0
$\begingroup$

I already search the net with regards to the significance of the baseline variable. One of the sources that I've read is you will not know whether a baseline variable is significant to the regression model.However, people would ask whether the response in the baseline variable is significant or not. Let say I have a data (see below), where:

1=Elementary Graduate, 2=High School Graduate, and 3=College Graduate.

Supposed factor(sex)1 is significant.

How will I know whether mean response for the categorical variable not represented by dummy in the regression model (1=Elementary Graduate) is significant? How will I quantify it since it does not have coefficient?

I understand that we can change the base category but it always leave the unrepresented category without decision whether its mean response is significant and what its estimated coefficient. How will I interpret the results in the analysis including the baseline variable?

$\endgroup$
  • $\begingroup$ It does have a coefficient, it is the intercept. $\endgroup$ – mdewey May 24 '17 at 10:38
  • $\begingroup$ To get predicted values for the baseline category, just set all the dummys to zero. $\endgroup$ – kjetil b halvorsen May 24 '17 at 10:45
  • $\begingroup$ But if you have more than 1 categorical variable in your model, then the intercept will confound the different reference/baseline categories - And there is nothing you can do about it. $\endgroup$ – Umka May 24 '17 at 10:49
  • $\begingroup$ @mdewey with control for no other factors, yes that's true. But after control for, say, age, it gets hairy. You can use prediction at the means for a continuous response to get a nice interpretation (the same as centering age) $\endgroup$ – AdamO Apr 2 '18 at 21:43
2
$\begingroup$

From your post, it seems like you are confused about what actually constitues an effect in your model. Let's say that y stands for income (thousands od dollars), a stands for age and educ stands for education level.

For instance, the effect of the predictor variable a on the response variable y controlling for the effect of educ is estimated as 0.67439. This effect can be interpreted as follows:

Among those with the same education level, we estimate that each 1-  
year increase in age is associated with an increase in the mean income 
of 0.67439 thousands of dollars. 

The effect of the predictor variable educ on the response variable y controlling for age can be described by a collection of two separate effects, estimated as 0.63322 and -0.07283. Here is the interpretation of these two effects:

The difference in mean incomes between those with a high school   
education and those with an elementary education is estimated to be 
0.63322 thousand dollars (assuming they share the same age). 

The difference in mean incomes between those with a college graduate 
education and those with an elementary education is of -0.07283 
thousand dollars (assuming they share the same age). 

(It doesn't make sense that college graduates would earn less than those with an elementary education, but this example is made up.)

As you can see, effects quantify the change in the mean value of the response variable y associated with a change in the values of the predictor variables a and educ. If you don't change the value of educ, you can't really speak of an effect for it.

$\endgroup$
0
$\begingroup$

Well I think you already the know the answer to your question (i.e., "you will not know whether a baseline variable is significant"). The effect for the reference (baseline) category is constrained to be null, and what you estimate with the other dummy variables are relative effects, that is the effect of moving from the reference category to the target one (in your case: Elementary school to High school). This is really important!

Do not re-estimate your model with a different reference/baseline category to obtain an estimate for the initially "missing" effect (in your case: elementary school) - This would not make sense at all!

$\endgroup$
  • $\begingroup$ To follow on my post: The model intercept would capture the baseline effect (i.e., effect when all predictors (X) are set to 0) - In the simplistic case of a single categorical predictor (Y ~ X), the model intercept would tell you something about the baseline category, but this type of modelling would have a limited interest as simple descriptive analysis of (Y) would be more than enough. $\endgroup$ – Umka May 24 '17 at 10:39

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.