Good timing! I'm in the process of making a series on Mixed Models. It won't be specific to fMRI, but more for using lmer in R. Hope that is helpful in answering your question! I'm specifically covering how things are handled in lmer when subjects have different amounts of data.
@@mumfordbrainstats Awesome. In general, I'm really struggling with when to use nlme vs lme4 for this issue, as the former is the only one that has the ability to specify AR1. And yet I see a bunch of people doing LMER models on this type of data, presumably ignoring the potential for autoregressive covariance structures. It's really perplexing . Thanks again. Your videos are so great as a resource.
I noticed that these examples and one early in your mixed models series set REML to false. Will you be covering this in the mixed model series? I found differences in convergence as a function of whether or not I set REML to false and am curious how this is affecting the model.
@@laurenatlas Thanks, I'll try to figure out why that might happen. REML focuses on estimating the variance parameters while ML is focused on the ML. I need to double check, but I believe REML provides unbiased estimates of the variance paramters while ML provides unbiased fixed effects. Matuschek used REML=FALSE, which is why I did here, but it is interesting, because the focus is on testing the random effects...so it seems backwards. Either way, I'll work on something to explain REML vs ML.
Would you consider doing a video on repeated measures lme with heterogeneous time between time points by individuals
Good timing! I'm in the process of making a series on Mixed Models. It won't be specific to fMRI, but more for using lmer in R. Hope that is helpful in answering your question! I'm specifically covering how things are handled in lmer when subjects have different amounts of data.
@@mumfordbrainstats Awesome. In general, I'm really struggling with when to use nlme vs lme4 for this issue, as the former is the only one that has the ability to specify AR1. And yet I see a bunch of people doing LMER models on this type of data, presumably ignoring the potential for autoregressive covariance structures. It's really perplexing . Thanks again. Your videos are so great as a resource.
I noticed that these examples and one early in your mixed models series set REML to false. Will you be covering this in the mixed model series? I found differences in convergence as a function of whether or not I set REML to false and am curious how this is affecting the model.
I will add this to the list! Do you recall when the convergence was better? REML or ML?
@@mumfordbrainstats Yes - it wouldn't converge when REML was included (i.e. it converged when I set REML to false)
@@laurenatlas Thanks, I'll try to figure out why that might happen. REML focuses on estimating the variance parameters while ML is focused on the ML. I need to double check, but I believe REML provides unbiased estimates of the variance paramters while ML provides unbiased fixed effects. Matuschek used REML=FALSE, which is why I did here, but it is interesting, because the focus is on testing the random effects...so it seems backwards. Either way, I'll work on something to explain REML vs ML.
Can you please do a video on cost functions?
What type of cost functions?