Random Double Truncation Methods and Theory

• Random double truncation is a mechanism where data are observed only when a variable falls between two random bounds, introducing selection bias.
• It is central to applications in survival analysis, Gaussian latent models, and random matrix theory, with methods like IPW and NPMLE used to correct bias.
• Practical estimation strategies include nonparametric corrections, semiparametric kernel hazard estimation, and EM-based approaches for handling dependent truncation.

Random double truncation is an observation mechanism in which an underlying variable is retained only when it falls inside a two-sided window, while the window itself may be random or the truncation event may be random because it depends on latent variables. In the classical statistical formulation, one observes a triplet $(U,X,V)$ only if $U \le X \le V$, so the data are drawn from the conditional law given inclusion rather than from the target law. The same phrase also appears in Gaussian latent-variable models, where fixed bounds generate random truncation events because they are applied to random coordinates, and in random matrix theory, where independent row and column selections induce a two-parameter random truncation field. Across these formulations, the common themes are bias induced by selective observation, likelihoods defined on truncated support, inverse-probability correction, and asymptotic analysis of estimators or stochastic processes (Moreira et al., 2021, Anne et al., 2019, Donati-Martin et al., 2013, G et al., 2012).

1. Core definitions and principal formulations

In survival analysis and related sampling problems, random double truncation is defined through random truncation variables $(U^*,V^*)$, with observation of $(U^*,X^*,V^*)$ only when $U^* \le X^* \le V^*$. Under independent truncation, the sampling probability function is

$G(x)=P(U^* \le x \le V^*),$

and the observed density is

$$f^*(x)=\frac{G(x)f(x)}{\alpha}, \qquad \alpha=P(U^* \le X^* \le V^*)=\int G(t)F(dt).$$

Hence the observed distribution is biased unless $G$ is constant on the support of $X^*$. A particularly important special case is interval sampling, $V^*=U^*+\tau$, where each subject is observable only if the event occurs inside a calendar-time window of width $U \le X \le V$0 (Moreira et al., 2021, Uña-Álvarez, 2023).

The cited literature also uses the term for mechanisms that do not discard observations outright. In the zero-inflated Gaussian model, the latent vector $U \le X \le V$1 is transformed coordinatewise by

$U \le X \le V$2

so out-of-range latent values are replaced by zero rather than removed. The bounds $U \le X \le V$3 are fixed and known, but truncation is random overall because it depends on the random Gaussian coordinates. In the random matrix setting, a Haar-distributed matrix $U \le X \le V$4 is “doubly truncated” by independently retaining row $U \le X \le V$5 if $U \le X \le V$6 and column $U \le X \le V$7 if $U \le X \le V$8, leading to

$U \le X \le V$9

In multivariate normal theory, double truncation refers to conditioning a Gaussian vector on a rectangle $(U^*,V^*)$0, with some bounds possibly infinite (Anne et al., 2019, Donati-Martin et al., 2013, G et al., 2012).

Setting	Observation rule	Main inferential object
Survival and registry data	Observe iff $(U^*,V^*)$1	$(U^*,V^*)$2, $(U^*,V^*)$3, regression effects
Zero-inflated Gaussian data	$(U^*,V^*)$4	$(U^*,V^*)$5, $(U^*,V^*)$6, graph structure
Haar or orthogonal matrices	Keep row $(U^*,V^*)$7 if $(U^*,V^*)$8, column $(U^*,V^*)$9 if $(U^*,X^*,V^*)$0	Limit of truncated norm process
Doubly truncated MVN	Condition on $(U^*,X^*,V^*)$1	Truncated moments and marginals

2. Distributional structure and likelihoods on truncated support

The basic effect of random double truncation is that the observed law is not the target law. For the classical survival model,

$(U^*,X^*,V^*)$2

and the inverse-probability identity

$(U^*,X^*,V^*)$3

shows that estimation of $(U^*,X^*,V^*)$4 requires correction by the sampling probability. This identity underlies both nonparametric maximum likelihood and semiparametric weighting procedures (Moreira et al., 2021, Uña-Álvarez, 2023).

