get_predicted_ci for mlm models

[mattansb]

trafficstars

m <- lm(cbind(mpg, hp, drat) ~ factor(cyl) + wt, data = mtcars)

insight::get_predicted_ci(m, predict(m))
#> Error in mm %*% vcovmat: non-conformable arguments

(Once this is fixed, this can also be applied to afex_aov models. #400)

[DominiqueMakowski]

m <- lm(cbind(mpg, hp, drat) ~ factor(cyl) + wt, data = mtcars)
predict(m)
#>                          mpg        hp     drat
#> Mazda RX4           21.33650 121.76515 3.701569
#> Mazda RX4 Wag       20.51907 122.03216 3.642144
#> Datsun 710          26.55377  82.67225 4.062922
#> Hornet 4 Drive      19.42916 122.38818 3.562909
#> Hornet Sportabout   16.89262 208.62872 3.359606
#> Valiant             18.64379 122.64473 3.505814
#> Duster 360          16.47590 208.76485 3.329311
#> Merc 240D           23.76489  83.58325 3.860176
#> Merc 230            23.89311  83.54136 3.869497
#> Merc 280            18.70790 122.62378 3.510475
#> Merc 280C           18.70790 122.62378 3.510475
#> Merc 450SE          14.87309 209.28841 3.212790
#> Merc 450SL          15.96300 208.93239 3.292024
#> Merc 450SLC         15.80272 208.98474 3.280372
#> Cadillac Fleetwood  11.09046 210.52401 2.937800
#> Lincoln Continental 10.53269 210.70621 2.897251
#> Chrysler Imperial   10.78593 210.62348 2.915661
#> Fiat 128            26.93844  82.54660 4.090887
#> Honda Civic         28.81373  81.93403 4.227217
#> Toyota Corolla      28.10849  82.16440 4.175947
#> Toyota Corona       26.08896  82.82408 4.029131
#> Dodge Challenger    16.63618 208.71249 3.340963
#> AMC Javelin         16.90865 208.62349 3.360771
#> Camaro Z28          15.61038 209.04757 3.266389
#> Pontiac Firebird    15.59435 209.05280 3.265224
#> Fiat X1-9           27.78793  82.26911 4.152643
#> Porsche 914-2       27.13078  82.48377 4.104870
#> Lotus Europa        29.14070  81.82723 4.250987
#> Ford Pantera L      17.75814 208.34600 3.422527
#> Ferrari Dino        20.85566 121.92221 3.666613
#> Maserati Bora       16.47590 208.76485 3.329311
#> Volvo 142E          25.07919  83.15393 3.955723

^{Created on 2021-07-18 by the reprex package (v2.0.0)}

what should we do in this case? return one prediction vector for each response?

[DominiqueMakowski]

Now that get_predicted_ci is a method (get_predicted_ci.default), we could perhaps sapply it on mlm objects? I have no idea how mlm objects work

[mattansb]

I have no answers for you, just more questions...

[bwiernik]

Where do we use get_predicted_ci() internally?

[DominiqueMakowski]

https://github.com/easystats/insight/blob/34af319fa233d6611fa7e40013f899343a558293/R/get_predicted.R#L189-L194

https://github.com/easystats/insight/blob/34af319fa233d6611fa7e40013f899343a558293/R/get_predicted.R#L262

As a matter of fact I consider get_predicted_ci almost as an semi-internal function, since I don't expect much users to use it directly (unless they want to compute different CIs for the same predictions without re-running predictions or some specific other specific usecases)

[bwiernik]

Yeah, that was my understanding of it. I like that it's exported for that case (say to visualize 50, 80, 95, 99 intervals).

I see 3 options for a structure here:

• array--nope
• list of data frames, one for each response, each with same structure as a univariate response
• stacked data frame with an added response column

Which of the last two do you think would be most intuitive and useful for both us internally and users potentially using?

[mattansb]

3 can easily be turned into 2 w/ split, but you'd also need to have an index column for getting each multivariate prediction... Errrrr

[bwiernik]

What do you mean by index there (versus response)?

[bwiernik]

Oh you mean if keep_iter is TRUE? Yeah. That wouldn't be too bad?

[mattansb]

I don't know what keep_iter is... But I mean that each prediction is a multi-dimensional point:

m <- lm(cbind(mpg, hp) ~ cyl, mtcars)
predict(m) |> head()
#>                        mpg        hp
#> Mazda RX4         20.62984 140.69532
#> Mazda RX4 Wag     20.62984 140.69532
#> Datsun 710        26.38142  76.77876
#> Hornet 4 Drive    20.62984 140.69532
#> Hornet Sportabout 14.87826 204.61188
#> Valiant           20.62984 140.69532

^{Created on 2021-07-19 by the reprex package (v2.0.0)}

So if we re-shape this, we need to be able to tell which values from y1 go with which values from y2 ... yk.

[bwiernik]

I'm leaning toward a list, then we don't need to add any sort of index column, and each element can just be a typical univariate predict_ci object with appropriate classes.

[mattansb]

The question is, how will other functions that work with the results deal with a list?

[bwiernik]

Give it its own class, whose method is "lapply the generic over the list". We would need a new class for either approach