PRINCIPLES OF TOXICOLOGY

increased or decreased risk of disease. The odds ratio can be tested for statistical significance using
the χ 2 square test and a confidence interval.
21.9 BIAS
21.9 BIAS 519
Bias is the systematic difference in the study from reality. This can happen at any time during an
epidemiologic study. Bias is any trend in the collection, analysis, interpretation, or review of data that
can lead to conclusions that are systematically different from the truth. In general, bias affects the
comparability of the data.
Epidemiologic studies are inherently biased because they are studying human populations, often
based on data and records collected in the past. Much of bias can be prevented by careful study planning
or avoided by the use of various statistical analysis techniques. There are several types of bias which
are well described and should be evaluated in reviewing epidemiologic studies.
Selection bias is introduced during the beginning of a study when the study population is divided
into different groups (either exposed or diseased). If the diseased group has different criteria for exposure
than the control group, selection bias is present and the two groups are not comparable. Observation bias
is often seen in studies where the interviewer is aware of the disease or exposure state of the study subject
(i.e., not blinded); this can lead to differing amounts of information being obtained from the two groups,
and thus, the two groups are not comparable. Recall bias, seen particularly in case–control studies, is due
to the increased recall of past events (including exposures) by persons who have a disease as compared to
the recall of those who do not have the disease (i.e., the controls). For example, a woman who has had a
child with a birth defect is more likely to remember every medication taken during the pregnancy than a
woman whose child was born without a birth defect.
Another bias commonly seen in occupational epidemiology studies has been dubbed “the healthy
worker effect.” This occurs when rates of disease in working populations are compared with those in
the general population. Often, working populations in this context will show little or no increased risk
of disease. However, the general population is not an appropriate comparison population for working
populations. As was stated above, working populations are in general much healthier and younger than
the general population, which consists of people of all ages and stages of health. Therefore, the
appropriate comparison population for an exposed working population would be another unexposed
working population.
Confounding is a particular form of bias which is very common in epidemiologic studies.
Confounding results when the association between an exposure and a disease is really due to another
exposure. For example, smoking is associated with lung cancer and with heavy alcohol intake. If
exposure to smoking is not controlled for in either the study design or the analysis (see discussion
below), then it could appear that lung cancer is due to exposure to heavy alcohol intake. Smoking is a
common confounding variable, as well as age, sex, and socioeconomic class.
Although many factors are assumed to be confounders (age, sex, socioeconomic class, smoking,
etc.), an important component of the statistical analysis is to assess for the possibility of confounding.
If a particular variable is associated (i.e., has a rate ratio or its equivalent measure greater than or less
than one) separately with both the disease and the exposure of interest, then that variable is considered
to be a confounder.
There are a variety of ways to control for confounding in epidemiologic studies. One method, in
case control studies, is to match the cases and controls on well-known confounders, such as age and
sex; then, these cannot be confounding variables since the proportion of the population with the
confounding variable is now the same in both the cases and controls. Another method used in clinical
trials is the randomization of subjects to the treatment and nontreatment groups so that possible
confounding variables are randomly distributed. Both randomization and matching are methods to
control for confounding, which must be implemented in the design phase of the epidemiologic study.
Usually, confounding is controlled for during the statistical analysis at the end of the epidemiologic
study. One method is to stratify the data by the confounding variable, and then report the risk measure