. Author manuscript; available in PMC: 2017 Apr 1.

Published in final edited form as: Am J Med Genet B Neuropsychiatr Genet. 2015 Sep 3;171(7):948–957. doi: 10.1002/ajmg.b.32375

Table 5.

Latent Growth Model Fit in Females

Model	# Cl	# Params	LogLik	BIC	Entropy	Class Percentages
Mother Reports (Age 7–12)
1. No Random Effects	1	4	−25320	50675
1. No Random Effects	2	7	−23645	47353	.65	58, 42
	3	10	−23289	46668	.61	46, 37, 17
	4	13	−23250	46616	.54	42, 27, 21, 10
	5	16	−23232	46607	.54	41, 22, 19, 10, 8
	6	Failed to converge
2. Random Intercept	1	5	−23302	46648
	2	Failed to replicate likelihood
3. Random Intercept and Slope	1	6	−23301	46657
3. Random Intercept and Slope	2	Failed to replicate likelihood

Self-Reports (Age 14–18)
4. No Random Effects	1	4	−7254
4. No Random Effects	2	7	−6894	13843	.59	51, 49
	3	10	−6823	13724	.54	48, 27, 25
	4	13	−6812	13727	.52	37, 27, 25, 11
5. Random Intercept	1	5	−6824	13688
	2	15	−7600	15319	.36	59, 41
6. Random Intercept and Slope	1	6	−6819	13685
6. Random Intercept and Slope	2	17	−7586	15307	.40	62, 38

Piecewise Models (Age 7–18)
7. Random intercept	1	9	−34056	68294
	2	15	−34000	68136	.31	.61, .39
8^. Rater Mean* Difference = 0	1	8	−34190	68453

χ2=216.4, df=1, p < 1E-16

Latent class growth models for maternal and self-ratings separately, and jointly using a piecewise model. Models were estimated with an increasing number of classes (Cl). Shown are the number of estimated parameters in the model, the final log likelihood value, the Bayesian information criterion (BIC), Entropy, and class proportions. Models with multiple classes did not have significantly lower BIC than models with a single class (bold). For the self-reports, a likelihood ratio test showed that the random slope (6) did not improve fit, so model (5) was chosen. For the piecewise models, we selected a single class model because entropy was low indicating poor class assignment. Secondly, the BIC of the two class model fell within 1 SD of bootstrapped BIC for the single class model (BIC SD = 1264). Model 8 tested whether the second intercept was necessary to model the discontinuity between raters at ages 12 and 14. This was done by testing whether the mean of the second intercept could be set equal to 0. This model fit significantly worse than the unconstrained model, indicating a mean difference between raters.