In many real-world datasets, variables don't range from negative to positive infinity. Instead, they hit "walls." A student’s GPA cannot exceed 4.0; a patient’s viral load cannot be measured below the laboratory’s detection limit; and a firm’s innovation output cannot be less than zero.
When data piles up at these boundaries, standard statistical tools like Ordinary Least Squares (OLS) regression begin to break. They treat these "walls" as literal values rather than thresholds, leading to biased estimates and impossible predictions. To solve this, researchers across disciplines—from medicine to economics—rely on censored regression, most notably the Tobit model.
The Problem of the Artificial Floor
Imagine trying to measure the effect of a new drug on viral load. If the drug is highly effective, many patients will have viral levels so low they fall below the "limit of quantification." In a standard spreadsheet, these might be recorded as zero.
As Carolyn T. Bramante and colleagues demonstrate in their study of Metformin and SARS-CoV-2, treating these undetectable values as simple zeros biases the results. A linear model assumes the data could theoretically keep going into negative numbers. By forcing a straight line through a cluster of zeros, the model underestimates the true rate of decline.
This "floor effect" is a recurring challenge in healthcare. Xue Long and team encountered this when analyzing healthcare expenditures, where a significant portion of the population reports zero spending. Similarly, Jia Wang highlights that in dose-response modeling, accounting for this left-censoring is essential for calculating accurate "Excess Risk Related Effective Doses."
Why Linear Models Fail at the Boundary
The intuition behind why OLS fails is simple: it assumes the error terms are normally distributed around the prediction. But at a boundary, the distribution is "crushed."
If a model predicts a GPA of 4.2 for a top student, but the maximum possible is 4.0, the error is not random—it is systematically constrained. Daniel E. Rigobon and co-authors found that in the Fragile Families Challenge, standard algorithms often produced unrealistic predictions outside valid ranges for metrics like GPA or psychological grit.
By ignoring the boundary, OLS tries to "pull" the regression line toward the mass of points at the limit, distorting the slope for the rest of the data. This is why Amy Peng argues that Tobit models are necessary for modeling labor supply, where many participants may not work any hours at all.
The Tobit Solution: Latent Variables
The Tobit model, named after economist James Tobin, works by assuming there is an underlying, unobserved "latent" variable that is linear and normally distributed. We only see the "censored" version of this variable when it crosses a certain threshold.
Think of it like a submerged iceberg. We only see the tip (the data above zero), but the Tobit model uses the shape of the visible part to infer the size and trajectory of the part underwater. This allows researchers to:
- Estimate the probability that an observation will be at the boundary.
- Estimate the linear relationship for those who are above the boundary.
This dual capability is vital for studying efficiency. In Western China’s healthcare allocation, Dongxue Zhao used Tobit regression to analyze efficiency scores that were capped at 1.0. Because many regions hit this "perfect" efficiency ceiling, a standard model would mask the environmental factors actually driving performance. Similar logic was applied by Jiazeng Li to port eco-efficiency and L.L. SUN to fishery operations.
Two-Limit and Longitudinal Complexity
In many cases, data is boxed in from both sides. Jiawei Wang utilizes two-limit Tobit modeling to study agricultural insurance structures where proportions are strictly constrained between 0 and 1.
The challenge becomes even steeper in longitudinal studies. When tracking disease progression, patients often reach a "ceiling" on clinical tests where they can no longer show improvement because they have maxed out the scale. Clemens Schächter and Alexander Schmidt-Richberg highlight how these ceiling effects in Alzheimer's and rare diseases can hide the true rate of decline.
To address this, Bin Huang and colleagues developed PCATS, an API that integrates truncated normal distributions into Bayesian frameworks. By natively embedding these bounds into the regression, the model ensures that predicted outcomes—like pain scores or cognitive levels—remain within the valid range of the measurement scale, providing a more honest reflection of human health and behavior.