Robust Inference Under Tobit Misspecification

Why Robust Inference Matters

Gaussian Tobit estimation uses a particular distributional assumption: conditional on the regressors, the latent outcome has a normal error. That assumption gives a convenient likelihood for both the censored observations and the positive outcomes. But if the true error distribution is not normal, the maximizer of the Gaussian likelihood can still be a useful statistical target. Its interpretation and its standard errors, however, require care.

Under misspecification, the estimator generally does not converge to a structural coefficient defined by the true data-generating process. It converges instead to the parameter that best fits the chosen Gaussian Tobit criterion on average. Sandwich standard errors account for the difference between the variability of the likelihood score and the curvature of the criterion. They do not make the Gaussian model true or turn its target into a structural parameter.

The Gaussian Tobit Criterion and Its Target

Consider a left-censored outcome with censoring point zero. Let $y_i^*$ be the latent outcome, $x_i$ a vector of regressors, and $y_i=\max(0,y_i^*)$ the observed outcome. Write $\mu_i=x_i'\beta$, let $\sigma>0$, and define $\delta_i=1$ when $y_i>0$ and $\delta_i=0$ when $y_i=0$. The Gaussian Tobit likelihood treats the latent error as normal with mean zero and variance $\sigma^2$.

Definition: The Gaussian Tobit log-likelihood contribution for observation $i$ is $$ \ell_i(\beta,\sigma) = \delta_i\left[-\log(\sigma)+\log\phi\left(\frac{y_i-\mu_i}{\sigma}\right)\right] + (1-\delta_i)\log\Phi\left(\frac{-\mu_i}{\sigma}\right), $$ where $\phi$ and $\Phi$ are the standard normal density and cumulative distribution functions. An uncensored observation contributes a density; a censored observation contributes the probability that the latent outcome is at or below zero.

The likelihood can be maximized even when the actual conditional distribution of $y_i^*$ is not normal, provided the optimization and regularity conditions are suitable. In that case, define $\theta=(\beta',\log\sigma)'$, and let $\ell_i(\theta)$ denote the same criterion expressed using $\log\sigma$.

Definition: Under misspecification, the pseudo-true parameter $\theta^*$ is the unique interior maximizer of the expected Gaussian Tobit log likelihood: $$ \theta^*=\arg\max_{\theta} E[\ell_i(\theta)]. $$ It is the parameter value that best fits this likelihood criterion in expectation under the actual data distribution. It need not equal a structural parameter, nor need its coefficient describe the true conditional mean or another familiar feature of the latent outcome.

Worked Example: Evaluating an Expected Tobit Criterion

Suppose the observed outcome has a hypothetical distribution with probability $0.5$ at zero and probability $0.5$ at one. Evaluate the expected Gaussian Tobit log-likelihood contribution at the candidate $\mu=0,\sigma=1$, with no varying regressors.

At zero, the contribution is $\log\Phi(0)=\log(0.5)$. At one, it is $\log\phi(1)$, because the uncensored normal density is evaluated at $y=1$. Thus $$ Q(0,1) = 0.5\log(0.5)+0.5\log\phi(1). $$

Using $\phi(1)\approx0.2419707$, the arithmetic is $$ Q(0,1) \approx 0.5(-0.693147)+0.5(-1.418939) = -0.346574-0.709470 = -1.05604. $$

This is the expected criterion at one candidate, not evidence that the candidate maximizes it. The pseudo-true parameter is found by maximizing the expected criterion over the admissible parameter values. The example illustrates why the censored probability and the uncensored density both enter that target.

Score, Curvature, and the Sandwich Covariance

For inference, let $s_i(\theta)=\partial\ell_i(\theta)/\partial\theta$ be the score vector. Define the expected negative Hessian $A$ and the score variance $B$, both evaluated at $\theta^*$:

A=-E\left[\frac{\partial s_i(\theta^*)}{\partial\theta'}\right], \qquad B=E[s_i(\theta^*)s_i(\theta^*)'].

At an interior pseudo-true parameter, $E[s_i(\theta^*)]=0$. Therefore $B$ is the covariance matrix of the score. Under correct likelihood specification, the information identity gives $A=B$. Under misspecification that equality need not hold, which is why the usual inverse-Hessian covariance alone may be wrong.

