Confounding

Confounds in Correlational Designs

Confounding variables are at the heart of the third-variable problem in correlational studies. In a correlational study, researchers examine the relationship between two variables. Even if two variables are correlated, it is possible that a third, confounding variable is responsible for the apparent relationship between the two variables. For example, if there were a correlation between icecream consumption and homicide rates, it would be a mistake to assume that eating ice cream causes homicidal rages or that murderers seek frozen treats after killing. Instead, a third [Page 221]variable—heat—is likely responsible for both increases in ice cream consumption and homicides (given that heat has been shown to increase aggression). Although one can attempt to identify and statistically control for confounding variables in correlational studies, it is always possible that an unidentified confound is producing the correlation.

Confounds in Quasi-Experimental and Experimental Designs

The goal of quasi-experimental and experimental studies is to examine the effect of some treatment on an outcome variable. When the treatment systematically varies with some other variable, the variables are confounded, meaning that the treatment effect is comingled with the effects of other variables. Common sources of confounding include history, maturation, instrumentation, and participant selection. History confounds may arise in quasi-experimental designs when an event that affects the outcome variable happens between pretreatment measurement of the outcome variable and its posttreat-ment measurement. The events that occur between pre- and posttest measurement, rather than the treatment, may be responsible for changes in the dependent variable. Maturation confounds are a concern if participants could have developed—cognitively, physically, emotionally—between pre- and posttest measurement of the outcome variable. Instrumentation confounds occur when different instruments are used to measure the dependent variable at pre-and posttest or when the instrument used to collect the observation deteriorates (e.g., a spring loosens or wears out on a key used for responding in a timed task). Selection confounds may be present if the participants are not randomly assigned to treatments (e.g., use of intact groups, participants self-select into treatment groups). In each case, the confound provides an alternative explanation—an event, participant development, instrumentation changes, preexisting differences between groups—for any treatment effects on the outcome variable.

Even though the point of conducting an experiment is to control the effects of potentially confounding variables through the manipulation of an independent variable and random assignment of participants to experimental conditions, it is possible for experiments to contain confounds. An experiment may contain a confound because the experimenter intentionally or unintentionally manipulated two constructs in a way that caused their systematic variation. The Illinois Pilot Program on Sequential, Double-Blind Procedures provides an example of an experiment that suffers from a confound. In this study commissioned by the Illinois legislature, eyewitness identification procedures conducted in several Illinois police departments were randomly assigned to one of two conditions. For the sequential, double-blind condition, administrators who were blind to the suspect's identity showed members of a lineup to an eyewitness sequentially (i.e., one lineup member at a time). For the single-blind, simultaneous condition, administrators knew which lineup member was the suspect and presented the witness with all the lineup members at the same time. Researchers then examined whether witnesses identified the suspect or a known-innocent lineup member at different rates depending on the procedure used. Because the mode of lineup presentation (simultaneous vs. sequential) and the administrator's knowledge of the suspect's identity were confounded, it is impossible to determine whether the increase in suspect identifications found for the single-blind, simultaneous presentations is due to administrator knowledge, the mode of presentation, or some interaction of the two variables. Thus, manipulation of an independent variable protects against confounding only when the manipulation cleanly varies a single construct.

...