Square-Root Transformations of Outcomes

Why Model the Square Root of an Outcome?

Some outcomes are nonnegative and have a relationship with a predictor that is curved on their original scale. A square-root transformation can make the mean relationship more nearly linear while keeping zero outcomes available for analysis. But a regression fitted to \(\sqrt{Y}\) answers questions on the transformed scale first. To report results in the units of \(Y\), we must translate the fitted values carefully.

The square root changes both the units and the interpretation of a coefficient. Squaring a fitted value returns a quantity in the original units, but it does not automatically return the conditional arithmetic mean of the outcome. The distinction matters whenever predictions or changes are reported in context.

The Model and Its Scale

Let \(Y_i\) be a nonnegative outcome for individual \(i\), and let \(x_i\) be a predictor. Define the transformed response \(W_i=\sqrt{Y_i}\). A simple square-root outcome regression is

W_i=\beta_0+\beta_1x_i+\varepsilon_i, \qquad E(\varepsilon_i\mid x_i)=0.

Thus the conditional mean of the transformed response is \(\eta(x)=E(\sqrt{Y}\mid x)=\beta_0+\beta_1x\). The fitted regression estimates this mean, not \(E(Y\mid x)\). For standard regression inference, the usual working assumptions are that the conditional mean has the stated linear form, errors are independent with mean zero and constant variance, and errors are approximately normal when small-sample \(t\)-based inference is used. Residual plots and the study design help assess whether these assumptions are reasonable. The outcome itself must be nonnegative for its square root to be defined.

Definition: In a square-root outcome model, the response used in regression is \(W=\sqrt{Y}\), where \(Y\geq 0\) is the outcome of interest. A fitted value \(\hat{\eta}(x)\) estimates the conditional mean of \(\sqrt{Y}\) at predictor value \(x\). Its square, \(\hat{\eta}(x)^2\), is an original-scale plug-in prediction, not automatically an estimate of the conditional arithmetic mean \(E(Y\mid x)\).

The intercept \(\beta_0\) is the expected square root of the outcome when \(x=0\), if zero is meaningful and within the relevant predictor range. The slope \(\beta_1\) is the change in the expected square root of the outcome for a one-unit increase in \(x\). Its units are the square root of the outcome units per unit of \(x\). It is not a constant change in \(Y\) per unit of \(x\).

Worked Example: Interpreting a Fitted Value

Suppose a hypothetical analysis of daily water use fits the model \(\widehat{\sqrt{Y}}=3.0+0.40x\), where \(Y\) is water use in liters per day and \(x\) is hours of a specified activity per day. Assume the transformed response has a roughly linear conditional mean over the observed range, the residuals show no serious pattern or change in spread, observations are independent, and the fitted value considered is positive.

At \(x=5\) hours, the fitted square-root outcome is

\hat{\eta}(5)=3.0+0.40(5)=5.0\ \sqrt{\text{liters per day}}.

Squaring gives the original-scale plug-in prediction:

\hat{\eta}(5)^2=(5.0)^2=25.0\ \text{liters per day}.

The slope \(0.40\) means that each additional hour per day is associated with an estimated increase of \(0.40\) square-root liters per day in the conditional mean of \(\sqrt{Y}\), under this fitted model. It does not mean water use increases by \(0.40\) liters per day. Also, \(25.0\) liters per day is the squared fitted value; without information about variation around the fitted transformed mean, it should not be described as the estimated conditional arithmetic mean of water use.

Translating Changes Back to the Outcome Scale

The original-scale change between two predictor values is found by calculating a fitted transformed value at each point, squaring each one, and then taking their difference. Because squaring is nonlinear, this change depends on the fitted values at both points. Even if the transformed-scale regression has a constant slope, the corresponding original-scale change per predictor unit generally varies along the range of \(x\).

Result: For two predictor values \(x_1\) and \(x_2\), the difference between the squared fitted values is

\hat{\eta}(x_2)^2-\hat{\eta}(x_1)^2 = \bigl[\hat{\eta}(x_2)-\hat{\eta}(x_1)\bigr] \bigl[\hat{\eta}(x_2)+\hat{\eta}(x_1)\bigr].

If the fitted transformed mean is linear, \(\hat{\eta}(x)=\hat{\beta}_0+\hat{\beta}_1x\), its squared value is a quadratic function of \(x\). Where the fitted mean is positive, the local rate of change of the squared fitted value is \(2\hat{\eta}(x)\hat{\beta}_1\) original outcome units per predictor unit. This is a rate of change in the squared fitted value; it is not necessarily the rate of change in the conditional arithmetic mean.

Worked Example: Comparing Two Predictor Values

Use the hypothetical fitted model \(\widehat{\sqrt{Y}}=2.0+0.50x\), where \(Y\) is a nonnegative outcome measured in counts and \(x\) is a predictor measured in units. Suppose diagnostics support the linear mean and roughly constant residual spread over the observed range, observations are independent, and the fitted transformed means below are positive. Compare \(x=2\) with \(x=4\).

At \(x=2\), the fitted square-root count is \(2.0+0.50(2)=3.0\), so the squared fitted value is \(3.0^2=9.0\) counts. At \(x=4\), it is \(2.0+0.50(4)=4.0\), giving \(4.0^2=16.0\) counts. The change in squared fitted values is therefore

16.0-9.0=7.0\ \text{counts}.

