA1c measured over a 5 years period. Using this outcome, we compared parameter estimates from the proposed random effects meta regression (REMR) with estimates based on simple random sampling and VISN (Veterans Integrated Service Networks) based stratified sampling of the full data. Our results indicate that REMR provides parameter estimates that are less likely to be biased with tighter confidence intervals when the VISN level estimates are homogenous.ConclusionWhen the interest is to fit REM in repeated measures data with very large sample size, REMR can be used as a good alternative. It leads to reasonable inference for both Gaussian and non-Gaussian responses if parameter estimates are homogeneous across VISNs.

Big Data and Clinicians: A Review on the State of the Science

“…One of the major uses for clinical big data is in analysis of the prevalence or trends of a disease or phenotype among different populations. An early big data study evaluated a cohort consisting of 890,394 US veterans with diabetes followed from 2002 through 2006 [ 43 ]. Bermejo-Sanchez et al [ 44 ] observed 326 of the birth defect Amelia among 23 million live births, stillbirths, and fetal anomalies from 23 countries and 4 continents, and found the trend of higher prevalence of Amelia among younger mothers.…”

“…Gebregziabher et al [ 43 ] stated that the datasets generated through many translational research projects to answer questions of public health interest are not self-explanatory due to complexity and inadequate description/documentation of the dataset's parameters and associated metadata. The methodologies for interpreting the data can therefore be subject to all sorts of philosophical debate.…”

Harnessing electronic healthcare data for wound care research: Wound registry analytic guidelines for less‐biased analyses

“…In wound care, common examples include the unit of analysis (e.g., multiple wounds per patient) or patients treated at different investigational sites or settings or by different clinicians. 37,38 Any time significant clustering is encountered in a dataset, not adjusting for the clustering effects is likely to lead to overestimation of the effect size of the intervention or parameter under study. Mixed models in which such clusters are treated as random effects are the most frequently methods used for adjustment, but these methods are computationally intensive, and may be impossible in very big datasets due to available computing power or memory.…”

“…Mixed models in which such clusters are treated as random effects are the most frequently methods used for adjustment, but these methods are computationally intensive, and may be impossible in very big datasets due to available computing power or memory. 38 Standard of care (SOC) is likely to vary considerably in large datasets but nevertheless can have a large impact on wound healing. The TIME algorithm, developed from a meeting of wound care experts, outlines an ideal SOC that includes tissue management, infection, moisture imbalance, and edge of the wound.…”

Fast Moment-Based Estimation for Hierarchical Models

“…In general situations, one can partition the data between multiple processors, compute separate parameter estimates for each chunk and then combine the results (Huang and Gelman, ; Gebregziabher et al ., ; Khanna et al ., ; Scott et al ., ). These splitting strategies often require the same total computational cost, but they split the costs between K processors, reducing wall clock time by a factor of K .…”