`estimate_contrasts()` problem with p-value adjustments

[andieich]

I would like to use the modelbasedpackage for an introductory statistics class. To show post-hoc tests, I am using the estimate_contrasts() function.
When I keep the default setting, i.e. no p-value adjustment and run the following test:

library(palmerpenguins)
library(modelbased)
m_spec <- lm(body_mass_g ~ species, data = penguins)
estimate_contrasts(m_spec)

I get:

We selected `contrast=c("species")`.
Marginal Contrasts Analysis
Level1    | Level2    | Difference |    SE |             95% CI | t(330) |      p
---------------------------------------------------------------------------------
Chinstrap | Adelie    |      26.92 | 67.65 | [-106.16,  160.01] |   0.40 |  0.691
Gentoo    | Adelie    |    1386.27 | 56.91 | [1274.32, 1498.22] |  24.36 | < .001
Gentoo    | Chinstrap |    1359.35 | 70.05 | [1221.55, 1497.15] |  19.41 | < .001

Variable predicted: body_mass_g
Predictors contrasted: species
p-values are uncorrected.

When I use Tukey adjustments:

estimate_contrasts(m_spec, p_adjust = "tukey")

I get:

We selected `contrast=c("species")`.
Marginal Contrasts Analysis

Level1    | Level2    | Difference |    SE |             95% CI | t(330) |      p
---------------------------------------------------------------------------------
Chinstrap | Adelie    |      26.92 | 67.65 | [-106.16,  160.01] |   0.40 | < .001
Gentoo    | Adelie    |    1386.27 | 56.91 | [1274.32, 1498.22] |  24.36 |  0.691
Gentoo    | Chinstrap |    1359.35 | 70.05 | [1221.55, 1497.15] |  19.41 | < .001

Variable predicted: body_mass_g
Predictors contrasted: species
p-value adjustment method: Tukey

Why are the results so different? Why are the p-values exactly the same, but for different comparisons? How can a t-value of 0.40 (Chinstrap vs. Adelie with Tukey adjustment) lead to a p-value of < .001?

packageVersion("modelbased")
[1] ‘0.13.1’

[andieich]

I noticed that this bug only occurs when I use the default backend = "marginaleffects". With backend = "emmeans", the p-values seem correct.

[strengejacke]

Thanks for reporting. Should be fixed now