Score-Based Testing Methods

• Score-based tests are hypothesis testing methods that compute the score under the null to verify first-order optimality conditions without fitting the full alternative.
• They utilize observed or expected information to generate chi-squared approximations, addressing negative curvature issues by modifying rejection rules and restoring test power.
• Extensions include approaches for high-dimensional, penalized, and function-valued parameters, integrating orthogonalization and recalibration to ensure robust performance in diverse scenarios.

A score-based test is a hypothesis test that evaluates whether a null-constrained model already satisfies a first-order optimality condition, using a score vector, score process, or score functional computed under the null rather than fitting the full alternative. In regular parametric problems, the classical score test based on either expected or observed information is asymptotically equivalent under $H_0$ to the likelihood-ratio and Wald tests and has a $\chi^2$ null limit; in more recent work, the same logic has been extended to high-dimensional regression, function-valued parameters, mixed models, goodness-of-fit, forecast calibration, change detection, and early stopping for boosting (Karavarsamis et al., 2018).

1. Classical score-test formulation

In the standard formulation, the parameter is partitioned into null-constrained and unrestricted components, and the score is evaluated at the null-constrained maximum likelihood estimate. In the two-sample occupancy problem with imperfect detection, the full parameter is

$\theta=(\psi_1,p_1,\psi_2,p_2)^T,$

and the null hypothesis is

$H_0:\psi_1=\psi_2=\psi,$

so under $H_0$ the parameter reduces to

$\theta_S'=(\psi,p_1,p_2)^T.$

Let $U_0(\theta_S')$ be the score vector under the null, and let $\widehat\theta_S'$ solve $U_0(\widehat\theta_S')=0$. The observed-information score statistic is

$$T_O(\widehat\theta_S')=U_0(\widehat\theta_S')^T J(\widehat\theta_S')^{-1}U_0(\widehat\theta_S'),$$

while the expected-information version is

$\chi^2$0

Under $\chi^2$1, both satisfy

$\chi^2$2

so the usual rule rejects for large values above a $\chi^2$3 cutoff (Karavarsamis et al., 2018).

A closely related classical instance is Rao’s score test for covariance structure. For testing

$\chi^2$4

with nuisance mean replaced by $\chi^2$5, the statistic reduces to

$\chi^2$6

and under classical fixed-$\chi^2$7 asymptotics it is asymptotically $\chi^2$8 with $\chi^2$9 degrees of freedom. This form makes explicit that many score-based tests are quadratic forms in score-like residual quantities, with calibration determined by an information matrix or its analogue (Jiang, 2015).

These formulations share the same inferential logic: only the null model is fitted, and the test measures whether the local derivative information at the null is too large to be attributed to sampling fluctuation. This suggests why score-based methods remain attractive when the alternative is difficult to fit or even undefined.

2. Observed information, negative statistics, and consistency

The main complication emphasized in the occupancy-model study is that the observed-information score statistic need not remain nonnegative under the alternative. The statistic

$\theta=(\psi_1,p_1,\psi_2,p_2)^T,$0

is a quadratic form. If $\theta=(\psi_1,p_1,\psi_2,p_2)^T,$1 is positive definite, then $\theta=(\psi_1,p_1,\psi_2,p_2)^T,$2; if $\theta=(\psi_1,p_1,\psi_2,p_2)^T,$3 is indefinite, then $\theta=(\psi_1,p_1,\psi_2,p_2)^T,$4 is also indefinite and the quadratic form can take either sign. In the occupancy simulations, as the departure from the null increases, the smallest eigenvalue of the observed information matrix becomes negative, and the observed score statistic can itself become negative (Karavarsamis et al., 2018).

