Square-Root Predictors for Diminishing Returns

Why Use a Square-Root Predictor?

Some relationships change quickly at low values of a predictor and more slowly at higher values. For example, an increase in an input might be associated with a substantial outcome change when the input is small, but a smaller change when the input is already large. A straight-line model in the original predictor assigns the same fitted change to every equal-sized increase. A square-root predictor allows the fitted change to diminish as the predictor grows.

This is a particular way to represent curvature, not a guarantee that the scientific relationship has that shape. The predictor must be nonnegative, and the model should be used only over a range with data support. A fitted association alone does not show that changing the predictor causes the outcome to change.

The Square-Root Predictor Model

Definition: A square-root predictor regression model has the form \(Y_i=\beta_0+\beta_1\sqrt{x_i}+\varepsilon_i\), where \(x_i\geq 0\), \(Y_i\) is the outcome for observation \(i\), \(\beta_0\) and \(\beta_1\) are regression coefficients, and \(\varepsilon_i\) is the error. Its fitted equation is \(\hat{y}=\hat{\beta}_0+\hat{\beta}_1\sqrt{x}\).

The model is linear in the transformed predictor \(\sqrt{x}\), so ordinary least squares can be used by creating a transformed predictor \(z=\sqrt{x}\) and regressing \(Y\) on \(z\). The assumptions concern the model’s conditional mean and errors: the mean outcome at each predictor value follows the stated linear form in \(\sqrt{x}\); errors have mean zero, are independent, and have constant variance. Normal errors are additionally used for exact small-sample \(t\)-based inference. The transformed model does not require the outcome itself to be transformed.

The intercept \(\beta_0\) is the model’s mean outcome at \(x=0\), if zero is meaningful and within the relevant range. The coefficient \(\beta_1\) describes the change in fitted outcome for a one-unit increase in \(\sqrt{x}\), not for a one-unit increase in \(x\). If \(x\) is measured in units \(U\) and \(Y\) in units \(V\), then \(\beta_1\) is measured in \(V/\sqrt{U}\).

Worked Example: Fitting the Model After Transforming the Predictor

Suppose a hypothetical set of four plots has nonnegative fertilizer amounts \(x\), in kilograms per hectare, and yields \(y\), in kilograms per plot. The observations are \((0,10), (1,14), (4,17), (9,23)\). Fit a least-squares line using the square root of fertilizer amount.

\(x\) (kg/ha)	\(z=\sqrt{x}\) (\(\sqrt{\text{kg/ha}}\))	\(y\) (kg/plot)
0	0	10
1	1	14
4	2	17
9	3	23

For a least-squares line of \(y\) on \(z\), the slope is \(S_{zy}/S_{zz}\), where \(S_{zy}=\sum(z_i-\bar{z})(y_i-\bar{y})\) and \(S_{zz}=\sum(z_i-\bar{z})^2\). Here \(\bar{z}=1.5\) and \(\bar{y}=16\). Thus,

S_{zz}=(-1.5)^2+(-0.5)^2+(0.5)^2+(1.5)^2=5

S_{zy}=(-1.5)(-6)+(-0.5)(-2)+(0.5)(1)+(1.5)(7)=21

Therefore \(\hat{\beta}_1=21/5=4.2\), and \(\hat{\beta}_0=\bar{y}-\hat{\beta}_1\bar{z}=16-(4.2)(1.5)=9.7\). The fitted equation is \(\hat{y}=9.7+4.2\sqrt{x}\). For example, at \(x=4\) kg/ha, the fitted yield is \(9.7+4.2(2)=18.1\) kg/plot.

The residuals for the four observations are \(0.3, 0.1, -1.1,\) and \(0.7\) kg/plot, respectively. All predictor values are nonnegative, as required. These few observations cannot establish independence, constant variance, or normality; for this small calculation, treat the model assumptions as working assumptions and do not use the fit to claim precise inference.

