Hierarchical regression for multiple comparisons in a case-control study of occupational risks for lung cancer

Our plain-English summary of Hierarchical regression for multiple comparisons in a case-control study of occupational risks for lung cancer, by M. Corbin, L. Richiardi, R. Vermeulen, et al (PloS One, 2012).

The study

• Authors: M. Corbin, L. Richiardi, R. Vermeulen, H. Kromhout, F. Merletti, S. Peters, L. Simonato, K. Steenland, N. Pearce, M. Maule
• Journal: PloS One
• Published: 2012;7(6):e38944
• Cited by: 7 articles, as counted by NIH iCite
• Read it: PubMed · Publisher · Free full text

Abstract

Background: Occupational studies often involve multiple comparisons and therefore suffer from false positive findings. Semi-Bayes adjustment methods have sometimes been used to address this issue. Hierarchical regression is a more general approach, including Semi-Bayes adjustment as a special case, that aims at improving the validity of standard maximum-likelihood estimates in the presence of multiple comparisons by incorporating similarities between the exposures of interest in a second-stage model.

Methodology/principal Findings: We re-analysed data from an occupational case-control study of lung cancer, applying hierarchical regression. In the second-stage model, we included the exposure to three known lung carcinogens (asbestos, chromium and silica) for each occupation, under the assumption that occupations entailing similar carcinogenic exposures are associated with similar risks of lung cancer. Hierarchical regression estimates had smaller confidence intervals than maximum-likelihood estimates. The shrinkage toward the null was stronger for extreme, less stable estimates (e.g., "specialised farmers": maximum-likelihood OR: 3.44, 95%CI 0.90-13.17; hierarchical regression OR: 1.53, 95%CI 0.63-3.68). Unlike Semi-Bayes adjustment toward the global mean, hierarchical regression did not shrink all the ORs towards the null (e.g., "Metal smelting, converting and refining furnacemen": maximum-likelihood OR: 1.07, Semi-Bayes OR: 1.06, hierarchical regression OR: 1.26).

Conclusions/significance: Hierarchical regression could be a valuable tool in occupational studies in which disease risk is estimated for a large amount of occupations when we have information available on the key carcinogenic exposures involved in each occupation. With the constant progress in exposure assessment methods in occupational settings and the availability of Job Exposure Matrices, it should become easier to apply this approach.

In plain English

This study re-analysed an Italian case-control study of lung cancer (Turin and eastern Veneto, 1990–1992) in men: 951 cases and 1,237 controls. It tested hierarchical regression, a statistical method for estimating risk across 129 occupations. A job-exposure matrix rated each occupation's asbestos, chromium and silica exposure none, low or high. Adjusted for age, area and smoking, the lung cancer odds ratio versus unexposed men was 1.43 (95% CI 1.18–1.73) for ever low and 1.62 (1.01–2.61) for ever high asbestos exposure. Hierarchical regression narrowed confidence intervals and pulled unstable extreme estimates toward expected values. For ships' engine-room ratings, highly exposed to asbestos (8 cases, 2 controls), the odds ratio was 5.88 (95% CI 0.9–36.71); hierarchical regression gave 1.78 (0.9–3.32) to 2.43 (0.7–7.46), depending on settings, all intervals including 1. Semi-Bayes, a simpler adjustment, shrank every raised estimate toward no effect; hierarchical regression moved some away from it, taking railway engine drivers and firemen from 0.97 to 1.35 or 1.47 (Semi-Bayes: 1.01). The authors concluded the method could be valuable when information on the key carcinogenic exposures in each occupation is available.

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 re-analysis of an existing population-based case-control study, comparing standard (maximum-likelihood) estimates with hierarchical regression and Semi-Bayes adjustment
• Who was studied: Men from a population-based case-control study in Turin and eastern Veneto, Italy, 1990–1992. 951 lung cancer cases and 1,237 controls were included in the analyses (Table 1). Risks were estimated for 129 occupations (3-digit ISCO codes held by at least ten subjects).
• Exposure: Occupations, plus exposure to asbestos, chromium and silica assigned to each occupation by a job-exposure matrix (DOM-JEM) scoring 0 = no exposure, 1 = low exposure, 2 = high exposure
• Main results:
  • Ever being exposed to each of the three carcinogens was associated with increased lung cancer risk, with higher risks at high exposure levels (adjusted for age, study area and smoking)
  • Asbestos, compared with unexposed (429 cases/682 controls): ever low exposure 477 cases/512 controls, OR 1.43 (95% CI 1.18–1.73); ever high exposure 45 cases/43 controls, OR 1.62 (95% CI 1.01–2.61)
  • Hierarchical regression estimates had smaller confidence intervals than standard maximum-likelihood estimates
  • Ships' engine-room ratings, rated highly exposed to asbestos (8 cases/2 controls), had these estimates. Standard: OR 5.88 (95% CI 0.9–36.71). Semi-Bayes: OR 1.54 (0.6–3.73). Hierarchical regression: OR 2.43 (0.7–7.46) at t = 0.76, 2.16 (0.8–5.73) at t = 0.59, 1.93 (0.8–4.29) at t = 0.41 and 1.78 (0.9–3.32) at t = 0.23. All of these confidence intervals include 1.
  • Railway engine drivers and firemen, rated highly exposed only to asbestos: standard OR 0.97, Semi-Bayes OR 1.01, hierarchical regression OR 1.35 and 1.47 under two settings
  • Semi-Bayes adjustment shrank all increased estimates toward the null, while hierarchical regression pulled some increased estimates away from the null
• The authors' conclusion: The authors concluded that hierarchical regression could be a valuable tool in occupational studies that estimate disease risk for many occupations, when information on the key carcinogenic exposures in each occupation is available
• Limitations the authors state:
  • If the job-exposure matrix is weighted too heavily, high risks for occupations classified as unexposed to the three carcinogens (but possibly exposed to others) could always be pushed down
  • Including many second-stage covariates can cause collinearity problems, so the analysis was restricted to three carcinogens
  • The advantage of hierarchical regression over Semi-Bayes depends on the prior exposure information being reasonably valid

Topics

The National Library of Medicine indexes this study under subjects our site covers: Lung cancer.

Medical terms in this study, explained in our Lexicon: Case-Control Study, Lung Cancer, Carcinogenic, confidence interval.

Related studies

• Analysis of the effect of occupational exposure to asbestos based on threshold regression modeling of case-control data — A. Chambaz, D. Choudat, et al. (2014)
• Lung Cancer Risks Associated with Occupational Exposure to Pairs of Five Lung Carcinogens: Results from a Pooled Analysis of Case-Control Studies (SYNERGY) — A. Olsson, L. Bouaoun, et al. (2024)
• A Bayesian approach to strengthen inference for case-control studies with multiple error-prone exposure assessments — J. Zhang, S.R. Cole, et al. (2013)
• Parental occupational exposure and risk of childhood central nervous system tumors: a pooled analysis of case-control studies from Germany, France, and the UK — C. Huoi, A. Olsson, et al. (2014)
• Estimation and inference for case-control studies with multiple non-gold standard exposure assessments: with an occupational health application — H. Chu, S.R. Cole, et al. (2009)
• Comparison of exposure assessment methods for occupational carcinogens in a multi-centre lung cancer case-control study — S. Peters, R. Vermeulen, et al. (2011)

Browse more research in PloS One, or return to medical research.