constant term highly significant

[lisa carnicelli]

I am using a difference in difference model, and when I ran the regression the constant term is highly significant like the t-value is 968.72. Is that normal or am I doing something wrong ? and what does it mean ?. thanks you very much for your help

[Bruce Weaver]

The constant is the fitted value of Y when all explanatory variables are equal to 0. What are your explanatory variables? I suspect that for at least one of them, 0 is an extreme (if not impossible) value. For that reason, the t-test on the constant is usually not of much interest. HTH.

[lisa carnicelli]

I am checking the impact that the introduction of a legislation had on certain variables. My model is this:

y = treated + post + treated * post

treated and post are dummies. Treated = 1 for the treatment group and 0 for the control group while post = 1 for the period after the introduction of the legislation and 0 for the period prior.

[Carole J. Wilson]

The magnitude of the constant is dependent on what Y is measuring. If it is dollars or centimeters or weight, then you would expect the constant (the expected value of Y for the control group prior to the introduction of the legislation) to reflect that. The coefficients on Treated, Post, and Interaction are essentially differences from the value of the constant.

[Bruce Weaver]

As Carole noted, the constant is just the mean of the group with both treat = 0 and post = 0. You have not shown the code for your analysis, but try something like the following.

* Generate some fake data, as none was provided.
clear *
set seed 20180805
set obs 100
generate y = rnormal(50,10)
generate byte treat = _n > 50
generate byte post = inrange(_n,26,50) | inrange(_n,76,100)
* Your analysis likely looks something like the following:
regress y treat##post
margins treat#post
* Notice that the constant from the regression model is
* equal to the mean of group with both treat and post = 0.
contrast treat@post, nowald effects
contrast post@treat, nowald effects

Finally, remember that the t-test on the coefficient is testing the null hypothesis that the group with treat=0 and post=0 is a random sample from a population with mean of 0. Given your design, that hypothesis is almost certainly of no importance or interest.

HTH.