Marginal Effect on the Original Predictor Scale

Because the model is curved when expressed in \(x\), the effect of a small increase in \(x\) depends on the current value of \(x\). For \(x>0\), differentiate the fitted mean function with respect to \(x\):

\frac{d\hat{y}}{dx} =\frac{d}{dx}\left(\hat{\beta}_0+\hat{\beta}_1\sqrt{x}\right) =\frac{\hat{\beta}_1}{2\sqrt{x}}

This derivative is the fitted marginal effect: the approximate change in outcome for a small increase in the original predictor, at a specified \(x\). If \(\hat{\beta}_1>0\), the fitted effect is positive and decreases in magnitude as \(x\) increases. If \(\hat{\beta}_1<0\), the fitted effect is negative but becomes less negative as \(x\) increases. In either case, the second derivative is

\frac{d^2\hat{y}}{dx^2} =-\frac{\hat{\beta}_1}{4x^{3/2}},\qquad x>0.

Thus a positive coefficient gives a concave-down fitted curve, the usual diminishing-positive-returns shape. A negative coefficient gives a concave-up curve and a diminishing negative association. At \(x=0\), the derivative formula is undefined; this does not prevent calculating fitted outcomes or finite changes involving zero.

Key result: For any \(0\leq x_1<x_2\), the fitted outcome change is \(\hat{\beta}_1(\sqrt{x_2}-\sqrt{x_1})\). The average fitted change per unit of \(x\) over that interval is \(\hat{\beta}_1/(\sqrt{x_2}+\sqrt{x_1})\). This finite-interval calculation also works when \(x_1=0\).

To obtain the average-change expression, subtract the two fitted values. Then use the identity \(x_2-x_1=(\sqrt{x_2}-\sqrt{x_1})(\sqrt{x_2}+\sqrt{x_1})\):

\frac{\hat{y}(x_2)-\hat{y}(x_1)}{x_2-x_1} = \frac{\hat{\beta}_1(\sqrt{x_2}-\sqrt{x_1})} {(\sqrt{x_2}-\sqrt{x_1})(\sqrt{x_2}+\sqrt{x_1})} = \frac{\hat{\beta}_1}{\sqrt{x_2}+\sqrt{x_1}}.

The denominator grows as the interval moves to larger predictor values. For a positive coefficient, equal-sized increases in \(x\) therefore produce smaller fitted increases at higher values. The derivative is a local rate; the finite difference is the exact model-based change over a specified interval.

Worked Example: Comparing Marginal Effects at Low and High Values

Suppose a hypothetical fitted model relates fertilizer amount \(x\), in kilograms per hectare, to crop yield \(y\), in kilograms per plot: \(\hat{y}=32+5\sqrt{x}\). Assume \(x\) is within the observed range of 0 to 120 kg/ha, and that the model’s residual and design checks support using this form for description. Compare the fitted marginal effects at 4 and 100 kg/ha.

At \(x=4\), the derivative is \(5/(2\sqrt{4})=5/4=1.25\). Near 4 kg/ha, a small increase of 1 kg/ha corresponds to an approximate fitted increase of 1.25 kg/plot. At \(x=100\), it is \(5/(2\sqrt{100})=5/20=0.25\), or approximately 0.25 kg/plot per additional kg/ha near that value. Both are positive, but the fitted marginal effect at 100 kg/ha is smaller.

For an exact finite comparison, consider an increase from 4 to 9 kg/ha. The fitted change is \(5(\sqrt{9}-\sqrt{4})=5(3-2)=5\) kg/plot. The average fitted change is \(5/5=1\) kg/plot per kg/ha over this interval. This average is between the local marginal effects at its endpoints, \(1.25\) and \(5/6\approx0.8333\), consistent with a declining marginal effect.

Calculating Changes Over Equal Increases

A useful way to communicate diminishing returns is to compare equal-sized increases at different starting points. The square-root curve does not say that every increase at a high value has no effect. It says that, for a positive \(\beta_1\), the fitted increase associated with the same added amount of \(x\) gets smaller as the starting value rises. The exact fitted change is preferable to multiplying a single derivative by a large change in \(x\), because the derivative varies across the interval.