The application concerns a two-sample occupancy model with imperfect detection. For a site visited $\theta=(\psi_1,p_1,\psi_2,p_2)^T,$5 times,

$\theta=(\psi_1,p_1,\psi_2,p_2)^T,$6

and for $\theta=(\psi_1,p_1,\psi_2,p_2)^T,$7,

$\theta=(\psi_1,p_1,\psi_2,p_2)^T,$8

The inferential target is whether occupancy differs between two regions,

$\theta=(\psi_1,p_1,\psi_2,p_2)^T,$9

In the paper’s standard configuration,

$H_0:\psi_1=\psi_2=\psi,$0

with $H_0:\psi_1=\psi_2=\psi,$1 and $H_0:\psi_1=\psi_2=\psi,$2, the naive rule $H_0:\psi_1=\psi_2=\psi,$3 showed power that first increased and then dropped sharply as $H_0:\psi_1=\psi_2=\psi,$4 grew. Around $H_0:\psi_1=\psi_2=\psi,$5, the median observed-information score statistic became negative, and the smallest eigenvalue of the observed information matrix became negative around the same point (Karavarsamis et al., 2018).

To address this, the paper proposes the modified rejection rule

$H_0:\psi_1=\psi_2=\psi,$6

denoted $H_0:\psi_1=\psi_2=\psi,$7. The reported findings are that power is restored, power is mostly higher than the other tests over the range of alternatives, consistency is largely regained, and inference is always possible, even when $H_0:\psi_1=\psi_2=\psi,$8 is negative. The appendix further shows that the expectation of the observed information matrix itself can become indefinite under the alternative, so negative score statistics are not pathological computation errors; they reflect a genuine loss of positive definiteness (Karavarsamis et al., 2018).

A common misconception is therefore that a negative observed-information score statistic must indicate numerical failure. In this setting, the negative value is structurally informative: it signals that the null model’s local curvature has changed qualitatively under the alternative.

3. High-dimensional and penalized score tests

In high dimensions, the difficulty is typically not the alternative model but the nuisance fit and the null calibration. One line of work replaces ordinary least squares in the null fit by penalized regression. For a feature $H_0:\psi_1=\psi_2=\psi,$9, write

$H_0$0

fit the null model by

$H_0$1

and define

$H_0$2

The tested hypothesis is

$H_0$3

With an $H_0$4 penalty, the full lasso selection rule satisfies

$H_0$5

so the lasso sparsity pattern is exactly a threshold on the penalized score statistic. With an $H_0$6 penalty, the statistic coincides with a score test in a mixed-effects model. The inferential target changes with $H_0$7: $H_0$8 recovers the multiple-regression coefficient, while large $H_0$9 approaches the simple-regression coefficient (Voorman et al., 2014).

A second line targets ultrahigh-dimensional blocks of coefficients. In the model

$\theta_S'=(\psi,p_1,p_2)^T.$0

the original score-function test uses

$\theta_S'=(\psi,p_1,p_2)^T.$1

Because high correlation between $\theta_S'=(\psi,p_1,p_2)^T.$2 and $\theta_S'=(\psi,p_1,p_2)^T.$3 can make the nuisance-estimation error term non-negligible, an orthogonalized score is introduced: $\theta_S'=(\psi,p_1,p_2)^T.$4 leading to the orthogonal score

$\theta_S'=(\psi,p_1,p_2)^T.$5

The corresponding test debiases the score and reduces the asymptotic variance, with asymptotic relative efficiency

$\theta_S'=(\psi,p_1,p_2)^T.$6

This is especially important when the tested and nuisance covariates are strongly correlated (Yang et al., 2022).

For covariance-structure testing, the classical $\theta_S'=(\psi,p_1,p_2)^T.$7 calibration fails when $\theta_S'=(\psi,p_1,p_2)^T.$8 is comparable to or larger than $\theta_S'=(\psi,p_1,p_2)^T.$9. The corrected Rao score test standardizes

$U_0(\theta_S')$0

as

$U_0(\theta_S')$1

with $U_0(\theta_S')$2, $U_0(\theta_S')$3, $U_0(\theta_S')$4, and $U_0(\theta_S')$5. The paper emphasizes that this corrected Rao test remains valid even in the ultra-high-dimensional regime $U_0(\theta_S')$6 (Jiang, 2015).

