0

in the model below, can I say that 2021 G2 is not significant because it is calculated from a non-significant interaction ?

  • These are my variables:
CONT_Y = scores 
YEAR = A/B (2020 or 2021)
MY_GROUP = test1 (G1) or test2 (G2)
  • I have this model: Y ~ MY_GROUP * YEAR
modInteraction <- lmer(Y ~  MY_GROUP * YEAR + (1|PARTICIPANTS), data = data, REML = FALSE)
confint(boot_modInteraction, type = "norm")
# A tibble: 4 x 6
  term                estimate  lower  upper type  level
  <chr>                  <dbl>  <dbl>  <dbl> <chr> <dbl>
1 (Intercept)           17.6   16.7   18.5   norm   0.95 sig
2 GROUPG2               0.915  0.117  1.69  norm   0.95    sig
3 YEAR2021              1.14   0.180  2.13  norm   0.95 sig
4 GROUPL2:YEAR2021     -0.602 -1.77   0.572 norm   0.95    ns
  • From which I get:

G1 2020 = the intercept = 17.6
G1 2021 = B0 + B2 = 17.6 + 1.14 = 18.74
G2 2020 = B0 + B1 = 17.6 + 0.915 = 18.51
G2 2021 = B0 + B1 + B2 + B3 = 17.6 + 0.915 + 1.14 + -0.60 = 19.05

I know from paired t-tests that:

test A) G1 2020 < G2 2020 sig
test B) G1 2020 < G1 2021 sig
test C) G1 2021 < G2 2021 ns
test D) G2 2020 < G2 2021 ns

  • Questions:

  • Q1: Is it ok to say that tests C and D are not significant because G2 2021 score derived from a non-significant interaction ? Thanks.

Obs: more details of the model and the dataset can be seen here

Larissa Cury
  • 741
  • 4
    To compare levels of your terms, you probably want to use the emmeans() function in the emmeans package. Likewise, I imagine you want to get the p-value for the effects from an anova-like table. For this you could use the joint_tests() function from the emmeans package, although it's probably more common to use the lmerTest package. – Sal Mangiafico Oct 15 '22 at 19:36
  • hi, thank you, shouldn't I be able to get the four results from the 4 t-tests only by interpreting the regression coefficients and significance of the betas? I mean, I've never seen a "post hoc" for regression...so, is it necessary, then? – Larissa Cury Oct 15 '22 at 19:39
  • 1
    I assume you are using t-tests conducted on the raw data and not on estimated values derived from the fitted model. Rather than doing this, I think you'd be better off comparing the estimated marginal means, which are estimated from the model you are fitting. Why fit a model, and then ignore it when comparing groups ? – Sal Mangiafico Oct 15 '22 at 21:12
  • 1
    The way I would unpack the results: the interaction being non significant means that the effect of year on outcome does not differ between groups, and the effect of group does not differ between years. Thus, after getting the non-significant interaction, you are only "warranted" to look at the main effects of group and year more closely. – Sointu Oct 18 '22 at 07:28

0 Answers0