Theorem: Suppose observations are independent and identically distributed; the expected criterion has a unique interior maximizer; the score has finite second moments; the relevant derivatives and laws of large numbers and central limit theorem apply; and $A$ is nonsingular. Then the maximum-likelihood estimator converges to $\theta^*$, and its asymptotic covariance is $$ \operatorname{Var}(\hat\theta)\ \approx\ \frac{1}{n}A^{-1}BA^{-1}. $$ The expression $A^{-1}BA^{-1}$ is the sandwich covariance per observation.

This result follows by expanding the sample score around $\theta^*$. The estimator solves $\sum_i s_i(\hat\theta)=0$. A first-order Taylor expansion gives

0 \approx \frac{1}{n}\sum_{i=1}^n s_i(\theta^*) + \left\{\frac{1}{n}\sum_{i=1}^n \frac{\partial s_i(\theta^*)}{\partial\theta'}\right\} (\hat\theta-\theta^*).

The derivative average converges to $-A$. Rearranging and multiplying by $\sqrt n$ yields

\sqrt n(\hat\theta-\theta^*) \approx A^{-1}\frac{1}{\sqrt n}\sum_{i=1}^n s_i(\theta^*).

The central limit theorem gives the score sum an asymptotic covariance $B$. Premultiplication and postmultiplication by $A^{-1}$ therefore give $A^{-1}BA^{-1}$, establishing the sandwich form.

Score Contributions for Left-Censored Tobit

To see what enters the calculation, use $\tau=\log\sigma$, so $\sigma=e^\tau$, and write $a_i=-\mu_i/\sigma$ and $\lambda(a)=\phi(a)/\Phi(a)$. Differentiating the likelihood gives the score for $\beta$ and for $\tau$:

s_{\beta i} = x_i\left[ \delta_i\frac{y_i-\mu_i}{\sigma^2} - (1-\delta_i)\frac{\lambda(a_i)}{\sigma} \right], \qquad s_{\tau i} = \delta_i\left[-1+\left(\frac{y_i-\mu_i}{\sigma}\right)^2\right] + (1-\delta_i)\lambda(a_i)\frac{\mu_i}{\sigma}.

The censored score depends on the probability of being at or below the censoring point, while the uncensored score depends on the normal density contribution. For a fitted model, a practical empirical sandwich estimate uses the observed scores and Hessians at $\hat\theta$:

\widehat A=-\frac{1}{n}\sum_{i=1}^n \frac{\partial s_i(\hat\theta)}{\partial\theta'}, \qquad \widehat B=\frac{1}{n}\sum_{i=1}^n s_i(\hat\theta)s_i(\hat\theta)', \qquad \widehat{\operatorname{Var}}(\hat\theta) = \frac{1}{n}\widehat A^{-1}\widehat B\widehat A^{-1}.

This version uses the outer product of individual scores for $\widehat B$. For independent observations, it estimates the score variability without imposing the Gaussian information identity. Dependence between observations requires an appropriate covariance estimator for that dependence; the independent-observation formula does not automatically account for clustering or serial correlation.

Worked Example: A Scalar Robust Standard Error

Consider a hypothetical sample $y=(0,0,1,2)$ from a left-censored Tobit with an intercept only and known $\sigma=1$. Estimate the intercept $\mu$, and use the empirical sandwich formula to illustrate its standard error. This four-observation calculation is for illustrating the arithmetic, not for claiming reliable small-sample inference.

For an uncensored observation the intercept score is $y_i-\mu$; for a censored observation it is $-\lambda(-\mu)$. The likelihood score equation is therefore $$ (1-\mu)+(2-\mu)-2\lambda(-\mu)=0. $$ Solving this equation gives approximately $\hat\mu=0.4182$, with $\lambda(-\hat\mu)\approx1.0818$. The four fitted scores are approximately $-1.082,-1.082,0.582,1.582$; their sum is approximately zero, as required by the score equation.

