Good evening everyone, the attached document contains the literature on group-based
trajectory modeling.
I can't understand a few things:
1) Is occ >5 mandatory? I ask because it's present in section 3.1, but not in the model specification (section 3.2). So, is occ >5 a necessary or recommended condition?
2) Model validation. So, if I assume 5 groups for a timepoint and run an ANOVA, is it possible, for example, that there's no difference between groups 2 and 3? Or do I have to correct it, and how?
Model Specification:
. First, Settle for a starting polynomial order (the same for all ks) and maximum number of trajectories k. These calls are partly arbitrary, partly theory driven, depending also on the nature of the data (a large number of time points allows for a higher order of polynomials to be explored, whereas large sample sizes give the researcher the freedom to play around with a larger number of clusters). Keeping the order of polynomials constant, run several models for different ‘ks’; . Second, among all models, select the one yielding the best BIC, with corroboration by the supportive roles of APPA (above 0.7), and the 1% rule for ^j. The indices themselves are seldom provided in publications; . Third, keep the selected k constant, conduct significance tests of polynomials terms keeping a watch on the BIC;
2. Model validation: Lastly, in the overwhelmingly vast majority of papers, a classify–analyseii strategy follows, either for latent classes profiling (latent classes’ comparisons based, e.g. on ANCOVA/ANOVA), or for predictive validity in terms of the association between trajectories and a distal outcome (e.g. logistic regression Klijn et al. 2427 for dichotomous outcome). At this stage, a prominent feature of GBTM, which is the uncertainty of cluster membership, is taken out of the equation. A closer discussion on model validation is beyond the scope of the present work.
thanks to everybody
trajectory modeling.
I can't understand a few things:
1) Is occ >5 mandatory? I ask because it's present in section 3.1, but not in the model specification (section 3.2). So, is occ >5 a necessary or recommended condition?
2) Model validation. So, if I assume 5 groups for a timepoint and run an ANOVA, is it possible, for example, that there's no difference between groups 2 and 3? Or do I have to correct it, and how?
Model Specification:
. First, Settle for a starting polynomial order (the same for all ks) and maximum number of trajectories k. These calls are partly arbitrary, partly theory driven, depending also on the nature of the data (a large number of time points allows for a higher order of polynomials to be explored, whereas large sample sizes give the researcher the freedom to play around with a larger number of clusters). Keeping the order of polynomials constant, run several models for different ‘ks’; . Second, among all models, select the one yielding the best BIC, with corroboration by the supportive roles of APPA (above 0.7), and the 1% rule for ^j. The indices themselves are seldom provided in publications; . Third, keep the selected k constant, conduct significance tests of polynomials terms keeping a watch on the BIC;
2. Model validation: Lastly, in the overwhelmingly vast majority of papers, a classify–analyseii strategy follows, either for latent classes profiling (latent classes’ comparisons based, e.g. on ANCOVA/ANOVA), or for predictive validity in terms of the association between trajectories and a distal outcome (e.g. logistic regression Klijn et al. 2427 for dichotomous outcome). At this stage, a prominent feature of GBTM, which is the uncertainty of cluster membership, is taken out of the equation. A closer discussion on model validation is beyond the scope of the present work.
thanks to everybody
Comment