Brief Summary of Chapter 12

In Chapter 12, we begin by explaining why the pure between subjects design has limited power:

  1. it only collects one observation per participant, and
  2. between subjects differences can create enough error variance to hide even moderately strong treatment effects

Therefore, you may want to look at ways to get more power than you would have with a pure between subjects design.

We start by looking at the matched pairs design. It tries to reduce individual differences by matching participants (but it still establishes internal validity by using random assignment).

Then, we discussed the within-subjects experiment, which could be considered an extreme form of the matched pairs design (rather than matching a participant with similar other, we match the participant with him/herself!). The within-subjects design has even more power than the matched pairs design because

  1. it eliminates, rather than just reduces, random error due to differences between participants
  2. it allows you to collect more than one observation per participant.

Unfortunately, within-subjects designs are not perfect. Results that seem to be due to the treatment may really be due to order effects (participants scoring differently on the first trial than on the last).

Order efffects may due to

  1. practice
  2. fatigue
  3. carry-over (an earlier treatment affecting performance on a later trial)
  4. sensitization (participants figuring out the hypothesis during the latter part of the study).
In addition to using logic to avoid these threats (e.g., allow plenty of time between treatments to avoid carry-over effects), we can use counterbalancing to balance out the effects of order. You may be most familiar with counterbalancing from sporting events. In football, for example, to balance out the effect of wind, teams switch sides every quarter. Thus, if the wind is consistent, each side goes into the wind the same amount of time.

Using a counterbalanced design complicates your experiment. In the simplest case, one group gets the sequence Treatment A--then Treatment B, whereas the other group gets the sequence Treatment B--then Treatment A. Treatment is a within-subjects factor (each participant gets Treatment A and Treatment B) and sequence is a between subjects factor.

The point is that counterbalancing is fairly easy to do, but you may get confused when you start looking at your results. You have to avoid confusing a between groups sequence effect (which means that Group 1 scored differently than Group 2) with an order effect (which would mean that participants scored differently on the first trial than on the second trial). To avoid being confused, study pages 390-397.

Another problem with counterbalancing is that it doesn't work if the order effects are inconsistent. For example, what if, at the football game, a big wind storm occurred during the fourth quarter? Then, the counterbalancing would not have balanced out the effect of wind. In psychological experiments, order effects may be uneven. For instance, suppose our IV is brain surgery/no surgery. The group getting brain surgery then no surgery may experience tremendous carry-over. The group getting no surgery then brain surgery may experience very little treatment carry-over. Thus, we have failed to balance out carry-over effects. Consequently, if we were to study the effect of a certain brain surgery, we should probably use a between subjects design.

Sometimes, you will find that you want to study one variable within-subjects and another between subjects. You can do both in a single experiment by using a mixed design.

There are only two complications in using a mixed design. First, you should make the right choice about whether a variable should be within or between. Table 12-12 (p. 400) will help you make that decision.

Second, you need to make sure that the computer has correctly analyzed your experiment. The computer might analyze your between factor as a within (or vice-versa). Making sure the degrees of freedom make sense (e.g., only 1 df for your between groups treatment if it had 2 levels, your between groups df adding up to one less than your number of participants) and seeing that your error terms make sense (the within should be smaller than the between) can help you find an error.

Back to Chapter 12 Menu