Biological assays, environmental sensors, and economic auctions frequently hit physical or practical boundaries. When an instrument cannot measure a value below its lower limit of detection, the resulting data is left-censored: we do not know the exact value, only that it falls somewhere below the threshold. This occurs widely when tracking viral loads (Modélisation longitudinale de données incomplètes : exemple de la charge virale plasmatique du VIH), environmental toxins (Effect of plasma heavy metals on testosterone levels among male partners of the couples investigated for subfertility), or antibody responses (Antibody responses and correlates of protection in the general population after two doses of the ChAdOx1 or BNT162b2 vaccines).
The traditional, intuitive workaround is simple substitution—imputing a fixed constant such as zero, the detection limit, or half the detection limit for all unobservable points. However, as Rodolphe Thiébaut et al. demonstrate in Mixed models for longitudinal left-censored repeated measures, this naive approach introduces severe statistical bias, artificially compresses variance, and distorts longitudinal trends. Clamping values to an artificial floor makes a declining trajectory look significantly flatter and slower than it is in reality.
To recover the true underlying dynamics, a robust framework must shift from treating censored observations as missing points to treating them as interval constraints within a full likelihood function.
The Core Mechanism: From Density to Mass
In standard maximum likelihood estimation, every exact observation $y_i$ contributes its probability density function (PDF), $f(y_i | \theta)$, to the total likelihood. This represents the relative likelihood of observing that exact value under the model parameters $\theta$.
When a data point is censored at a lower threshold $L$, we lose its exact position but retain a powerful piece of information: $Y_i \le L$. Instead of guessing a point value within this range, likelihood-based adjustments incorporate the probability mass of the entire unobservable region. This is achieved by substituting the PDF with the cumulative distribution function (CDF):
$$\mathcal{L}_i(\theta) = \begin{cases} f(y_i | \theta) & \text{if } y_i > L \text{ (Observed)} \\ F(L | \theta) = \int_{-\infty}^{L} f(y | \theta) dy & \text{if } y_i \le L \text{ (Censored)} \end{cases}$$
By maximizing a joint likelihood that blends densities for exact measurements and cumulative probabilities for censored intervals, the model respects the physical boundaries of the assay without inventing data. As Rodolphe Thiébaut et al. show in Estimation of dynamical model parameters taking into account undetectable marker values, integrating the likelihood over the censored range reduces parameter estimation bias to less than 2%, even when a substantial portion of the dataset is undetectable.
Standard Likelihood (Observed Point):
Value observed at y_i --> Contributes PDF density: f(y_i)
|
---------------------------*---------------------------
Censored Likelihood (Below Threshold L):
Value is hidden below L --> Contributes CDF mass: Area under curve from -∞ to L
=======[ Integrated Area ]=======|
---------------------------------L---------------------
Preserving Trajectories in Dynamical Systems
Accounting for this interval information is especially critical in longitudinal studies and differential equation modeling. For instance, when tracking how a patient's viral load responds to an antiviral drug, the viral count often crashes rapidly below the detection limit.
In Maximum Likelihood Estimation in Dynamical Models of HIV, Jérémie Guedj et al. apply the full-likelihood framework to non-linear mixed models of viral clearance. If the data points below the threshold are omitted or flattened by constant imputation, the estimated clearance rate is severely underestimated. By evaluating the CDF over the unobservable region, the model recognizes that the true trajectory must be steep enough to plunge into the sub-threshold zone, thereby accurately recovering the true efficacy of the treatment.
Generalizing Across High Dimensions and Machine Learning
The principle of interval-constrained likelihoods extends naturally beyond classical regression into high-dimensional statistics and machine learning architectures.
Probabilistic Dimensionality Reduction
In high-throughput single-cell qPCR data, many target genes are unexpressed or fall below the instrument's detection limit. To visualize this data without the artifacts caused by zero-imputation, Victoria Moignard et al. developed a modified version of Principal Component Analysis in Probabilistic PCA of censored data: accounting for uncertainties in the visualization of high-throughput single-cell qPCR data. By introducing a probit noise model, the algorithm uses the normal distribution's CDF to capture the inherent uncertainty of sub-threshold measurements, yielding undistorted low-dimensional representations of single-cell expression landscapes.
Regularization and Non-Parametric Regression
When dealing with high-dimensional predictors alongside censored outcomes, standard regularization techniques must be adapted. Perrine Soret et al. address this in Lasso regularization for left-censored Gaussian outcome and high-dimensional predictors by pairing Lasso regularization with the Buckley-James algorithm, iteratively estimating censored values based on expected conditional distributions. For non-linear relationships, the authors of Generalized multi-output Gaussian process censored regression integrate a censored-Gaussian likelihood into a multi-output Gaussian Process framework, allowing the model to share strength across correlated outputs while properly bounding unobserved ranges.
Truncation vs. Censoring
While censoring preserves the knowledge that a sample exists (and provides its threshold), truncation occurs when samples outside a specific range are completely omitted from the dataset, leaving no record of their existence.
To overcome this harsher selection bias, Constantinos Costis Daskalakis et al. in Efficient Truncated Linear Regression with Unknown Noise Variance maximize the conditional log-likelihood of the observed samples, scaling the density by the total probability mass of the visible window. Similarly, for clustering applications, Sai Ganesh Nagarajan et al. introduce a gradient-truncated Expectation-Maximization algorithm in Mean Estimation of Truncated Mixtures of Two Gaussians: A Gradient Based Approach. These methods compute expectations strictly within the boundaries of the truncation set, allowing the optimization routine to recover the true distribution parameters without needing to reconstruct the missing portions of the space.
Neighborhood Constraints
In applications like real-time ad bidding, losing an auction provides a right-censored bound: you only know that the winning price was higher than your bid. However, assuming an infinite upper range can cause models to overestimate extreme outliers. In Arbitrary Distribution Modeling with Censorship in Real-Time Bidding Advertising, Xu Li et al. introduce a Neighborhood Likelihood Loss. This approach concentrates the probability mass within a localized window directly above the observed threshold, preventing outlier distortion while strictly respecting the mathematical boundary imposed by the censorship.
Key Foundational Work
Mixed models for longitudinal left-censored repeated measures (2003) — contributed 4 concept elements to this direction.
Modélisation longitudinale de données incomplètes : exemple de la charge virale plasmatique du VIH (2004) — contributed 3 concept elements to this direction.