I would like to direct expert's attention to this question: stats.stackexchange.com/questions/123997/…

It should be relatively easy to answer. If not, please let me know and I will improve it. I would need "a peer-review" for the method I "made up" before submitting a paper.

It makes sense @Mikko, and those are one type of standardized effect size. For a reference off-hand Long talks about them in Regression Models for Categorical and Limited Dependent Variables - but I'm sure references exist for linear models.

There are many ways to standardize effect sizes though. For two examples see stat.columbia.edu/~gelman/research/published/ape17.pdf and stat.columbia.edu/~gelman/research/unpublished/…

@AndyW

thanks!

Now that I post those examples Gelman and Hill's book probably has discussion of the standardized effect sizes, but I don't have it at my current desk to check for sure.

Also your pred3 panels in the first plot look a little strange. Are there some 0 observations mixed in that are hard to see? It would be easier to see their distributions if you zoomed into the range around 30+.

I typically default to line charts for such things, I don't like multiple bars like that unless it is for a histogram.

True. I could do that (and certainly have done it for the paper). The plot in that question is just for overview to see the data (bars are to illustrate that some variables have gaps). In the paper I used bars for the growth rate and lines for predictor variables.

To answer your question: there are no 0 observations in data, but there are missing weeks for some of the individuals. In the example the missing weeks are at the beginning and end of the time series

So the model formulations make the correct contrasts between predictor variables? i.e. running the model separately for each predictor variable for point 1 is a reasonable approach? and removing the intercept and using random effects intercepts in point 2 makes sense too?

Sorry was at a meeting for the last few minutes - what is the motivation for removing the intercept? You can calculate the standardized effect sizes with it included.

I kind of understood that the intercept should be removed when calculating effects (based on this question: stats.stackexchange.com/questions/117641/…). Anyways, the random effects will make intercepts. There is one more term in the model, which could change the effects / contrasts

Your talk about contrasts only makes sense if pred1, pred2 and pred3 were categorical variables, which they aren't in your example.

Ok, so I should leave the intercept.

Thanks.

Russ's suggestion in your second linked question is the only case I have encountered where it makes sense - when the model includes all the categorical dummy variables for one category.

It seems that including the intercept makes no large difference at least for my data. The CIs are about 1-2 % wider with intercept

@whuber, thanks! I didn't know about the other statisticians. Pressumably Roger Koenker stayed in memory as an outlier. It's a shame that he doesn't visit regularly. There are many more questions on quantile regressions that could do with his input.