Taken together, these developments show that high-dimensional score testing is less about preserving the exact classical formula than about preserving the score-test principle under regularization, orthogonalization, or random-matrix recalibration.

4. Function-valued parameters, variance components, and instability

Score-based methods extend naturally to infinite-dimensional and mixed-model settings once the parameter of interest is represented as a risk minimizer. For a function-valued parameter

$U_0(\theta_S')$7

the restricted score test considers a prespecified candidate $U_0(\theta_S')$8 and tests

$U_0(\theta_S')$9

Since $\widehat\theta_S'$0 must satisfy $\widehat\theta_S'$1 in every admissible direction, the restricted null is

$\widehat\theta_S'$2

with statistic

$\widehat\theta_S'$3

Under $\widehat\theta_S'$4, $\widehat\theta_S'$5 converges to a Gaussian functional, and a multiplier bootstrap consistently estimates its distribution. The framework yields simultaneous confidence regions and bands without requiring direct inference on a slower-than-parametric estimator of $\widehat\theta_S'$6 (Hudson et al., 2021).

In functional linear concurrent regression,

$\widehat\theta_S'$7

the null

$\widehat\theta_S'$8

is rewritten as a variance-component problem after a spline-based random-effects reformulation: $\widehat\theta_S'$9 The resulting one-sided score statistic is

$U_0(\widehat\theta_S')=0$0

The paper reports that the procedure has the right levels asymptotically under null, higher power than a bootstrapped $U_0(\widehat\theta_S')=0$1 test, and good performance even when the data are sparsely observed and the covariate is contaminated with noise (Ghosal et al., 2018).

A related but distinct use appears in linear mixed models, where score-based fluctuation tests assess whether a parameter is stable across clusters ordered by an upper-level auxiliary variable. The cumulative score process is

$U_0(\widehat\theta_S')=0$2

The statistics $U_0(\widehat\theta_S')=0$3, $U_0(\widehat\theta_S')=0$4, and $U_0(\widehat\theta_S')=0$5 summarize the extent to which this process drifts away from zero. Because each parameter has its own score contributions, the method can distinguish a genuine cross-level interaction in a fixed effect from heterogeneity in random-effect variances, covariances, or residual variance, and the fluctuation plot indicates where instability occurs across the ordered auxiliary variable (Wang et al., 2019).

These formulations indicate that the score-based paradigm survives well beyond finite-dimensional likelihood theory: what matters is the first-order optimality condition, not the dimensionality of the parameter.

5. Goodness-of-fit and calibration

In goodness-of-fit testing, score-based methods are attractive because they only require fitting the null model once. One recent development shows that this classical viewpoint can be re-expressed through exponentially tilted alternatives: $U_0(\widehat\theta_S')=0$6 Differentiating at $U_0(\widehat\theta_S')=0$7 yields the score

$U_0(\widehat\theta_S')=0$8

and maximizing over a function class $U_0(\widehat\theta_S')=0$9 gives the integral probability metric

$$T_O(\widehat\theta_S')=U_0(\widehat\theta_S')^T J(\widehat\theta_S')^{-1}U_0(\widehat\theta_S'),$$0

Under this view, Kolmogorov–Smirnov, Wasserstein-1, and maximum mean discrepancy can all be reinterpreted as score-based tests arising from different choices of $$T_O(\widehat\theta_S')=U_0(\widehat\theta_S')^T J(\widehat\theta_S')^{-1}U_0(\widehat\theta_S'),$$1. The semiparametric KSD test then specializes this equivalence to a Stein-based function class, yielding a universally consistent test with Pitman efficiency and a generic parametric bootstrap procedure (Huang et al., 23 Dec 2025).

