Performing repeated measures ordinal logistic regression - output interpretation

Hi there,

for my research i have performed a repeated measures ordinal logistic regression:
Ordinal variable(value): 0, 1, 2, 3
time(variable): t0, t1
Group: IG, CG

Thats my code:

> Call:
> polr(formula = as.factor(value) ~ variable * as.factor(group), 
>     data = Ordinaldf, Hess = TRUE)
> Coefficients:
>                                          Value Std. Error t value
> variableweighting_t1                     0.577     0.2381   2.424
> as.factor(group)IG                      -1.099     0.2296  -4.787
> variableweighting_t1:as.factor(group)IG  1.394     0.3158   4.416
> Intercepts:
>                          Value   Std. Error t value
> unknown|know from seeing -0.3083  0.1699    -1.8146
> know from seeing|known    1.9679  0.1923    10.2324
> known|friend              3.2828  0.2453    13.3827
> Residual Deviance: 1324.547 
> AIC: 1336.547


z test of coefficients:

                                        Estimate Std. Error z value  Pr(>|z|)    
variableweighting_t1                     0.57703    0.23805  2.4240   0.01535 *  
as.factor(group)IG                      -1.09929    0.22963 -4.7872 1.692e-06 ***
variableweighting_t1:as.factor(group)IG  1.39448    0.31576  4.4163 1.004e-05 ***
unknown|know from seeing                -0.30832    0.16991 -1.8146   0.06958 .  
know from seeing|known                   1.96793    0.19232 10.2324 < 2.2e-16 ***
known|friend                             3.28275    0.24530 13.3827 < 2.2e-16 ***
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

I am a little bit confused why the Estimate for Group is negative. Can anyone give me please an explanation for this result. Or maybe an interpretation of the output.

Thank you very much in advance.


If I understand what you are doing, the negative value of as.factor(group)IG just means that the effect of IG is negative relative to CG. When fitting with factors, it is standard to pick one level as the baseline and R picks the first level alphabetically. If you recoded your group so that IG became AA, the sign of the coefficient would change and it would be labeled as.factor(group)CG since AA would now be the baseline.

Hello FJCC,
thanks for your explanation.
I will explain the design a bit more: I wanted to analyze if more people got to know each other better over time. We had two different groups. I want to know if the measure in the intervention group is conducive to people getting to know each other compared to the control group.

The output tells me that time and interaction are positive predictors of value. The value has increased, or in other words, people have gotten to know each other or even consider each other friends. The intervention is beneficial.

But the group factor tells me that IG is negative compared to CG, but at what point in the beginning of the study?
Both group didn't start from the level at baseline.

This sort of study is far outside my expertise, so I do not want to say anything definitive. The coefficient on IG says that membership in that group lowers the outcome response. BUT, I see there is a significant interaction between IG and varaibleweighting_t1. That means you cannot talk about the effect of IG by itself but must consider the other variable at the same time.
An example I like is the effects of methane and lighted candles on health. Having some methane in your home is not good, but the health effects are not serious. Lighted candles in the home have no effect on health. Methane and a lighted candle causes an explosion!

I do not understand your variables, but does using IG in the fit make sense if other variables that designate the treatments applied, necessarily imply which group a subject is in? I am not implying an answer to that question, I am suggesting you think about it.