Written and maintained by CASRAI Editorial Board
Last updated
A Tobit model is the right tool when the outcome variable itself has a real floor or ceiling — not when data are missing. If every unit in your sample is observed, but some of them hit a genuine lower or upper limit on the outcome (spending that cannot go below $0, a lab measurement below an instrument’s detection limit, a top-coded survey income), that is censoring, and it calls for a Tobit model. If instead certain units never entered your dataset at all because they fell outside a range — you only sampled people who were readmitted to hospital, say, and never recorded the ones who weren’t — that is truncation, a different problem with a different model (truncated regression, not Tobit). Confusing the two is the most common Tobit misapplication, and running ordinary least squares (OLS) directly on censored data without correcting for it produces a biased, inconsistent estimate every time. This guide works through both distinctions with a worked numeric example showing exactly how large that OLS bias is, and what a Tobit model recovers instead.
Censoring vs. truncation: the distinction that decides the tool
James Tobin introduced the model in 1958 to handle household spending on durable goods — an outcome that cannot be negative, so a household that would, if it could, spend a negative amount on (say) a new car simply reports $0 instead (Tobin, “Estimation of Relationships for Limited Dependent Variables,” Econometrica, 26(1), 1958). That example fixes the definition precisely:
- Censoring: the independent variables are observed for every unit in the sample, but the dependent variable’s true value is known only within a restricted range. Outside that range, the value is recorded at the boundary, not left blank. A survey that top-codes household income at “$250,000+” still has a row for that household — the true income just isn’t fully known. An assay that reports “below detection limit” for a low concentration still logged that sample; the true concentration is somewhere below the limit, not missing.
- Truncation: whole observations outside the range are absent from the dataset — neither the outcome nor the predictors were ever recorded for them. A study built entirely from hospital readmission records only contains patients who were readmitted; patients who weren’t simply don’t appear anywhere in the data, not even as a $0 or boundary row.
The practical test: if you could, in principle, add a column to your dataset holding the true unobserved value for every affected row, and every row still has one, you have censoring and Tobit is the right family of model. If entire rows are absent from the sample by construction, you have truncation, and a Tobit model applied to it is misspecified — a truncated regression model, which does not use the boundary-coded observations Tobit relies on (because there aren’t any), is the correct tool instead.
A related but distinct trap is the corner solution: some outcomes pile up at zero not because a true negative value got clipped, but because zero is a genuine, meaningful choice with its own decision process — a household that decided not to buy a car at all this year, as opposed to one that would have spent a small negative amount if it could. Tobit assumes the same latent process drives both the zeros and the positive values, just clipped at the boundary. When the zero-generating decision is fundamentally different from the how-much decision, a two-part (hurdle) model — a participation-probability stage plus a separate positive-outcome stage — usually fits the data-generating process better than a single-index Tobit. Whether that distinction matters for a given dataset is a substantive judgment about the outcome, not something the software can decide for you.
Why OLS on censored data is the wrong move
Running an ordinary regression on an outcome that has been floored or capped treats every boundary observation as if it were the literal, true value of y. It isn’t — it’s only known to be at or beyond that boundary. Because the observations most likely to be pushed to the floor are exactly the ones with low predicted values of x, censoring mechanically compresses the variance of y at the low end of x and flattens the fitted line. The direction of the resulting bias is well established: OLS applied directly to censored data is an inconsistent estimator, with a downward-biased slope coefficient and an upward-biased intercept. The more observations sit at the boundary, the worse the distortion — a dataset that is 5% censored barely notices; one that is 40% censored can have its slope cut by a third or more, which is exactly what the worked example below shows.
A worked example: quantifying the bias
The numbers below are a small, constructed illustrative dataset — not real research data, and not drawn from any actual study — built specifically to isolate the mechanism and show its rough size, not to represent a real finding. Suppose x is a 1–10 workload-pressure index for a group of researchers, and the outcome is hours of unpaid overtime logged in a week. Unpaid overtime cannot be negative, but the underlying willingness to work extra hours — call it the latent variable y* — can be: a researcher under very low pressure might, if it were possible, want to log negative overtime (i.e., work noticeably less than a full week). What actually gets recorded is y = max(0, y*):
| Workload index (x) | Latent willingness, y* (unobservable) | Observed overtime, y |
|---|---|---|
| 1 | −4.80 | 0.00 |
| 2 | −2.10 | 0.00 |
| 3 | −1.90 | 0.00 |
| 4 | 0.80 | 0.80 |
| 5 | −0.50 | 0.00 |
| 6 | 2.20 | 2.20 |
| 7 | 4.40 | 4.40 |
| 8 | 4.10 | 4.10 |
| 9 | 6.30 | 6.30 |
| 10 | 6.00 | 6.00 |
Four of the ten rows (40%) are censored at zero. Fitting three different models to this same dataset:
- OLS on the latent y* (the benchmark — what you’d get if the true willingness were fully observable, which it never actually is in practice): intercept −5.20, slope 1.209.
- Naive OLS on the observed, censored y (what a researcher gets by running a standard regression without accounting for the floor): intercept −2.01, slope 0.798 — roughly a third below the true slope, exactly the downward attenuation the mechanism above predicts.
