Hello Vito,

For Likert-type, summated scales, you'd likely be better served using common factor analysis in lieu of principal components analysis to start with. The reasons are that: (a) PCA assumes all observed variance is common variance, and no unique, specific, or error variation exists, whereas FA only attempts to partition the common variance, and explicitly allows for unique, specific, and error variation; (b) unless the number of variables being considered is large, PCA tends to yield upward biased estimates of loadings.

In general, I'd always assume that oblique rotation is more realistic, especially among factors/constructs that deal with human beings. Only if you find that estimated correlations are negligibly different from zero would it make sense to adopt a simpler, orthogonal solution.

Now, for your question!

First, I've seen no evidence that a specific, single value for either method works best for all data sets. In other words, your mileage may vary.

For promax: start with a value of 4 (and choose even integers higher or lower).

For oblimin: start with a value of 0 (default), then, if you want to restrict the degree of obliqueness, enter successively smaller negative values.

In either case, it's likely a judgment call as to which solution best fits your specific research aims.

Good luck with your work.

Hello Vito,

For Likert-type, summated scales, you'd likely be better served using common factor analysis in lieu of principal components analysis to start with. The reasons are that: (a) PCA assumes all observed variance is common variance, and no unique, specific, or error variation exists, whereas FA only attempts to partition the common variance, and explicitly allows for unique, specific, and error variation; (b) unless the number of variables being considered is large, PCA tends to yield upward biased estimates of loadings.

In general, I'd always assume that oblique rotation is more realistic, especially among factors/constructs that deal with human beings. Only if you find that estimated correlations are negligibly different from zero would it make sense to adopt a simpler, orthogonal solution.

Now, for your question!

First, I've seen no evidence that a specific, single value for either method works best for all data sets. In other words, your mileage may vary.

For promax: start with a value of 4 (and choose even integers higher or lower).

For oblimin: start with a value of 0 (default), then, if you want to restrict the degree of obliqueness, enter successively smaller negative values.

In either case, it's likely a judgment call as to which solution best fits your specific research aims.

Good luck with your work.

David Morse Thank you very much for the exhaustive answer.

How do I know whether the solution is improved when I enter other values?

Thannks again.

Hello Vito,

Unfortunately, when the population structure is unknown, there is no way to be certain as to what the "best" solution may be. Only people running simulation studies from a known generating structure would have that luxury.

However, if you agree that a chosen solution makes sense conceptually and theoretically, and no simpler solution (which would work equally well or better at reproducing the observed inter-variable relationships) exists, then stick with it. Observing that the resultant structure holds/works with subsequent data sets can certainly give you more confidence in its generalizability.

Good luck with your work.