Jas Wer posted on Monday, August 17, 2015 - 5:46 am
I'm fitting an ACE twin model, similar to User's Guide example 5.21. The syntax is:
MODEL: [y1-y2] (1); y1-y2 (var); y1 WITH y2 (covmz); MODEL dz: y1 WITH y2 (covdz);
MODEL CONSTRAINT: NEW(a c e h s n); var = a**2 + c**2 + e**2; covmz = a**2 + c**2; covdz = 0.5*a**2 + c**2; h = a**2/(a**2 + c**2 + e**2); s = c**2/(a**2 + c**2 + e**2); n = e**2/(a**2 + c**2 + e**2);
I'm using this with the MLR estimator, because my variable is very skewed. It all runs fine and the estimates match the prediction based on twin correlations. However, for one of the estimates, the lower CI limit is negative. This is theoretically impossible as there can't be a negative genetic or environmental influence. Is this a sign of a problem in the model or just the result of Mplus providing symmetrical intervals? Can anything be done about this, e.g. specify that the lower CI limit is zero in this case?
It could be the results of symmetric confidence intervals. You can try bootstrapping and Cinterval(bootstrap). Or, use Bayes which offers non-symmetric posterior distributions of the estimates.
Jas Wer posted on Tuesday, August 18, 2015 - 8:14 am
Thank you for your advice. I tried bootstrapping with Cinterval(bootstrap) and it worked. The finding is the same (estimate non-significant) and the lower CI limit is now zero. I also wanted to use Bayes to get a feel for the range of results but I can't seem to use this with a multiple group model (my groups are MZ and DZ twins).
Now I'm just not sure whether the approach of using bootstrapping instead of the MLR estimator is theoretically correct. My sample is large (approx. pairs each of 500 MZ and DZ pairs) and I've only read of bootstrap being used to obtain more precise SEs in small samples. Is it justified to use bootstrapping to take into account non-normal distributions, i.e. is this a useful alternative to using the MLR estimator?
Bootstrapping is not only for small samples, but also for non-normal distributions of the estimates. If your estimate (not outcome) distribution is non-normal, bootstrap CIs is the way to go. You can tell how non-symmetric the BS CI is by computing the differences from the mean to the limits. If quite non-symmetric, the ML(R) CI is not reliable.
Bayes multiple-group analysis is done via Type=Mixture Knownclass.
Jas Wer posted on Thursday, August 20, 2015 - 6:40 am
Brilliant, thank you so much. Based on your advice I found that the BS CIs were quite non-symmetric, so I will go ahead and use these instead of using the MLR CIs.
One last question: what is the rule of thumb of how many bootstrap samples to use? The advice I have found so far was quite varied so I'm wondering whether you are able to recommend a paper or have a general rule.
We have some references in the user's guide which may address this. You could try 1000 and then do less and more to see if it changes anything.
Jas Wer posted on Thursday, August 20, 2015 - 9:27 am
Thank you so much. I just realised that this is also the answer to another problem I had, when conducting a bivariate Cholesky decomposition (Linda, I emailed you about this ages ago). So useful. Thank you again.