In the Gaussian zero-inflation model, the full likelihood of the observed vector $(U^*,X^*,V^*)$5 involves $(U^*,X^*,V^*)$6 zero/nonzero patterns and multiple integrals up to order $(U^*,X^*,V^*)$7, so the analysis proceeds through bivariate marginals. For each pair $(U^*,X^*,V^*)$8, the single-observation likelihood is written as a sum of four contributions, $(U^*,X^*,V^*)$9, $U^* \le X^* \le V^*$0, $U^* \le X^* \le V^*$1, and $U^* \le X^* \le V^*$2, corresponding respectively to both coordinates truncated, only $U^* \le X^* \le V^*$3 truncated, only $U^* \le X^* \le V^*$4 truncated, and both observed. The terms $U^* \le X^* \le V^*$5, $U^* \le X^* \le V^*$6, and $U^* \le X^* \le V^*$7 integrate the latent Gaussian density over complements of the truncation intervals, which makes the likelihood explicitly dependent on the truncation geometry (Anne et al., 2019).

For a doubly truncated multivariate normal vector $U^* \le X^* \le V^*$8, the conditional density on a rectangle is

$U^* \le X^* \le V^*$9

with $G(x)=P(U^* \le x \le V^*),$0. Using a Tallis-type mgf argument, explicit formulas are obtained for truncated moments and for bivariate marginal densities $G(x)=P(U^* \le x \le V^*),$1. These marginals enter directly into the second-moment formulas, so the doubly truncated MVN is treated not merely as a conditional density but as a fully computable moment system (G et al., 2012).

In the random matrix model, the centered truncated field admits the decomposition

$G(x)=P(U^* \le x \le V^*),$2

where $G(x)=P(U^* \le x \le V^*),$3 is built from the empirical processes of the row and column selectors and $G(x)=P(U^* \le x \le V^*),$4 captures the residual contribution of matrix-entry fluctuations. This decomposition isolates the two distinct sources of randomness: the truncation mask itself and the underlying Haar matrix (Donati-Martin et al., 2013).

3. Estimation strategies

A central estimator under random double truncation is the Efron–Petrosian nonparametric maximum likelihood estimator. Writing $G(x)=P(U^* \le x \le V^*),$5 for the empirical cdf of the observed $G(x)=P(U^* \le x \le V^*),$6's and $G(x)=P(U^* \le x \le V^*),$7 for the estimated sampling probability, the estimator has the inverse-probability form

$G(x)=P(U^* \le x \le V^*),$8

with

$G(x)=P(U^* \le x \le V^*),$9

This representation makes explicit that doubly truncated observations are reweighted by $$f^*(x)=\frac{G(x)f(x)}{\alpha}, \qquad \alpha=P(U^* \le X^* \le V^*)=\int G(t)F(dt).$$0 to recover the target distribution (Moreira et al., 2021).

Hazard estimation under random double truncation is built on this IPW representation. The nonparametric kernel hazard estimator smooths the cumulative hazard induced by $$f^*(x)=\frac{G(x)f(x)}{\alpha}, \qquad \alpha=P(U^* \le X^* \le V^*)=\int G(t)F(dt).$$1, while the semiparametric version replaces $$f^*(x)=\frac{G(x)f(x)}{\alpha}, \qquad \alpha=P(U^* \le X^* \le V^*)=\int G(t)F(dt).$$2 by a parametric $$f^*(x)=\frac{G(x)f(x)}{\alpha}, \qquad \alpha=P(U^* \le X^* \le V^*)=\int G(t)F(dt).$$3 derived from a model $$f^*(x)=\frac{G(x)f(x)}{\alpha}, \qquad \alpha=P(U^* \le X^* \le V^*)=\int G(t)F(dt).$$4 for the truncation law. The semiparametric kernel hazard estimator is

$$f^*(x)=\frac{G(x)f(x)}{\alpha}, \qquad \alpha=P(U^* \le X^* \le V^*)=\int G(t)F(dt).$$5

This formulation makes the truncation correction explicit and was introduced to avoid the nonexistence, nonuniqueness, and high-variance issues of the fully nonparametric NPMLE in severe truncation regimes (Moreira et al., 2021).

For Cox regression with left, right, or double truncation under dependent truncation, the key assumption is conditional independence,

$$f^*(x)=\frac{G(x)f(x)}{\alpha}, \qquad \alpha=P(U^* \le X^* \le V^*)=\int G(t)F(dt).$$6

The conditional likelihood of the observed sample includes the selection term

$$f^*(x)=\frac{G(x)f(x)}{\alpha}, \qquad \alpha=P(U^* \le X^* \le V^*)=\int G(t)F(dt).$$7

and is maximized by an EM algorithm. The E-step constructs expected counts for latent truncated event times, and the M-step solves a weighted Cox score equation on a pseudo-dataset with standard coxph-style software. This procedure avoids modeling the joint distribution of $$f^*(x)=\frac{G(x)f(x)}{\alpha}, \qquad \alpha=P(U^* \le X^* \le V^*)=\int G(t)F(dt).$$8 and directly adjusts the risk sets through the conditional likelihood (Rennert et al., 2018).

