McDonalds omega: warnings in R

Question

I'm computing omega for several different scales; and get different warning messages for different scales with different omega functions in R. My questions are regarding how to interpret these warnings and if it is safe to report the retrieved omega statistics.

When I'm using the following function from the article "From alpha to omega: A practical solution to the pervasive problem of internal consistency estimation"

 ci.reliability(subscale1, interval.type="bca", B=1000)

I get these warnings:

 1: In lav_object_post_check(lavobject) :
   lavaan WARNING: some estimated variances are negative
 2: In lav_object_post_check(lavobject) :
   lavaan WARNING: observed variable error term matrix (theta) is not positive definite; use inspect(fit,"theta") to investigate.

And it can be many of them!

What do they mean? I still receive omega statistics; can they be interpreted or not?

When I use the function:

  psych::omega(subscale1)

I get this warning:

 Warning message:
 In GPFoblq(L, Tmat = Tmat, normalize = normalize, eps = eps, maxit =      maxit,  :
   convergence not obtained in GPFoblq. 1000 iterations used.

Again, What does it mean; and can I use the omega-statistics that I get?

Note that these warnings appear on different subscales; so one subscale can be computed using one of the function but not the other and vice versa.

EDIT: If it helps: Subscale1 encompasses 4 items; the sample includes N>300. Also, I can run a CFA analysis on these 4 items in lavaan (Chi2=11.8, p<.001; CFI=0.98; RMSEA=0.123).

Northumbria

Welcome to CV. Note that your username, identicon, & a link to your user page are automatically added to every post you make, so there is no need to sign your posts. In fact, we prefer you don't. I've added the [reliability] tag; note that purely software-specific questions are off-topic here - see our help center - but the error messages here are statistical in nature, so I think the question should be on-topic. – Edgewise

I think this means you have a problem of some sort. It's often caused by small sample size - what's your N? The lavaan errors are not fatal, you still get a result that is OK, but it means that your result is based on an underlying model that must be wrong (because it cannot exist). The non-convergence error is more problematic - I probably wouldn't trust that. You probably need to explore your data more to work out the precise issue. – Pion

Answer 1

That particular article to which you are referring seems to be the British Journal of Psychology (2014), 105, 399–412© 2013 by Dunn, Baguley and Brunsden. The omega coefficient they discuss is actually what Rick Zinbarg and I refer to as omega_total. (McDonald developed two omega coefficients which has led to this confusion.)

You are having problems using omega in my psych package. The omega function in psych is meant to find omega_hiearchical as well as omega_total. Thus, it tries (by default) to extract three lower level factors and then, in turn, factor the resulting correlations of those factors. However, with only 4 variables in your sub scale, it can not find a meaningful 3 factor solution. You can specify that you want to find two factors:

omega(subscale1,2) 

and it will work. However, omega_h is not particularly meaningful for 4 items.

Contrary to the suggestion of sample size, it is actually due to the number of items.

I think you might find the tutorial for finding omega_h using psych helpful:

[http://personality-project.org/r/psych/HowTo/R_for_omega.pdf]

Carpel