Post-Screening Inference (PSI)

• Post-Screening Inference (PSI) is a framework for valid statistical inference performed after data-driven screening, adjusting for the randomness introduced by selection.
• It distinguishes between fixed inferential targets and those altered by screening, employing methods like conditioning on the selection event and simultaneous protection over candidate models.
• Key PSI frameworks such as PoSI and conditional selective inference provide theoretical guarantees and practical techniques to mitigate biases in regression, causal discovery, and high-dimensional settings.

Post-Screening Inference (PSI) denotes inference carried out after a data-dependent screening step has used the same data to determine which variables, models, regions, graphs, or effects will be examined further. Its central concern is that classical tests and confidence intervals are justified when the inferential target is fixed a priori, whereas after screening the selected object is random and the null law is altered by the screening event. The modern PSI literature therefore studies procedures that remain valid either by conditioning on the realized selection event, by protecting simultaneously over all candidate submodels, or by enlarging reported sets so that validity survives screening and, in some settings, arbitrary stopping times (Berk et al., 2013, Lee et al., 2014, Toyoda et al., 20 Aug 2025).

1. Definition, scope, and inferential targets

In the cited literature, “post-selection inference” is the broader label, while “post-screening inference” is especially apt when the first stage is a screening rule such as marginal feature ranking, graph discovery, top-$m$ retention, or region selection (Yamada et al., 2016, Chang et al., 2024). The common structure is two-stage: first, data are used to reduce a candidate space; second, inference is performed on the retained object. What fails is the assumption that the inferential target and the testing rule were fixed independently of the observed sample.

A central refinement of PSI is that screening can affect not only the estimator but also the parameter being estimated. In Berk, Brown, Buja, Zhang, and Zhao’s PoSI formulation, the parameter attached to a selected regression submodel $M$ is the submodel-specific projection

$\beta_M=(X_M^T X_M)^{-1}X_M^T\mu,$

rather than a single coefficient vector shared by all models (Berk et al., 2013). In causal discovery, this distinction becomes sharper: one may target the true fixed population causal effect $\beta(G)$, or instead a selected functional such as $\beta(\widehat G)$, whose meaning changes with the discovered graph (Chang et al., 2024). In feature-screening PSI based on HSIC or MMD, the post-screening null is typically feature-wise and marginal, for example

$$H_{0,m}:\mathrm{HSIC}(X_m,Y)=0 \mid \{\text{feature }m\text{ selected}\},$$

or its MMD analogue, rather than a partial-regression coefficient in a joint model (Yamada et al., 2016, Yamada et al., 2018).

This suggests that PSI is defined not only by a screening event but also by a choice of target. The literature repeatedly distinguishes inference for a fixed population quantity, inference for a selected-model parameter, and inference for a selected or data-dependent functional. Much of the technical and conceptual variation across PSI methods is a consequence of which of these targets is being protected.

2. Foundational frameworks

One major framework is PoSI, or post-selection inference by simultaneous inference. Berk et al. reduce the problem of arbitrary data-driven model selection to simultaneous protection over all coefficients in all candidate submodels. Their key constant is

$$K(X,\mathcal M,\alpha,r)=\inf\left\{k:\mathbb P\!\left[\max_{M\in\mathcal M}\max_{j\in M}|t_{j\cdot M}|\le k\right]\ge 1-\alpha\right\},$$

which yields

$$\mathbb P\!\left[\beta_{j\cdot \hat M}\in CI_{j\cdot \hat M}(K)\ \forall j\in \hat M\right]\ge 1-\alpha \qquad \forall \hat M.$$

Under fixed design, Gaussian homoscedastic errors, and an independent variance estimator, this guarantee is finite-sample exact and valid for arbitrary selection procedures, including informal or partly unspecified ones (Berk et al., 2013).

A second framework is conditional selective inference. In the Gaussian polyhedral setting, the selection event is written as

$Az\le b,$

and inference is conditioned on that event. For marginal screening in linear regression, the event that a set of variables with given signs was selected is a polyhedral event in the response vector, which yields an exact truncated-Gaussian law for a selected contrast (Lee et al., 2014). The same logic underlies later selective-inference developments for lasso-style procedures, generalized linear models, and top-$k$ screening rules.

The distinction between these frameworks is substantive. PoSI does not condition on the realized selection event and instead buys “simultaneity insurance” over all possible submodels (Berk et al., 2013). Conditional selective inference conditions on the realized event and is usually less conservative, but it requires an explicit model of the selection rule. Tibshirani et al. sharpen this viewpoint by arguing that valid post-selection inference should condition only on the information actually used in hypothesis generation; conditioning on more, such as unused sign information, can produce unnecessarily wide intervals (Liu et al., 2018).

A related PoSI-style construction appears in post region selection. There the selected object is a spatial interval or rectangle, the selection rule may be arbitrary, and validity is obtained by simultaneous coverage over the full admissible region class rather than by conditioning on a particular region-selection event (Bontemps et al., 13 Jun 2025). This broadens PSI beyond regression submodels to geometric screening problems.

