[Marie Baron]

Thank you Andrew,

It works well with the code you propose.

Thank you everyone for your help on this. I will report it to tech support.

[Jeff Pitblado (StataCorp)]

In the 06nov2017 update, whatsnew item number 21 states:

gsem with option lclass() to specify models with
categorical latent variables no longer implies option
listwise. The old behavior is preserved under version
control.

I assume that Marie Baron's dataset contains missing values. These
missing values make it difficult (I think impossible) to compute the
saturated likelihood that is comparable to the likelihood for the fitted
LCA model. Hence, no LR test from estat lcgof.

If Marie is willing to throw out observations with any missing values,
by specifying the listwise option, then estat lcgof should
produce the LR test.

[Weiwen Ng]

I am glad we uncovered this little mystery. Jeff, thanks to you and Bingsheng, who corresponded with me.

[Marie Baron]

Hi,

Indeed, I have missing values. Thank you for the explanation, it does make sense now. And thank you for the suggestion.

[Paul Lambe]

Hi all with a estat lcgof problem. On my personal computer, Stata 15.0 after running gsem latent class model with categorical variables and entering command estat lcgof I am rewarded with the loglikelihood -ratio chi sq statistic and associated p value, and BIC and AIC, essential for model selection. However, working on a project with remote access to a secure database with no admin rights using Stata 15.1 and running gsem latent class model with categorical variables and entering command estat lcgof, I only get AIC and BIC. Any suggestions on how to resolve this warmly welcome. BW Paul

[Weiwen Ng]

Hi all with a estat lcgof problem. On my personal computer, Stata 15.0 after running gsem latent class model with categorical variables and entering command estat lcgof I am rewarded with the loglikelihood -ratio chi sq statistic and associated p value, and BIC and AIC, essential for model selection. However, working on a project with remote access to a secure database with no admin rights using Stata 15.1 and running gsem latent class model with categorical variables and entering command estat lcgof, I only get AIC and BIC. Any suggestions on how to resolve this warmly welcome. BW Paul

Paul, as the above discussion indicates, under version 15.0 Stata used listwise deletion for any missing values (i.e. if one indicator has a missing value, the whole observation is dropped). 15.1 will not do this (i.e. all pairwise present info is used), but this means that it can't calculate an exact likelihood for the fitted model, hence only AIC and BIC are available. You can read earlier to see how you can force Stata to use the version 15.0 behavior.

In the literature I'm familiar with in health services research, I don't think I've seen the LR chi-sq statistic used a lot. People have selected models via BIC alone or in combination with the bootstrap LR test (which Stata cannot do, I've tried to figure this out and it would have to be manually programmed, which is beyond my current ability) or the Lo-Mendell-Rubin LR test (this is a theoretically derived analytical correction to the straight up LR test).