Difference in P value in sjplot::tab_model() vs lmer summary()

Question

When I run my model:

BurstEn <- lmer(Burst_energy_norm_ms ~ StopType + (StopType | Speaker), data = AllData)

I get the following result:

Linear mixed model fit by REML. t-tests use Satterthwaite's method ['lmerModLmerTest']
Formula: Burst_energy_norm_ms ~ StopType + (StopType | Speaker)
   Data: AllData

REML criterion at convergence: 496.7

Scaled residuals: 
    Min      1Q  Median      3Q     Max 
-1.1787 -0.3863 -0.1817  0.0391  8.7832 

Random effects:
 Groups   Name         Variance Std.Dev. Corr 
 Speaker  (Intercept)  0.03533  0.18797       
          StopType+A-F 0.00981  0.09905  -1.00
 Residual              0.34991  0.59153       
Number of obs: 270, groups:  Speaker, 5

Fixed effects:
             Estimate Std. Error       df t value Pr(>|t|)  
(Intercept)   0.27765    0.09213  4.14476   3.014   0.0376 *
StopType+A-F -0.19928    0.08678  8.52753  -2.297   0.0488 *
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Correlation of Fixed Effects:
            (Intr)
StopTyp+A-F -0.407
optimizer (nloptwrap) convergence code: 0 (OK)
boundary (singular) fit: see help('isSingular')

But if I run sjPlot:: tab_model(BurstEn)

then my p-values change to 0.003 (Intercept) and 0.022 (StopType+A-F).

Is there a reason for this? I don't fully understand the documentation.

I tried reading documentation on this, but I didn't really understand if the p-values were computed differently or if I just misunderstood how this works.

Answer 1

These p-values cannot be derived exactly for coefficient estimates of a linear mixed-effects model because the degrees of freedom cannot be derived exactly. Thus, you need to approximate the degrees of freedom and by default the lmerTest package uses Satterthwaite's method as the output tells you (cf. https://www.jstatsoft.org/article/view/v082i13).

help("tab_model") indicates (see entry describing the df.method parameter) that you can use different approximations for the degrees of freedom in tab_model. By default a simple Wald test is used because of performance reasons. This is an inferior approach and you should change this default if possible.

library(lmerTest)
fm1 <- lmer(Reaction ~ Days + (Days | Subject), sleepstudy)
summary(fm1)$coef

tab_model(fm1, digits.p = 6, p.style = "scientific", df.method = "satterthwaite")