Lagging exposure information in cumulative exposure-response analyses

Our plain-English summary of Lagging exposure information in cumulative exposure-response analyses, by D.B. Richardson, S.R. Cole, H. Chu, et al (American Journal of Epidemiology, 2011).

The study

• Authors: D.B. Richardson, S.R. Cole, H. Chu, B. Langholz
• Journal: American Journal of Epidemiology
• Published: 2011;174(12):1416-22
• Cited by: 40 articles, as counted by NIH iCite
• Read it: PubMed · Publisher · Free full text

Abstract

Lagging exposure information is often undertaken to allow for a latency period in cumulative exposure-disease analyses. The authors first consider bias and confidence interval coverage when using the standard approaches of fitting models under several lag assumptions and selecting the lag that maximizes either the effect estimate or model goodness of fit. Next, they consider bias that occurs when the assumption that the latency period is a fixed constant does not hold. Expressions were derived for bias due to misspecification of lag assumptions, and simulations were conducted. Finally, the authors describe a method for joint estimation of parameters describing an exposure-response association and the latency distribution. Analyses of associations between cumulative asbestos exposure and lung cancer mortality among textile workers illustrate this approach. Selecting the lag that maximizes the effect estimate may lead to bias away from the null; selecting the lag that maximizes model goodness of fit may lead to confidence intervals that are too narrow. These problems tend to increase as the within-person exposure variation diminishes. Lagging exposure assignment by a constant will lead to bias toward the null if the distribution of latency periods is not a fixed constant. Direct estimation of latency periods can minimize bias and improve confidence interval coverage.

In plain English

This was a statistical methods paper about 'lagging' exposure: leaving out the most recent years of exposure to allow for the delay between exposure and disease when relating cumulative exposure to disease. Using mathematical derivations and computer simulations based on data from a study of uranium miners exposed to radon, the authors compared three approaches: choosing the lag that gives the largest effect estimate, choosing the lag that best fits the data, and estimating the latency period together with the effect. In the simulations, the largest-estimate approach gave estimates above the true value, and the best-fit approach gave confidence intervals that were too narrow. Joint estimation showed little bias. As an illustration, the authors analyzed lung cancer deaths among 3,072 South Carolina asbestos textile workers (198 lung cancer cases, 7,505 controls). The excess rate ratio per 100 fiber-years/mL rose from 1.07 with a 1-year lag to 1.32 with a 20-year lag. A model allowing latency to vary between people gave 1.15 (95% CI: 0.43, 2.25). The authors concluded that this approach performed reasonably well.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Statistical methods paper using analytical derivations, computer simulations (based on Colorado Plateau uranium miners data on radon), and an illustrative nested case-control analysis of an asbestos textile worker cohort
• Who was studied: Illustrative example: 3,072 South Carolina asbestos textile workers employed in production for at least 1 month between January 1, 1940, and December 31, 1965, with vital status followed through December 31, 2001; the analysis included 198 lung cancer cases and 7,505 controls. Simulations used 10,000 hypothetical studies of radon exposure and lung cancer.
• Exposure: Cumulative asbestos exposure in fiber-years per milliliter, calculated from time in each job and the estimated asbestos exposure rate for that job, analyzed under lags of 1 to 20 years. The article does not name the fiber type. The simulations concerned radon exposure.
• Main results:
  • Asbestos textile workers: the 20-year lag gave the largest estimate, an excess rate ratio per 100 fiber-years/mL of 1.32 (95% CI: 0.45, 2.62); the estimate rose from 1.07 under a 1-year lag.
  • Asbestos textile workers: the best-fitting lag was 9 years, with an excess rate ratio of 1.15. Its 95% CI was 0.43, 2.22 if the lag were treated as chosen in advance, and 0.42, 2.33 when the lag search was taken into account.
  • Asbestos textile workers: a model assuming latency periods follow a lognormal distribution gave an excess rate ratio of 1.15 (95% CI: 0.43, 2.25), with an estimated most common latency of 8.47 years (95% CI: 0.66, 24.53) and a coefficient of variation of 0.148.
  • Simulations with a fixed 5-year latency and a true value of 1: picking the lag with the largest estimate gave average estimates of 1.33, 1.31 or 1.24 across the three exposure-variability scenarios.
  • Simulations with a fixed 5-year latency: picking the best-fitting lag gave average estimates of 1.04, 1.03 and 1.02, but 95% confidence intervals covered the true value only 62%, 68% and 85% of the time. Jointly estimating latency and effect gave 1.01, 0.99 and 1.01, with coverage of 88%, 88% and 92%.
  • Simulations with latency varying between people (lognormal distribution): with the best-fitting-lag approach, bias increased as variation in latency increased (average estimates 0.99, 0.97, 0.94), and coverage fell to 81%, 78% and 65%.
• The authors' conclusion: The authors concluded that choosing the lag with the largest estimate can bias the effect estimate, and that choosing the best-fitting lag, while unbiased when latency is the same for everyone, gives confidence intervals that are too narrow. They proposed a likelihood approach allowing latency to vary between persons and reported that it performed reasonably well.
• Limitations the authors state:
  • The proposed method assumes latency periods follow a lognormal distribution; the authors state that if this assumption is wrong, bias may occur.

Topics

Medical terms in this study, explained in our Lexicon: Cumulative Exposure, Latency, confidence interval, Asbestos Exposure, Lung Cancer.

Related studies

• Potential pitfall in using cumulative exposure in exposure-response relationships: demonstration and discussion — M.M. Finkelstein (1995)
• Hierarchical latency models for dose-time-response associations — D.B. Richardson, R.F. MacLehose, et al. (2011)
• The potential impact of epidemiology on the prevention of occupational disease — D.H. Wegman (1992)
• Design and conduct of occupational epidemiology studies: IV. The analysis of case-control data — N. Pearce, H. Checkoway, et al. (1989)
• Estimating the effect of cumulative occupational asbestos exposure on time to lung cancer mortality: using structural nested failure-time models to account for healthy-worker survivor bias — A.I. Naimi, S.R. Cole, et al. (2014)
• Joint modeling of multiple longitudinal patient-reported outcomes and survival — L.A. Hatfield, M.E. Boye, et al. (2011)

Browse more research in American Journal of Epidemiology, or return to medical research.