3
$\begingroup$

I have created a model for which it was necessary to scale my predictor values by subtracting by the mean and dividing by the standard deviation of the X values. This resulted in variables centered around a mean of 0 and standard deviation of 1, and allowed my model to converge (the values prior to scaling were quite large).

My model includes quadratic (e.g., distance^2 to roads) and associated linear terms (e.g., distance to roads), and I'm interested in determining the maximum values (i.e., the vertex) for each quadratic term where b > 0. I believe my steps should be as follows, but confirmation would be quite useful:

1) Determine the vertex (v) of the quadratic variable by using the term v = -b/2a where b is the coefficient/beta of the quadratic term and a is the beta of the linear term.

2) Back transform this maximum value v by: v * σ + μ where σ is the standard deviation of the unstandardized, squared values of X and μ is the mean of the unstandardized, squared values of X.

3) Take the square root of the resulting backtransformed maximum value to convert from distance^2 to distance.

Is this correct? My results, while interpretable in the context of my study, make me question the validity of this approach. I appreciate any insight.

$\endgroup$
3
$\begingroup$

Your model appears to take the form

$$\mathbb{E}(Y) = \alpha + \beta Z + \gamma Z_2.$$

$Z$ and $Z_2$ are standardized versions of an explanatory variable $X$ and its square $X^2$, respectively, so that $\sigma Z + \mu = X$ and $\sigma_2 Z_2 + \mu_2 = X^2$, with $\sigma$ and $\sigma_2$ the (necessarily positive) standard deviations of $X$ and $X^2$ in the dataset.

The coefficients in the model are $\alpha$, $\beta$, and $\gamma$ which are estimated by $a$, $b$, and $c$, respectively: the latter two are the "betas" of $X$ and $X^2$ when $Y$ itself has been standardized. Finally, $a$ depends on all other variables. In particular, it is not constant in the dataset if there are any other non-constant variables in the regression. We will see that does not matter.

Re-expressing the fit in terms of $X$ and $X^2$ gives

$$\eqalign{ \mathbb{E}(Y) &= a + b\left(\frac{X-\mu}{\sigma}\right) + c\left(\frac{X^2-\mu_2}{\sigma_2}\right) \\ &= \left(a - \frac{b\mu}{\sigma}-\frac{c\mu_2}{\sigma_2}\right) + \left(\frac{b}{\sigma}\right)X + \left(\frac{c}{\sigma_2}\right)X^2. }$$

(Because varying $X$ does not change any of the other variables, $a$ disappears altogether from these expressions.)

The vertex occurs where the variation in the expectation with respect to $X$ is zero:

$$0 = \frac{d\mathbb{E}(Y)}{dX} = \frac{b}{\sigma} + \frac{2 c}{\sigma_2}X,$$

with the unique solution

$$X^{*} = -\left(\frac{b}{\sigma}\right) / \left(\frac{2 c}{\sigma_2}\right) = -\left(\frac{b}{2c}\right)\left(\frac{\sigma_2}{\sigma}\right)$$

provided $c\ne 0$. This estimate of the vertex will be a minimum when the second derivative $2 c/\sigma_2$ is positive: that is, necessarily $c\gt 0$.

As a double-check, $b$ and $c$ have the same units of measure as $Y$ (because $Z$ and $Z_2$ are unitless); $\sigma$ is in the units of $X$ (such as a distance); and $\sigma_2$ is in the squared units of $X$. Thus the units of $X^{*}$ are those of $\sigma_2/\sigma$: the same units as $X$, as it should be.


In this analysis it did not matter what values $\sigma$, $\sigma_2$, $\mu$, or $\mu_2$ might have had. In particular, if you chose to standardize using non-standard estimates of dispersion or location, the formula for the vertex would remain unchanged.

Edit

Comments suggest that $Z_2 = Z^2 = ((X-\mu)/\sigma)^2$. In this case

$$0 = \frac{d\mathbb{E}(Y)}{dX} = \frac{b}{\sigma} + \frac{2 c}{\sigma^2}(X-\mu),$$

with solution

$$X^{*} = \left(\frac{-b}{2c}\right)\sigma + \mu.$$

Little else changes in the analysis. It is no longer the case that $c$ is the beta (i.e., the standardized coefficient) for $X^2$, but that matters little. $X^{*}$ is still a vertex if and only if $c$ is positive.

$\endgroup$
  • 1
    $\begingroup$ In this situation, would you recommend standardizing $Z$ and then using the square of the standardized $Z$ in the model, or also standardizing the square as Beth has done? $\endgroup$ – Matthew Drury Jun 18 '15 at 20:24
  • $\begingroup$ @whuber - That clears things up a lot; thank you. To clarify - X* is the estimate of the vertex, but the second derivative (while not the estimate of the vertex) indicates whether the estimate is a minimum or maximum? $\endgroup$ – Beth S. Jun 18 '15 at 20:26
  • 1
    $\begingroup$ Yes, Beth, that is correct. @Matthew It matters little: once $Z$ is standardized, its square normally won't be far from standardized. If one's software is blind to the fact that $Z^2$ is actually the square of $Z$, and treats it as just one more variable, then it will separately standardize each variable, as I have assumed here. The possibility of this nuance is precisely why I started out this answer with the cautious "appears to take." I know from your many excellent answers here that you would find it easy to carry out this analysis in either case, but for completeness I will include it. $\endgroup$ – whuber Jun 18 '15 at 20:39
  • 1
    $\begingroup$ If both $b$ and $c$ are negative, then (1) $X^{*}$ is positive and (2) it is a maximum rather than a minimum. The only way you can obtain a negative estimate for a (minimum) vertex is when $c$ is positive and $b$ is positive. These conclusions have to be modified according to the edit when $Z_2=Z^2$: then you have to consider whether the vertex is to the right or left of $\mu$ rather than positive or negative, respectively. $\endgroup$ – whuber Jun 18 '15 at 20:50
  • 1
    $\begingroup$ When $c$ is negative, the vertex will exist and be a maximum. The sign of $c$ tells us whether the parabola is concave up ($c$ positive) or down ($c$ negative). When $c=0$, it degenerates to a line. The sign of $b$ is the sign of the slope when $X$ is equal to its mean value. By sketching some pictures you can identify all the constraints on $b$, $c$, and $X^{*}$. For instance, when $b\lt 0$ and $c\gt 0$, the curve must be descending when $X$ is at its mean, reach a minimum at some greater value $X=X^{*}$, and then go back up again. $\endgroup$ – whuber Jun 18 '15 at 21:35

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.