Chapter 8: Exploring Personality-Behavior Relationships
This exercise illustrates multiple regression analysis, correlation, and content aggregation. 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.
This exercise continues Chapter 7, which several methods of choosing behaviors to aggregate into a scale. One statistical technique which is useful for exploring how various behaviors are related to a personality measure is multiple regression analysis (MRA, or MRC, for Multiple Regression/Correlation). MRA allows the researcher to correlate a large number of variables together at one time, predicting one "criterion" variable (sometimes called the dependent variable) with several "predictor" variables (sometimes called the independent variables). Usually, MRA is performed with correlational data, that is, data that have no manipulation and control, and so have no real "independent" or "dependent" variables in the experimental sense.
To use MRA for our data, we will need to be flexible in our thinking and change our usual approach of "predicting behavior from personality." Since we are interested in exploring how several behaviors might relate to a single trait (e.g., outgoingness), we will treat the behaviors (e.g., wbo1-wbo43) as the "predictor" variables, and the trait measure (JPISPT & CPISY, used separately in 1 analysis each) as the "criterion" variable. Conceptually, this means that we are "predicting personality from behavior!" However, since both variables are just measured (reported) characteristics of the participants, and thus we have no experimental independent and dependent variables, this is not a problem. Further, there are behaviorally-oriented psychologists who would find nothing wrong with this, interpreting personality or any other measured characteristic of persons to be descriptions of their behavior.
When using MRA, it actually does not matter if some predictors are reverse-scored or not. If they are, you will notice negative coefficients for those variables to describe their relationship with the criterion. We will regress the 43 weekly behaviors for outgoingness as predictors on the JPI outgoingness scale (jpispt) as the criterion variable. Unfortunately, SPSS uses the words "Independent" for predictors and "Dependent" for criterion, but everything noted above about these terms still applies.
Use the Analyze>Regression>Linear menu.
Independent: wbo1, wbo2, wbo3, Ö. etc., through wbo43
Click on the Statistics button and select "Part and partial correlations"
Run the procedure with OK
This performs a "Stepwise" regression analysis, which means that highly correlated predictors (our behaviors) are chosen from the list of 43 to build an optimal "model" that best correlates with our criterion variable (jpispt). The predictors are "entered in" the model on the basis of correlations with the criterion, and overlap with any predictors already in the model. As the computer builds the model by adding predictors, predictors which become too weak are dropped. The final model contains those behaviors correlate maximally with jpispt, yet have little overlap with each other, so that they all contribute something unique to the prediction of (i.e., correlation with) jpispt. SPSS uses a criterion of "statistical significance," set by default to the ".05 level" to determine whether predictors should be entered or dropped from the model. Refer to your statistics course for a basic understanding of "statistical significance." For more advanced information, use the SPSS help buttons, glossary, and information provided when you point to something and click on the right (rather than left) mouse button.
The output will provide a history of the variables entered and removed as the model was constructed; a Model Summary showing the multiple correlation (R) for each "model" (step in the model building process); an ANOVA table showing the statistical significance of each model; regression coefficients for the predictors included in the model at each step; and information about variables that were excluded from the model.
What do these results show? There are amazingly only 2 behaviors in the final model: wbo7 ("I chatted with others in my spare time) and wbo27 ("I selected a solitary rather than a social activity for myself"). Note that the coefficients for wbo27 are appropriately negative, since this is a reverse-worded item. The multiple correlation for these two variables together with jpispt is R=.615. This is larger than any of the correlations we have examined so far for these weekly outgoingness measures. Apparently, these two behaviors are key indicators of outgoingness as measured by the JPI. Try repeating the MRA, using "cpisy" instead of jpispt, to see if the optimal behavioral indicators for outgoingness are different for the other personality inventory.
You can use this technique to identify key behaviors, to better understand what aspects of behavior are tapped by a particular personality inventory, and certainly to address the question, "How well can personality trait measures predict behavior?" You can use this MRA technique as an exploratory step to identify behaviors to aggregate. If you create a simple 2-behavior aggregate from wbo7 and wbo27 (remember to reverse this one before aggregating), and correlate it with jpispt, what do you find? Finally, another thing to try is using the global trait ratings that participants also made at the beginning of the study. These overall trait measures function like the jpi or cpi measures, and can be used this way in analyses. These are listed in the codebook under "General Self-Ratings on Trait Adjectives." The aggregated rating "GOOSTL," for example, is an average of each participantís overall rating of themselves on 8 different outgoingness adjectives.