Uncovering steroidopathy in women with autism: a latent class analysis

Table 3 Information criteria for multi-group latent class analysis models

Model	-log likelihood	AIC	BIC
Unconstrained	-3,390.623	6,879.246	7,105.367
Semiconstrained	-3,410.932	6,875.863	7,000.461
Semiconstrained, PCOS posterior probability variable	-3,404.259	6,866.517	7,000.344
Fully constrained	-3,420.879	6,893.758	7,013.74
Fully constrained, PCOS posterior probability variable	-3,417.608	6,891.216	7,020.428

The lowest values on the information criteria indicate the best model (shown in bold), weighing absolute model fit and number of estimated parameters. In the unconstrained model, both conditional probabilities and latent class prevalences are allowed to vary by diagnostic group; in the semi-constrained model, only latent class prevalences are allowed to vary by group; in the fully constrained model, both diagnositic groups are forced to have the same latent class prevalences and the same item-response conditional probabilities. Further discussion of model comparison can be found in Additional file1, and 5-fold cross validated loglikelihood, AIC, and BIC values for these models can be found in Additional file2. AIC, Akaike Information Criterion; BIC, Bayesian Information Criterion; df, degrees of freedom; PCOS, polycystic ovary syndrome.

ISSN: 2040-2392