Shrinkage of Selected Treatment-Effect Estimates

Why Shrink a Selected Estimate?

In a multi-arm, multi-stage trial, treatment effects are often estimated for several arms, while only arms meeting interim or final selection rules receive attention. An arm selected partly because its estimated effect was unusually large will tend to have an estimate that overstates its true effect. This is the winner’s curse: selection makes positive estimation errors more likely among the reported winners.

Shrinkage addresses that exaggeration by moving a selected estimate partway toward a reference value. It is a transparent compromise between reporting the estimate unchanged and replacing it with the reference. The amount of shrinkage is controlled by a factor chosen in advance for the estimator being described here. Shrinkage can reduce selection-driven exaggeration when the reference is appropriate, but it is not, by itself, a general selection-adjustment procedure.

Definition: Let $\hat{\theta}_j$ estimate the treatment effect $\theta_j$ for arm $j$, and let $m$ be a reference value. For a fixed shrinkage factor $0\leq\lambda\leq1$, define the shrinkage estimator $$ \tilde{\theta}_j=m+\lambda(\hat{\theta}_j-m). $$ Equivalently, $\tilde{\theta}_j=\lambda\hat{\theta}_j+(1-\lambda)m$. When $\lambda=1$, there is no shrinkage; when $\lambda=0$, the estimate is replaced by the reference.

The reference $m$ might be the null effect, often zero, if effects are expressed as a difference from control. It could instead be another scientifically justified value. The interpretation depends on the effect scale and units: shrinking a difference in mean response toward zero is different from shrinking a log odds ratio toward zero, even though both use the same algebra. The reference and shrinkage factor should be specified as part of the estimation rule; a data-dependent choice requires additional justification.

Conditional Bias and Error

Let $A$ denote a specified selection event—for example, that an arm is selected at an interim analysis or that a particular full selection path occurs. The same calculations apply whether $A$ refers to one selection decision or a more detailed path. Write $\mu_A=E(\hat{\theta}_j\mid A)$ and $V_A=\operatorname{Var}(\hat{\theta}_j\mid A)$. These conditional quantities describe the estimate among repetitions of the trial that satisfy $A$.

Since $m$ and $\lambda$ are fixed, taking a conditional expectation of the estimator gives

E(\tilde{\theta}_j-\theta_j\mid A) =\lambda(\mu_A-\theta_j)+(1-\lambda)(m-\theta_j).

Thus the conditional bias is a weighted average of the original conditional bias and the reference’s displacement from the true effect. In the special case $m=\theta_j$, the shrinkage estimator’s conditional bias is exactly $\lambda$ times the original bias. If selection has pushed the expected estimate above the truth, this pulls the bias toward zero. If $m$ is far from $\theta_j$, however, the second term can offset or exceed that improvement.

Result: Conditional on $A$, the mean squared error of the shrinkage estimator is $$ E\{(\tilde{\theta}_j-\theta_j)^2\mid A\} =\lambda^2 V_A+ \left[\lambda(\mu_A-\theta_j)+(1-\lambda)(m-\theta_j)\right]^2. $$ This follows from the identity “mean squared error = variance + squared bias,” because $\operatorname{Var}(\tilde{\theta}_j\mid A)=\lambda^2 V_A$.

When the reference equals the true effect, the squared conditional bias is also multiplied by $\lambda^2$, and the conditional variance by $\lambda^2$. Conditional mean squared error is then multiplied by $\lambda^2$. In practice the true effect is not known, so this special case explains the mechanism rather than supplying a usable tuning rule. For a reference that differs from the truth, the bias term can grow even while the variance falls.

Worked Examples

Worked Example: Shrinking a Selected Estimate Toward Zero

Suppose a hypothetical MAMS trial reports a selected treatment-effect estimate of $2.40$ units, with standard error $0.80$ units. Consider a fixed shrinkage factor $\lambda=0.65$ and a zero-effect reference $m=0$. This calculation describes the point estimate; it does not make an interval or p-value selection-adjusted.

