easystats
make `parameters()` show fixed effects restricted to 0
Sorry if I am missing something, but I can't find a way to include the fixed effects solutions that are set to 0 due to restrictions/constraint. Here is an example where SAS does it:
As far as I can tell, paramters::parameters() (and stats::coef()) will always drop them from the table:
levels(PlantGrowth$group)
#> [1] "ctrl" "trt1" "trt2"
m <- lm(weight ~ group, PlantGrowth)
parameters::parameters(m)
#> Registered S3 method overwritten by 'parameters':
#> method from
#> format.parameters_distribution datawizard
#> Parameter | Coefficient | SE | 95% CI | t(27) | p
#> ------------------------------------------------------------------
#> (Intercept) | 5.03 | 0.20 | [ 4.63, 5.44] | 25.53 | < .001
#> group [trt1] | -0.37 | 0.28 | [-0.94, 0.20] | -1.33 | 0.194
#> group [trt2] | 0.49 | 0.28 | [-0.08, 1.07] | 1.77 | 0.088
#>
#> Uncertainty intervals (equal-tailed) and p values (two-tailed) computed
#> using a Wald t-distribution approximation.
Created on 2022-05-17 by the reprex package (v2.0.1)
Yet, I sometimes want an additional line group [ctrl] with just a 0 for Coefficient and NA for everything else in my parameters table. Is there a way to do this with {parameters}?
I can see the value of an include_reference_level argument to include the reference for factors
So it would show up something like:
#> Parameter | Coefficient | SE | 95% CI | t(27) | p
#> -----------------------------------------------------------------------
#> (Intercept) | 5.03 | 0.20 | [ 4.63, 5.44] | 25.53 | < .001
#> group [ref: ctrl] | | | | |
#> group [trt1] | -0.37 | 0.28 | [-0.94, 0.20] | -1.33 | 0.194
#> group [trt2] | 0.49 | 0.28 | [-0.08, 1.07] | 1.77 | 0.088
#>
I would leave all of columns aside from Parameter blank. Tricky thing might be detecting when there is a reference level (eg, only for treatment or SAS contrasts with an Intercept included)
This would only work for treatment coding, so that would need to be tested.
Personally, I don't see the merit of adding all these 0s?
Well I'd be happy with an include_reference_level argument that leads to all blanks and no 0s, too. (I just found 0s to be intuitive when teaching and I am used to it from SAS.)
Personally, I don't see the merit of adding all these 0s?
This is mainly for completeness. Alternatively, you could add a footnote indicating the reference levels. And it's not that unusual to add the "estimate" (i.e. 0 for linear, or 1 for OR etc.), sometimes there's just "Ref." in the estimate column.
Here's an example of a recent paper:
The idea is to have a table that is completely self-explaining, so you don't need to read the methods section to remember all levels of categorical variables.
But then the intercept(s) are omitted, to have a table of slopes only?
This is less intuitive to me (I would prefer a clear label of the intercept instead), but I can understand why someone would want this (especially if they are accustomed to it).
But then the intercept(s) are omitted, to have a table of slopes only?
No, not necessarily. This is more common in my field, where we're more interested in the strength of the associations instead of the predicted outcome. That's why we often omit the intercept in tables.
I just saw in two other of my recent papers, intercepts are included in the tables ;-)
My (controversial) view: This will add code complexity and convey little information (literally 0s and dots). I understand why we may want to add empty rows for presentation purposes in "finished" regression tables intended for publication, but that's not quite the job of parameters.
parameters is the package that provides regression parameters tables that people display in publications
Haha, yeah, sorry. I guess I always only ever use/see the markdown in console, so I lost sight ;)
Actually, we already do something similar for grouping parameters: https://easystats.github.io/parameters/articles/model_parameters_print.html#group-parameters
Maybe we just add a special option "reference" to that argument that adds the reference to factors?
And maybe allow a subset of factors or combining reference and grouping by, if the argument is given a list with a slot called "reference", the reference formatting is applied to the stated factors?
