0
$\begingroup$

I want to move from using repeat measure ANOVAs to linear mixed models (LMM). However, where I have good intuitions about ANOVAs, LMMs are new to me. I'm using python's StatsModels as my package. Here's the form of my data:

participant_ID  Condition_1 Condition_2 dependent_var
1               1           1           0.71
1               2           1           0.43
2               1           1           0.77
2               2           1           0.37
3               1           1           0.58
3               2           1           0.69
4               2           1           0.72
4               1           1           0.12
26              2           2           0.91
26              1           2           0.53
27              1           2           0.29
27              2           2           0.39
28              2           2           0.75
28              1           2           0.51
29              1           2           0.42
29              2           2           0.31

As you can seen, this is a classic repeat-measures ANOVA design, with fixed effects nested in participants. What I wish to do is establish (1) the independent effects of Condition_1 and Condition_2, and (2) the effect of their interaction, all on dependent_var. My statsmodels code is as follows: 

md = smf.mixedlm("dependent_var ~ C(Condition_1)+C(Condition_2) + C(Condition_1):C(Condition_2)", toy_data, groups=toy_data["participant_ID]).fit()

This outputs the following summary.

Summary Table

Allowing that this data is contrived, and p values are meaningless, etc, etc, am I correct to read this as saying that neither variable is significant as a main effect, and neither is their interaction?

I appreciate that LMMs aren't ANOVAs and I should avoid translating them into ANOVAs, but my actual data was arranged for an ANOVA design, and I wish to be confident in my interpretation.

Improve this question
$\endgroup$
10
  • 2
    $\begingroup$ You mention that "fixed effects are nested in participants", but in the data excerpt, it each subject appears to be uniquely assigned to one of the two values of Condition_2, suggesting that this factor is between-group. Can you clarify? $\endgroup$
    – Ous
    Jun 27, 2019 at 16:25
  • $\begingroup$ Sorry, yes, Condition_2 is between-group. I probably used the wrong terminology in describing it as nested. $\endgroup$
    – Lodore66
    Jun 27, 2019 at 16:31
  • 2
    $\begingroup$ Generally speaking, repeated-measures ANOVAs are simply a specific type of linear mixed model (LMM). In other words, repeated-measured ANOVAs are a subset of linear mixed models, so it doesn't really make sense to say you want to move from one to the other (unless you mean you want your model to be more general). But there are some important differences in the terminology. See here for example: theanalysisfactor.com/… $\endgroup$
    – StatsStudent
    Jun 27, 2019 at 16:40
  • $\begingroup$ I accept that they're versions of the same thing in the statistical sense; when I say 'move' I mean simply that LMMs are the model I select, largely for reasons of their ability to better deal with missing data than ANOVAs. My question centres on whether the specific model I implement in my code is appropriate to the structure of my data. $\endgroup$
    – Lodore66
    Jun 27, 2019 at 16:46
  • 1
    $\begingroup$ @Lodore66 They aren't two versions of some overarching model, a repeated measures ANOVA is a (type of) linear mixed model, so there is no reason why it would have different proporties like the ability to deal with missingness. Second, if there is indeed an interaction effect, you cannot then estimate the independent/main effects of the variables in the interaction as this violates the principle of marginality. $\endgroup$
    – Frans Rodenburg
    Jun 28, 2019 at 3:59

1 Answer 1

Reset to default
4
$\begingroup$

Starting with an additive "variance components" model, I think the Python/Statsmodels code you want is like this:

# df is your "toy data"
df["groups"] = 0                                                                                           

fml = "dependent_var ~ 1"                                                                                  
vcf = {"participant": "0 + C(participant_ID)", 
       "cond1": "0 + C(Condition_1)",                              
       "cond2": "0 + C(Condition_2)"}                                                                      
model = sm.MixedLM.from_formula(
           fml, vc_formula=vcf, 
           groups="groups", data=df)                             
result = model.fit(method='powell')

Since your Condition_1 and Condition_2 are crossed, you need to put everyone in a single group and use the variance components argument to specify all the random effects.

I get the results below:

           Mixed Linear Model Regression Results
===========================================================
Model:            MixedLM Dependent Variable: dependent_var
No. Observations: 16      Method:             REML         
No. Groups:       1       Scale:              0.0459       
Min. group size:  16      Likelihood:         0.4363       
Max. group size:  16      Converged:          Yes          
Mean group size:  16.0                                     
-----------------------------------------------------------
                   Coef. Std.Err.   z   P>|z| [0.025 0.975]
-----------------------------------------------------------
Intercept          0.531    0.054 9.917 0.000  0.426  0.636
cond1 Var          0.000    0.133                          
cond2 Var          0.000                                   
participant Var    0.000    0.089                          
===========================================================

For comparison, the R results are below, they are the same.

> lmer(dependent_var ~ (1|participant_ID) + (1|Condition_1) + (1|Condition_2), data=df)
singular fit
Linear mixed model fit by REML ['lmerMod']
Formula: dependent_var ~ (1 | participant_ID) + (1 | Condition_1) + (1 |  
    Condition_2)
   Data: df
REML criterion at convergence: -0.8726
Random effects:
 Groups         Name        Std.Dev.
 participant_ID (Intercept) 0.0000  
 Condition_1    (Intercept) 0.0000  
 Condition_2    (Intercept) 0.0000  
 Residual                   0.2143  
Number of obs: 16, groups:  participant_ID, 8; Condition_1, 2; Condition_2, 2
Fixed Effects:
(Intercept)  
     0.5312  
convergence code 0; 1 optimizer warnings; 0 lme4 warnings

Both packages struggle with the optimization since your variance parameters are on the boundary. I had to use the non-default "powell" option in Statsmodels to get a converged result.

I'm not sure if you care about this data set, or if you just are using this as an example. But the interpretation here would be that there is no evidence for additive effects of participant, Condition_1, or Condition_2 in relation to your dependent variable, at least in an additive sense.

Improve this answer
$\endgroup$
1
  • $\begingroup$ This is brilliant, thank you! Don't know why I only saw it now; SE never notified me of it. But great, perfect answer. $\endgroup$
    – Lodore66
    Jul 16, 2019 at 13:39

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.