Copas-Jackson Bound in Meta-Analysis

• The Copas-Jackson bound is a worst-case bias bound that quantifies publication bias in random-effects meta-analysis using a nonparametric selection model.
• It employs both analytic closed-form expressions and generalized numerical methods via Monte Carlo simulation to address various selection mechanisms.
• The bound guides robust sensitivity analyses by contrasting bias estimates under monotone standard error and t-statistics driven selection procedures.

Searching arXiv for papers on the Copas-Jackson bound and closely related selection-model work in meta-analysis. arXiv search query: "Copas-Jackson bound publication bias meta-analysis" The Copas-Jackson bound (C-J bound) is a worst-case bound for publication-bias-induced bias in random-effects meta-analysis. In the formulation revisited in recent work, the meta-analytic outcome satisfies $y \sim N(\mu,\sigma^2)$ with $\sigma^2=s^2+\tau^2$, and publication occurs according to a selection probability $p(y,\sigma)$ (Hu et al., 25 Aug 2025). The original C-J construction provides an analytic envelope for the bias of the naive observed-data estimator over a nonparametric class of selection models defined by monotonicity of the marginal publication probability in study standard error. More recent work extends this construction to a broader class of selection mechanisms, especially those driven by $t$-statistics or statistical significance, while separate work on Copas’ parametric selection method emphasizes that publication-bias adjustment can be highly sensitive to the assumed selection mechanism (Hu et al., 25 Aug 2025, Almalik et al., 2020).

1. Statistical formulation

In the C-J framework, the random-effects meta-analysis model is

$$y \sim N(\mu,\sigma^2), \qquad \sigma^2=s^2+\tau^2,$$

where $s$ is the within-study standard error and $\tau^2$ is the between-study heterogeneity (Hu et al., 25 Aug 2025). Publication bias is represented through a selection probability $p(y,\sigma)$, and the marginal publication probability given $\sigma$ is defined by

$$p(\sigma)=\Pr(\text{selected}\mid \sigma)=\int \Pr(\text{selected}\mid y,\sigma)f(y\mid \sigma)\,dy.$$

The key structural restriction in the original C-J bound is Assumption $\sigma^2=s^2+\tau^2$0: $\sigma^2=s^2+\tau^2$1 is a non-increasing function of $\sigma^2=s^2+\tau^2$2 (Hu et al., 25 Aug 2025). Its interpretation is that studies with smaller standard errors, typically larger studies, are more likely to be published.

This formulation is explicitly nonparametric at the level of the publication mechanism. Rather than specifying a fully parametric selection model, it bounds the bias induced by publication selection over a class of admissible selection functions satisfying the monotonicity condition. That feature distinguishes the C-J bound from parametric Copas-style sensitivity analyses, which posit a latent selection equation with fixed sensitivity parameters and then vary those parameters over a grid (Almalik et al., 2020).

2. Original analytic bound

Under Assumption $\sigma^2=s^2+\tau^2$3, Copas and Jackson derived an analytic worst-case bound for the publication-bias-induced bias $\sigma^2=s^2+\tau^2$4 of the naive observed-data estimator (Hu et al., 25 Aug 2025). In the notation reproduced in the later paper, the bound is

$\sigma^2=s^2+\tau^2$5

where $\sigma^2=s^2+\tau^2$6 is the overall marginal publication probability, $\sigma^2=s^2+\tau^2$7 is the population density of study standard errors, and $\sigma^2=s^2+\tau^2$8 denotes expectation among published studies.

With observed standard errors $\sigma^2=s^2+\tau^2$9, the empirical C-J bounds are given by

$p(y,\sigma)$0

The practical significance of the original C-J result is that it yields a closed-form worst-case expression. This allows publication-bias sensitivity analysis without fully specifying the selection process. The bound is symmetric around zero in the reported formulation, and its width depends on the published-study distribution of $p(y,\sigma)$1 and on the assumed overall publication probability $p(y,\sigma)$2 (Hu et al., 25 Aug 2025).

3. Scope of the original model class

The main limitation of the original C-J bound is the restrictiveness of Assumption $p(y,\sigma)$3. In the later generalization, this restriction is described as allowing only selection mechanisms whose marginal publication probability is monotone in standard error (Hu et al., 25 Aug 2025). That accommodates some intuitive size-based publication preferences, but it does not cover all practically relevant publication processes.

