Search tips
Search criteria 


Logo of oenvmedOccupational and Environmental MedicineVisit this articleSubmit a manuscriptReceive email alertsContact usBMJ
Occup Environ Med. 2007 December; 64(12): e4.
PMCID: PMC2095361

Epidemiological methods

003 Biologically‐based models in exposure assessment and epidemiology

D. K. Kriebel1, T. J. Smith2. 1Department of Work Environment, University of Massachusetts Lowell; 2Department of Environmental Health, Harvard School of Public Health

ObjectivesWe hypothesise that better quantification of environmental risks is possible through better linkage of epidemiology and exposure assessment. This synthesis can be based on explicit biologic models of the dynamic processes leading from exposures through pathogenic processes to disease.

MethodsWe propose that two elements of any disease process are critical when designing exposure‐response studies: (1) whether the process appears reversible or irreversible; and (2) whether the response is proportional to the tissue concentration or alternatively occurs in a discrete step – from absent to present. Combining these two characteristics, there are four distinct models which can be used as the starting point for studying most environmental diseases. These model types define profoundly different time courses and exposure–response relationships, but their basic features can often be identified using readily observable information about the adverse effects. While a very large number of pathologic processes occur, their temporal behaviour is typically limited by one or a few slow “rate limiting” steps. Thus often the observed behaviours of many different processes follow similar patterns. Starting from a description of the temporal features of the outcome and a hypothesis about the environmental agent, one can propose an explicit disease process model. The approach is illustrated for all four disease processes: type I – reversible proportional; type II – irreversible proportional; type III – reversible discrete; and type IV – irreversible discrete.

ResultsThe use of explicit temporal models resolves many ambiguities of exposure assessment – choosing what agent to study, how to measure it, what time scale for averaging, and what final dose metric to calculate. The model is an explicit hypothesis about the relationship between exposure and outcome. If the model does not fit the data, it can be rejected or revised.

ConclusionExplicit but simple biological process models can be used to integrate exposure assessment with epidemiology. Basic choices in the design, measurements, and analysis of an epidemiologic study can be guided by explicit hypotheses, about the time course of the adverse processes leading to the outcome.

Key wordsdose modelling; exposure assessment; disease processes

004 Understanding attitudes towards health research, or who can phone you at dinnertime?

K. Teschke, A. Harris, J. K. C. Tsui, S. Marino, S. A. Marion. University of British Columbia

ObjectivesPrivacy legislation in Canada and elsewhere has increasingly limited options for recruiting subjects into health studies. Policy changes are motivated by assumptions about public attitudes towards participation, yet surveys of attitudes have rarely been done. This study's purpose was to investigate people's willingness to participate in health research and to examine how their willingness was affected by the method of selection and the mode of contact.

MethodsThe survey was conducted in two groups on the west coast of Canada: a random sample of 3000 subjects selected from the telephone directory; and samples of 833 potential cases and 329 controls from an ongoing study of Parkinson's disease, selected from government databases. Mailed self‐administered questionnaires included 24 questions related to the method of selection of study subjects, the organisation making the contact, and factors that would affect willingness to participate. Data analyses included summary statistics for all questions, as well as cross‐tabulations according to the demographics of the participants, their health status, and their mode of selection for the survey.

ResultsThe participation rate was 48%. Analyses of preliminary data indicate that 90% of potential Parkinson's disease cases and 86% of those selected without reference to disease status were willing to participate in health research. The method of selection impacted willingness to participate with more people unwilling if they were selected from the telephone directory (25%) rather than via an administrative database without regard to health status (8%). The organisation making the contact was more influential: only 7% of people were unwilling to participate if contacted by a university, but more were unwilling when contacted by hospitals (12%), the government (18%), or private research firms (44%). Factors most positively influencing choice to participate included future health benefits to society (92%) and to oneself (88%).

ConclusionParticipation in health research appears to be viewed favourably by members of the public, and participation may be highest when university‐based researchers are able to contact subjects directly using information from administrative databases. A limitation of this study is the low participation rate; analyses will examine whether responses differed according to the number of contacts needed to elicit a response.

Key wordsprivacy legislation; epidemiological methods; survey

005 Bias analysis in regression models with individual‐ and group‐based exposure assessment

H. M. Kim1, I. Burstyn1, D. Loomis2, D. Richardson3. 1University of Alberta; 2University of Nevada; 3University of North Carolina

ObjectivesIn occupational epidemiology, exposure assessment is often conducted via group‐based strategies in which each subject is assigned the mean exposure of their group. In an individual‐based strategy, in contrast, exposure is assigned based on each subject's own measurements. We aimed to develop models of bias in exposure–disease analyses for fixed and random group‐based assessment and compare them to individual‐based assessment in linear, logistic and Cox proportional hazard models when disease is rare.

MethodsNormal distribution of exposure within each of the five groups was assumed; between‐group group distribution was also assumed to be normal for random group‐based assessment. Mathematical expressions for bias were derived where possible and all considered scenarios were illustrated in simulation studies of a large occupational cohort; true association parameter was fixed to 0.4; measurements of 100 subjects per group were used to estimate group means. We further illustrate our findings in re‐analyses of associations between exposure to magnetic fields and cancer mortality among electric utility workers.

