Hello, I am new to growth modeling and my questions may sound rather too basic. I am running a linear growth model, as in example 6.1, except I only have three non-equidistant time points and, in addition, I am regressing the intercept (i) and slope (s) on an outcome y. When I run the model without covariates (for the regression part of it, ie y ON i s) only the intercept is significant; however, when I include covariates (in the regression, eg y ON i s x1 x2), the estimates for the slope and intercept as such change and the slope becomes significant. I am not quite sure I understand why. Does it mean i and s are conditional upon these covariates? Many thanks
The answer is yes - just like in regular regression the regression coefficients are partical coefficients (the effect of one variable holding other constant). Note also that when you include covariates the model may not fit well even if the unconditional model did fit well, in which case the y on i s x1 x2 estimates are not trustworthy.
Jungmeen Kim posted on Wednesday, January 21, 2009 - 9:45 am
Dear Linda and Bengt,
I have tested conditional growth curve models and as I interpret the results, I have got to question whether the "prevalence" or the mean degree of the covariates affect their "power" to influecne the growth factors. For instance, I have emotio(SA) as covariates. In my sample of maltreated children, there are a lot of children experienced EM whereas only a few children experienced SA. I assigned 0 = no experience and 1~5 as severity of the subtype (EM or SA) The results showed significant effects of EM but not SA. Is there a statistically based reason to expect that even though the impact of SA on an individual should be severe(in terms of its effects on mental health) than EM, my data were not "powerful" enough to detect the SA effect due to the small cell size of children with SA experience?
You words of wisdom and advice will be greatly appreciated. Thank you!
Thank you Bengt, for your guidance for lit search!
csulliva posted on Friday, October 02, 2009 - 4:41 pm
Is it possible to include both a time stable and time varying component of the same covariate in a GMM or LCGA model? This model would condition the intercept and slope on a deviation score at each point in time, but also include the covariate's mean across those time points as a predictor of the latent class variable. This parameterization would be similar to one sometimes used pooled time series random effects models.