The Informal Model: Pile Sorts, Triad Tests, and Rank Ordered Data

When you have ordinal or ratio or rank-ordered data, including data from pile sorts and triad tests, use the informal model of consensus analysis (Weller 2007) (box 16.7).

BOX 16.7

LUMPERS AND SPLITTERS

Be careful, though: With free pile sorts, some people make just a few piles and others make many. Lumpers and splitters are technically not responding to the same cues when you ask them to sort items freely into piles. This means you can test whether there is a single culture—whether most people see the relations among a set of items in a pile sort similarly—and you can look for informants who are most representative of the culture, but you wouldn't put much stock in individual factor scores or use those scores as input into any other analysis.

In this model, you create a people-by-people similarity matrix from the original people-by-item test and factor the similarity matrix. You can do this in most statistical packages, but ANTHROPAC and UCINET will do it automatically. Here’s an example.

Adam Kis (2007) found that people in the village of Njolomole, Malawi had stopped going to every funeral because, with AIDS, there were just too many to go to. Kis asked 23 people to: ‘‘Name all of the reasons you can think of for attending a funeral.’’ He listed the 12 reasons cited most often on a piece of paper and asked 30 people: ‘‘If there were too many funerals in your village so that you could not attend each one, which of the following reasons would be the most important in helping you decide which funerals to attend?’’ Then he asked people to mark the next most important reason, and the next, and so on down the list. Table 16.11 shows his data.

These are rank-ordered data, so the informal consensus model is appropriate. To do this, correlate all pairs of rows in table 16.11 and turn it into a 30-by-30, people-bypeople similarity matrix (any statistical package will do). The result for the first 10 rows is shown in table 16.12.

Read table 16.12 as follows: Informants 4 and 9 are highly and positively correlated (0.741); informants 6 and 2 are hardly correlated at all (0.070); informants 2 and 3 are weakly and negatively correlated (-0.364); and so on. The negative correlation for informants 2 and 3 means that they tended to rank the reasons for going to a funeral in some opposite ways. For example, looking across rows 2 and 3 of table 16.11, informant 2 ranked HEL (helping the family of the deceased with funeral preparations) last on his list, and informant 3 ranked it third.

Next, factor analyze the 30-by-30 matrix of informant agreements using a variant of factor analysis called minimal residuals (or MINRES) or maximum likelihood (ML). I did this with SYSTAT, but you can use SPSS or any major statistical package. The results are in figure 16.22. The first factor is large, relative to the second (the ratio is 3.157), which means that there is a single culture at work, despite the differences in the way people ranked their reasons for going to a funeral. There is a nice range of scores—from 0.07 (for informant 2) to 0.92 (for informant 30)—and there are no negative scores, but some people (like informants 10, 11, 22, 25, and 30) are clearly more knowledgeable about the shared reasons for going to a funeral than others are.

Kis interviewed these knowledgeable informants in depth and asked them why reciprocity (attending someone’s funeral so that his or her family will attend your family’s funerals) was the runaway most-important reason given and why carrying the coffin and partaking of the traditional funeral feast were ranked so low. It turned out that carrying coffins was only for young men and his sample of 30 informants did not have that many young men in it. If he’d had more young men in his sample, this might have resulted in a higher overall ranking of eating as a reason for attending a funeral, since coffin bearers get generous portions of food.