2
$\begingroup$

A random intercept logistic regression is performed to assess the association between $Y$: Disease (Yes/No) and Standardized Predictor($X_1$) adjusting for control variables ($X_2$, $X_3$) based on clustered survey data. A $X_1^2$ term is considered in the model to explore the nonlinear relationship. Results:

               coef     p-value
intercept     0.240     <0.001
    X1        0.285     <0.01
    I(X1)^2  -0.084     <0.01
    X2        0.114     <0.05
    X3        0.210     0.345

I'm trying to interpret the $X_1$ and $X_1^2$ as follows: "A unit increase in $X_1$ (standardized) is associated with $exp(0.285)$ higher odds of disease when $X_1$ (standardized) is at its mean, each additional level of $X_1$ is associated with $exp(-0.084)$ decrease in the likelihood of disease." Is this appropriate? Does anybody have any thoughts on this?

Thank you.

$\endgroup$

1 Answer 1

2
$\begingroup$

A few things:

1) If you have a regression of the form $y = \beta_{0} + \beta_{1}x_{1} + \beta_{2}x^2_{1}$, then taking a derivative of $y$ with respect to $x_{1}$ will return $dy/dx_{1} = \beta_{1} + 2\times\beta_{2}x_{1}$. So $\beta_{1}$ should be interpreted as the expected change in $y$ from a one-unit change in $x_{1}$ when $x_{1} = 0$. Even though you standardized your $x_{1}$, the $0$ part does not change, so you should interpret the coefficient on the linear part of $x_{1}$ as the effect of the variable when it is at $0$.

2) If you look again at the derivative formula above ($dy/dx$), you'll see that each additional one-unit increase in $x_{1}$ is associated not with a $\beta_{2}$ change in $y$, but with a $\beta_{1} + 2\times\beta_{2}x_{1}$ change in $y$. In other words, the change in $y$ is not constant, but varies depending on what the value of $x_{1}$ is. So you should not interpret the two coefficients separately, but simply say what each additional one-unit increase in $x_{1}$ does using values of both $\beta_{1}$ and $\beta_{2}$.

3) You might as well calculate and report the values of odds by exponentiating the coefficients. If you do that with $0.285$, you will get $1.330$, and you can then say that a one-unit change in $x_{1}$ is associated with a $33\%$ increase in odds of having the disease (when $x_{1}$ is at $0$). Exponentiating $-0.084$ will give you a number smaller than $1$ and you can use it together with the $33\%$ value to say what an additional one-unit change in $x_{1}$ is associated with.

4) You should use "on average" and "holding other independent variables constant" when discussing the effect of an explanatory variable on the dependent variable.

$\endgroup$
2
  • $\begingroup$ Thank you very much. For $X_1^2$ do I need to interpret this as exp(0.285+(2*(-0.084)))= 1.124, that means additional one-unit change in x1 is associated with a 12% increase in odds of having the disease? $\endgroup$
    – JRK
    May 21, 2019 at 10:50
  • 1
    $\begingroup$ Sorry, I should have been more clear. $exp(\beta_{1} + 2 \times \beta_{2}x_{1})$ is the odds ratio for a one-unit change in $x_{1}$. You still need to keep the $x_{1}$ in the expression. So an additional one-unit change in $x_{1}$ is associated with a $exp(0.285 - 2 \times 0.084x_{1})$ change in odds of having the disease. $\endgroup$
    – AlexK
    May 22, 2019 at 0:26

Your Answer

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

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