Why the variation between CIs computed with emmeans and confint are so high?

Question

I have some strangeness going on when calculating CIs for my model. I've found 2 ways to do this, and I decided to try them both (emmeans and confint).

I have 6 timepoint and 2 groups. My model looks something like this:

model<-lmer("responseVAR~factor(timepoint)*factor(group) + (timepoint|subject) + sex", data)

This is what I get with emmeans:

emmeans(model, "timepoint", "group")

group = 1:
 timepoint emmean    SE    df lower.CL upper.CL
        1   8.38 0.486 110.8     7.42     9.34
        2   6.91 0.536 154.8     5.85     7.97
        3   5.75 0.581 163.8     4.60     6.90
        4   5.21 0.656 156.5     3.92     6.51
        5   4.74 0.692 113.4     3.37     6.11
        6   5.16 0.966 111.4     3.24     7.07

group = 2:
 timepoint emmean    SE    df lower.CL upper.CL
        1   8.28 0.692 106.5     6.91     9.66
        2   7.85 0.799 168.8     6.27     9.42
        3   7.08 0.832 167.9     5.44     8.72
        4   6.84 0.921 159.0     5.02     8.66
        5   6.84 0.887  92.5     5.08     8.60
        6   4.14 1.219 111.2     1.72     6.56

And this is what I get with confint

confint(model)
                                       2.5 %     97.5 %
.sig01                             0.99893416  2.9788424
.sig02                            -0.81362432  1.0000000
.sig03                             0.03442704  0.7801731
.sigma                             2.01005448  2.570744
(Intercept)                        4.77889981  8.6825788
factor(timepoint)2                -2.48547074 -0.4602945
factor(timepoint)3                -3.72496655 -1.5462229
factor(timepoint)4                -4.40019534 -1.9419390
factor(timepoint)5                -4.92791619 -2.3636326
factor(timepoint)6                -5.02534365 -1.3708410
factor(group)2                    -1.60598418  1.4010876
Sex                               -0.35709146  2.5673636
factor(timepoint)2:factor(group)2 -0.78898029  2.8635743
factor(timepoint)3:factor(group)2 -0.47127777  3.3355089
factor(timepoint)4:factor(group)2 -0.38316365  3.8446059
factor(timepoint)5:factor(group)2  0.11266374  4.3063505
factor(timepoint)6:factor(group)2 -3.91956691  1.9927936

Also, adding covariates (e.g. Sex) seems to affect a lot the results of the confint method because lower and upper confidence levels get very far apart. emmeans results seem to not vary that much.

What method should I trust more? I would have used emmeans, but I didn't find how to calculate CIs for the covariates.

How can I calculate CIs for covariates with emmeans?

Answer 1

You are comparing apples and oranges. The emmeans() results are predictions from the model, or averages thereof. The confint() results are estimates of the fixed-effect regression coefficients.

In notational terms, you have a model that says something like

$$ y = \beta_0 + \beta_1 x_1 + \beta_2 x_2 + \ldots + e $$

Your model is a lot more sophisticated than this, but this is the basic idea. The emmeans() results are estimates of predictions for $y$ at various combinations of the $x_j$, and the confint() results are estimates of $\beta_0, \beta_1, \beta_2, \ldots$.