The same calculation using the difference-of-squares identity gives \((4.0-3.0)(4.0+3.0)=1.0(7.0)=7.0\) counts. Across the two-unit increase in \(x\), this is an average change of \(7.0/2=3.5\) counts per predictor unit in the squared fitted values. The transformed-scale slope is \(0.50\) square-root counts per predictor unit, which is not the same quantity. Neither original-scale calculation alone establishes a change in conditional arithmetic means unless the predictions account for the conditional variation in \(\sqrt{Y}\).

Why Squaring Does Not Generally Recover the Mean

A key distinction follows from the conditional variance identity. Write \(\eta(x)=E(\sqrt{Y}\mid x)\). Since \(Y=(\sqrt{Y})^2\), the conditional mean of \(Y\) is the conditional second moment of \(\sqrt{Y}\), not merely the square of its conditional mean. Expanding the definition of variance gives a short justification:

\begin{aligned} \operatorname{Var}(\sqrt{Y}\mid x) &=E\!\left[\bigl(\sqrt{Y}-\eta(x)\bigr)^2\mid x\right]\\ &=E(Y\mid x)-2\eta(x)E(\sqrt{Y}\mid x)+\eta(x)^2\\ &=E(Y\mid x)-\eta(x)^2. \end{aligned}

Rearranging yields

E(Y\mid x)=\eta(x)^2+\operatorname{Var}(\sqrt{Y}\mid x).

The added variance is nonnegative. Therefore, when the conditional variance is positive, squaring the transformed conditional mean gives a value below the conditional arithmetic mean. If the variance is constant, an estimate of that variance can be added to the squared fitted value for an arithmetic-mean prediction under the model. If variance changes with \(x\), the correction must also depend on \(x\). Simply squaring the fitted value does not supply that correction.

Worked Example: An Arithmetic-Mean Correction

Suppose a hypothetical regression estimates the conditional mean of \(\sqrt{Y}\) at a particular predictor value as \(4.0\) square-root units. Suppose the conditional variance of \(\sqrt{Y}\) at that value is estimated as \(0.36\) squared square-root units. Assume the transformed-scale conditional mean and variance estimates are appropriate at this predictor value, and that the observations are independent with no important residual pattern. The squared fitted value is

(4.0)^2=16.0\ \text{units of }Y.

Using the conditional-moment identity, the estimated arithmetic mean on the original scale is

16.0+0.36=16.36\ \text{units of }Y.

Thus \(16.0\) is the squared transformed-scale fitted mean, while \(16.36\) is the arithmetic-mean estimate after accounting for the stated conditional variance. The variance estimate has squared square-root units, which are units of \(Y\), so it can be added to the squared fitted value. In practice, uncertainty in both fitted quantities also matters; this calculation is a point-prediction correction, not an interval.

Practical Interpretation and Common Pitfalls

A square-root model can be useful when the transformed conditional mean is reasonably linear and the residual behavior is more suitable for regression than it is on the original scale. The original-scale interpretation should nevertheless be reported at predictor values that are scientifically relevant. For a change, give the two predictor values, the corresponding fitted transformed values, their squares, and the difference in original units.

Pay particular attention to the fitted transformed mean. Since \(\sqrt{Y}\) cannot be negative, a negative fitted value is incompatible with the response's range. Squaring a negative fitted value produces a positive number, but that arithmetic does not make the prediction sensible: the square would conceal the invalid transformed-scale fit. Negative fitted values or predictions far outside the observed predictor range are reasons to question the model's use there, rather than to rely on squaring as a repair.

Mistake: “The slope is \(0.50\), so \(Y\) increases by \(0.50\) units for each one-unit increase in \(x\).” Correction: The slope describes change in the conditional mean of \(\sqrt{Y}\). Original-scale changes must be calculated from squared fitted values at specified predictor values.
Mistake: “The squared fitted value is the conditional arithmetic mean of \(Y\).” Correction: The conditional arithmetic mean also includes \(\operatorname{Var}(\sqrt{Y}\mid x)\). Squaring alone gives a plug-in prediction based on the transformed mean.
Mistake: “A negative fitted square-root value is fine because its square is positive.” Correction: The transformed outcome cannot be negative. The resulting squared number does not resolve a model fit that is inappropriate at that predictor value.

Always attach units to the result. If \(Y\) is measured in liters, then \(\sqrt{Y}\) is measured in square-root liters, and squaring a fitted value returns liters. A transformed-scale coefficient and an original-scale difference are different kinds of quantities, even though both come from the same regression.

Check Your Understanding

Answer each question using the distinction between the transformed response and the original outcome.

A fitted square-root outcome is \(6\) square-root units. What is its squared plug-in prediction on the original scale, and what additional quantity is needed to estimate the conditional arithmetic mean?
In a square-root outcome regression, what does a slope of \(0.3\) describe? Why is it not generally a constant change in the original outcome?
A fitted transformed mean is \(3\) at \(x_1\) and \(5\) at \(x_2\). Calculate the difference between the squared fitted values.
Why can a negative fitted value for \(\sqrt{Y}\) not be made acceptable merely by squaring it?
If \(\operatorname{Var}(\sqrt{Y}\mid x)\) is larger at one predictor value than another, why might adding one constant correction to every squared fitted value be inappropriate?