For a censored observation, the derivative of its score with respect to $\mu$ is $-\lambda(a)\{a+\lambda(a)\}$, where $a=-\mu$. Thus the total negative Hessian at the fitted value is approximately $$ 2+2(1.0818)(-0.4182+1.0818)\approx3.436, \qquad \widehat A=\frac{3.436}{4}\approx0.859. $$

The empirical score variance is $$ \widehat B = \frac{2(1.0818^2)+0.5818^2+1.5818^2}{4} \approx1.295. $$ Consequently, $$ \widehat{\operatorname{Var}}(\hat\mu) = \frac{1}{4}\frac{1.295}{0.859^2} \approx0.4388, \qquad SE_{\mathrm{robust}}(\hat\mu)\approx\sqrt{0.4388}=0.6624. $$

For comparison, the inverse-Hessian standard error would be $1/\sqrt{4(0.859)}\approx0.5395$, which implicitly uses the information identity. The difference here illustrates the separate roles of curvature and score variability; with only four observations, neither standard error should be treated as a dependable large-sample approximation.

Inference and Interpretation

In practice, calculate the sandwich covariance, take square roots of its diagonal entries for robust standard errors, and use the relevant normal or Wald approximation when the sample is sufficiently large and regularity conditions are credible. For example, an approximate 95% interval for component $\theta_j^*$ is $\hat\theta_j\pm1.96\,SE_{\mathrm{robust}}(\hat\theta_j)$. The interval targets the pseudo-true parameter under the working Gaussian Tobit criterion.

Worked Example: A Robust Interval for a Tobit Coefficient

Suppose a hypothetical fit has $n=200$, and its estimated per-observation matrices, with the coefficient of interest first, are $$ \widehat A= \begin{pmatrix}4&1\\1&2\end{pmatrix}, \qquad \widehat B= \begin{pmatrix}9&2\\2&5\end{pmatrix}. $$ Assume the coefficient estimate is $\hat\beta_1=0.300$, measured in latent-outcome units per one-unit increase in its regressor.

Since $\widehat A^{-1}=\frac{1}{7}\begin{pmatrix}2&-1\\-1&4\end{pmatrix}$, the first diagonal element of the per-observation sandwich is $$ \frac{1}{49} \begin{pmatrix}2&-1\end{pmatrix} \widehat B \begin{pmatrix}2\\-1\end{pmatrix} = \frac{33}{49} \approx0.6735. $$ Dividing by $n$ gives an estimated variance of $0.6735/200=0.003367$, so the robust standard error is approximately $0.05803$.

The approximate interval is $$ 0.300\pm1.96(0.05803)=[0.186,\,0.414]. $$ Under the stated large-sample conditions, this interval is for the pseudo-true Gaussian Tobit coefficient: the coefficient that maximizes expected Gaussian Tobit log likelihood. It is not automatically an interval for a true latent conditional-mean effect. Its units are latent-outcome units per one-unit increase in the regressor.

What Robust Inference Does—and Does Not—Fix

The sandwich changes the estimated sampling uncertainty to reflect score variation under misspecification. It does not correct a misspecified conditional mean, guarantee identification, eliminate bias in a finite sample, or make the pseudo-true coefficient answer a different scientific question. Nor does it provide a dependence-robust covariance unless the meat of the sandwich is constructed to match the dependence structure.

A common mistake is to say, “The robust standard error makes the Tobit model robust to nonnormality, so its coefficient is the true effect.” The correction is more limited: the coefficient still estimates the pseudo-true parameter of the Gaussian Tobit criterion, while the sandwich can provide asymptotically valid uncertainty for that target under suitable conditions. Whether that target is substantively useful must be argued separately.

Key takeaway: Under distributional misspecification, Gaussian Tobit maximum likelihood is an estimator of a pseudo-true criterion maximizer. Robust inference uses both likelihood curvature and score variability; it does not change the target or repair every form of model error.

Check Your Understanding

Use the definitions and results above to answer the following questions.

In a left-censored Tobit likelihood, why does an observation at zero contribute a probability rather than a normal density evaluated at zero?
What does the pseudo-true parameter maximize, and why need it not be a structural coefficient?
State what the matrices $A$ and $B$ represent in the sandwich covariance.
Under correct likelihood specification, what relation holds between $A$ and $B$, and why can it fail under misspecification?
A researcher reports a sandwich standard error and claims this proves the Gaussian Tobit coefficient is the true latent-outcome effect. What is the error in that interpretation?