Tobit Regression: Formula, Interpretation, Marginal Effects, Diagnostics and Worked Results

import numpy as np
from scipy import optimize, stats

def neg_loglik(theta, X, y, lower=0.0):
    beta = theta[:-1]
    sigma = np.exp(theta[-1])
    mu = X @ beta
    cens = y <= lower
    ll = np.empty_like(y, dtype=float)
    ll[~cens] = stats.norm.logpdf((y[~cens] - mu[~cens]) / sigma) - np.log(sigma)
    ll[cens] = stats.norm.logcdf((lower - mu[cens]) / sigma)
    return -ll.sum()

# X must include an intercept column.
# Start values can come from OLS; log(sigma) keeps sigma positive.
result = optimize.minimize(neg_loglik, start, args=(X, y, 0.0), method="BFGS")
beta = result.x[:-1]
sigma = np.exp(result.x[-1])

mu = X @ beta
z = (0.0 - mu) / sigma
p_censored = stats.norm.cdf(z)
expected_observed = mu * (1 - p_censored) + sigma * stats.norm.pdf(z)

library(censReg)

fit <- censReg(
  G3 ~ G1 + G2 + studytime + failures + absences + age + Medu + Fedu,
  left = 0,
  right = Inf,
  data = dat
)

summary(fit)

# Observed-mean effect for predictor j at profile x:
# beta_j * pnorm((mu - left_limit) / sigma)

* Prepare the identical complete-case dataset and verify the limit.
DESCRIPTIVES VARIABLES=G3 G1 G2 studytime failures absences age Medu Fedu.
FREQUENCIES VARIABLES=G3.

* Base REGRESSION fits OLS, not Tobit.
* Use the installed censored-regression extension or Python integration
* that maximizes the same left-censored normal likelihood at c = 0.
* Verify N=649, censored=15, the coefficient vector, sigma,
* and log likelihood before interpreting the output.

Latent mean (mu):
=Intercept + SUMPRODUCT(Predictor_Row, Coefficient_Row)

z:
=(Lower_Limit - mu) / Sigma

P(censored):
=NORM.S.DIST(z, TRUE)

Expected observed mean:
=Lower_Limit*NORM.S.DIST(z,TRUE)
 +mu*(1-NORM.S.DIST(z,TRUE))
 +Sigma*NORM.S.DIST(z,FALSE)

Observed residual:
=Observed_G3 - Expected_Observed_Mean

The classical Type I Tobit Regression model uses one latent normal equation and one deterministic censoring rule. The same coefficient vector determines both the latent mean and, through its distance from the limit, the probability of being censored. This economy is useful when the observation mechanism is genuinely a measurement or reporting limit. It becomes restrictive when a boundary value is generated by a separate behavioral decision. In the worked grade model, the assumption means G1, G2, studytime, failures, absences, age, Medu, and Fedu all influence the latent grade and the chance of appearing at zero through the same linear index. A researcher should defend that common-process assumption rather than treating it as a default created by software.

Tobit Regression retains censored cases at the boundary. Truncated regression analyzes a sample from which cases beyond the threshold are absent. The distinction changes the likelihood: a censored case contributes a cumulative probability, while a truncated density is normalized by the probability of entering the observed region. Confusing the two can bias slopes and scale because the model uses the wrong information about missing or boundary observations. In the worked analysis, all 649 records remain present and 15 are treated as boundary values. That is a censoring specification. If students with latent grades below zero had been omitted entirely, the appropriate likelihood would be truncated rather than censored.

The worked model is left-censored at zero. Right-censored Tobit Regression is used when values above an upper limit are recorded at that ceiling. A two-limit model handles both lower and upper censoring. The model form must follow the data-collection rule. A grade scale may have both a floor and a ceiling even when the current sample does not contain observations at the upper endpoint. If the upper limit is an active reporting mechanism, a one-sided specification is incomplete. If no values are compressed at the ceiling, adding a right limit can be unnecessary. The correct choice depends on design knowledge, not only the observed minimum and maximum.

A Tobit Regression coefficient is a partial slope for the latent response y*. G2 = 0.918706 means that, conditional on the other seven predictors, a one-unit increase in G2 is associated with a 0.918706-unit increase in the latent mean. It does not automatically mean the recorded G3 rises by exactly that amount. Near the censoring limit, some latent movement changes the probability of escaping the boundary rather than appearing entirely as movement in the observed mean. Far from the limit, the observed and latent effects become similar. Stating the latent scale prevents a common reporting error and makes the later marginal-effect calculation transparent.

For a standard left-censored normal model, the derivative of the unconditional expected observed outcome is βj multiplied by the probability of being uncensored at the evaluation profile. This probability varies across cases, so an observed-mean marginal effect is generally not constant even though the latent coefficient is constant. The worked table evaluates effects at the vector of predictor means, where the uncensoring probability rounds to 1.000. That makes the displayed observed-mean effects numerically equal to the latent slopes. Average marginal effects would instead calculate the derivative for each of the 649 rows and then average those values. The two summaries should be labeled separately.

