2

I am trying to conduct an exploratory factor analysis R using the psych package. My data consists of items from different scales- they are all very skewed, but differ in their number of response catgories (ranging from 4 to 7). Which factor analysis method should I use? Minres or WLS? After conducting the EFA I want to conduct CFAs using the lavaan package. Here the estimator MLM seems appropriate for my data. In MPlus there is a MLM method available for exploratory and confirmatory factor analysis. Do you know which factor analysis method most likely corresponds in psych to the MLM estimator? Or what do you think about transforming the data? Thank you for your help, kind regards A.

A. Bi.
  • 21
  • 2

2 Answers2

2

For categorical items ranging from 4-7 categories, I would use the cor="poly" option in fa and then run either the default (minres) or the WLS option.

When finding the correlation matrix with the cor="poly" option, it will give a warning that the number of categories are not the same, but this can be ignored.

For complete data, the cor="poly" option produces correlation matrices that are identical to the lavann polychoric matrices. I have not compared them to MPlus.

I have found that the minres (default) option in fa is more robust than WLS, but your mileage may vary.

William Revelle
  • 1,200
  • 8
  • 15
  • I have two related questions for Dr. Revelle: – Larry Hunsicker Dec 22 '19 at 19:58
  • I have two related questions for Dr. Revelle: First, if one uses the fa() function, how does one identify the polychoric variables? The mixedCor arguments c, p, and d don't seem to be accepted by fa. Second, irrespective of selected factor extraction method, I get a warning running my data: " The estimated weights for the factor scores are probably incorrect. Try a different factor extraction method." The results of the various methods other than "ml" give very similar results. Though this is only a warning, it suggests that the analyses aren't valid. Do I need to be concerned? – Larry Hunsicker Dec 22 '19 at 20:07
0

Using psych_1.9.12 you can specify cor="mixed" and mixedCor will figure out which variables are dictotmous, which require polychorics, etc.

It gives a message that mixed.cor is deprecated. Ignore that.

The factoring is probably ok. the problem is in estimating the weights needed to find factor scores. That particular message is also not very helpful because it turns out that the weights used in that calculation are always just regression weights, no matter what you specify. That is a bug. However, the factoring is good, just the estimation of the factor indeterminancy is messed up.

William Revelle
  • 1,200
  • 8
  • 15