In the zero-inflated Gaussian graphical model, estimation is two-step. First, each latent covariance entry is estimated by maximizing the corresponding bivariate marginal log-likelihood,

$$f^*(x)=\frac{G(x)f(x)}{\alpha}, \qquad \alpha=P(U^* \le X^* \le V^*)=\int G(t)F(dt).$$9

Second, the estimated covariance matrix is plugged into graphical lasso,

$G$0

so that sparsity of the latent precision matrix $G$1 encodes the recovered graph (Anne et al., 2019).

For doubly truncated multivariate normals, the estimation problem is often analytic rather than iterative. The truncated mean satisfies

$G$2

and explicit second-moment formulas involve the bivariate truncated marginals $G$3. These formulas are implemented in the tmvtnorm package, while numerical evaluation of the required Gaussian probabilities relies on mvtnorm::pmvnorm() (G et al., 2012).

4. Asymptotic theory and limit behavior

The cited literature shows that random double truncation does not preclude rigorous asymptotic theory, but it changes the form of the empirical process and often inflates constants or introduces additional regularity conditions. For kernel hazard estimation, the practical estimators $G$4 and $G$5 are first-order equivalent to the oracle estimator

$G$6

Under the stated regularity assumptions, $G$7 is pointwise consistent and asymptotically normal, the bias is

$G$8

and the variance is

$G$9

The resulting AMISE-optimal bandwidth remains of order $X^*$0, with truncation entering through $X^*$1 in the constant (Moreira et al., 2021).

For the Cox model under dependent truncation, the EM-based MLE is strongly consistent: $X^*$2 converges almost surely to $X^*$3, and $X^*$4 converges almost surely and uniformly on $X^*$5 to $X^*$6. Under the regularity assumptions stated in the paper, $X^*$7 converges weakly to a mean-zero tight Gaussian process, and practical inference is based on the bootstrap (Rennert et al., 2018).

In the zero-inflated Gaussian graph problem, the covariance estimator has an elementwise sup-norm rate under assumptions bounding correlations away from $X^*$8 and imposing a Morse condition on the population log-likelihood. Specifically, if $X^*$9 is large enough, then with high probability

$V^*=U^*+\tau$0

Combined with standard graphical lasso theory under incoherence and beta-min conditions, this yields elementwise consistency of $V^*=U^*+\tau$1 and consistency of graph recovery (Anne et al., 2019).

In the Haar-matrix model, deterministic truncation and random double truncation lead to different functional limits. Deterministic truncation yields a tied-down bivariate Brownian bridge, whereas random double truncation produces, after centering and normalization by $V^*=U^*+\tau$2,

$V^*=U^*+\tau$3

with $V^*=U^*+\tau$4 and $V^*=U^*+\tau$5 independent one-dimensional Brownian bridges. The matrix-driven term $V^*=U^*+\tau$6 vanishes at the $V^*=U^*+\tau$7 scale, so the leading fluctuations are generated by the random row and column selectors rather than by the fine dependence structure of the Haar entries (Donati-Martin et al., 2013).

5. Identifiability, ignorability, and structural invariance

Random double truncation is not synonymous with unavoidable correction. Under independence between $V^*=U^*+\tau$8 and $V^*=U^*+\tau$9, the sampling bias is ignorable exactly when

$U \le X \le V$00

equivalently $U \le X \le V$01, or equivalently $U \le X \le V$02. In that case the naive empirical cdf $U \le X \le V$03 is consistent and fully efficient for $U \le X \le V$04, and under the null the NPMLE coincides with $U \le X \le V$05. To test this null, a Kolmogorov–Smirnov-type statistic is proposed:

$U \le X \le V$06

with bootstrap approximation of its null distribution. Under the stated conditions, $U \le X \le V$07 converges to the supremum of a centered Gaussian process, while under fixed alternatives $U \le X \le V$08 (Uña-Álvarez, 2023).

