The prefix “quasi” means, in essence, “sort of.” So a quasi-experiment is a “sort of” experiment. Specifically, a quasi-experiment is a study that includes a manipulated independent variable but lacks important controls (e.g., random assignment), or a study that lacks a manipulated independent variable but includes important controls. So a quasi-experiment has some features of a well conducted experiment but not others.
Types of Quasi-Experiments
There are many types of quasi-experiments. Here we discuss just a few of the more common ones.
Non-Equivalent Groups Design
A non-equivalent groups design includes an existing group of participants who receive a treatment and another existing group of participants to serve as a control or comparison group. Participants are not randomly assigned to conditions, but rather are assigned to the treatment or control conditions along with all the others in their existing group.
For example, imagine that we wanted to do a study to compare student performance in a cooperative learning section of Psych 144 with student performance in a standard lecture section. Imagine further that we scheduled two sections of the course, let students sign up for which one they wanted, and then taught one using cooperative learning and the other using standard lecture. Note that this study includes a manipulated independent variable, but it lacks random assignment of participants to conditions. The problem with this approach, of course, is that there might be differences between the two groups of students other than the style of teaching to which they were exposed. Perhaps the students who signed up for the earlier section are more “gung ho.” Or perhaps the students who signed up for the evening section are more likely to be working adults. Or perhaps the students in the 1:00 p.m. section tend to be drowsy after lunch. It is possible that differences in the dependent variable could have been caused by these differences rather than differences in teaching style.
You might wonder what the difference is between this and a plain old correlational study. The answer is “not much.” In fact, if you referred to this as a correlational study, we think few people would fault you. In general, however, non-equivalent groups are usually chosen to be as similar as possible to each other, which helps to control extraneous variables. For example, we probably would not use a daytime class as our cooperative learning group and an evening class as our standard lecture group. Instead, we would use another daytime class as our control group.
In a pretest-posttest design, a single group of participants is measured on the dependent variable both before and after the manipulation of the independent variable. Imagine that a group of 100 sixth graders is given a test of their attitudes toward drugs. This is the pretest. Then, a week later, a police officer comes to school and presents an anti-drug program (complete with “cool” decorated car and performing police dog). This is the treatment. Then, in another week, the students are given another test of their attitudes toward drugs. This is the posttest. Obviously, the substantive question here is whether the students’ attitudes toward drugs change after being presented with the anti-drug program.
The problem with pretest-posttest designs is that you cannot be completely sure that a change in the dependent variable was caused by the manipulation of the independent variable. For example, it is possible that something other than the anti-drug program might have occurred between the pretest and the posttest that influenced the students’ attitudes. It is even possible that having taken the pretest influenced their scores on the posttest.
The basic pretest-posttest design can be augmented by adding a control group. One way to do it would be to add a non-equivalent control group, for example, 100 sixth graders at another school who are not presented with the anti-drug program. Note that if this group also showed the same change in attitudes, it would not make sense to conclude that the change was due the program. Another way to do it would be to start with a larger sample and then randomly assign participants to either a treatment group or a true control group. A true control group is better than a non-equivalent control group because one can assume greater similarity between the control group and the treatment group. Note also that a pretest-posttest design with a true control group is an experiment and not a quasi-experiment.
Interrupted Time-Series Designs
A time series is simply a set of measurements of a variable taken at various points in time. For example, we could measure the moods of the students in our class each day throughout the semester, and we could see how people’s moods changed (or did not change) over time. In an interrupted time-series design, a time series like this (the dependent variable) is interrupted (usually near the middle) by the manipulation of the independent variable. For example, if we were interested in the effect of the color of the classroom on students’ moods, we could start measuring your moods each day from the beginning of September through Halloween. Then we could paint the room yellow (hoping to raise your moods) and continue measuring your moods each day from the beginning of November to the end of the semester. In analyzing the data, we would want to see whether there was an increase in your moods shortly after the painting of the room that continued to the end of the semester. The figure below shows some results that would confirm our hypothesis.
Note that this design is like a pretest-posttest design but with multiple pretests and multiple posttests. The advantage of this approach is that it provides greater confidence that the change in the dependent variable was caused by the manipulation and is not just a random fluctuation. For example, if students’ moods bounced around from week to week—high some weeks and low some weeks—then the change between Weeks 7 and 8 might just be one of these random fluctuations. However, the fact that students’ moods were consistently around 6.00 and then jumped right after the manipulation to be consistently around 7.00 makes it clear that this the jump is not just a random fluctuation. The interrupted time series design also allows us to see how long the effect of the manipulation lasts. Like the pretest-posttest design, the interrupted time series design can be augmented with either a non-equivalent control group or a true control group.
Sometimes, the independent variable in an interrupted time series design is not manipulated; it just changes naturally. For example, research has shown that the number of suicides in the general population increases right after a particularly prominent suicide (e.g., the suicide of a famous person). This has been shown by looking at the number of suicides for several weeks or months before the prominent suicide and then seeing how it changes immediately after the prominent suicide. Note that this design includes no manipulation and also does not include any attempt to control extraneous variables. Yet if the number of suicides suddenly increases after a prominent suicide, this seems like strong evidence of a causal connection—especially if this same result can be shown after many different prominent suicides.
Many researchers (including us) see the difference between correlational studies, quasi-experiments, and experiments as one of degree rather than as one of kind. At one end of this continuum are ideal experiments in which only the independent variable differs across condition, so that it is perfectly clear that changes in the dependent variable were caused by the independent variable. But as we move from ideal experiments to less ideal ones to quasi-experiments to correlational studies, there are more and more variables that differ across conditions, which makes it more and more difficult to see whether it was the independent variable that was responsible for changes in the dependent variable.
Researchers actually have a name for this continuum: internal validity. To the extent that a study allows one to conclude that the independent variable affected the dependent variable, we say that it has good internal validity. So an ideal experiment has perfect internal validity, experiments usually have good internal validity, quasi-experiments have somewhat less internal validity, and correlational studies often have poor internal validity.