Commentary: Does the spectre of ecologic bias haunt epidemiology?

Thomas Webster

Department of Environmental Health (T2E), Boston University School of Public Health, 715 Albany St, Boston, MA 02118–2526, USA. E-mail: twebster{at}bu.edu

Beginning students of epidemiology learn that non-differential exposure misclassification biases results (on average) toward the null. Ecologic studies are one of the most interesting exceptions. As shown by Brenner et al. in 1992, ecologic studies are biased away from the null when exposure is non-differentially misclassified with the same sensitivity and specificity in each group.1 In this issue, Björk and Strömberg2 extend this result in an interesting and disconcerting direction: ecologic bias in an important type of occupational study.

In an individual-level study, investigators measure outcome, exposure and covariates for each subject. In contrast, ecologic studies measure these variables at the group level. Mixed designs are also possible. For example, one might measure exposure at the group level but outcome and covariates at the individual level. Such partially ecologic studies are actually rather common. In occupational and environmental epidemiology measurement of exposure for each individual is often infeasible. Instead, exposure is assigned based on group membership, for example job title or pollution as measured at central monitors. These types of partially ecologic studies have been designated semi-individual by some authors.3,4

Are partially ecologic studies subject to ecologic bias? More precisely—since ecologic bias can occur in a number of different ways5—are partially ecologic studies subject to some or all of the biases that can occur when making inferences about individuals based on group-level data? Opinions vary. For example, Loomis et al.6 argued that semi-individual air pollution studies are not subject to ecologic biases: ‘Instead, these studies should be viewed as individual level studies in which exposure is measured with error’. As I have argued elsewhere,4,7 it is certainly possible to think of partially ecologic studies as individual-based studies with exposure measurement errors. But this does not mean they are exempt from ecologic-like biases. I analysed cohort studies, using weighted least squares regression to estimate risk differences. Under the ‘Brenner scenario’—non-differential misclassification of binary exposure within groups, constant sensitivity and specificity across groups—partially ecologic studies are biased away from the null just like completely ecologic studies.

Björk and Strömberg extend this result in important ways. They examine occupational case-control studies with exposure determined via probabilistic job exposure matrices.8 Such studies are partially ecologic since exposure is assigned at the group level. They estimate the odds ratio using a linear odds model and iterative weighted least squares. Despite the mathematical complications introduced by this approach, the authors show that the bias of the odds ratio caused by exposure measurement error can often be nicely approximated by applying a general error model9 to the slope (ß coefficient) component of the model. Under the ‘Brenner scenario’, the odds ratio is once again biased away from the null. Since this type of error structure may not be very realistic in practice, Björk and Strömberg analyse two other scenarios, examining both bias and precision. Using these other error structures, they show that non-differential exposure misclassification does not always bias the results of partially ecologic studies away from the null.

Completely ecologic studies are subject to a number of biases, some potentially quite severe. But most epidemiological research of these biases has been theoretical. Do these biases occur in the real world? How big are they? Side-by-side comparisons of individual and ecologic analyses of the same data provide a good way of examining these questions, but relatively few studies have done so. Björk and Strömberg compare partially ecologic and individual-level analyses of the same data, an important contribution to the literature. Exposure was assigned via a probabilistic job exposure matrix for the partially ecologic design, via phone interviews/occupational hygienists for the individual-level formulation. The partially ecologic study was biased away from the null and less precise. Although the upward bias may be due to ecologic-like bias associated with non-differential exposure misclassification, other types of ecologic bias were not ruled out. For example, adjustment for confounders does not perform correctly in partially ecologic studies unless there is no confounding within groups.2,7,10

Bias away from the null under the ‘Brenner scenario’ of non-differential exposure misclassification has become almost a hallmark of ecologic bias. Its appearance in partially ecologic studies warns us to look for other ecologic-like biases, for example confounding and effect modification by group. For these effects to occur, exposure on the underlying individual level need not be binary. Instead, these biases are related to the loss of information on the joint distribution of outcome, exposure and covariates that takes place during aggregation.4,7

Some degree of aggregation is common throughout epidemiology. Might ecologic bias be lurking in other places? Probably, but rather than treating ecologic bias as something to be avoided at all cost, perhaps we should try to domesticate it: treat it as just another type of bias that needs to be minimized in design where possible and evaluated during analysis.

References

1 Brenner H, Savitz D, Jöckel K-H, Greenland S. Effect of nondifferential exposure misclassification in ecologic studies. Am J Epidemiol 1992;135:85–95.[Abstract]

2 Björk J, Strömberg U. Effects of systematic exposure assessment errors in partially ecologic case-control studies. Int J Epidemiol 2002;31: 154–60.[Abstract/Free Full Text]

3 Künzli N, Tager IB. The semi-individual study in air pollution epidemiology: a valid design as compared to ecologic studies. Environ Health Perspect 1997;105:1078–83.[ISI][Medline]

4 Webster T. Can semi-individual studies have ecologic bias? Epidemiology 2000;11:S95.

5 Greenland S. Divergent biases in ecologic and individual-level studies. Statist Med 1992;11:1209–23.[ISI]

6 Loomis D, Hertz-Picciotto I, O'Neill M. Comments on ‘PM2.5 mortality in long-term prospective cohort studies: cause-effect or statistical association?’ Environ Health Perspect 1999;107:A392–93.[ISI][Medline]

7 Webster T. Bias in Ecologic and Semi-individual Studies. DSc dissertation, Boston University School of Public Health, 2000.

8 Bouyer J, Hémon D. Comparison of three methods of estimating odds ratios from a job exposure matrix in occupational case-control studies. Am J Epidemiol 1993;137:472–81.[Abstract]

9 Wacholder S. When measurement errors correlate with truth: surprising effects of nondifferential misclassification. Epidemiology 1995;6:157–61.[ISI][Medline]

10 Webster T. Confounder adjustment in semi-individual studies. Epidemiology 2001;12:S82.