Independent and Dependent Variables

In much of our research, particularly when we're interested in theory-building, we want to discover causal relationships between variables. This is intrinsic to the task of explaining why things happen the way they do.

In the broadest sense, the independent variable in a study is the proposed cause, while the dependent variable is the effect. One would hope that being in therapy would lead to some kind of improvement, i.e. that therapy would be the cause of improvement. A study could be designed in which half the participants are placed in a therapy group, and half in a no-treatment control group. At the end of 10 sessions, the two groups are compared on their scores on a depression inventory. In this case, "group" (therapy vs. no therapy) would be the independent variable, and depression would be the dependent, on the supposition that it is the therapy that accounts for the variability across the groups in depression scores. One can say that the level of depression (the dependent variable) depends on which group they were assigned to (the independent variable).

This also illustrates another, more technical way to distinguish between the independent and dependent variable. In experimental design, the independent variable is manipulated, while the dependent variable is observed. In the study described above, participants are not naturally in or out of therapy--they have to be assigned to a group by the experimenter. Thus the level of the independent variable is manipulated by the researcher according to the research design. Once this manipulation has occurred, however, the level of depression is simply measured or observed.

In part, this is why "correlation is not causation"--in a correlational study (e.g., a survey where all variables are in effect measured at the same time) there is no experimental manipulation, hence no true independent variable, and therefore no real "cause."

We recently had some mysterious problems with the lighting in our dining room and kitchen: every so often, the lights would flicker, sometimes rather dramatically. We called in an electrician to diagnose the problem. He examined the wiring and found that the construction crew that remodeled our kitchen years ago had wired the lighting and the switches in a way that was entirely out of compliance with the building code.

But to find the cause of the flickering required more than just this observation. He had to keep experimenting: if I change this (e.g., the wiring of the switches, the breaker at the electrical panel), will the flickering stop? In other words, accurately diagnosing the cause of the problem required some experimental manipulation--"I'll change this, and see what happens." Thus, from the experimentalist's point of view, an independent variable cannot be inferred to be the cause of variability in the dependent variable without such manipulation. (And yes, our lights are working fine now.)

Frankly, outside of a research course, you won't really hear the language of independent and dependent variables much anyway. But the place where it may become important is in determining which statistical procedure to use for an analysis. Textbooks often tell you that you must identify the independent and dependent variables, and which level of measurement is used for each. One-way analysis of variance, for example, requires that the independent variable for the analysis be nominal (a categorical, groups variable) and the dependent variable be continuous (interval level or higher). The research described above fits this description, and thus analysis of variance would be appropriate to those data (as would a t-test, since there are only 2 groups).

But what if you wanted to study the relationship between gender and depression? Theoretically, you'd say that gender would be the independent variable. It's possible that gender differences explain differences in depression scores, but nonsensical to say that depression can change your gender. Gender is therefore a nominal independent variable (cause), and presumably, depression would be measured as a continuous dependent variable (effect). And you would still use analysis of variance (or the t-test) for these data, even if there is no experimental manipulation (let's not talk about how one might manipulate gender, OK?).

Bottom line: distinguishing independent from dependent variables means distinguishing cause from effect in your proposed explanatory model. This, plus knowing the level of measurement of each variable, helps you to decide which statistical procedure is appropriate to the data. But we should always keep in mind why it is that the experimental method is considered the "gold standard" for inferences of causality: you need some form of experimental manipulation to make that inference stick.