3. Selective laws after screening

The central selective-inference lemma used across many PSI constructions is the truncated-normal pivot. If

$M$0

and the selection event is polyhedral,

$M$1

then for any contrast vector $M$2,

$M$3

where

$M$4

and

$M$5

These formulas are the basis of exact selective $M$6-values and interval inversion in the Gaussian case (Lee et al., 2014, Tsai et al., 2018).

Many screening rules encountered in practice are polyhedral. In top-$M$7 feature screening by a score vector $M$8, the event that every selected feature outranks every unselected feature can be written as $M$9, with one row per selected–unselected comparison (Yamada et al., 2016, Yamada et al., 2018). In GAN selection by minimum discrepancy, the event that generator $\beta_M=(X_M^T X_M)^{-1}X_M^T\mu,$0 has the smallest estimated $\beta_M=(X_M^T X_M)^{-1}X_M^T\mu,$1 is

$\beta_M=(X_M^T X_M)^{-1}X_M^T\mu,$2

again a polyhedral event (Tsai et al., 2018). In high-dimensional logistic regression after marginal screening, the event that the screened set and its sign pattern were selected is also affine, which permits an asymptotic truncated-normal pivot after suitable score-based linearization (Umezu et al., 2019).

The technical challenge outside Gaussian linear models is that the screening statistic is often not exactly Gaussian. Several PSI papers therefore build asymptotically normal “source statistics” that can replace the raw data in the polyhedral lemma. This move is explicit in kernel PSI based on block HSIC, in incomplete-U MMD PSI, and in selective inference for logistic classification after marginal screening (Yamada et al., 2016, Yamada et al., 2018, Umezu et al., 2019). A plausible implication is that much of modern PSI can be read as a search for statistics whose joint law is close enough to multivariate normality to support a truncated-normal selective pivot.

4. Statistical constructions for screened data

Kernel-based PSI extends selective inference beyond linear correlation and scalar Gaussian responses. In hsicInf, each feature is screened by empirical HSIC, the top $\beta_M=(X_M^T X_M)^{-1}X_M^T\mu,$3 features are retained, and the vector of block-HSIC statistics is asymptotically multivariate normal because it averages i.i.d. blockwise terms with fixed block size $\beta_M=(X_M^T X_M)^{-1}X_M^T\mu,$4. This allows selective inference for nonlinear dependence, multivariate regression, and multi-class outputs through kernels, while the hypotheses remain feature-wise marginal independence statements (Yamada et al., 2016).

An analogous construction underlies post-screening inference with MMD. For two-sample feature screening, one computes a feature-wise score vector

$\beta_M=(X_M^T X_M)^{-1}X_M^T\mu,$5

selects the top $\beta_M=(X_M^T X_M)^{-1}X_M^T\mu,$6 features, and conditions on that ranking event. The obstacle is that the complete unbiased MMD U-statistic is degenerate under the null and therefore ill-suited to Gaussian selective inference. The incomplete U-statistics estimator

$\beta_M=(X_M^T X_M)^{-1}X_M^T\mu,$7

is introduced precisely because, under regimes such as $\beta_M=(X_M^T X_M)^{-1}X_M^T\mu,$8, it is asymptotically normal and computationally $\beta_M=(X_M^T X_M)^{-1}X_M^T\mu,$9, which makes the Lee-style PSI machinery applicable (Yamada et al., 2018). The same incomplete-MMD construction is then used to select the lowest-discrepancy GAN and test that selected model against real data after conditioning on the winner-selection event (Tsai et al., 2018).

Selective inference after marginal screening for high-dimensional classification adapts this logic to logistic regression. Variables are screened by $\beta(G)$0, the screening event is affine, and the post-selection logistic MLE is approximated by an asymptotically Gaussian score-based statistic

$\beta(G)$1

Under conditions including $\beta(G)$2, this yields asymptotically valid selective $\beta(G)$3-values for individual selected coefficients (Umezu et al., 2019).

Semiparametric propensity-score analysis provides a different extension. There the selection step is lasso-type variable selection in an inverse-probability-weighted objective, and the paper conditions not only on the selected model but also on treatment assignments and covariates to recover a tractable truncated-normal pivot. Confidence intervals for selected confounding variables are then obtained with asymptotic guarantees, without requiring modeling of nonparametric outcome regression functions in the main construction (Ninomiya et al., 2021).

Sequential screening introduces yet another inferential regime. Sequential Correct Screening outputs a nested sequence of random subsets $\beta(G)$4 that, with high probability, always contain the true top-$\beta(G)$5 set. Post-screening intervals are then reported at a stopping time $\beta(G)$6 using

$\beta(G)$7

and the target guarantee is not per-coordinate selective validity but stopped false coverage rate control,

$\beta(G)$8

for any adapted stopping time (Toyoda et al., 20 Aug 2025). This identifies a distinct PSI tradition built from confidence sequences and e-processes rather than truncated Gaussian laws.

