3
$\begingroup$

Having read through a few posts, I still couldn't find an answer to my question.

I'm trying to investigate for the effect of covariate C on a longitudinal dataset. I have two linear mixed effect models given below:

A.lme <- lme( A ~ B + C, data = data1, random = ~ 1 | id)

B.lme <- lme( A ~ B*C, data = data1, random = ~ 1 | id)

I just want to be sure that I'm interpreting these two the right way. I believe in order to investigate for covariate C, I should be analysing B.lme.

B represents time whilst A represents immune cells in the body whilst C represents a viral infection status.

The summary and anova for B.lme suggests that C has no significant effect on both intercept and slope as given below:

>summary(B.lme)
Linear mixed-effects model fit by REML
Data: data1 
   AIC      BIC    logLik
4238.806 4270.106 -2113.403

Random effects:
Formula: ~1 | id
         (Intercept)  Residual
StdDev:   0.9242001 0.9692625

Fixed effects: A ~ C + B + C:B 
              Value     Std.Error   DF   t-value p-value
(Intercept)  -3.0675750 0.6212136 1118 -4.938036  0.0000
C             0.7364624 0.6264595  244  1.175595  0.2409
B             0.2200117 0.1988966 1118  1.106161  0.2689
C:B           0.0131436 0.2000672 1118  0.065696  0.9476
Correlation: 
             (Intr)    C       B   
C            -0.992              
B            -0.849  0.842       
C:B           0.844 -0.844 -0.994

Standardized Within-Group Residuals:
    Min          Q1         Med          Q3         Max 
-8.51192452 -0.38169972  0.05365992  0.47695927  7.43457534 

Number of Observations: 1366
Number of Groups: 246 
anova(B.lme)
              numDF denDF  F-value p-value
(Intercept)     1  1118 811.5700  <.0001
C               1   244   3.7171  0.0550
B               1  1118 117.6260  <.0001
B:C             1  1118   0.0043  0.9476

When I had a closer look at A.lme, the summary/anova suggests that variable C is significant.

>summary(A.lme)
Linear mixed-effects model fit by REML
Data: data1 
   AIC      BIC    logLik
4235.429 4261.517 -2112.715

Random effects:
Formula: ~1 | id
         (Intercept)  Residual
StdDev:   0.9228998 0.9690801

Fixed effects: A ~ B + C 
             Value      Std.Error  DF   t-value   p-value
(Intercept) -3.1021904 0.3332309  1119  -9.309431 0.0000
B            0.2330059 0.0214786  1119  10.848303 0.0000
C            0.7713974 0.3352298  244   2.301100  0.0222
Correlation: 
          (Intr)   B   
B         -0.171       
C         -0.971  0.034

Standardized Within-Group Residuals:
    Min          Q1         Med          Q3         Max 
-8.51328019 -0.38179254  0.05385169  0.47724088  7.43658227 

Number of Observations: 1366
Number of Groups: 246 

anova(A.lme)
              numDF denDF F-value p-value
(Intercept)     1  1119 813.4873   <.0001
B               1  1119 116.1162   <.0001
C               1   244   5.2951   0.0222

My question is which of the two models is more suitable for investigating C as a covariate? My second question is how important is the significance of the p-value of C in A.lme-this seems to suggest to me that C has a significant impact on the slope and intercept but not when combined with B (C:B). Can I safely conclude that C is not significant in B.lme? I'm using the nlme package in R.

Any help would be highly appreciated.

$\endgroup$
2
  • 1
    $\begingroup$ You forgot the ANOVA of B.lme. Or rather, you have an ANOVA of A.lme but it looks like it's for a B.lme. $\endgroup$
    – John
    May 18 '14 at 1:00
  • $\begingroup$ @John, well spotted. I've edited my original post with the correct anova results. $\endgroup$
    – John_dydx
    May 18 '14 at 6:48
4
$\begingroup$

Your model B.lme shows that $C$ is not a significant predictor of the slope--this is what the interaction $B \cdot C$ tells you. In other words, the effect of $B$ on the dependent variable is constant across the values of $C$, or conversely, the effect of $C$ on the dependent variable is constant across the values of $B$. Because $B$ and $C$ are involved in an interaction in B.lme, the "main effects" are actually not main effects, but merely the effect of the variable when the other variable is 0. For example, $0.736$ in B.lme is the effect of $C$ on the dependent variable when $B=0$.

For this reason, having determined that $C$ is not a significant predictor of the slope, it's a good idea to remove the non-significant interaction so that the effects are truly interpretable as main effects. The model A.lme shows that $C$ is a significant predictor of the dependent variable controlling for (holding constant) $B$, and similarly, that $B$ is a significant predictor of the DV controlling for $C$.

$\endgroup$
7
  • $\begingroup$ This is all true but doesn't quite get to answering the question. $\endgroup$
    – John
    May 18 '14 at 15:34
  • $\begingroup$ @Patrick Coulombe, thanks so much for that thorough explanation-really helpful. However as John already alluded to, I'm still wondering which of the two models would be the preference for investigating C as a covariate? I ask this because one model shows C is significant but the other is saying its significant. Thanks once again. $\endgroup$
    – John_dydx
    May 18 '14 at 15:51
  • 2
    $\begingroup$ I'm not sure why this does not answer the question. I mention that B.lme tests for $C$ as a predictor of the slope (the interaction), which is not significant, and which is a test of $C$ only at $B=0$. To get the overall test of $C$ (i.e. not conditional on the value of $B$), you need to look at A.lme. Again, the test for $C$ in B.lme is only testing at $B=0$, which may or may not be a useful (or possible) value for $B$. Because the interaction is not significant in B.lme, you should remove it to get overall tests for $B$ and $C$, so A.lme should be your final model. $\endgroup$ May 18 '14 at 16:53
  • 1
    $\begingroup$ And of course, the difference in significance for $C$ between A.lme and B.lme is due to the fact that they're not testing the same thing: B.lme tests for the effect of $C$ when $B=0$, whereas A.lme tests for the effect of $C$ for any value of $B$ (the latter being what you're interested in). If this doesn't answer your questions, let me know which part isn't answered, and I'll try to comment/edit. $\endgroup$ May 18 '14 at 17:14
  • 1
    $\begingroup$ Perhaps if you couched some things in terms of assessment of covariates and analysis of covariance it would be a more direct answer. Or perhaps state how an assessment in terms ore multiple regression might be a better way to go in this case than merely considering C a covariate (are B and C correlated?). That's all I meant. As I said, the answer is correct, just not quite answering the question. (BTW, this is a different John, not the one who posed the question) $\endgroup$
    – John
    May 18 '14 at 18:00
1
$\begingroup$

Fit model with interaction first. Since interaction term is not significant, remove interaction term only and fit again. Then look for main effect whether it is significant or not. If it is, keep it and this will be your final model. Hope it helps.

$\endgroup$
-1
$\begingroup$

If you'd like to know if your second model improves the first one, you could also statitically compare your models. In R you could do it with the function anova(model1,model2)

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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