A predictor can also be interpreted through its effect on P(y > c | x). For a left-censored normal model, the derivative is βj φ((μ − c)/σ)/σ. The effect is largest for profiles near the limit, where the normal density is high, and becomes small when a case is already far above or below the boundary. Because most fitted profiles in this analysis have negligible censoring probability, the probability effects will be small for many students even when the latent G2 coefficient is large. This is another reason that one Tobit Regression coefficient cannot stand in for every substantive effect.

A third target is E(y | y > c, x), the expected outcome among observations that are uncensored. Its derivative differs from both the latent coefficient and the unconditional observed-mean effect because conditioning changes the distribution through an inverse-Mills-ratio term. Analysts sometimes call every post-estimation derivative a “marginal effect” without naming the target. That practice creates conflicting numbers that are all mathematically valid but answer different questions. A clear Tobit Regression report should define the target population and scale before presenting any derivative.

Only 2.311% of the worked sample is treated as censored. A small censoring fraction often makes Tobit Regression resemble ordinary least squares because most rows contribute ordinary normal-density terms. It also means the lower-tail mechanism is estimated from limited direct information. The model may be sensitive to a few boundary cases, and complex heteroskedastic or nonlinear censoring behavior can be difficult to identify. Small censoring does not make the method wrong when the observation rule is real, but it reduces the practical difference from OLS and increases the importance of sensitivity analysis.

The OLS comparison has slightly lower in-sample RMSE and MAE and lower displayed AIC and BIC. That result is unsurprising with only 15 boundary cases. However, OLS and Tobit Regression encode different meanings for zero. OLS treats it as exact; Tobit treats it as incomplete information about an underlying value. Predictive error cannot by itself decide which scientific statement is correct. A defensible comparison therefore combines fit, residual evidence, boundary calibration, and the substantive recording mechanism. When zero is exact, OLS or another bounded-outcome model may be preferable. When zero is censored, the Tobit likelihood respects the observation rule even if its in-sample RMSE is not smaller.

AIC and BIC compare penalized likelihoods, but meaningful comparison requires compatible response definitions, observation sets, and likelihood constants. The attached table places Tobit Regression and OLS criteria side by side. The lower OLS values are informative as a descriptive comparison, yet they do not prove that censoring should be ignored. If the two models treat boundary observations as different kinds of information, the scientific interpretation still matters. Information criteria help select among defensible candidates; they do not validate an indefensible observation mechanism.

Standard Tobit Regression assumes normal latent errors. The Q-Q plot shows strong lower-tail departure, and the histogram is asymmetric. These patterns can affect maximum-likelihood estimates and especially their uncertainty. A few extreme low outcomes may create more probability mass in the tail than a Gaussian model allows. Possible responses include robust sensitivity checks, alternative latent distributions, censored quantile methods, or explicit mixture structures. The correct reaction is not to discard the model automatically, but to report the departure and test whether substantive conclusions persist under reasonable alternatives.

The classical model uses one sigma for every predictor profile. If latent variance changes with grade history, studytime, failures, or another covariate, both the probability of censoring and the expected observed outcome can be misspecified. Residual-versus-fitted plots may reveal changing spread, but formal heteroskedastic Tobit Regression lets log sigma depend on predictors. Such flexibility increases complexity and requires enough boundary information. With only 15 censored cases, a richly parameterized variance equation could be unstable. Parsimony and sensitivity analysis are therefore important.

A heteroskedastic specification replaces constant sigma with a positive function such as exp(z′γ). This can separate changes in the latent mean from changes in latent dispersion. It is useful when residual spread clearly varies across groups or predictor levels. Interpretation becomes more complex because a predictor in the scale equation changes censoring probabilities and observed means even when it is absent from the mean equation. The worked output does not estimate a scale equation, so heteroskedastic effects should be described as a possible extension rather than a result.

The worked Tobit Regression enters every predictor linearly. Grade relationships may contain diminishing returns, thresholds, or ceiling effects. Residual curvature, partial-residual plots for uncensored cases, and substantive theory can motivate polynomial terms, splines, or transformations. Any nonlinear term affects the latent mean and therefore the censoring probability. Marginal effects then depend on both the nonlinear derivative and the uncensoring probability. A better functional form can matter more than choosing between OLS and Tobit when the linear predictor is misspecified.

An interaction allows the latent effect of one predictor to depend on another. For example, the relationship between G2 and final grade might differ by prior failures or studytime. In Tobit Regression, an interaction coefficient is not the complete observed-scale interaction effect because censoring introduces additional nonlinearity. Researchers should compute predicted observed means or marginal effects across meaningful combinations rather than interpreting the product-term coefficient alone. No interactions are included in the worked model, so the reported slopes are average additive associations.