I found that {broom.helpers} does what I was looking for and {ggally} is making use of that, too:
library(dplyr)
m <- lm(weight ~ group, PlantGrowth)
broom.helpers::tidy_plus_plus(model = m) %>%
select(term, contrasts:conf.high)
#> # A tibble: 3 x 12
#> term contrasts contrasts_type reference_row label n_obs estimate std.error
#> <chr> <chr> <chr> <lgl> <chr> <dbl> <dbl> <dbl>
#> 1 groupct~ contr.tr~ treatment TRUE ctrl 10 0 NA
#> 2 grouptr~ contr.tr~ treatment FALSE trt1 10 -0.371 0.279
#> 3 grouptr~ contr.tr~ treatment FALSE trt2 10 0.494 0.279
#> # ... with 4 more variables: statistic <dbl>, p.value <dbl>, conf.low <dbl>,
#> # conf.high <dbl>
GGally::ggcoef_model(
model = m,
add_reference_rows = TRUE,
categorical_terms_pattern = "{level} (ref: {reference_level})"
)
Created on 2022-07-12 by the reprex package (v2.0.1)
See examples (and maybe further discussion) here: https://github.com/easystats/parameters/pull/902
Use add_reference = TRUE in the print() method.
library(parameters)
data("fish")
m1 <- glmmTMB::glmmTMB(
count ~ child + camper + zg + (1 | ID),
ziformula = ~ child + camper + (1 | persons),
data = fish,
family = glmmTMB::truncated_poisson()
)
print(model_parameters(m1, effects = "fixed"), add_reference = TRUE)
#> # Fixed Effects
#>
#> Parameter | Log-Mean | SE | 95% CI | z | p
#> ---------------------------------------------------------------
#> (Intercept) | 1.41 | 0.18 | [ 1.06, 1.75] | 8.02 | < .001
#> child | -0.53 | 0.12 | [-0.77, -0.29] | -4.40 | < .001
#> camper [0] | 0.00 | | | |
#> camper [1] | 0.58 | 0.10 | [ 0.39, 0.78] | 5.93 | < .001
#> zg | 0.13 | 0.04 | [ 0.05, 0.21] | 3.17 | 0.002
#>
#> # Zero-Inflation
#>
#> Parameter | Log-Odds | SE | 95% CI | z | p
#> ---------------------------------------------------------------
#> (Intercept) | -0.39 | 0.65 | [-1.67, 0.89] | -0.60 | 0.551
#> child | 2.05 | 0.31 | [ 1.45, 2.66] | 6.63 | < .001
#> camper [0] | 0.00 | | | |
#> camper [1] | -1.01 | 0.32 | [-1.64, -0.37] | -3.12 | 0.002
#>
#> Uncertainty intervals (equal-tailed) and p-values (two-tailed) computed
#> using a Wald z-distribution approximation.
#>
#> The model has a log- or logit-link. Consider using `exponentiate =
#> TRUE` to interpret coefficients as ratios.
data(mtcars)
mtcars$gear <- as.factor(mtcars$gear)
m <- glm(vs ~ wt + gear, data = mtcars, family = "binomial")
print(model_parameters(m, exponentiate = TRUE, drop = "(Intercept)"), add_reference = TRUE)
#> Parameter | Odds Ratio | SE | 95% CI | z | p
#> -------------------------------------------------------------
#> wt | 0.07 | 0.09 | [0.00, 0.52] | -2.05 | 0.040
#> gear [3] | 1.00 | | | |
#> gear [4] | 3.21 | 3.98 | [0.27, 41.36] | 0.94 | 0.348
#> gear [5] | 0.03 | 0.07 | [0.00, 1.47] | -1.41 | 0.159
#>
#> Uncertainty intervals (profile-likelihood) and p-values (two-tailed)
#> computed using a Wald z-distribution approximation.
Created on 2023-09-11 with reprex v2.0.2