- Tobit maximum-likelihood estimate on the same observed y, modeling the censoring explicitly: intercept −4.57, slope 1.127 — within about 7% of the true slope, versus naive OLS’s roughly 34% understatement.
The mechanical difference is what each method does with the four zero rows. OLS treats each zero as a real, literal outcome value equal to 0 and lets it pull the fitted line down. The Tobit likelihood instead treats each censored row as contributing the probability that the latent value falls at or below zero, given x — using the cumulative normal distribution for those rows and the normal density for the uncensored rows — and finds the intercept, slope, and error variance that jointly make the observed pattern of zeros and positive values most likely. That is the entire mechanical difference between the two: OLS spends no information on which observations were censored, only on their reported value; Tobit’s likelihood uses the fact of censoring itself as data.
Interpreting Tobit coefficients — they are not the marginal effect on y
A Tobit coefficient describes the effect of a predictor on the latent variable y*, not directly on the observed, censored outcome y. Reporting a Tobit slope as if it were “the effect of x on hours of overtime actually logged” overstates the effect for anyone near the floor, because part of any latent increase for those units is absorbed by moving them off zero rather than by raising an already-positive value. Most statistical packages that fit Tobit models (Stata’s tobit with a post-estimation margins call, R’s censReg or AER::tobit) can report the marginal effect on the observed outcome separately from the latent-index coefficient — check which one a table is reporting before interpreting or citing a Tobit result, since the two numbers are routinely different and papers do not always label which they used.
When a Tobit model is the wrong tool
Three situations call for something other than a standard Tobit model:
- The zeros are a corner solution with their own decision process (see above) — a two-part/hurdle model, estimated as a separate participation stage and positive-outcome stage, fits better than forcing both through Tobit’s single latent index.
- You have truncation, not censoring — entire out-of-range observations are absent from the dataset, not merely capped. Use a truncated regression model instead; it is built for exactly that sampling structure and does not assume the boundary-coded rows Tobit needs.
- Selection into observability correlates with unobserved factors that also affect the outcome — for example, a wage regression that can only observe wages for people who chose to work. That is a sample-selection problem, and a Heckman-style two-stage or full-information selection model, which explicitly models the participation decision alongside the outcome, is the standard alternative to reach for.
Frequently asked questions
Is a Tobit model just OLS with a dummy variable for the censored observations?
No. Adding a censoring indicator as a covariate to an OLS model does not correct the bias described above — it changes what the model is estimating without addressing the underlying problem, which is that OLS uses the wrong likelihood for boundary-coded data. Tobit replaces the likelihood function itself: a probability mass (from the cumulative distribution) for censored rows, a density (from the probability distribution) for uncensored ones, estimated jointly by maximum likelihood.
What is the difference between a Tobit model and a Heckman selection model?
Both handle a boundary/missingness problem in the outcome, but they assume different data-generating processes. Tobit assumes one latent index drives both whether an observation is censored and, conditional on not being censored, its value — the same coefficients govern both. A Heckman selection model explicitly separates the two into a selection equation and an outcome equation, which can have different predictors and a correlated but distinct error structure. Use Heckman when you have genuine reason to think selection into observability and the outcome level are driven by related but separable processes; use Tobit when the boundary is a mechanical floor or ceiling on a single underlying quantity.
Can an outcome be censored from above as well as below?
Yes — right-censoring (a ceiling, such as a top-coded income or an instrument’s upper measurement range) uses the identical Tobit logic with the inequality direction reversed, and some outcomes are censored on both sides at once (a two-limit Tobit). The likelihood construction is the same idea: a probability mass at each affected boundary, a density everywhere in between.
Does Tobit require the errors to be normally distributed?
The standard Tobit model does assume normally distributed errors with constant variance (homoskedasticity), because the maximum-likelihood estimator is built directly on the normal density and CDF. Both assumptions are checkable and both matter more here than in ordinary OLS: violating them (heteroskedasticity in particular) is well documented to bias Tobit’s maximum-likelihood estimates, unlike OLS where heteroskedasticity mainly affects standard errors rather than the point estimate itself. Robustness checks and heteroskedasticity-robust or semiparametric variants exist for exactly this reason.
Is the Tobit model still used, or have two-part models mostly replaced it?
Both remain in active use, and the choice is a modeling decision about the outcome rather than a matter of one having superseded the other. Tobit is the right choice when a single underlying process plausibly generates both the boundary values and the positive ones; two-part and hurdle models are preferred when theory or the data suggest the zero/participation decision is genuinely separate from the how-much decision. Neither is a default — the censoring-vs-corner-solution question above is what should decide it for a given dataset.
Related CASRAI guides
For the standard uncensored-outcome regression this guide contrasts against, see Regression Analysis: Assumptions, Interpretation, and How to Report It. For a binary-outcome limited dependent variable model built on the same maximum-likelihood logic, see Logistic Regression (the Logit Model). For diagnosing whether OLS assumptions like homoskedasticity hold in an uncensored model, see How to Read a Residual Plot. Right-censoring shows up under a different name and a different model family in time-to-event data — see How to Read a Kaplan-Meier Curve and The Cox Proportional Hazards Model for that version of the concept. For another outcome type with its own distributional constraints, see Poisson Distribution. Browse the full Research Methods & Statistics hub for related quantitative-analysis guides.
