Asing several questions related to the lectures on day 6


I would like to ask several questions related to the lectures on day 6.

  1. For the lecture " [Biometrical Age-Based Latent Growth Curve Modeling].", there is one point in the last slide which stated that “REALLY IMPORTANT: Know difference between genetic variance estimate and heritability”, I would like to ask whether you could clarify the difference between genetic variance and heritability. I know the formula of heritability which is the proportion of the total phenotypic variance attributed by the genetic variance. What else dfferences do they have?

  2. This question is about the variance and covariance. In LGC for example, we fit latent intercept and latent slope , and could obtain means and variance and covariance for both of them. So variance means that the intercept/slope varies among different individuals, and is there any criteria that we could use to determine how big the variance/covariance is or not? If the variance of intercept is 0.001, could we say that is no variation for the intercept of LGC among different individuals? Another thing is to confirm that, the means and variances of intercept and slope in LGC using SEM are fixed and random effects of intercept and slope in the mixed effect model, is that correct?

Really very nice lectures and thank you so much.



1 Like

@ xueyingqin Great questions!

I’m going to let Mike Neale answer the first question.

I believe determining whether you should include a growth factor (e.g. linear or quadratic) or a variance/covariance (r_lin_quad) is an empirical question. If the mean of a latent growth factor is null, then it is possible to test whether you should include that factor. Similarly if the variance of a factor is small it may not improve the fit of your model when you include it. This can be easily tested with standard model fitting methods, such as likelihood ratio tests. This is relatively common with smaller datasets and/or with higher order factors.

A few caveats:

First, if you don’t have any variance in a factor, you should question whether allowing that factor to covary makes empirical sense.

Relatedly, just because your factor doesn’t vary significantly, does not mean that you need to drop if from your model. You can set the variance and the covariance of the factor to zero and effectively fit a fixed effect. If a factor doesn’t vary and it has a null mean, then it probably isn’t explaining variation in your data.

Finally, yes the SEM version of a latent growth curve could alternatively be fit with other random effect models such as an LMM.

Thank you, Brad.