03-H

Conducting an Experiment

In any experiment, there are two different types of variables that are going to be involved. An independent variable (IV) is something we are going to manipulate (change). A dependent variable (DV) is something we will measure after manipulating the independent variable. Our goal is to determine if changing the IV CAUSES a change in the DV.

Note the key difference between an experiment and a correlational design. If we surveyed people to ask how often they play violent video games, and then observed how they react when provoked by a stranger, we might find that those who play more violent games react more aggressively. However, as we discussed previously, we cannot conclude that the games caused the aggression from the positive correlation alone. We don't know if...

  • The games cause aggression.
  • Aggressive people tend to choose more violent games to play.
  • Something else, like testosterone levels, independently cause differences in both game selections and aggressive responses.

In an experiment, we manipulate exposure to violence (the IV) in the lab and then measure aggression (the DV).

For our purposes, we will consider two different types of experimental designs...

Between-Subjects Experimental Designs

One approach is to have different groups of participants who experience different "levels" of the independent variable. For example, Bartholow and Anderson (2002) had participants play one of two different games -- Mortal Kombat (high violence) or PGA Golf (low violence) -- and told them that this was to determine whether playing games influences reaction times. Afterward, they then played a competitive reaction time game. If they won the round, they could decide how loud of a sound blast the other person was punished with. Thus, they operationalized the conceptual dependent variable of aggression as the intensity of punishments that the participant selected to inflict on another person. The researchers found that the group who played the violent fighting game (IV) tended to use more intense punishments on average (DV).

Of course, we can only conclude that the manipulation of the IV caused a change in the DV if we can account for the potential effect of of all other variables that might also influence how aggressive they responded. For example, some people have more aggressive personalities than others, some people might be in a bad mood because of some stressful life event, and others might not have had breakfast that day. We need groups that are, on average, equivalent to each other on all these variables in order to make a fair comparison between them. Thus, in a between-subjects experiment, we use random assignment to create the different groups. By randomly assigning participants to the various levels of the IV, all of those other variables should be (on average) equally distributed within the groups. That is, we'll have a few people in each group that happen to be particularly aggressive people, so if we do observe a difference when we measure the DV, we can be reasonably sure that personality difference between the two groups cannot be the explanation. The only thing that systematically varied between the two groups is which game they played (IV), so that must be the cause of the difference we observe in aggression (DV).

Within-Subjects Designs

Another approach is to have participants experience both levels of the IV and measure the DV twice. For example, imagine that we want to examine whether sleep deprivation affects attention spans. Participants will spend three nights in the lab, and the first night everyone will get a full-night's sleep to ensure we start out without a pre-existing deprivation. For the two remaining nights, participants will be allowed to sleep one night and kept awake the other (IV), and each morning they will be given a series of cognitive tests to measure how able they are to pay attention for long periods of time (DV). Thus, each participant is tested twice -- once after having slept well, and once after being deprived of sleep.

In within-subject designs, the groups we need to create now involve the order in which they experience the two levels of the IV. Imagine that we let everyone sleep before the first test, and then deprived them all of sleep before the second. If they tended to preform worse on the second attention tests, it could be because of the sleep deprivation... but it could also be that they tended to get bored by their third day in the lab and were not as motivated to try. On the other hand, if we deprive them all before the first test and find they preformed better on the second test, it could be because they had practice on the tests the day before. To rule out the possibility that order effects are biasing the results, we have to counterbalance the order using random assignment. Half of the participants will randomly be assigned to sleep before the first test and be deprived before the second. The other half will experience the two levels of the IV in the opposite order. Thus, when we combine all the data and look for difference in performance on the DV, any overall differences we find between the time they slept and the time they were kept awake (IV) can be attributed to the IV and not the order of the testing.

Regardless of whether we use a between-subjects or within-subjects design, the essential element of the methodology that allows us to draw causal conclusions is random assignment.

Random Assignment vs. Random Sampling

Many students confuse random assignment (which is the essential part of any experiment) with random sampling (which is how a group of people is selected to participate in a study). Despite the fact that they both include the word “random,” these two concepts have nothing to do with each other. Regardless of how a sample of participants was recruited, you need to use random assignment if you want to establish causality in your experiment. Furthermore, even if you used non-random sampling (say, a sample of local college students) in your study, you can still conduct an experiment and draw causal conclusions, but only if you used random assignment. See this 2x2 chart that helps explain things further.