A second strand treats calibration as conditional goodness-of-fit. For multivariate forecast distributions, the null notion is auto-calibration,

$$T_O(\widehat\theta_S')=U_0(\widehat\theta_S')^T J(\widehat\theta_S')^{-1}U_0(\widehat\theta_S'),$$2

The framework first reduces $$T_O(\widehat\theta_S')=U_0(\widehat\theta_S')^T J(\widehat\theta_S')^{-1}U_0(\widehat\theta_S'),$$3 to a scalar $$T_O(\widehat\theta_S')=U_0(\widehat\theta_S')^T J(\widehat\theta_S')^{-1}U_0(\widehat\theta_S'),$$4, then applies either a PIT-based or a mean-zero test. When $$T_O(\widehat\theta_S')=U_0(\widehat\theta_S')^T J(\widehat\theta_S')^{-1}U_0(\widehat\theta_S'),$$5 is a proper scoring rule, two specific tests arise. The generalized Box transform uses

$$T_O(\widehat\theta_S')=U_0(\widehat\theta_S')^T J(\widehat\theta_S')^{-1}U_0(\widehat\theta_S'),$$6

while the entropy test uses

$$T_O(\widehat\theta_S')=U_0(\widehat\theta_S')^T J(\widehat\theta_S')^{-1}U_0(\widehat\theta_S'),$$7

With the log score or the energy score, these procedures provide formal calibration tests that are invariant to the ordering of variables and workable even when the forecast is available only through simulation (Knüppel et al., 2022).

For general probabilistic models with well-defined scores, the Kernel Calibration Conditional Stein Discrepancy test defines

$$T_O(\widehat\theta_S')=U_0(\widehat\theta_S')^T J(\widehat\theta_S')^{-1}U_0(\widehat\theta_S'),$$8

and estimates it with a $$T_O(\widehat\theta_S')=U_0(\widehat\theta_S')^T J(\widehat\theta_S')^{-1}U_0(\widehat\theta_S'),$$9-statistic. The key computational innovation is a family of score-based kernels such as

$\chi^2$00

which can be estimated without probability density samples. This yields a non-parametric calibration test with type-I-error control that remains applicable to unnormalized probabilistic models (Glaser et al., 16 Oct 2025).

The broader implication is that score-based testing now spans both derivative-based and discrepancy-based methodologies. The unifying feature is that the test statistic is constructed from a null-model optimality or compatibility condition rather than from explicit fitting of a rich alternative.

6. Sequential, algorithmic, and machine-learning uses

In modern machine learning, score-based tests increasingly appear as control procedures embedded inside algorithms. For binary simple hypothesis testing with intractable likelihoods, one such procedure replaces the log-likelihood ratio by a difference of average Hyvärinen scores: $\chi^2$01 The paper derives Chernoff-style upper bounds on Type I and II errors and proves that their large-deviation exponents are asymptotically tight for simple null and alternative hypotheses. For Gaussian models with equal covariance, the type-I exponent is available in closed form (Diao et al., 2024).

For gradient boosted decision trees, early stopping can itself be cast as a functional score test. At iteration $\chi^2$02, the null is that the current predictor is the population risk minimizer. For a direction $\chi^2$03,

$\chi^2$04

with asymptotic null law

$\chi^2$05

The stopping rule is

$\chi^2$06

Because the statistic is scale-invariant in the update direction and uses gradients rather than loss values, the same construction applies to implicit losses such as LambdaRank and data-dependent losses such as Cox regression via influence functions (Hines et al., 1 Jun 2026).

Streaming monitoring of learning machines admits an analogous construction. If parameters are estimated by empirical risk minimization, a changepoint at $\chi^2$07 can be tested with

$\chi^2$08

and the unknown changepoint is handled by maximization over $\chi^2$09. The paper develops a linear test for dense changes, a scan test for sparse changes, and an auto-test combining the two. Under the null, fixed-$\chi^2$10 statistics converge to $\chi^2$11 laws; under fixed alternatives, the procedures are consistent in power; and the implementation is designed for automatic differentiation in differentiable programming frameworks (Liu et al., 2021).

These algorithmic uses do not abandon the classical score-test idea. They repurpose it. The score becomes a device for deciding whether a model should keep learning, whether its environment has changed, or whether a likelihood-free model still supports reliable hypothesis testing.