5. Application domains

Causal discovery turns PSI into a post-screening problem over graphs. Constraint-based algorithms such as PC or PC-stable execute sequences of conditional independence tests and output a CPDAG, which in turn determines which covariates are adjusted for in a causal effect estimator. The corresponding post-screening challenge is that naive reuse of the same data for discovery and estimation invalidates confidence intervals, and even sample splitting may protect only a selected functional $\beta(G)$9 if the discovered graph is wrong. The resampling-and-screening method of post-selection inference for causal effects after causal discovery addresses this by perturbing intermediate test statistics, rerunning graph discovery many times, retaining valid graphs, and taking a union of graph-based intervals to obtain asymptotic coverage for the true fixed causal effect $\beta(\widehat G)$0 (Chang et al., 2024).

Region selection produces a geometric form of PSI. In one dimension, the object of interest is the average signal over an interval $\beta(\widehat G)$1,

$\beta(\widehat G)$2

and the problem is to construct intervals for $\beta(\widehat G)$3 after a data-driven segmentation or scanning procedure has selected $\beta(\widehat G)$4. Rather than conditioning on a specific selected region, the region-selection paper controls

$\beta(\widehat G)$5

shows convergence to a Brownian-motion functional $\beta(\widehat G)$6, and derives PoSI-style intervals valid for any selected admissible interval or model of disjoint intervals (Bontemps et al., 13 Jun 2025).

Group testing yields a latent-response version of PSI. In regression for group-testing data, one observes pooled, error-prone test outcomes instead of true individual binary responses, performs lasso-type selection through an EM algorithm, and then seeks inference on the selected covariate effects. The paper extends polyhedral-lemma methodology to this partially observed logistic setting and reports that its post-selection procedure is more reliable than naive inference using the same data for selection and estimation (Shen et al., 16 Apr 2025).

A/B testing supplies a more pragmatic perspective. There the screening rule is often informal—teams focus on statistically significant experiments, segments, or metrics—and the paper studies post-selection correction by conditional maximum likelihood, empirical Bayes, and supervised learning with experiment splitting. Its target is not exact selective validity in the polyhedral sense but reduction of post-selection bias and improvement of interval coverage after significance-based filtering (Deng et al., 2019). This suggests that PSI, in practice, includes both formal conditional-inference methods and calibration methods designed for large experimentation systems.

6. Guarantees, limitations, and related distinctions

The guarantees provided by PSI methods vary sharply. PoSI offers finite-sample exact universal coverage under Gaussian linear-model assumptions, but it is generally conservative because it protects against all possible model selection procedures (Berk et al., 2013). Polyhedral selective inference can be exact and less conservative, but only when the selection rule is explicitly modeled and the relevant statistic is Gaussian (Lee et al., 2014). Kernel PSI, incomplete-MMD PSI, logistic-screening PSI, and semiparametric propensity-score PSI rely instead on asymptotic normality, so their selective validity is approximate in finite samples (Yamada et al., 2016, Yamada et al., 2018, Umezu et al., 2019, Ninomiya et al., 2021).

A recurrent limitation is that screening may alter the estimand, not merely the estimator. This point is explicit in both regression PoSI and causal discovery. Different adjustment sets imply different regression parameters $\beta(\widehat G)$7 in Berk et al., and an incorrect learned graph may redirect inference from the true causal effect $\beta(\widehat G)$8 to a graph-dependent functional $\beta(\widehat G)$9 (Berk et al., 2013, Chang et al., 2024). A common misconception is therefore that sample splitting automatically solves PSI; it can remove direct reuse of data, but it does not by itself guarantee that the post-screening target remains the scientifically intended one.

Another distinction separates sure-screening theory from PSI proper. ExSIS analyzes when correlation-based screening preserves all active variables with high probability, often allowing dimension reduction to about $$H_{0,m}:\mathrm{HSIC}(X_m,Y)=0 \mid \{\text{feature }m\text{ selected}\},$$0 under favorable coherence or sub-Gaussian assumptions. This is highly relevant to downstream PSI because it protects the preprocessing stage, but it does not provide post-screening-valid $$H_{0,m}:\mathrm{HSIC}(X_m,Y)=0 \mid \{\text{feature }m\text{ selected}\},$$1-values or intervals (Ahmed et al., 2017). Support containment is thus a prerequisite for many workflows, not a substitute for inference after screening.

Computational and modeling issues remain central. Covariance estimation for score vectors can be nontrivial; this is explicit in kernel PSI and incomplete-MMD PSI (Yamada et al., 2016, Tsai et al., 2018). Complex or nonlinear screening rules may not yield tractable polyhedral descriptions. Universal procedures can be wide, while conditional procedures can become fragile when the assumed selection event is misspecified. Sequential settings add the further difficulty that fixed-time guarantees need not survive optional stopping, which motivates stopped-FCR control and confidence-sequence-based constructions (Toyoda et al., 20 Aug 2025).

Taken together, these developments define PSI as a family of methods rather than a single theorem. Some variants condition on the realized screening event; some protect simultaneously over all candidates; some enlarge unions of intervals over screened models; and some target average false coverage under arbitrary stopping. What unifies them is the insistence that inference after screening must account for how the reported object became reportable in the first place.