Two model families are particularly informative in this regard. The later paper examines the Copas-Heckman selection model and several $p(y,\sigma)$4-statistics type selection models, including 1-logit, mlogit, exp, 2-probit, and 2-logit (Hu et al., 25 Aug 2025). It shows that the Copas-Heckman model satisfies $p(y,\sigma)$5 because its marginal selection probability is monotone in $p(y,\sigma)$6 when $p(y,\sigma)$7. By contrast, at least some $p(y,\sigma)$8-statistics type models do not satisfy $p(y,\sigma)$9. For the 2-probit model, the derived marginal representation can have a turning point, so it is not globally non-increasing in $t$0.

A concise comparison is useful.

Feature	Original C-J bound	Generalized C-J-type bound
Structural assumption	$t$1: $t$2 non-increasing in $t$3	$t$4
Model classes covered	Copas-Heckman-type mechanisms	Copas-Heckman-type and $t$5-statistics type mechanisms
Form of bound	Closed-form analytic bound	Monte Carlo + nonlinear programming

This restriction matters because many publication processes are driven more directly by statistical significance or $t$6-statistics than by standard error alone. The later paper therefore treats the original C-J class as too narrow for a broad robustness analysis (Hu et al., 25 Aug 2025).

4. Generalization via the $t$7-representation

To weaken the monotonicity restriction, the generalized framework rewrites the random-effects model as

$t$8

and introduces the $t$9-representation of the selection model (Hu et al., 25 Aug 2025). For a selection function $$y \sim N(\mu,\sigma^2), \qquad \sigma^2=s^2+\tau^2,$$0, it defines

$$y \sim N(\mu,\sigma^2), \qquad \sigma^2=s^2+\tau^2,$$1

followed by

$$y \sim N(\mu,\sigma^2), \qquad \sigma^2=s^2+\tau^2,$$2

The relaxed structural condition is Assumption $$y \sim N(\mu,\sigma^2), \qquad \sigma^2=s^2+\tau^2,$$3: there exists $$y \sim N(\mu,\sigma^2), \qquad \sigma^2=s^2+\tau^2,$$4 such that either $$y \sim N(\mu,\sigma^2), \qquad \sigma^2=s^2+\tau^2,$$5 is non-increasing in $$y \sim N(\mu,\sigma^2), \qquad \sigma^2=s^2+\tau^2,$$6 for $$y \sim N(\mu,\sigma^2), \qquad \sigma^2=s^2+\tau^2,$$7 and non-decreasing for $$y \sim N(\mu,\sigma^2), \qquad \sigma^2=s^2+\tau^2,$$8, or the reverse monotonicity pattern holds (Hu et al., 25 Aug 2025). The paper characterizes this as a single turning point structure.

The associated model class is

$$y \sim N(\mu,\sigma^2), \qquad \sigma^2=s^2+\tau^2,$$9

The generalized theory then considers the class

$s$0

The importance of this extension is that it captures the behavior of many $s$1-statistics type selection rules. The paper proves that all $s$2-statistics type selection models in its list satisfy $s$3, stated as

$s$4

This broadens the admissible selection mechanisms from monotone standard-error selection to a family that can encode threshold-like or turning-point behavior around a latent mean parameter (Hu et al., 25 Aug 2025).

5. Bias representation and numerical construction

Under the generalized model class, the overall marginal selection probability becomes

$s$5

and the bias is represented by an integral expression involving $s$6, $s$7, and the latent-normal variables $s$8 and $s$9 (Hu et al., 25 Aug 2025). The published-study density of $\tau^2$0 is

$\tau^2$1

Unlike the original C-J bound, this generalized bias expression does not simplify to a closed analytic bound (Hu et al., 25 Aug 2025). The paper therefore develops a computational procedure based on Monte Carlo approximation and constrained optimization. The inner integrals are approximated with draws $\tau^2$2 and $\tau^2$3, and the extremal bias is characterized through an optimization problem $\tau^2$4 with constraints:

• (C1) $\tau^2$5
• (C2) averaging identities from the definitions of $\tau^2$6 and $\tau^2$7
• (C3) monotonicity structure implied by Assumption $\tau^2$8
• (C4) overall selection probability

Because unpublished studies are unavailable, the paper replaces (C4) with

$\tau^2$9

obtaining a relaxed problem $p(y,\sigma)$0, justified by a Cauchy–Schwarz argument (Hu et al., 25 Aug 2025). The optimization is solved using nonlinear programming with linear constraints, implemented with software such as the OPTMODEL procedure in SAS, with $p(y,\sigma)$1 in the reported examples.

