Hello,
I was wondering how one can conduct simple slope tests for a two-way interaction term. Especially, I'm using a social relations modeling to look at how person A's feeling toward B influences A'
s behavior toward B across many dyads (e.g., A-B, A-C, A-D and B-A, B-C, B-D, etc.) and this may vary depending on characteristic of A or B; that is, cross-level interactions.
Any advice will be appreciated.
Simple slope test
Re: Simple slope test
There are a number of ways that you can evaluate ‘significance’ in MLwin.
1 if the model does not have a discrete response as is estimated by maximum likelihood or restricted maximum likelihood you can do a Likelihood Ratio Test – essentially calculate the difference in the likelihood for the two models and the difference is distributed as chi-square with degrees of freedom given by the and the change in the number of parameters – this is the recommended procedure
2 If you are using quasi likelihood for discrete outcome then you cannot perform a LRT so you need to use the Intervals and test window to carry out a Wald test, this can be done for either the fixed part of the random part.
3 If you are using MCMC estimation - you can store the model estimates and tick on extra MCMC information and you will get the Bayesian p values ( in version 2.30) of the fixed part and the 95% credible intervals of the fixed and the random part. You can also compare a sequence of models via the Deviance information Criterion and these will be automatically penalised for model complexity – the smaller the DIC, the ‘better’ the model
4 If there are interactions involved, I find it useful to use the customised predictions procedure (see Manual Supplement) to plot the predicted result and put confidence intervals around the estimated relations
You can download this from Research gate which considers testing and storage in Chapter 6, customised predictions with interactions in Chapter 8 and the DIC in Chapter 10 (I have not updated it yet to include Bayesian p values)
Kelvyn Jones, SV Subramanian (2014) Developing multilevel models for analysing contextuality, heterogeneity and change using MLwiN, Volume 1
https://www.researchgate.net/publicatio ... ev=prf_pub
Hope this helps
1 if the model does not have a discrete response as is estimated by maximum likelihood or restricted maximum likelihood you can do a Likelihood Ratio Test – essentially calculate the difference in the likelihood for the two models and the difference is distributed as chi-square with degrees of freedom given by the and the change in the number of parameters – this is the recommended procedure
2 If you are using quasi likelihood for discrete outcome then you cannot perform a LRT so you need to use the Intervals and test window to carry out a Wald test, this can be done for either the fixed part of the random part.
3 If you are using MCMC estimation - you can store the model estimates and tick on extra MCMC information and you will get the Bayesian p values ( in version 2.30) of the fixed part and the 95% credible intervals of the fixed and the random part. You can also compare a sequence of models via the Deviance information Criterion and these will be automatically penalised for model complexity – the smaller the DIC, the ‘better’ the model
4 If there are interactions involved, I find it useful to use the customised predictions procedure (see Manual Supplement) to plot the predicted result and put confidence intervals around the estimated relations
You can download this from Research gate which considers testing and storage in Chapter 6, customised predictions with interactions in Chapter 8 and the DIC in Chapter 10 (I have not updated it yet to include Bayesian p values)
Kelvyn Jones, SV Subramanian (2014) Developing multilevel models for analysing contextuality, heterogeneity and change using MLwiN, Volume 1
https://www.researchgate.net/publicatio ... ev=prf_pub
Hope this helps