Choice of exposure scores for categorical regression in meta-analysis: a case study of a common problem
Objective: Reporting categorical relative risk estimates for a series of exposure levels versus a common reference category is a widespread practice. In meta-analysis, categorical regression estimates a dose–response trend from such results. This method requires the assignment of a single score to each exposure category. We examined how closely meta-analytical categorical regression approximates the results of analysis based on the individual-level continuous exposure.
Methods: The analysis included five studies on tea intake and outcomes related to colorectal cancer. In addition, we derived categorical mean and median values from published distributions of tea consumption in similar populations to assign scores to the categories of tea intake when possible. We examined whether these derived mean and median values well approximate the individual-level results.
Results: In meta-analytical categorical regression, using the midrange scores approximated the individual-level continuous analyses reasonably well, if the value assigned to the uppermost, open-ended category was at least as high as the lower bound plus the width of the second-highest category. Categorical mean values derived from the published distributions of regular tea (in the US) and green tea (in Japan) well approximated the slope obtained from individual-level analysis.
Conclusion: Publication of both the categorical and the continuous estimates of effect in primary studies, with their standard errors, can enhance the quality of meta-analysis, as well as providing intrinsically valuable information on dose–response.
Keywordscolorectal cancer dose–response relationships epidemiology epidemiologic method meta-analyses tea.
confidence limit ratios
Greenland, S 1998Basic methods for sensitivity analysis and external adjustmentRothman, KJGreenland, S eds. Modern EpidemiologyLippincott-Raven PublishersPhiladelphia343358Google Scholar
Greenland, S, Longnecker, MP 1992Methods for trend estimation from summarized dose–response data, with applications to meta-analysisAm J Epidemiol13513011309PubMedGoogle Scholar
Berlin, JA, Longnecker, MP, Greenland, S 1993Meta-analysis of epidemiologic dose–response dataEpidemiology4218228PubMedGoogle Scholar
Brown, KG 1999Lung cancer and environmental tobacco smoke: occupational risk to nonsmokersEnviron Health Perspect107885890PubMedGoogle Scholar
Villanueva, CM, Fernandez, F, Malats, N, Grimalt, JO, Kogevinas, M 2003Meta-analysis of studies on individual consumption of chlorinated drinking water and bladder cancerJ Epidemiol Community Health57166173PubMedGoogle Scholar
Ford, ES, Smith, SJ, Stroup, DF, Steinberg, KK, Mueller, PW, Thacker, SB 2002Homocyst(e)ine and cardiovascular disease: a systematic review of the evidence with special emphasis on case–control studies and nested case-control studiesInt J Epidemiol315970PubMedGoogle Scholar
Reynolds, K, Lewis, B, Nolen, JD, Kinney, GL, Sathya, B, He, J 2003Alcohol consumption and risk of stroke: a meta-analysisJAMA289579588PubMedGoogle Scholar
Norat, T, Lukanova, A, Ferrari, P, Riboli, E 2002Meat consumption and colorectal cancer risk: dose–response meta-analysis of epidemiological studiesInt J Cancer98241256PubMedGoogle Scholar
Spitzer, WO, Faith, JM, MacRae, KD 2002Myocardial infarction and third generation oral ontraceptives: aggregation of recent studiesHum Reprod1723072314PubMedGoogle Scholar
Bergstrom, A, Hsieh, CC, Lindblad, P, Lu, CM, Cook, NR, Wolk, A 2001Obesity and renal cell cancer – a quantitative reviewBr J Cancer85984990PubMedGoogle Scholar
Shi, JQ, Copas, JB 2004Meta-analysis for trend estimationStat Med23319PubMedGoogle Scholar
Mantel, N 1963Chi-Square Tests with One Degree of Freedom: Extensions of the Mantel-Haenszel ProcedureJ Am Stat Assoc53690700Google Scholar
Rothman, KJ 1986Modern epidemiologyLittle, BrownBostonGoogle Scholar
Peters, U, Poole, C, Arab, L 2001Does tea affect cardiovascular disease? A meta-analysisAm J Epidemiol154495503PubMedGoogle Scholar
Baron, JA, Gerhardsson, D, Ekbom, A 1994Coffee, tea, tobacco, and cancer of the large bowelCancer Epidemiol Biomarkers Prev3565570PubMedGoogle Scholar
Baron, JA, Greenberg, ER, Haile, R, Mandel, J, Sandler, RS, Mott, L 1997Coffee and tea and the risk of recurrent colorectal adenomasCancer Epidemiol Biomarkers Prev6710PubMedGoogle Scholar
Il’yasova, D, Arab, L, Martinchik, A, Weisgerber, UM, Sdvizhkov, A, Urbanovich, L 2003Black tea consumption and risk of rectal cancer in Moscow populationAnn Epidemiol13405411CrossRefPubMedGoogle Scholar
Kono, S, Shinchi, K, Ikeda, N, Yanai, F, Imanishi, K 1991Physical activity, dietary habits and adenomatous polyps of the sigmoid colon: a study of self-defense officials in JapanJ Clin Epidemiol4412551261PubMedGoogle Scholar
Su, LJ, Arab, L 2002Tea consumption and the reduced risk of colon cancer – results from a national prospective cohort studyPublic Health Nutr5419425PubMedGoogle Scholar
Poole, C 2001Low P-values or narrow confidence intervals: which are more durable?Epidemiology12291294PubMedGoogle Scholar
US Department of Health and Human Services (DHHS). Second National Health and Nutrition Examination Survey I Epidemiologic Followup Study, 1982–1984. NCHS. 2001. NCHS.Google Scholar
Inoue, M, Tajima, K, Hirose, K, et al. 1997Epidemiological features of first-visit outpatients in Japan: comparison with general population and variation by sex, age, and seasonJ Clin Epidemiol506977PubMedGoogle Scholar
U.S. Environmental Protection Agency1986Guidelines for Carcinogen Risk AssessmentFed. Reg.513399234003Google Scholar
Hertz-Picciotto, I 1995Epidemiology and quantitative risk assessment: a bridge from science to policyAm J Public Health85484491PubMedGoogle Scholar
Shore, RE 1995Epidemiologic data in risk assessment – imperfect but valuableAm J Public Health85474476PubMedGoogle Scholar
Rohan, TE, McMichael, AJ 1988Alcohol consumption and risk of breast cancerInt J Cancer41695699PubMedGoogle Scholar
Cook, RJ, Brumback, BB, Wigg, MB, Ryan, LM 2001Synthesis of evidence from epidemiological studies with interval-censored exposure due to groupingBiometrics57671680PubMedMathSciNetGoogle Scholar
Bagnardi, V, Zambon, A, Quatto, P, Corrao, G 2004Flexible meta-regression functions for modeling aggregate dose–response data, with an application to alcohol and mortalityAm J Epidemiol15910771086PubMedGoogle Scholar
Hertz-Picciotto, I, Smith, AH 1993Observations on the dose–response curve for arsenic exposure and lung cancerScand J Work Environ Health19217226PubMedGoogle Scholar
Poole, C, Greenland, S 1999Random-effects meta-analyses are not always conservativeAm J Epidemiol150469475PubMedGoogle Scholar