The resulting bounds are denoted $p(y,\sigma)$2 and $p(y,\sigma)$3. The paper’s practical recommendation is the extended bound

$p(y,\sigma)$4

which combines the original analytic C-J bound with the broader numerical bound over $p(y,\sigma)$5 (Hu et al., 25 Aug 2025).

6. Empirical performance and interpretive consequences

The simulation evidence in the generalized-bound paper is organized around two data-generating families: Copas-Heckman and 2-probit selection models (Hu et al., 25 Aug 2025). For Copas-Heckman scenarios, the original C-J bound performed well and covered the true bias in most cases, especially when $p(y,\sigma)$6. For 2-probit selection, the original C-J bound often failed badly: in one setting, the bias exceeded the C-J bound in over 90% of simulated datasets, and in another, in over 60%. The proposed extended bound had much better coverage; failures were mostly limited to extremely low $p(y,\sigma)$7 values, and the bound length was typically about 1 to 3 times the C-J bound length.

The same paper reports two real-data applications. In the prophylactic corticosteroids meta-analysis, with 14 RCTs and essentially no between-study heterogeneity ($p(y,\sigma)$8), several standard $p(y,\sigma)$9-statistics type sensitivity analyses produced adjusted estimates not covered by the C-J bound, whereas the extended bound almost always covered the adjusted estimates across $\sigma$0 down to $\sigma$1 (Hu et al., 25 Aug 2025). In the Clopidogrel meta-analysis, with 12 studies and small heterogeneity ($\sigma$2), the C-J bound again failed to cover several sensitivity-analysis estimates, while the extended bound covered almost all adjusted estimates for realistic $\sigma$3.

A related, but methodologically distinct, line of evidence comes from work on Copas’ method, which is a selection-model approach for correcting publication bias in an aggregated-data meta-analysis (Almalik et al., 2020). That paper does not derive or analyze a Copas-Jackson Bound explicitly; instead, it studies the broader Copas selection-model framework. Its simulations show that Copas’ method performs best when the data are generated under Copas selection, becomes less effective when selection is based on significance, and performs poorly when selection is based on standardized effect size $\sigma$4, especially with heterogeneity. The paper’s stated practical message is that Copas’ method is not robust to publication-bias mechanisms other than the one it assumes and recommends improving it “to make it more robust against different forms of publication bias” (Almalik et al., 2020).

Taken together, these results support a precise interpretation. The original C-J bound is most credible when the publication process is plausibly monotone in standard error, while the generalized C-J-type bound is intended for a broader class that includes $\sigma$5-statistics-driven selection (Hu et al., 25 Aug 2025). A plausible implication is that publication-bias sensitivity analysis should be read as mechanism-conditional unless the admissible selection class is itself sufficiently broad.

7. Terminological ambiguity and unrelated uses of “Copas-Jackson”

The expression “Copas-Jackson bound” refers, in the meta-analysis literature, to the publication-bias bound associated with Copas and Jackson and its later generalizations (Hu et al., 25 Aug 2025). It should not be confused with unrelated uses of the surname Jackson in other fields.

A prominent example is the comment on Jackson and Sonnenschein’s linked-decisions mechanism, which corrects a bound in the definition of approximate truthfulness (Ball et al., 2022). In that mechanism-design setting, the original bound-based notion is

$\sigma$6

whereas the corrected notion is

$\sigma$7

The paper shows that permutation-truthful strategies imply the loosened inequality, not the original exact equality, and that this weaker bound is still sufficient because the fraction of lies vanishes as $\sigma$8 (Ball et al., 2022).

That correction is conceptually separate from publication-bias sensitivity analysis. It concerns incentive compatibility, permutation-truthfulness, and asymptotic efficiency in linked decisions, not worst-case bounds for publication selection in meta-analysis. The overlap in surnames can therefore generate a nomenclatural confusion that is substantive only at the level of terminology, not theory (Ball et al., 2022).