In my textbook I often see a quadratic term in the regression. For example let's say I have the regression model: $\log(\mathrm{wage})=\beta_0+\beta_1\mathrm{educ}+\beta_2\mathrm{exp}+\beta_3\mathrm{exp}^2+\varepsilon$. Why is it so common to put the squared term in there? Here's a quote from my book:

It often makes sense to add quadratic terms of any significant variables to a model.

Why is that? And if this is true, then if I'm studying endogeneity and a quadratic variable isn't in there is it likely to be that case the the squared term is endogenous if it's omitted?

  • 3
    $\begingroup$ Well, for starters, frequently relationships are not linear. Unless I had a priori reason to, I'd usually only do it if the residual analysis suggested the linear model was clearly inadequate. On the other hand if I had no a priori sense of what form the relationship might take I'd fit something more general from the outset. $\endgroup$
    – Glen_b
    Nov 26, 2012 at 1:10

1 Answer 1


Often the relationship between y and x is nonlinear. There are a variety of solutions. One solution is to add polynomial terms and the first one to look at is usually $x^2$. But you should first look at a scatterplot of x and y; you should also look at the residuals from the linear model without the quadratic term. But it turns out that many relationships are pretty well fit by $y \sim b_0 + b_1x + b_2x^2$ (plus any other x variables, of course).

It is also possible to add cubic, quartic and even higher order terms, but such models quickly become hard to interpret. Another possibility is to look at a spline regression.

I don't really understand your last bit about endogeneity. A variable can be endogenous or exogenous to a model, but that doesn't seem to relate to this ... at least, not in a way that's clear to me.

  • $\begingroup$ I can see why you would be confused. I appologize for that, but a friend explained it to me. Thanks for you explanation. $\endgroup$
    – Kyle
    Nov 26, 2012 at 4:54
  • 1
    $\begingroup$ Adding arbitrary polynomials to "account for nonlinearity" is almost as arbitrary as asserting linearity. Best to use a nonparametric smoothing regression in the absence of explicit theory regarding functional relationship, right? And if a parametric estimate is needed, following that with an algorithmically determined fractional polynomial, or appropriate custom nonlinear least squares, or something similar second, yes? A quadratic or cubic function generally does a poor job of representing linear threshhold or saturation effects, for example. $\endgroup$
    – Alexis
    Dec 10, 2018 at 0:59
  • $\begingroup$ Let me try asking how I interpreted his bit about endogeneity: suppose that the true relationship is quadratic, but we omit the quadratic term. Do we now have an endogeneity problem? $\endgroup$
    – user106860
    May 14, 2020 at 0:37

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.