Chapter 7: Selecting Behaviors to Aggregate
This exercise illustrates 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 builds on Chapter 6, which illustrated the creation of "scales" of multiple behaviors to better correlate with personality trait inventories. Trait measures refer to behaviors at a general, rather than specific level. Aggregating different, relevant behaviors into a single "scale" score thus produces a behavioral measure that is comparable in breadth and generality to a trait measure (a personality "scale").
In Chapter 6, the first 10 weekly behaviors for outgoingness were aggregated into a single behavioral measure, and correlated with personality trait inventories (JPI and CPI). This was done for ease of illustration; there were actually 43 weekly behaviors measured for outgoingness. Does that mean that all of them should be used? If only some are selected for use, which ones are selected and why? This exercise addresses the question of how to select behaviors to aggregate.
Many strategies exist to aggregate multiple behaviors into a scale. Some of them are: (a) personally select some small subset of the behaviors, (b) use all the behaviors that have been measured for the trait, (c) use the behaviors which are considered most exemplary and prototypical of the trait, (d) use behaviors which form a consistent conceptual group or factor. This exercise will explore the first three options. The fourth option involves Factor Analysis or other empirical data clustering technique, and is beyond the scope of this exercise.
In the Chapter 6, the first 10 weekly behaviors were used. This is an example of option (a) above, and it has little to recommend it. Why only 10? Why these 10? You can try creating aggregates of other weekly outgoingness behaviors this way—the possibilities are almost endless. Try creating other aggregates if you want, and correlating them with the JPISPT and CPISY, if you would like.
The second option is the easiest—use them all. There are 43 weekly behaviors for outgoingness, nine of them reverse-scored. See the codebook for which of these items are reversed and suggestions for recoding (examining the actual questionnaire items should confirm that these are reverse-worded). When summing different behavior measures into an aggregate, it is important that all items are coded in the same "direction," otherwise, the reverse-scored item will detract from, rather than add to, the aggregate sum. First, recode those weekly outgoingness behaviors which are reverse-scored. Since there are several, it might be easier to use the Transform>Recode>Into Same Variables procedure. If you do this, you need to remember that you have already recoded these variables if you use them later in some other analysis. The nine items (see Recoding suggestions in codebook) can all be selected at once, and old and new values entered which will apply to all of them. For old and new values, use 0=7, 1=6, 2=5, 3=4, 4=3, 5=2, 6=1, 7=0.
Once you have recoded the reversed items, create a new aggregate with Transform>Compute. Target Variable: wbo43tot (or something similar)
Numeric Expression: sum(of wbo1 to wbo43)
When you have created this aggregate, correlate it with JPISPT and CPISY. You should find that the results are not much different than using the first 10. Why might that be? Is there a point of "diminishing returns" when adding more and more behaviors to the aggregate? One possibility is that as the number of behaviors added increases, the overlap between them does too.
Option (c) is to select the "best," or most representative behaviors for the trait. One way of doing this is to collect data on the perceived "prototypicality" or relevance of each behavior to the trait being measured. Then, the most prototypical behaviors (by consensus) can be selected for aggregation. See Buss and Craik (1984) for further explanation of prototypicality of behaviors for a trait. The PERS dataset has ratings of subjective relevance of each behavior for its trait. They appear in the codebook table of contents as "Subjective relevance of behaviors" under "VARIABLES MEASURED ONCE AT THE BEGINNING OF THE STUDY." For the weekly outgoingness behaviors, these ratings are variables srelo1 to srelo43. Get descriptive statistics for each of these to determine which behaviors were considered most relevant to outgoingness: Analyze>Descriptive Statistics>Descriptives. If you want, click on the Options button and select "Descending Means" for "Display Order," and the output will display the items in order from highest to lowest, according to mean subjective relevance.
You should find that srelo28 has the highest mean rating (5.68) and srelo31 has the lowest mean rating (3.51). If we select those items with a mean relevance rating of 5 or more, we have 14 items: 28, 2, 9, 21, 18, 24, 17, 33, 19, 36, 29, 34, 7, 30. So, we will create an aggregate of these "most prototypical" behaviors for outgoingness. Remember that although these ratings use variables srelo1-srelo43, our aggregate will use the behavioral measures (wbo1-wbo43). Use Transform>Compute.
Target Variable: wbo14rel
Numeric Expression: sum(wbo28,wbo2,wbo9,wbo21,wbo18,wbo24,wbo17,wbo33,
Correlate this aggregate (wbo14rel) with JPISPT and CPISY. You can try creating another aggregate of the LEAST relevant behaviors, and compare results. What do you find? Why do you think behavior prototypicality does (or doesn’t) make a difference?
Buss, D. M. & Craik, K. H. (1984). Acts, dispositions, and personality. Progress in Experimental Personality Research, 13, 241-301.