The shrinkage rule is $$ \tilde{\theta}=m+\lambda(\hat{\theta}-m). $$ Substitution gives $$ \tilde{\theta}=0+0.65(2.40-0)=1.56\text{ units}. $$ The estimate is pulled down by $2.40-1.56=0.84$ units. It remains positive, but the reported magnitude is 35% smaller than the unshrunk estimate because $1-\lambda=0.35$. Whether that is a useful correction depends on how well zero represents the selected arm’s true effect; this arithmetic alone cannot establish that the estimate is closer to the truth.

Worked Example: Expected Winner’s-Curse Reduction in Two Arms

For a simple hypothetical illustration, suppose two arms have equal true effects of zero and their estimates $X_1$ and $X_2$ are independent, each distributed as $N(0,1)$ in standardized effect units. Select the arm with the larger estimate, so the selected estimate is $M=\max(X_1,X_2)$. Shrink it toward zero with $\lambda=0.60$. The equal-effect, independent, normal assumptions make this a deliberately simple illustration; shared controls or unequal standard errors would change the distribution.

The expected selected estimate is positive, despite both true effects being zero. To calculate it, use $$ \max(X_1,X_2)=\frac{X_1+X_2+|X_1-X_2|}{2}. $$ Since $E(X_1+X_2)=0$, and $X_1-X_2\sim N(0,2)$, the mean absolute value is $E|X_1-X_2|=2/\sqrt{\pi}$. Therefore $$ E(M)=\frac{1}{2}\frac{2}{\sqrt{\pi}}=\frac{1}{\sqrt{\pi}}\approx0.5642. $$ The unshrunk winner’s expected exaggeration is $0.5642$ standardized units. The shrunk estimate is $0.60M$, so its expected exaggeration is $$ E(0.60M)=0.60(0.5642)=0.3385. $$ The expected exaggeration falls by $0.5642-0.3385=0.2257$, or 40%. This is a reduction, not elimination: the expected estimate remains above the true effect of zero. For a realized pair of estimates $1.10$ and $0.30$, the selected estimate is $1.10$, and its shrunk value is $0.60(1.10)=0.66$ standardized units.

Worked Example: When the Reference Is Too Far From the Truth

Suppose, hypothetically, that conditional on selection an arm’s true effect is $3.0$ units, and its estimate has conditional mean $3.2$ units and conditional variance $0.64$ square units. The small positive conditional bias is consistent with modest upward selection. Now use the zero reference with $\lambda=0.50$.

The unshrunk conditional bias is $3.2-3.0=0.2$ units. The shrunk estimate has conditional mean $$ 0+0.50(3.2-0)=1.6\text{ units}, $$ so its conditional bias is $1.6-3.0=-1.4$ units. The reference has pulled the estimate past the truth, creating substantial downward bias.

The unshrunk conditional mean squared error is $$ 0.64+(3.2-3.0)^2=0.64+0.04=0.68\text{ square units}. $$ For the shrunk estimate, the conditional variance is $0.50^2(0.64)=0.16$ square units, and its squared bias is $(-1.4)^2=1.96$ square units. Its conditional mean squared error is therefore $0.16+1.96=2.12$ square units, which is larger than $0.68$. Here the variance reduction does not compensate for the error introduced by an unsuitable reference.

What Shrinkage Does—and Does Not—Correct

Shrinkage is a rule for modifying an estimate, not a guarantee of conditional unbiasedness. For example, if $m=\theta_j$, the conditional bias becomes $\lambda$ times the original conditional bias; it is zero only if the original conditional bias is zero or $\lambda=0$. If $m\ne\theta_j$, the conditional bias combines the selection-driven bias with the reference’s error. Consequently, the phrase “shrinkage corrects the winner’s curse” is too strong unless the assumptions and target of the correction are stated.

The selection event matters. An estimate’s distribution conditional on being the winner need not match its distribution conditional on a complete multi-stage selection path. The earlier material on conditioning on the winner or the full selection path distinguishes these targets. The bias and error formulas above remain valid for either target, but $\mu_A$ and $V_A$ must correspond to the event actually being considered. A calculation that uses unconditional moments to claim conditional correction can be misleading.

