Dear R experts,
As you know, by default the reference level in an R model (e.g. lmer or glmer model) is alphabetically or numerically arranged. But sometimes you need to re-level the default/reference level. Does this cause any difference in results? Put another way, when should levels be left alphabetically arranged and when should they be re-levelled?
The default reference level is not automatically the "lowest" alphabetical value - it depends on what you specified when creating your factor variable:
dummy
Many thanks, Daniel. This is clear except the last point. What do you mean by "Re-levelling a factor should not happen because of alphabetical order, but for good reasons which reference category is most meaningful." Suppose we have the variable 'gender' with males and females and we test the role/effect of gender on dropping out from school (DV). How can I know whether males or females are "most meaningful"? Does this make sense?
Best
If you don't know which level is most meaningful then neither level is most meaningful. The idea of what's "most meaningful" depends on your own hypotheses and predictions.
For a factor with only 2 levels then this doesn't matter, because either way it's going to be making the same comparisons; the only difference will be whether the coefficient is positive or negative. Choice of reference is more important when you're analyzing a polytomous (more than 2 levels) factor.
Just a note:
If there is no meaningful "reference level" there are possibly more appropriate ways to code the factor levels. E.g. the "deviation coding" will give the coefficients as differences to the grand mean:
http://www.ats.ucla.edu/stat/r/library/contrast_coding.htm
Thinking about coding is not really relevant if you have only two levels in your factor (i.e. a binary or dichotomous predictor variable). As Stephen said, the default "dummy coding" (I would call it "reference coding") takes on level as the reference, and in your gender example the coefficients are either (male-female) or (female-male), depending which one is selected as the reference level. The deviation coding would give you two coefficients (one for male and one for female), each being 0.5x the coefficient value of the dummy-coded model, but with opposite signs. That's simple to calculate manually, anyway.
"Most meaningful", as mentioned by Stephen, mostly applies for factors with more than two levels. If your factor levels have a "natural order", say: educational level of a person with levels "low", "mid" and "high" - then it might make sense to use "low education" as reference, because you then might see an effect with "increasing" educational level".
On the other hand, if you have a depression status with levels low, mid and high, and you want to know if low depression has a lower effect and high depression has a higher effect than the "average" depression, "middle depression" might be a good reference category.
If you have levels with arbitrary order (car brands, countries), your choice of reference might be arbitrary as well.
Many thanks again, Daniel. Sorry, I am quite new to stats. What does depression here mean?
I think this has already been explained multiple times. Can you specify what is still unclear?
I found this thread in pursuit of trying to figure out this "releveling." I have a treatment that has four levels (constant food, pulsed with food, pulsed without food, control). I first ran a model with the control as the reference level. The other levels of the factor were significantly different. But then I wanted to know is constant food different from pulsed food. I releveled the factors, re-ran the model, and noticed an unrelated covariate, date, became more significantly (from p=0.8 to p=0.06). Why did that happen?
Did "date" interact with treatment? What was your full model specification?
Yes, I did try to run the model with a date*treatment interaction
The model is as follows:
#birds ~ treatment*julian + treatment + minutes after sunrise + snow cover + wind speed + (1|site)
I've tried to fit this with family=poisson and =nbinom1.
It's a zero-inflated model with ziformula=~1.
Used glmmTMB() to run it.
Then what you're describing is normal. If you think through how the dummy-coding works (easiest way to do this is often to work out the regression equation on paper), you will see that the coefficient for "date" (which I assume is the variable "julian") in your model summary is only saying the slope of that line for the reference level of 'treatment'. In fact this is a case of the more general rule that when you have interactions with dummy-coded variables, the model coefficient you see only represents what's happening at the baseline level, it's not a main effect. For explanation see e.g. http://www.mypolyuweb.hk/~sjpolit/coding_schemes.html (the example being worked through here is with categorical variables, but the same logic applies to continuous variables) or a textbook on regression.
- Badges
- Science topic