I'm new to conducting CFA, and would be appreciative of any feedback users could provide. I've seen a few questions on here in the past that are similar but not quite the same to what I'll pose. I'm trying to implement an approach used in a paper by Biderman, Nguyen, Cunningham & Ghorbani (2011) where they examined the structure of the big five personality traits. In sum, they created a CFA model with 3 method bias factors (1 for all items, 1 for positive items, and 1 for negative items) in addition to the 5 personality factors. I'd like to attempt something similar with my data. Below is what I have so far where g is the general factor, method is the negative bias factor, method_p is the positive bias factor, and f1, f2, f3 are the three group factors.

Overall, I have a few questions. (1) Broadly, does how I've structured things below seems reasonable given my aim to model the 3 group and 3 method bias factors? (2) I've fixed the variances to 1 and made the factors orthogonal. Is doing so any different than adding orthogonal = "FALSE" to the CFA function? Finally, I'm using MLS as an estimator per Rhemtualla, Brosseau-Liard & Savalei (2012) as the scale items contain five categories; however, I've seen a lot of recent papers using DWLS when any sort of ordinal indicators are used...has any sort of consensus been reached with respect to what estimator should be used for ordinal data?

m1<- ' g =~Item3+ Item1+Item6+Item9+Item11+Item12+Item14+Item2+Item4+Item5+Item7+Item8+Item10+Item13+Item15+Item16+Item17+Item18+Item19+Item20
          f1  =~ Item1 + Item4 + Item5 + Item9 + Item14 + Item18 + Item19
          f2 =~ Item2 + Item8 + Item10 + Item13 + Item15 + Item17 + Item20
          f3   =~ Item3 +Item6+ Item7 + Item11+ Item12 + Item16
          method =~ Item2+Item4+Item5+Item7+Item8+Item10+Item13+Item15+Item16+Item17+Item18+Item19+Item20
          method_p =~ Item3+ Item1+Item6+Item9+Item11+Item12+Item14

          f1 ~~ 0*f2
          f1 ~~ 0*f3
          f1 ~~ 0*g
          f1 ~~ 0*method
          f1 ~~ 0*method_p
          f2 ~~ 0*f3
          f2 ~~ 0*method
          f2 ~~ 0*method_p
          f2 ~~ 0*g
          f3 ~~ 0*method_p
          f3 ~~ 0*method
          f3 ~~ 0*g
          method ~~ 0*g
          method ~~ 0*method_p
          method_p ~~ 0*g

          f1 ~~ 1*f1
          f2 ~~ 1*f2
          method ~~ 1*method
          method_p ~~ 1*method_p
          g ~~ 1*g
          f3 ~~ 1*f3'

   fit1<- cfa(m1, data=cfa_complete[,2:21],std.lv=TRUE, estimator = "MLS")
  • $\begingroup$ What's your sample size? $\endgroup$ Jun 5, 2017 at 16:30
  • $\begingroup$ Hello @JeremyMiles, n=950 $\endgroup$
    – AD21
    Jun 5, 2017 at 17:06
  • $\begingroup$ Is MLS a typo? For WLS? Looks fine to me. When you run it are there any problems or red flags? $\endgroup$ Jun 5, 2017 at 18:28
  • $\begingroup$ It is indeed, no flags when it runs...thanks for the feedback $\endgroup$
    – AD21
    Jun 30, 2017 at 13:41

1 Answer 1


When you have ordinal data, and a relatively complex model the traditional mls often will not converge. If that is what is happening use dwls.

  • 1
    $\begingroup$ Welcome to the site, @DonRock. This is rather sparse by our standards, can you expand on it in any way? Why does DWLS work better here, eg? $\endgroup$ Jun 3, 2017 at 22:43
  • $\begingroup$ @Don Your answer is being automatically flagged as low quality because of its extreme brevity. Please expand your answer to include additional explanation $\endgroup$
    – Glen_b
    Jun 3, 2017 at 23:35
  • $\begingroup$ I recommend an article by Mindrila,D. (2010)., ML &DWLS estimation procedures: A comparison of estimation bias with ordinal and multivariate non=normal data. International J. of Digital s $\endgroup$
    – Don Rock
    Jun 4, 2017 at 14:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.