Different needs in patients with schizophrenia spectrum disorders who behave aggressively towards others depend on gender: a latent class analysis approach

Table 1 Summary of different LCA model fit criteria

Number of classes	Number of estimated parameters	Residual degrees of freedom	Maximum log-likelihood	AIC	BIC	scBIC	Entropy	Number of times solution was found
1 (without covariate)	96	274	− 11195	22582	22957	23221	–	500/500
2 (with covariate)	193	177	− 10677	2741	22496	23024	0.8806	491/500
2 (without covariate)	194	176	− 1674	21737	22496	23027	0.8827	488/500
3 (without covariate)	290	80	− 10475	21530	22665	23460	0.889	32/500

BIC is considered the most relevant criterion for model selection according to which the two-class model indicated best model fit (highlighted with bold type). In the two-class model with covariate, initialisation of the priors was based on the potential class-predictor gender, but did not show any relevant differences to the two-class model without covariate. For the purpose of this study, subsequent results were based on the two-class model with covariate. Higher values of maximum log-likelihood indicate a better model fit but favour overfitting. Information criteria penalise the number of estimated parameters to prevent overfitting: AIC, Akaike’s Information Criterion; BIC, Bayesian Information Criterion; scBIC, sample-size-corrected Bayesian Information Criterion; Lower AIC, BIC and scBIC values indicate a good and parsimonious model fit. Entropy, measure of classification uncertainty with higher numbers indicating a better class separation; number of times solution was found = number of times solution was found out of 500 random initializations of prior probabilities to avoid local extrema, with higher numbers indicating a more unambiguous result

ISSN: 1744-859X