2
$\begingroup$

I have coefficients of a logistic regression model. I know from the statistical package used that the estimates are log odds. I am interested in combining two variables, beta 5 (B5) and beta 6 (B6) and then converting this combined log odds into a probability (as B5 and B6 are both measuring the same variable; B6 is an interaction term with B5 in the regression).

  1. Can you add two log odds together, and then convert them to make a combined conditional probability of, in this case, B5+B6?

  2. Also, I am following a paper published in a journal that does this and then converts B5+B6 to probability by p = (exp -(B5+B6)) - 1.
    This is not the way I have read how to convert log odds to probabilities. Hence I am wondering if the conversion process (from log odds to probability) changes when using combined coefficients?

$\endgroup$
3
$\begingroup$

I'm a little confused by your notion of coefficients interacting (I think you mean covariates, $x$'s). But I'll add some clarity around what it means to add coefficients.

Can you add two log odds together, and then convert them to make a combined conditional probability of, in this case, B5+B6?

We don't really say the "probability of $\beta_5 + \beta_6$", since these are coefficients. We can however ask about the change in log odds associated with a 1 unit increase in covariates $x_5$ and $x_6$, for example.

Let's say you have the following logistic regression model:

$$\text{logit} (P(Y=1|X=x)) = \alpha + \beta_1x_1 + \beta_2x_2 + \beta_3x_1x_2$$

You want to know if you can add $\beta_2 + \beta_3$? It's important to understand what the coefficient represents:

Let's say $X_2$ increased by 1 unit, the new log odds $P(Y=1|X)$ are:

$$\text{logit} (P(Y=1|X_1=x_1, X_2=x_2+1)) = \alpha + \beta_1x_1 + \beta_2(x_2+1) + \beta_3x_1(x_2 +1)$$

So the change in log odds associated with a 1 unit increase in $X_2$ is:

$$\frac{\text{logit} (P(Y=1|X_1=x_1, X_2=x_2+1))}{\text{logit} (P(Y=1|X_1=x_1, X_2=x_2))} = \frac{\alpha + \beta_1x_1 + \beta_2(x_2+1) + \beta_3x_1(x_2+1)}{\alpha + \beta_1x_1 + \beta_2x_2 + \beta_3x_1x_2}$$

$$ = \beta_2 + \beta_3x_1$$

So you can add the coefficients if you know your $x_1$ doesn't change. Of course, if it never changed then you should be modeling it as part of the intercept, $\alpha$.

Also, I am following a paper published in a journal that does this and then converts B5+B6 to probability by p = (exp -(B5+B6)) - 1. This is not the way I have read how to convert log odds to probabilities. Hence I am wondering if the conversion process (from log odds to probability) changes when using combined coefficients?

I'm also not familiar with this notation. In fact if

  • $\beta_5 + \beta_6 > 0$ then $e^{-(\beta_5+\beta_6)}-1$ is negative
  • $\beta_5 + \beta_6 < -\ln(2)$ then $e^{-(\beta_5+\beta_6)}-1$ is greater than 1

both of which are quite odd for a probability.

$\endgroup$

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.