Assessment of bias associated with person-year grouping methods in individual and pooled epidemiological cohorts from the Million Person Study

Abstract

Radiation epidemiology has progressed from analyses of accidental and wartime exposures to today’s studies of low-dose exposures resulting primarily from occupational settings and medical interventions. In the age of low-dose, chronic radiation exposure studies, it is imperative that cohorts are large enough to detect radiation-related health outcomes with precision. Statistical software capabilities for studies of this type have recently been improved to support cohorts with over 50 million person-years of data. Prior to these advancements, follow-up records in large datasets had been grouped into uniform intervals, reducing the size to accommodate software limitations. For example, if grouping into two-record intervals, an individual’s history consisting of with 30 years of annual follow-up would be represented by 15 records. The resulting data are less granular and potentially misrepresent exposure and outcome measures. Enhanced software capabilities permit survival analysis regressions on unprecedentedly large datasets. Full-size and grouped datasets can now be rapidly processed and results can be compared in an effort to characterize grouping-related bias. Three of the largest Million Person Study cohorts are analyzed. Individual cohort, pooled, and meta-analysis risk estimates are presented for ungrouped and grouped data. Grouping is performed three different ways to identify where biases are most likely introduced. Additionally, grouping bias is assessed for simulated data where the expected risk estimate is known. By utilizing these various methods, generalizations can be made on the expected magnitude of grouping bias. Findings reveal observable bias associated with person-year grouping, dependent on grouping method and cohort characteristics. The impact of these findings is assessed and recommendations are made for ongoing radiation epidemiology studies.