1
$\begingroup$

I have data from 3 ethnicities and two genders. One of my supervisors wants me to stratify the data and train models separately for each group then compare them with t-tests. My other supervisor says that it is more statistically efficient to train one model and include the groups as factors and study the interactions with predictors. There are many predictors but just to keep things simple I've only included "time spent in front of a screen during weekdays" as an example (weekdayscreen). We are using a mixed effects model the rest of the terms are just nuisance variables we need to control for.

    fml <- lmer(temp[,306]~1+(1|mri_info_deviceserialnumber) +   #this models random effect for site
 (1|mri_info_deviceserialnumber:rel_family_id)+#this  models  
   #random effect by nesting family
                 fml <- lmer(temp[,306] ~ 1 + 
     (1|mri_info_deviceserialnumber) +  
#this models random effect for site
             (1|mri_info_deviceserialnumber:rel_family_id) +
#this  models random effect by nesting family
            interview_age + rel_relationship + ehi_y_ss_scoreb + 
            scale(smri_vol_scs_intracranialv)+anthroheightcalc +
            sex + race_ethnicity +
#terms of interest start here, above are controlledttg variables
            weekdayscreen + sex:weekdayscreen +
            race_ethnicity:weekdayscreen +    sex:race_ethnicity:weekdayscreen,
          data = temp)

One things I am confused by is how I am supposed to get the reference levels of the fixed effects out of the model to report their coefficients. If I run coef(summary(fml)) I just get the estimates of the males and the ethnic groups 1 and 2 not for group 1.

> coef(summary(fml))
                                       Estimate Std. Error        df     t value  Pr(>|t|)
(Intercept)                        4174.2811183 71.2052162 7860.7811 58.62324896 0.0000000
interview_age                        -0.7325074  0.5304626 8902.8100 -1.38088408 0.1673493
rel_relationship                     -2.9555884  4.6849766  220.2807 -0.63086513 0.5287827
ehi_y_ss_scoreb                      -1.9601947  5.0425805 9049.0643 -0.38872849 0.6974861
scale(smri_vol_scs_intracranialv)   280.7137489  3.5027028 9072.5860 80.14204098 0.0000000
anthroheightcalc                      0.3422240  1.2843885 8991.9368  0.26644902 0.7898995
sexM                                 -1.4189458 16.5856861 9077.9730 -0.08555243 0.9318241
race_ethnicity2                       9.8555099 26.2152289 8434.3119  0.37594598 0.7069665
race_ethnicity3                      15.7084886 20.8602725 5236.3528  0.75303372 0.4514635
weekdayscreen                         6.2736929  5.9826176 8972.3845  1.04865351 0.2943659
sexM:weekdayscreen                   -6.1093124  7.4598324 9077.8342 -0.81896107 0.4128301
race_ethnicity2:weekdayscreen        -5.2071097  9.6925825 9061.5719 -0.53722624 0.5911245
race_ethnicity3:weekdayscreen        -9.5271568  8.7451980 9070.2193 -1.08941579 0.2759995
sexM:race_ethnicity2:weekdayscreen   -1.4281289  6.9047424 9077.8507 -0.20683305 0.8361449
sexM:race_ethnicity3:weekdayscreen   10.5974174  6.5771912 9077.9033  1.61123754 0.1071627

So for the interaction terms how am I supposed to know what the estimate for interaction terms for females in group 1 are?

$\endgroup$

1 Answer 1

1
$\begingroup$

Your other supervisor is correct. Splitting the data and doing multiple t tests is a bad idea as you lose statistical power and also encounter multiple testing issues.

One things I am confused by is how I am supposed to get the reference levels of the fixed effects out of the model to report their coefficients. If I run coef(summary(fml)) I just get the estimates of the males and the ethnic groups 1 and 2 not for group 1.

The reference levels are included in the intercept. So the individual estimates are contrasts between the estimate in question, and the reference level.

$\endgroup$
2
  • $\begingroup$ is it more efficient because we just do less testing? What causes the loss of power? $\endgroup$ Jul 17, 2021 at 3:14
  • 1
    $\begingroup$ Sorry only just noticed this comment. The loss of power is caused by the decrease in sample size. Power increases with sample size. $\endgroup$ Aug 17, 2021 at 11:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.