I'm decently familiar with mixed effects models (MEM), but a colleague recently asked me how it compares to latent growth models (LGM). I did a bit of googling, and it seems that LGM is a variant of structural equation modelling that is applied to circumstances where repeated measures are obtained within each level of at least one random effect, thus making Time a fixed effect in the model. Otherwise, MEM and LGM seem pretty similar (eg. they both permit exploration of different covariance structures, etc).

Am I correct that LGM is conceptually a special case of MEM, or are there differences between the two approaches with respect to their assumptions or capacity to evaluate different types of theories?

  • 2
    $\begingroup$ Terms random effects, fixed effects, latent growth might mean different things in different contexts. Concerning the second one Andrew Gelman had a blog post with examples of several definitions. So it would be great if you provided the links to the definitions of these models. In general I think you are right in your assumption. Time trends are usually treated separately, since the usual assumption that variance of the regressors is bounded does not hold, so you have to show that for time trend this does not actually change anything in terms of model estimation and interpretation. $\endgroup$ – mpiktas Apr 26 '11 at 20:35

LGM can be translated to a MEM and vice versa, so these models are actually the same. I discuss the comparison in the chapter on LGM in my multilevel book, the draft of that chapter is on my homepage at http://www.joophox.net/papers/chap14.pdf

| cite | improve this answer | |
  • $\begingroup$ Thank you for your reply and welcome to our site! (For the reasons why I removed the closing remarks in your reply, please visit our FAQ.) $\endgroup$ – whuber Jul 12 '11 at 13:26

Here is what I found when looking into this topic. I'm not a stats person so I tried to summarise how I understood it using relatively basic concepts :-)

These two frameworks treat “time” differently:

  • MEM requires nested data structures (e.g. students nested within classrooms) and time is treated as an independent variable at the lowest level, and the individual on the second level
  • LGM adopts a latent variable approach and incorporate time via factor loadings (this answer elaborates more on how such factor loadings, or "time scores", work).

This difference leads to different strengths of both frameworks in handling certain data. For example, in MEM framework, it is easy to add more levels (e.g. students nested in classrooms nested in schools), whilst in LGM, it is possible to model measurement error, as well as embed it in a larger path model by combining several growth curves, or by using growth factors as predictors for outcome variables.

However, recent developments have blurred differences between these frameworks, and they were termed by some researchers as the “unequal twin”. Essentially, MEM is a univariate approach, with time points treated as observations of the same variable, whereas LGM a multivariate approach, with each time point treated as a separate variable. The mean and covariance structure of the latent variables in LGM correspond to the fixed and random effects in MEM, making it possible to specify the same model using either framework with identical results.

So rather than considering LGM as a special case of MEM, I see it as a special case of factor analysis model with factor loadings fixed in such a way, that the interpretation of the latent (growth) factors is possible.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.