Chapter 10: Comparing Behavior Across Situations
This exercise illustrates correlation, aggregation, and 2-Way Mixed Model (Between/Within) ANOVA. It assumes prior familiarity with the basics of Factorial ANOVA. It uses the PERS dataset, consisting of 90 cases and 968 variables. The variables represent measures of traits and relevant behaviors for the dimensions of extraversion (outgoingness) and conscientiousness, reported each week for three weeks by a group of undergraduate psychology students.
If you could measure the same behaviors for a group of people in two different situations, how consistent do you think it would be? Do you think that your results would be different if you repeated the measurements again the following week? How much do you think human behavior is a reflection of personality characteristics, situational constraints, or variations over time?
This exercise illustrates the comparison of the same behaviors across two different situations, using both a correlational and ANOVA approach. The overarching question addressed by the PERS data is "How consistent is human behavior, and is the consistency predictable from human personality?" Over the course of the prior exercises, this question has been examined by intercorrelating different behaviors which presumably are products of the same personality trait, and by using personality trait measures to predict various relevant behaviors. The inherent unreliability of behaviors measured on a single occasion was remedied by aggregating across time (see Chapters 2, 3, 4). The narrow specificity of individual behaviors was broadened by aggregation across behaviors (see Chapters 5, 6), creating a behavioral measure comparable in generality to a personality trait scale. This current exercise takes a different approach to the consistency question, while incorporating the processes of temporal and content aggregation employed successfully in the previous exercises. This exercise examines directly the question "Is behavior consistent across different situational contexts?" This question has been historically at the very heart of a psychological debate known as the "person/situation controversy," (see Epstein & O’Brien, 1985, for a brief history) which seeks to understand how much of our behavior is a function of internal factors (e.g., personality), and how much is a function of external factors (e.g., the particular situational context at the time). To examine this requires measuring the same behaviors in at least two different situational contexts. Further, given the unreliability of single-occasion measures of behavior, the measures would need to be taken over several occasions.
The PERS dataset includes cross-situational measures as described above: the same several behaviors measured in each of two different, specific situations, repeated on the same day each week for three weeks. For outgoingness, several behaviors related to (a) an in-person conversation and (b) a telephone conversation, were measured each week. For conscientiousness, several behaviors related to (a) the participant’s most important class and (b) the participant’s least important class, were measured each week.
Several methodological and definitional questions may be raised about these measures. For example, it surely is not the same conversations which are being reported on each week, so how do these represent the same situations measured repeatedly over several occasions? When viewed in this manner, of course they are not. But then, every situation in our lives from moment to moment would be treated as different, and the issue of cross-situational consistency would be meaningless. Insofar as there is some continuity to our behavior when talking in person or on the telephone, we can speak of these as similar "situations." We would certainly expect variation over occasions on each of these, and that is why they are measured repeatedly over time. The purist can avoid the issue by only using the data for a single-week’s measurements. This is not advised, however (see Chapters 2, 3, 4). Another obvious question about the measures concerns the comparison of the "most important" versus "least important" classes. This is not the same as comparing "Psychology 1A" with "English Literature 5;" in fact, it means that most participants are measured in different classes from each other. Aside from the logistical difficulty (impossibility?) of finding two courses in common for all the participants to compare, the problem is one of psychological comparability. It seemed a clearer and more definitive situational contrast to use each student’s own highest priority and lowest priority courses, rather than specific courses which might not be perceived much differently by some or many students.
There are two analytic strategies to use in comparing the behavior across situations: correlational, or ANOVA. In either case, we will use the 3-week average measures (aggregates), and we will analyze both the individual behaviors and a behavioral aggregate (see Chapter 6). Examine the measures in the codebook in the section "Cross-situational Measures (3-Week Average)." We will use the outgoingness measures in this exercise (OSP1-OSP6; OST1-OST6). None of these variables need to be recoded, though the OSP6 and OST6 may be too different from the others, and so we will not use them. Create two new aggregate measures across the five behaviors:
osp5tot sum(osp1 to osp5)
ost5tot sum(ost1 to ost5)
The correlational approach to comparing behavior cross-situationally is simple: correlate osp1 with ost1, osp2 with ost2, etc., including osp5tot with ost5tot. Look at your results and explain them, given what you know about aggregation and the kinds of correlations obtained in other exercises between behaviors and with personality traits. You should find them to be lower than many of the other correlations we examined in previous exercises. Why do you think that is? What do you think would be the results if you used the measures for only Week#1 instead of the 3-week averages?
The ANOVA approach is more complex, but it yields some interesting information about the interaction between personality and situation influences, so it is favored by some researchers. Recall from Chapter 9 that ANOVA compares conditions or groups, so we must divide our participants into groups according to their personality characteristics. This was done in Chapter 6; if you don’t still have the "extgrp" variable on your datafile, go back to that exercise and recreate it. Our situations already represent two "conditions" under which measurements were taken.
We will use a "2-Way ANOVA." There are two "factors" in our analysis of variance: outgoingness level (low/average/high) and situation (in-person/telephone conversation), and so we have a 3x2 (or 2x3, if you prefer the reverse order) ANOVA. Since participants are classified in only 1 of the 3 outgoingness groups, this is a "between-subjects" factor. Since participants were measured in both situations, this is a "within-subjects" factor. This particular analysis, then, is called a "mixed-model," or "between/within" ANOVA. We need to perform separate ANOVAs on each variable; we will use the 5-item aggregate as an example here.
Since there are repeated measures (situations; our "within-subjects" factor), we will use:
Analyze>General Linear Model>Repeated Measures
Within-Subject Factor Name: converse
Number of Levels: 2
(click on Add, then on Define; you will get a new, larger dialog box)
Within-Subjects Variable (converse): osp5tot, ost5tot
Between-Subjects Factor(s): extgrp
Click on OK to run the ANOVA
The output is daunting, because the design is complex and SPSS provides many multivariate statistics even when they are not needed. There are several tables of output; we will look at the third table, called "Tests of Within-Subjects Effects." This table gives 4 different computations for each of 3 different effects. We will only examine the lines labeled "Sphericity Assumed." The F-ratio for Converse is 12.945, and is highly statistically significant (.001 in the "Sig." Column means less than a .001 probability of occuring by chance). The interaction effect ("Converse*Extgrp") is not statistically significant. Now move down to the last table, "Tests of Between-Subjects Effects." The line for "Extgrp" shows that it is statistically significant (F=8.034; probability less than .001 of occuring by chance). So the analysis shows that there is significant variability in the aggregated behaviors as a function of situational context (in-person versus telephone conversation), and as a function of personality trait level (low versus average versus high outgoingness). There does not appear to be any significant interaction between the personality factors and the situational factors in influencing behavior. Since significant effects were found, you would want to re-run the analysis and include Options for descriptive statistics (giving you means for the different conditions) and Post-hocs (giving you specific statistical comparisons between conditions). Also, since this only analyzed the aggregated measures (osp5tot & ost5tot), you will need to repeat the analysis for each of the separate behaviors (e.g., osp1 & ost1) to compare to the correlations from the first analysis.
Compare the two analyses (correlational and ANOVA). Which do you prefer and why? You can try the same analyses with the conscientiousness measures on your own.
Epstein, S. & O’Brien, E. J. (1985). The person-situation debate in historical and current perspective. Psychological Bulletin, 98, 513-537.