To assure the validity of epidemiologic studies confounding needs to be controlled. Whereas selection bias occurs primarily in the design stage of a study and measurement bias in the data collection stage, confounding takes place mainly in the analysis phase. Confounding can be controlled, however, in either the design or analysis stages of a study. In the design stage, restriction, matching, or random allocation may be used to reduce the potential for confounding. In the analysis stage, various statistical procedures can be applied to adjust or control for potential confounding after the fact, as mentioned previously. These methods include stratification and multivariable analysis and are discussed in subsequent chapters dealing with the specific epidemiologic study designs.
Restriction involves limiting the study subjects to those with certain characteristics. For example, if the sample is limited to women only, then differences in sex can not confound an association. Similarly, if a study of the effect or cigarette smoking on motor vehicle injuries is limited to nondrinkers, then alcohol consumption can not possibly confound an association between these variables. Other factors, such as risk taking, however, may still be confounders. Obviously, restriction as a method of controlling confounding has its limits. If many potential confounders exit, the study would have to be so restricted that it might be difficult to find a sufficient number of participants. Also, restriction can limit the external validity of a study. A study restricted to women, for example, may not be generalizable to populations including men.
Matching attempts to produce study and comparison groups that rare similar with regard to potential confounders. Pair matching, for example, uses study and comparison subjects who have the same characteristics on selected variables. This is illustrated in a case-control study that pair matches for age, sex, and smoking status. For each case, a corresponding control is selected who is also between 20 and 24 years old, a male, and a nonsmoker, then a control is selected who is also between 20 and 24 years old, a male, and a nonsmoker, and so on. The individually matched cases and controls are treated as pairs in the analysis phase.
Like restriction, matching has its limits. Matching on more than four five characteristics can become very tedious because many potential subjects may have to be considered in finding ones who meet the criteria. It would be very difficult, for instance, to locate many 55-59-year-old males who have type AB negative blood and live at home taking care of triplets. Another disadvantage of matching is a loss of information. For example, if one subject in a matched pair does not respond to a study questionnaire, the other paired subject has to be excluded from the analysis. Also, the variables for which matching wan performed can not be examined for their relationship to the study outcome. Thus, if one matches for age, its role can not be examined in the study.
Another form of matching, known as frequency matching, relies on obtaining similar frequencies of the matched variables in the study and comparison groups. For example, if 25% of the subjects in the exposed group in a prospective cohort study are females, and if sex is considered a potential confounder, then the investigator seeks to obtain 25% females in the unexposed group. The goal of frequency matching is to assure that the study and comparison groups are similar with respect to the distribution of potential confounders. Frequency matching, however, does not assure that the study and comparison groups are similar with respect to the distribution of potential confounders. Frequency matching, however, does not assure as precise a comparison as pair matching, and some subgroups created during the analysis phase may differ substantially in the frequencies of potential confounders; therefore, it is still important to adjust for potential confounding in studies using frequency matching.
In randomized controlled trials, and some community trials, confounding is controlled in the design stage by random allocation. Random allocation refers to the assignment of study subjects to either experimental or control groups using random methods. This technique tends to eliminate confounding by increasing the probability that the groups being compared are similar with regard to the distribution of incidental factors that might confound an association.