I was just thinking this morning about how it has been a hot minute since you dropped a vid. Glad you realized that too and dropped one. Commented even before I watched it. I was also enjoying that maximum likelihood series you were doing. That guest lecturer you had dumbed it down perfectly for my 9 year old brain. Now I feel ready for more.
I like the advice to subsample the random effects for visualizations, especially when the number of random effects is large. Another option is to plot a histogram of the estimated random effects for each parameter. Taking random slopes for example. If the distribution of estimated random slopes is narrow then that means that the estimated random slopes have a tendency to be 'close' to parallel. Histograms come with the problems of bin size selection and binning error, but when you are working with datasets with a lot of labels you can have sometimes have tens of thousands of random effects (I have experienced this). This histogram visualization approach may also help accentuate multimodality (e.g. a tendency for there to be two slopes that most of the random slopes are close to, which can indicate a latent structure to be modelled).
Great content as always! Really enjoyed your simplistics classes in the past year, but wondering if you plan on publishing any more content on random forests. In my own work and working with graduate students, I'm finding it to be the rule now that exploratory an RF methods are a much better fit for questions a lot of psych people pose. Your paper is great on this as well, but some questions remain (e.g., quantitative methods for choosing important variables, use of multiple outcomes and repeating the procedures over and over, sample size requirements for it if any etc.). Any plans to do more content on EDA methods?
@@QuantPsych amazing! Thank you for considering more content on this. If you do have the time and space for it, I think a very specific issue that I often come up against is when students have some kind of group comparison either in a clinical or experimental study and they run a T test or anova And they just repeat it for many different outcomes to see what the groups differ on. So if I suggest an exploratory method you end up having a random forest procedure with one categorical predictor, and many outcomes. Sometimes you also have covariates but in this particular scenario where you only really care about a single variable, it would be great to see your take on how that changes the application and interpretation of the variable importance indicators, as well as the idea of repeating a procedure like this for multiple outcomes rather than having many predictors that you’re trying to reduce. The issue of sample size requirements, if any, for an exploratory procedure like this also comes up often.
This could not have come at a better time - thank you, Dr. Fife!
I was just thinking this morning about how it has been a hot minute since you dropped a vid. Glad you realized that too and dropped one. Commented even before I watched it. I was also enjoying that maximum likelihood series you were doing. That guest lecturer you had dumbed it down perfectly for my 9 year old brain. Now I feel ready for more.
I like the advice to subsample the random effects for visualizations, especially when the number of random effects is large.
Another option is to plot a histogram of the estimated random effects for each parameter. Taking random slopes for example. If the distribution of estimated random slopes is narrow then that means that the estimated random slopes have a tendency to be 'close' to parallel. Histograms come with the problems of bin size selection and binning error, but when you are working with datasets with a lot of labels you can have sometimes have tens of thousands of random effects (I have experienced this). This histogram visualization approach may also help accentuate multimodality (e.g. a tendency for there to be two slopes that most of the random slopes are close to, which can indicate a latent structure to be modelled).
Good idea!
Great content as always! Really enjoyed your simplistics classes in the past year, but wondering if you plan on publishing any more content on random forests. In my own work and working with graduate students, I'm finding it to be the rule now that exploratory an RF methods are a much better fit for questions a lot of psych people pose. Your paper is great on this as well, but some questions remain (e.g., quantitative methods for choosing important variables, use of multiple outcomes and repeating the procedures over and over, sample size requirements for it if any etc.). Any plans to do more content on EDA methods?
There is now :) I love that idea!
@@QuantPsych amazing! Thank you for considering more content on this. If you do have the time and space for it, I think a very specific issue that I often come up against is when students have some kind of group comparison either in a clinical or experimental study and they run a T test or anova And they just repeat it for many different outcomes to see what the groups differ on. So if I suggest an exploratory method you end up having a random forest procedure with one categorical predictor, and many outcomes. Sometimes you also have covariates but in this particular scenario where you only really care about a single variable, it would be great to see your take on how that changes the application and interpretation of the variable importance indicators, as well as the idea of repeating a procedure like this for multiple outcomes rather than having many predictors that you’re trying to reduce. The issue of sample size requirements, if any, for an exploratory procedure like this also comes up often.
Now if i could just visualise the random effects of my GENERALISED linear mixed effects model with flexplot....that would be great!
I'm pretty sure you can....maybe?