ResultsThe individual‐based assessment is based on a classical measurement error model while the group‐based assessment with either random or fixed grouping leads to a Berkson error structure, assuming moderately large samples from each group. The degree of attenuation in risk estimates as a result of measurement error depends on between‐subject standard deviation (SD) if the grouping is fixed, and both between‐group and between‐subject SD if the grouping is random. For example, in fixed grouping, as between‐subject SD increased from 1 to 2, attenuation increased from 0.01 to 0.03. In random grouping, as between‐group SD increased from 0.5 to 2, attenuation decreased from 0.03 to <0.01 when between‐subject SD was moderate (1); with larger between‐subject SD (2), as between‐group SD increased from 0.5 to 2, attenuation decreased from 0.06 to 0.02. In individual‐based analysis, larger within‐subject variability further increased attenuation compared to group‐based strategies, but this was counter‐acted when between‐subject variability was large.

ConclusionIf the between‐group variability is large, we recommended applying group‐based assessment; otherwise, an individual‐based exposure assessment is preferred, if feasible.

Key wordsmeasurement error; attenuation; exposure assessment

006 Bayesian bias adjustments of lung cancer SMRs in a cohort study of German carbon black production workers

P. Morfeld1, S. F. Buechte1, R. M. McCunney2, C. Piekarski1. 1Institute for Occupational Medicine, Cologne University; 2Massachusetts Institute of Technology, Boston

ObjectivesUncontrolled confounding is an important problem in epidemiology. Information about confounders is often only indirectly available and of considerable uncertainty. In this study a significantly elevated lung cancer SMR of about 2 was observed without an obvious link to carbon black exposures at a plant. Some information about smoking habits and exposures received before being hired at the plant were available. We tried to use this information to adjust the SMR.

MethodsA cohort study comprising 1528 carbon black production workers, 1976–1998, showed a lung cancer SMR of 1.83 (50 cases, 95% CI 1.34 to 2.39). A nested case‐control study pointed at effects of smoking (OR 9) and prior exposures to silica (OR 2 or OR 5, adjusted for smoking; OR depended on the exposure assessment procedure applied). These point estimates and their CIs were used together with prevalence data on smoking and exposure (cohort and reference) to derive bias correction factors. A Bayesian approach was chosen applying non‐informative priors for the SMR (Steenland K, Greenland S. Am J Epidemiol 2004;160:384–92). Point estimates and credibility intervals for the posterior SMR were calculated by Markov Chain Monte Carlo (MCMC) methods programmed in R (Metropolis algorithm).

ResultsAfter a burn‐in phase (50 000 cycles), a Markov chain of length 1 000 000 returned a median posterior SMR of 1.30 (95% CI 0.67 to 2.0) or 0.97 (95% CI 0.2 to 3.3) depending on the method how prior exposures were assessed.

ConclusionBayesian bias adjustment is an excellent tool to combine data about confounders from different sources quantitatively. The method can take account of correlations between bias estimates and their different precisions. Even subjective guesses of the probable degree of distortion can be included. Obviously, the usually calculated lung cancer SMR statistic overestimated effect and precision when compared with the results from the MCMC approach in this study. Bayesian bias adjustment should become a regular tool in occupational epidemiology to overcome mere narrative discussions of potential distortions as usually presented in the last section of scientific publications.

Key wordsBayesian bias analyses; lung cancer; carbon black

007 Impact of including workers hired prior to start of follow‐up: a simulation study of the healthy worker survivor effect

K. M. Applebaum1, E. J. Malloy1, E. Eisen2. 1Harvard School of Public Health; 2University of California Berkeley

ObjectivesIncluding workers hired prior to the start of follow‐up may exaggerate the healthy worker survivor effect (HWSE) because long‐term workers tend to be healthier. To examine the impact of including workers hired prior to start of follow‐up, we used Monte Carlo simulations to examine the magnitude of the bias under different exposure–response relationships.

MethodsWe generated 1000 replicates of 3000 subjects hired over a 100‐year period (1900–2000) and followed from 1950 to 2050. We defined two subcohorts: prevalent hires were already employed by start of follow‐up and incident hires were hired afterwards. Subjects who left work prior to start of follow‐up are missing. Disease was modelled as a function of age and duration of work. HWSE was induced by forcing cases to truncate exposure 15 years prior to date of diagnosis, on average. Hazard ratios (HR) were estimated in Cox models as linear or smoothed functions of work duration using penalised splines. We compared the change in HR per year of exposure (slope) and the exposure–response curves between prevalent and incident hires, with and without HWSE, under both null and positive associations.

ResultsUnder the null, the dose–response parameters for incident and prevalent hires were similar with and without HWSE. When HWSE = 0, both slopes were 1.00 and the penalised splines curves were flat. When HWSE = 15, both slopes were biased downward, 0.96 and 0.92 for incident and prevalent hires, respectively, and both curves turned downward. For a positive association with no HWSE the slope was 2.77 and 2.80 for incident and prevalent hires, although the prevalent hires curve was below the incident hires curve. In the presence of HWSE, the slopes dropped to 2.39 and 2.25, and curves flattened at higher exposures, with predicted HR for prevalent hires consistently lower than for incident hires, at a given exposure.

ConclusionUnder the null, exposure–response curves are similar between incident and prevalent hires and HWSE produces equivalent downward bias in both. When the exposure–response association is positive, HWSE biases HR downward within each hire group. Including prevalent hires in occupational cohorts, in the presence of HWSE, will magnify downward bias.

Key wordshealthy worker effect; inception cohort; simulation study

Articles from Occupational and Environmental Medicine are provided here courtesy of BMJ Publishing Group