1
$\begingroup$

I would need a little clarification on this very basic issue. I'm trying to predict a continuous outcome based on a continuous predictor and a categorical variable with 3 levels using multiple linear regression. The fit tells me that when the continuous variable increases, also my dependent variable increases. Fair enough. On the other hand, it is also telling me that being in either of the k-1 categories decreases my dependent variable with respect to the baseline. If I check the mean value of the dependent variable in the three categories, nevertheless, I notice that, as a matter of fact it is higher in the two categories with respect to the baseline. What do I conclude from this? It sounds a little counter intuitive to me..

$\endgroup$
1

1 Answer 1

1
$\begingroup$

You fit 3 straight lines with the same slop and 3 different intercepts in the analysis. Let $$\hat{Y_1} = \hat\beta_{01} +\hat\beta_1 X \text{ for reference}$$ $$\hat{Y_2} = \hat\beta_{02} +\hat\beta_1 X $$ $$\hat{Y_3} = \hat\beta_{03} +\hat\beta_1 X $$ "being in either of the k-1 categories decreases my dependent variable with respect to the baseline" means $\hat\beta_{02}$ and $\hat\beta_{03}$ < $\hat\beta_{01}$.

"If I check the mean value of the dependent variable in the three categories, nevertheless, I notice that, as a matter of fact it is higher in the two categories with respect to the baseline." means $\bar{Y_1} < \bar {Y_2}$ and $\bar{Y_3}$, where $$\bar{Y_1} = \hat\beta_{01} +\hat\beta_1 \bar{X_1} \text{ for reference}$$ $$\bar{Y_2} = \hat\beta_{02} +\hat\beta_1 \bar{X_2} $$ $$\bar{Y_3} = \hat\beta_{03} +\hat\beta_1 \bar{X_3} $$

So although $\hat\beta_{02}$ and $\hat\beta_{03}$ < $\hat\beta_{01}$, given $\bar {X_1} < \bar {X_2}$ and $\bar{X_3}$ and the differences are large enough, you will get that $\bar{Y_1} < \bar {Y_2}$ and $\bar{Y_3}$ because $\beta_1 > 0$ as you noticed. So checking $\bar {X_1}, \bar {X_2}, \bar{X_3}$, you will find the answer.

($\bar X$ means the sample mean.)

$\endgroup$
3
  • $\begingroup$ Thank you, excellent explanation. In effect, I do preserve an higher outcome despite the negative coefficients for the k-1 categories. $\endgroup$ Commented Jun 1, 2017 at 10:13
  • $\begingroup$ I also noticed that by removing the continuous predictor, and leaving only the categorical ones, the negative coefficients disappear. Do I need to conclude for specificity in the way continuous predictor and dependent variable interact in my three groups? $\endgroup$ Commented Jun 1, 2017 at 10:19
  • $\begingroup$ No, it is not interaction. It seems from your data, you can get the following conclusion: "When continuous variable increases, dependent variable increases. The k-1 categories decreases in dependent variable comparing with to the baseline category (generally, we call it reference) when the continuous variable is controlled". $\endgroup$
    – user158565
    Commented Jun 1, 2017 at 16:44

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.