Tobit Regression does not solve predictor overlap. G1 and G2 are earlier grades and are likely to share substantial information. Multicollinearity can inflate uncertainty, destabilize signs, and make one coefficient appear small after another correlated predictor enters. The dominance of G2 should therefore be interpreted within the chosen eight-predictor specification. VIF-style diagnostics from the corresponding design matrix, condition indices, and specification comparisons can help assess stability, even though the final estimator is maximum likelihood rather than OLS.

Boundary cases can contribute strongly to the cumulative-probability part of the likelihood. With only 15 censored observations, one unusual predictor profile may noticeably affect beta or sigma. Case-deletion refits, likelihood-contribution plots, and bootstrap inclusion frequencies can identify that dependence. Influence should be assessed on coefficients, marginal effects, and predicted censoring probabilities. Removing a difficult case solely to improve diagnostics is not justified; the goal is to understand sensitivity and possible data-quality problems.

When the numerical Hessian is unreliable, a case-resampling bootstrap can estimate empirical uncertainty. Each bootstrap sample should resample complete records, refit the entire Tobit Regression model, and recompute the target coefficients or marginal effects. Failed or boundary solutions must be tracked rather than silently discarded. Percentile or bias-corrected intervals can then be reported with the number of successful refits. The attached analysis contains no bootstrap results, so no such intervals are claimed. Bootstrap uncertainty is an extension that could repair the current inferential gap.

Successful optimization means the algorithm found a parameter vector that satisfied its convergence criteria. It does not guarantee that the inverse Hessian is accurate or positive definite. Flat curvature, scaling problems, strong predictor overlap, or numerical approximation can prevent covariance recovery. The attached output preserves blank standard errors rather than converting failed calculations into zeros or nonsignificant tests. A careful workflow would rescale predictors, compare optimizers, calculate a more stable numerical Hessian, use profile likelihood, or bootstrap the estimates.

If G3 = 0 is an exact earned grade, standard Tobit Regression may impose a false latent-below-zero story. Ordinary regression with robust uncertainty, ordinal regression, bounded regression, or a model tailored to the grade-generating process may be more defensible. If zeros arise from a separate event such as nonparticipation, a two-part model can estimate the zero process separately from positive grades. The presence of zeros is a diagnostic clue, not a model-selection command.

A two-part model first estimates the probability of being above the boundary and then models the positive outcome among uncensored cases. Predictor effects may differ across parts. This flexibility is useful when crossing the boundary and determining positive magnitude are substantively distinct processes. The cost is additional parameters and a different estimand. Tobit Regression is more restrictive because one latent equation governs both pieces. Comparing the models can reveal whether the common-process assumption is plausible.

Censored quantile regression targets a conditional quantile of the latent or observed response rather than the conditional mean under normal errors. It can reveal whether predictor effects differ across the outcome distribution and may be more robust to heavy tails. Interpretation and estimation differ from the Tobit likelihood, and standard errors often rely on resampling. The strong lower-tail departure in the worked diagnostics makes quantile-based sensitivity analysis especially relevant, although no censored quantile results are available here.

A Heckman-style selection problem occurs when the outcome is observed only for a nonrandom subset and the selection mechanism has an unobserved component correlated with the outcome error. Tobit censoring instead assumes the outcome is transformed by a known limit. Both use latent-variable ideas and normal probabilities, which encourages confusion. The correct method follows the data mechanism: boundary recording calls for censoring; nonrandom missing outcomes call for a selection model. The 649-case worked dataset retains all records, so the fitted specification is censoring rather than sample selection.

The reported RMSE and MAE are in-sample summaries. A predictive Tobit Regression application should use a holdout set, cross-validation, or temporal validation. Evaluation should include overall observed-scale error, calibration of censoring probabilities, error among uncensored cases, and performance for boundary cases. Because only 15 observations are censored, random folds may contain very few boundary cases; stratified or repeated resampling can stabilize evaluation. No external validation is available in the attached analysis.

A high-quality Tobit Regression report states the censoring limit, number and percentage censored, latent equation, scale parameter, likelihood statistic, coefficient scale, marginal-effect target, diagnostics, and uncertainty method. It should also explain why censoring is scientifically credible. Avoid phrases such as “a one-unit increase causes a 0.92-unit observed increase” unless a causal design and the correct observed-scale derivative justify that statement. Transparent limits strengthen rather than weaken the analysis.

Reproducibility requires more than saving the final coefficient table. The same complete-case rows, variable coding, limit, optimizer settings, parameterization of sigma, and prediction equations must be preserved. A case-level ledger containing observed G3, latent μ, censoring probability, expected observed mean, and residual allows reviewers to trace charts and summary metrics. The downloadable Python report and worked workbook support that audit. A future R or SPSS reproduction should be reconciled against those quantities once the same specification is confirmed.