Covariates and slopes

[burrowsk]

Hi!
I'm having trouble marrying up the slopes to the average curve. The lines from one predicted period to another don't fall along the curve as expected.

The data for the predicted periods and the model fit are the same. I ended up playing with the prediction for the curve by setting my covariate to 0 (my covariate is coded 1/2) and the slopes matched much more closely:

Am I being paranoid here, doesn't the predicted periods used for deriving the slopes include covariates in the predictions? I know we specify "filter" in run_eggla_lmm() but I'm having trouble seeing how this is incorporated into predicting the slopes?

Thanks!

[mcanouil]

Short answer, the slopes are correct.

Long answer, currently in the function to draw the plots, the average curve is "independent"
Of the covariates while for the slopes, they are predicted from the mixed model.
For instance, you can see the formula to compute "y": https://github.com/mcanouil/eggla/blob/4538ff2d53424f94cca44f188c1006c952fd4929/R/plot_egg_slopes.R#LL70C9-L70C9

As you can see, it does not account for any covariates that could have been included.

For the figure to have curve and slopes to match, we would need to generalise that formula to include covariates effects.

If you feel you can make a PR, please do! 😉

[burrowsk]

Thanks very much! I thought about how the "predicted periods" in compute_slopes():
https://github.com/mcanouil/eggla/blob/4538ff2d53424f94cca44f188c1006c952fd4929/R/compute_slopes.R#LL104C1-L123C4
uses the coef (fixed + random effects) which of course is adjusted for covariates. Unfortunately, that popped into my head at 3:30 this morning :(

I didn't use the plot_slopes() function though you raise a good point about the curve not accounting for covariates so I can see how that would affect them matching up. However I predicted the curve using:

res_f <-readRDS("female-model-object.rds")
pred_f <- expand.grid(egg_ageyears = seq(min(res_f[["data"]][["egg_ageyears"]])+0.05,max(res_f[["data"]][["egg_ageyears"]])-0.05,0.01))
pred_f$sourceb <- 1
pred_f$bmi <-exp(predict(res_f,pred_f,level =0))

sourceb being my covariate. This is what goes towards creating the first plot where the slopes and curve do not match. However when I change to pred_f$sourceb <- 0 (sourceb is a binary variable of 1/2) it is much closer?

Thanks!

[mcanouil]

As done internally, it uses predict, but the coordinates for the segments representing the slopes are roughly defined by pred_period + end * slope which here use adjusted "end"/"x", thus does not have the fixed effect of any covariates, which lead to the vertical shift of the segments.
For an interval, we have:

• "pred_period", the corrected y value at the start of the interval.
• "end", the x value at the at the end of the interval.
• "slope" is the predicted slope in the interval.

We don't keep covariates effects in the results, thus, I don't see how from the slopes and predicted values we can put them back.

Note that this is only a visualisation issue.
Also, for the model we defined, the functions include "egg" in their names, such as egg_slopes() and plot_egg_slopes().

[mcanouil]

I could try to extract covariates effects and add them to the equation, to ensure the visualisation accounts for those, but I'll have to think about how to do that in a general way (interaction terms, etc.).

[burrowsk]

Sorry for lack of reply!, I'm out of action for the next two days but will pick this up on Friday :)

[mcanouil]

No worries, I will be off as well (theoretically).