Worked Example: Equal Predictor Increases Near Zero and at a High Value

In a hypothetical analysis of a productivity score \(y\) by years of practice \(x\), a fitted model is \(\hat{y}=20+6\sqrt{x}\). Suppose the observed practice range is 0 to 120 years and checks support using the model descriptively in that range. Compare the fitted changes for an increase of 9 years, starting at 0 years and at 100 years.

From 0 to 9 years, the fitted change is \(6(\sqrt{9}-\sqrt{0})=6(3-0)=18\) score points. The fitted value moves from 20 to \(20+6(3)=38\).

From 100 to 109 years, the fitted change is \(6(\sqrt{109}-\sqrt{100})=6(10.4403-10)\approx2.642\) score points. The fitted value moves from 80 to approximately 82.642. Both intervals add 9 years, but the model predicts a smaller score increase for the interval starting at 100 years.

The derivative at zero is not defined, so it would be incorrect to report a finite marginal effect there using \(6/(2\sqrt{0})\). The model nevertheless gives a well-defined fitted value at zero and a well-defined change from zero to 9 years. The interval calculation is the right tool for that comparison.

What the Model Does and Does Not Say

The square-root form is useful when the scientific context supports a nonnegative predictor and a relationship that changes at a decreasing rate. It can express a diminishing association with fewer parameters than a more flexible curve. But it is not automatically the right form because a scatterplot bends, and it does not force the fitted curve to level off at a fixed maximum: for a positive coefficient, \(\hat{y}\) continues to increase as \(x\) increases, though its marginal increase gets smaller.

Check whether the chosen transformation is meaningful, whether observations cover the values at which predictions or marginal effects will be reported, and whether residual patterns are reasonably consistent with the model assumptions. In particular, do not use the curve far beyond the data range simply because the formula can be evaluated there. The coefficient and marginal effect describe an association unless the study design and additional assumptions justify a causal interpretation.

Common Mistakes

Interpreting \(\hat{\beta}_1\) as the change per unit of \(x\). It is the change per unit of \(\sqrt{x}\). The local rate per unit of \(x\) is \(\hat{\beta}_1/(2\sqrt{x})\), for \(x>0\).
Reporting one marginal effect for every predictor value. The rate changes with \(x\); name the value at which it is evaluated or report a finite change over a specified interval.
Using the derivative at zero. The derivative is undefined at zero. Calculate a finite fitted change instead when an interval begins there.
Calling the curve a plateau. A positive square-root model keeps increasing, although more slowly. It does not approach a fixed upper limit.
Claiming causation from the fitted shape. A regression curve describes the fitted association; causal language requires suitable design and assumptions beyond the transformation itself.

Key takeaway: Transforming a nonnegative predictor to its square root makes the fitted mean linear in \(\sqrt{x}\), but nonlinear in \(x\). Interpret \(\hat{\beta}_1\) on the transformed scale, use \(\hat{\beta}_1/(2\sqrt{x})\) for the local marginal effect when \(x>0\), and use \(\hat{\beta}_1(\sqrt{x_2}-\sqrt{x_1})\) for an exact fitted change over an interval.

Check Your Understanding

Try these questions without looking back.

For \(\hat{y}=12+3\sqrt{x}\), what does the coefficient 3 measure, and what is the fitted marginal effect at \(x=9\)?
For a positive square-root coefficient, explain why equal increases in \(x\) give smaller fitted changes at higher starting values.
A fitted model is \(\hat{y}=8+4\sqrt{x}\). Calculate the fitted change from \(x=1\) to \(x=16\).
Why is the derivative at \(x=0\) not a suitable way to report the fitted change from zero to a positive predictor value?
Does a positive square-root predictor model approach a fixed upper limit as \(x\) increases? Explain.