I have the two following regression: $$y=X_1\beta_1+\varepsilon_1$$ and $$y=X_2\beta_2+\varepsilon_2$$

$x_1$ and $x_2$ are indicator variables and very similar. One expert claim that the two coefficients are the same. I want to know what kind of test I can perform to see if $β_1=β_2$.

  • $\begingroup$ Something like this? $Z=((β_1-β_2)/√(S_(β_1)^2+S_(β_2) ))^2 $ $\endgroup$
    – user150086
    Feb 21, 2017 at 0:03
  • $\begingroup$ Note that you can wrote $S_{\beta_1}^2$ in MathJax. Right-click on this notation, then chose "Show Math As" from the menu, then choose "TeX Commands", and you will see the code. $\endgroup$ Feb 21, 2017 at 3:10
  • $\begingroup$ You refer to $X_1$ and also to $x_1.$ Are those two different things? $\endgroup$ Feb 21, 2017 at 3:20
  • $\begingroup$ So far I'm the only one who's up-voted this question, although someone else has answered. That tends to be neglected. $\endgroup$ Feb 21, 2017 at 4:44
  • $\begingroup$ X1 and x1 are the same. $\endgroup$
    – user150086
    Feb 21, 2017 at 15:20

2 Answers 2


If the two predictors are orthogonal, an optimal test is to compare the model

$$ y_i \sim \beta_0 + \beta_1 X_1 + \beta_2 X_2 $$

with the submodel

$$ y_i \sim \beta_0 + \beta_1 (X_1 + X_2) $$

with the likelihood ratio test. The null hypothesis is that $\beta_1 = \beta_2$.

  • $\begingroup$ My main problem is that the two variable are highly correlated and I want to test if they are different and pick one of them in my model. $\endgroup$
    – user150086
    Feb 21, 2017 at 0:50
  • $\begingroup$ If they're highly correlated, you may just reduce it to a single variable with PCA. There's no need to do what you're asking. $\endgroup$
    – SmallChess
    Feb 21, 2017 at 1:23
  • 1
    $\begingroup$ @StudentT, Not sure there's any point in doing PCA on two binary variables. The combination can only take on four values. $\endgroup$ Feb 21, 2017 at 2:47
  • $\begingroup$ @user150086, if they are that closely related, the difference in effect size probably isn't large, so you will have low power to detect a statistically significant difference between their effects no matter what you do. You could just pick the one with a larger effect size, or the one that optimizes some predictive criteria. $\endgroup$ Feb 21, 2017 at 2:50
  • 2
    $\begingroup$ Yes @MichaelHardy. I assumed we were talking about the normal errors linear model, where you could use a LRT or an F test. $\endgroup$ Feb 21, 2017 at 4:31

You are analyzing a single model $y=X\beta+\varepsilon$. The regression model will return $y=\beta_0+X_1\beta_1$ with $\beta_0$ being the intercept. To determine the difference between the two models you need to look at both the slope $\beta_1$ and the intercept $\beta_0$. You can use the indicator variable as a regressor by itself and in combination with your other independent variables. If the indicator variable is $d$ it will be $d=0$ for the $X_1$s you include above and $d=1$ for the $X_2$s. The model becomes $y\sim d, X, (X*d)$. The regression will return $y=\beta_0+\beta_1X+\beta_2d+\beta_3Xd$. When $d=0$ the coefficients $\beta_2$ and $\beta_3$ vanish. When $d=1$ the intercept is $(\beta_0+\beta_2)$ and the slope is $(\beta_1+\beta_3)X$. Keep the $\beta$s with significant p-values and you end up with a model that tells exactly what and where is the difference between the two groups. See Ways of comparing linear regression interepts and slopes?.


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.