Inference and Diagnostics from Estimated Influence Functions

From an Influence Function to Uncertainty

An estimate alone does not show how precisely a target has been learned. In semiparametric analysis, the influence function links an estimator’s sampling variation to observation-level contributions. Once those contributions can be estimated, their empirical spread provides a standard error and supports confidence intervals and tests. The calculations are direct; their justification depends on the estimator having the right asymptotic behavior and on the contributions having sufficiently well-behaved moments.

Let \(O_1,\ldots,O_n\) be independent, identically distributed observations from a distribution \(P\), and let \(\theta(P)\) be the target parameter. Write \(\phi_P(O)\) for an influence function for this target. It has mean zero under \(P\). The influence functions developed for particular targets, such as means or missing-data means, provide examples of this general object.

The expansion says that, to first order, the estimation error is an average of independent contributions. If \(V\) is finite and positive, the central limit theorem gives \(\sqrt{n}\{\hat\theta-\theta(P)\}/\sqrt{V}\) an asymptotic standard normal distribution. Thus the approximate variance of \(\hat\theta\) is \(V/n\), and its standard error is \(\sqrt{V/n}\). In practice, \(P\), and therefore \(\phi_P\), is unknown. We use estimated influence values \(\hat\phi_i\), one for each observation, instead.

Estimating the Variance and Constructing Inference

The empirical variance of the estimated influence values estimates the variance of a single contribution. Centering is important: even when the population influence function has mean zero, estimated values need not sum to exactly zero in a finite sample. Define \(\bar{\hat\phi}=n^{-1}\sum_{i=1}^{n}\hat\phi_i\).

Under the assumptions below, an approximate two-sided \(100(1-\alpha)\%\) Wald confidence interval is \(\hat\theta\pm z_{1-\alpha/2}\widehat{SE}(\hat\theta)\), where \(z_{1-\alpha/2}\) is the corresponding standard normal quantile. For a null value \(\theta_0\), the Wald statistic is \((\hat\theta-\theta_0)/\widehat{SE}(\hat\theta)\); its two-sided approximate p-value is \(2\{1-\Phi(|T|)\}\), where \(\Phi\) is the standard normal cumulative distribution function. These are large-sample approximations, not exact finite-sample guarantees.

The empirical approximation condition is a useful way to express what is needed of estimated contributions; it does not by itself establish the estimator’s asymptotic linearity. For example, estimating nuisance functions may require additional conditions on their errors and on the estimator’s remainder. When nuisance predictions are obtained by cross-fitting, each observation’s contribution is evaluated using nuisance estimates trained without that observation. Whatever the construction, the standard error is justified only when the full estimator meets the asymptotic conditions, not merely because a software routine returns influence values.

Justification for the Plug-in Standard Error

The variance calculation follows from the asymptotic linear representation and consistency of the empirical variance. The reasoning also clarifies why checking the influence values matters.

Diagnosing Extreme Contributions and Unstable Variances

An influence value is a first-order contribution to estimation error, not a direct measure of an observation’s substantive importance or a guarantee that deleting it would produce a particular change. Nevertheless, the squared values determine the estimated variance. A small number of large absolute values can therefore dominate the standard error. Inspecting the empirical distribution and the contributions to the sum of squares can reveal instability that a single reported standard error hides.

Variance instability can also arise when the estimated influence values themselves are sensitive to fitted nuisance functions, or when a few observations have extreme weights or residuals. Useful diagnostics include plotting the values, reporting their range and selected quantiles, and calculating the fraction of the sum of squares contributed by the largest absolute values. Such summaries are diagnostics, not formal tests of the model assumptions. A finite-sample plot cannot establish that a population moment exists or does not exist.

Moment Conditions and Limits of Standard Inference

A finite second moment is central to the standard procedure. The central limit theorem used here requires a finite variance for the influence function. The empirical variance also needs to estimate a finite population variance. If the influence function has infinite variance, the usual square-root-\(n\) normal approximation and its plug-in Wald interval are not justified by this argument. A large observed influence value is a reason to investigate, but the absence of an extreme value in a modest sample cannot verify the required moment condition.

Common Pitfall and Practical Interpretation

A common mistake is to assume that calculating an empirical variance proves standard asymptotic inference is valid. The calculation only produces a number. Its interpretation depends on the estimator’s asymptotic linearity, the relevant moment conditions, and whether estimated influence values adequately approximate the true contributions. If these conditions are doubtful, the interval should not be presented as automatically reliable; explain the concern and investigate the estimator and data-generating features that could cause it.

Keep the units clear throughout. Influence values have the units of the target, their variance has squared units, and the standard error and confidence interval have the target’s units. An interval describes uncertainty under the stated sampling and asymptotic assumptions; it does not remove bias from a misspecified target, guarantee causal interpretation, or establish that the assumptions hold.