1
$\begingroup$

I am estimating latent curve models (btw, this is my first time using LCM) for multiple indices (n = 6) taken from the same sample of research subjects. For some of these indices, the fit measures TFI and IFI are indicating that the model is overfitted (TFI > 1.2 & IFI > 1.2), but for other indices they are telling me the model fit is quite good (TFI and IFI around 1.0). Also, none of the conditional models (predictors n = 3) indicate overfitting.

Can someone explain why, for some indices, the model is overfitted, yet for others it is not, even though the sample size nor the number of parameters does not change?

$\endgroup$
1
$\begingroup$

It's not overfit (necessarily). I've never heard people describe these indicators as overfitting.

Some fit indices can go over 1, others can't. You seem to have a typo in your question, as you report IFI being 1.0 and also being 1.2. I presume that by TFI you mean TLI - it's also called the NNFI, which stands for Non-Normed Fit Index - it's not normed because it can go over 1.

Check the equations for the indices - if your chi-square is lower than df, the the indices might go over 1 (or might be equal to 1).

I'm not sure what you mean by multiple indices.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ In Bollen & Curran (2005), they state that TLI (sorry for the typo) and IFI scores "much greater than 1 (e.g. 1.2) suggest the possibility of overfitting the data or having a model with too many parameters" (p. 46). For my study, I am creating multiple models for different indices (e.g. word frequency, word range, concreteness) taken from the same sample of subjects. The models for some of these indices are returning TLI and IFI values that are what Bollen and Curran (2006) claim to be evidence of overfitting. Models for some of the other indices, are not. Thanks. I'll check my equations. $\endgroup$ – jgarner17 Mar 14 '17 at 10:40
  • $\begingroup$ Thanks. I didn't remember that (and I've read that book!) $\endgroup$ – Jeremy Miles Mar 14 '17 at 15:34
  • $\begingroup$ I suspect your model has very few df - is that the case? $\endgroup$ – Jeremy Miles Mar 14 '17 at 15:35
  • $\begingroup$ The models returning TLI and IFI scores at or above 1.2 all have df = 8. I don't remember (from what I've read) if that's considered "very few". $\endgroup$ – jgarner17 Mar 15 '17 at 16:49
  • $\begingroup$ It's on the lower side. But it's not bad for LCMs that don't have anything complicated going on. $\endgroup$ – Jeremy Miles Mar 15 '17 at 23:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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