In real MAMS designs, estimates may have unequal standard errors, be correlated through a shared control group, or be combined across stages. The simple independent, equal-variance example should not be used to calculate the expected winner’s bias in those settings. The linear shrinkage formula itself still applies to a chosen estimate; what changes is the conditional distribution needed to assess its bias and error.

Approach	What it does	Main caution
No shrinkage	Reports $\hat{\theta}_j$ unchanged.	Can retain upward exaggeration caused by selection.
Fixed-factor shrinkage	Moves the estimate toward $m$ by a prespecified fraction.	Can leave bias or introduce substantial bias if $m$ is unsuitable.
Selection-adjusted estimation	Models or conditions on the selection mechanism to target a specified post-selection estimand.	Depends on correctly representing the selection process and its assumptions.

Selection-adjusted methods, including conditional likelihood and median-unbiased estimation, address the selection mechanism more directly. They may be preferable when a conditional inferential target is required. Fixed-factor shrinkage is simpler to calculate and communicate, and may be useful as a descriptive stabilization rule, but it should not be presented as a substitute for selection-aware inference. Nor does shrinking a point estimate automatically provide a valid confidence interval or p-value after selection.

Diagnostics and Common Pitfalls

Before interpreting a shrunk estimate, check which arm and which selection event define it, what effect scale and units are being used, and why the reference is defensible. If a reference is estimated from the same selected results, its uncertainty and dependence on selection may matter. The present fixed-reference calculation does not account for those complications. It is also useful to report the unshrunk and shrunk estimates together, so the amount of movement is visible rather than hidden.

Pitfall: treating shrinkage as a selection correction. Multiplying an estimate by $\lambda$ does not account for unequal standard errors, correlated estimates, or a multi-stage selection path. State the selection target and assess the conditional distribution if making a bias claim.
Pitfall: assuming a smaller estimate is automatically better. Shrinkage lowers variance when $0\leq\lambda<1$, but can add squared bias. As the third example shows, total mean squared error can increase.
Pitfall: using the null as a universal reference. Zero may be reasonable for some effect scales and purposes, but can be far from a genuinely effective treatment’s effect. Justify the reference scientifically and interpret the shrunken value in context.
Pitfall: interpreting the shrinkage factor as evidence. A factor such as $0.60$ states how the estimator is constructed; it is not a probability that the estimate is correct or a confidence level.

Key takeaway: A fixed shrinkage estimator is a weighted average of a selected estimate and a reference value. It reduces winner’s-curse exaggeration when the reference is suitably close to the truth and the selection-induced bias points away from it; it can leave bias or make error worse otherwise.

Check Your Understanding

Answer each question, then compare with the brief solutions below.

An estimate is $1.80$ response units, the reference is $0.20$ response units, and $\lambda=0.75$. What is the shrunken estimate?
If $m=\theta_j$, what happens to the conditional bias given a selection event $A$?
In the two-arm example, why is the expected selected estimate positive when both true effects are zero?
Give one reason that a zero reference could make a selected estimate worse rather than better.
Does shrinking a point estimate by itself make its usual confidence interval valid after arm selection? Explain briefly.

Solutions. (1) $\tilde{\theta}=0.20+0.75(1.80-0.20)=1.40$ response units. (2) The conditional bias is multiplied by $\lambda$: $E(\tilde{\theta}_j-\theta_j\mid A)=\lambda E(\hat{\theta}_j-\theta_j\mid A)$. (3) Selecting the larger estimate favors positive estimation errors; for two independent standard normal estimates its expectation is $1/\sqrt{\pi}$, not zero. (4) The true effect may be far from zero, so pulling toward zero can introduce substantial bias, as in the third worked example. (5) No. Shrinking the point estimate does not account for the selected sampling distribution or establish post-selection coverage; interval construction requires separate justification.

Approach	What it does	Main caution
No shrinkage	Reports \(\hat{\theta}_j\) unchanged.	Can retain upward exaggeration caused by selection.
Fixed-factor shrinkage	Moves the estimate toward \(m\) by a prespecified fraction.	Can leave bias or introduce substantial bias if \(m\) is unsuitable.
Selection-adjusted estimation	Models or conditions on the selection mechanism to target a specified post-selection estimand.	Depends on correctly representing the selection process and its assumptions.