A distinct but related misconception is that independence alone is enough to justify ordinary estimation. The survival papers make a sharper point: independence between $U \le X \le V$09 and $U \le X \le V$10 is the basis of the Efron–Petrosian framework, but ordinary empirical methods are valid only when the induced sampling function is constant. Conversely, the Cox-regression paper departs from unconditional independence and instead assumes conditional independence $U \le X \le V$11, showing that dependence induced by shared covariates can be accommodated without modeling the truncation distribution itself (Uña-Álvarez, 2023, Rennert et al., 2018).

The literature also identifies important existence and uniqueness issues. The NPMLE for doubly truncated data may fail to exist or be non-unique; in one formulation this is characterized through a directed graph on observed triplets, with unique NPMLE if and only if the graph is strongly connected. Severe truncation can also make the NPMLE highly variable even when it exists, which motivates semiparametric stabilization via $U \le X \le V$12 (Moreira et al., 2021).

In multivariate normal theory, double truncation preserves more structure than might be expected. When only a subset $U \le X \le V$13 of coordinates is truncated and the remaining subset $U \le X \le V$14 is untruncated, the inverse covariance matrix of the truncated distribution leaves unchanged the off-diagonal blocks linking truncated and untruncated variables and the block corresponding to relations within $U \le X \le V$15. The preservation of zero patterns in the precision matrix implies that conditional independence structure is unchanged by selection, and the stronger blockwise invariance shows that many precision elements involving untruncated variables are numerically invariant as well (G et al., 2012).

6. Empirical regimes, applications, and limitations

The applied papers show that the impact of random double truncation is highly regime-dependent. In hazard estimation, the semiparametric kernel estimator systematically attains smaller MISE than the nonparametric estimator in the reported simulations, especially for small $U \le X \le V$16 and narrow sampling windows, while both corrected estimators dominate the naive hazard when $U \le X \le V$17 is not flat. In the ACS age-at-diagnosis application, the estimated sampling function is essentially flat, so naive and corrected hazard curves are similar and the main difficulty is variance and NPMLE existence. In the AIDS incubation example, the estimated $U \le X \le V$18 is decreasing, longer incubation times are less likely to be observed, and the naive hazard lies above the corrected curves (Moreira et al., 2021).

The testing paper illustrates the same distinction from a distributional perspective. In the childhood cancer interval-sampling data, the statistic is $U \le X \le V$19 with bootstrap $U \le X \le V$20, so there is no evidence against ignorable sampling bias and the empirical cdf is substantially more efficient than the NPMLE. In the Parkinson’s disease study, the early-onset group yields $U \le X \le V$21 with $U \le X \le V$22, whereas the late-onset group yields $U \le X \le V$23 with $U \le X \le V$24, indicating strongly non-ignorable bias in the latter case (Uña-Álvarez, 2023).

In regression, the dependent-truncation EM estimator shows little bias and, in most practical situations reported in simulation, lower mean-squared error than the weighted estimators that assume unconditional independence. In the autopsy-confirmed Alzheimer’s disease cohort, the method yields statistically significant effects of age at onset, sex, and high occupational attainment, whereas the independence-based weighted analysis attenuates some of these effects (Rennert et al., 2018).

In zero-inflated Gaussian graph estimation, simulations compare the pairwise truncated-likelihood plus graphical-lasso procedure with naive graphical lasso applied directly to the truncated data. The reported findings are that naive Glasso suffers lower true-edge detection and more false positives, especially when zero inflation is heterogeneous or severe, whereas the likelihood-based correction remains close to graphical lasso on the untruncated latent data in several graph families (Anne et al., 2019).

The main limitations are model-specific. The survival and testing frameworks require independence between the target variable and truncation pair, except in the Cox model where only conditional independence is assumed; violations are not handled there. The fully nonparametric survival estimators depend on NPMLE existence and can be unstable under severe truncation. The zero-inflated Gaussian method assumes latent Gaussianity, known double-truncation bounds, and incurs $U \le X \le V$25 nonconvex one-dimensional optimizations. The Cox EM formulation does not straightforwardly accommodate time-varying covariates measured after study entry. The random-matrix analysis focuses on weak convergence rather than non-asymptotic concentration. These limitations indicate that “random double truncation” is not a single technique but a family of selection mechanisms whose treatment depends on whether the key object is a distribution, a hazard, a regression parameter, a covariance graph, or a stochastic process (Moreira et al., 2021, Rennert et al., 2018, Anne et al., 2019, Donati-Martin et al., 2013).