Testing Tobit Restrictions Against Two-Part Alternatives

Why Test the Tobit Link?

An outcome with many zeros can arise because some units do not enter an activity, while others have a positive amount. Tobit gives a specific account of both features: a single latent outcome determines whether the observed outcome is zero and, when positive, its size. A two-part model instead allows the zero-versus-positive process and the positive amounts to follow separate models. Testing the Tobit restrictions asks whether the data are consistent with the particular link Tobit imposes, rather than merely whether either model can describe the observations.

The distinction matters for interpretation and prediction. Under Tobit, covariates affect the probability of a positive outcome and the distribution of positive outcomes through the same latent index and error scale. A two-part model can relax that connection. However, not every two-part model is a nested alternative to Tobit, so a familiar likelihood-ratio test is not always valid.

Tobit and Its Two-Part Representation

Suppose \(y_i\) is observed for each unit, and zero is the censoring point. The Tobit model assumes an underlying outcome \(y_i^*=x_i'\beta+\varepsilon_i\), with \(y_i=\max(0,y_i^*)\). Here \(x_i\) is a vector of regressors, \(\beta\) is a coefficient vector, and the errors are independent across units with \(\varepsilon_i\sim N(0,\sigma^2)\), where \(\sigma>0\). The regressors are treated as given, and the conditional normal, constant-variance specification is assumed to be correctly stated for the standard Tobit likelihood.

An observation is zero when \(y_i^*\leq 0\), and positive when \(y_i^*>0\). Thus the probability of a positive outcome is determined by the latent index \(x_i'\beta\), measured relative to \(\sigma\).

For Tobit, the first part is a probit probability:

For the second part, conditional on a positive outcome, \(y_i\) has the latent normal distribution truncated below at zero. Its density is

The numerator is the normal density of the latent outcome; the denominator normalizes that density over the positive region. Both parts therefore involve the same \(\beta\) and \(\sigma\).

The Restrictions in the Likelihood

For a zero observation, the Tobit likelihood contribution is the probability that the latent outcome is at most zero. For a positive observation, it is the latent normal density evaluated at the observed value. Equivalently, the positive contribution can be written as the probability of being positive multiplied by the conditional positive density. This factorization shows exactly how Tobit fits into a particular two-part model.

To form a nested two-part alternative, specify a probit model for positivity and a truncated normal model for positive outcomes. For example, let the participation probability be \(\Phi(x_i'\gamma)\), and let the positive-outcome density be a normal distribution with location \(x_i'\delta\) and scale \(\tau>0\), truncated below at zero. The unrestricted two-part model estimates \(\gamma,\delta,\tau\) separately. Tobit is obtained by imposing

This equation represents one restriction for each regressor. Under it, the participation index is \(x_i'\delta/\tau\), precisely the index implied by the positive-outcome distribution. The restricted model can be reparameterized as Tobit by setting \(\beta=\delta\) and \(\sigma=\tau\). The alternative permits the participation relationship to differ from the one implied by the positive-outcome distribution.

Worked Examples of the Link and Its Tests

Likelihood-Ratio Tests for Nested Alternatives

Let \(\ell_R\) be the maximized log likelihood under the restricted Tobit model and \(\ell_U\) the maximized log likelihood under the unrestricted, nested two-part model. The likelihood-ratio statistic is

Under the null and standard regularity conditions, this statistic has an approximate chi-squared distribution with degrees of freedom equal to the number of independent restrictions, denoted \(q\). These conditions include correct likelihood specification under the null, an interior parameter value, and an identifiable model. The two models must be fitted to the same observations using likelihoods for the same observed outcome. A small p-value is evidence against the restrictions linking the two parts; it is not the probability that Tobit is true.

Wald Tests of the Same Restrictions

A Wald test evaluates the restrictions using estimates from the unrestricted model. Write the vector of restrictions as \(r(\theta)=\gamma-\delta/\tau\), where \(\theta=(\gamma,\delta,\tau)\). If \(\hat r=r(\hat\theta)\) and \(\widehat V_r\) estimates its covariance matrix, the Wald statistic is

The covariance matrix must account for estimation uncertainty in all parameters, including the scale \(\tau\); it can be obtained by the delta method from the unrestricted parameter covariance matrix. Under the null and regularity conditions, \(W\) is approximately chi-squared with \(q\) degrees of freedom. The Wald test can be convenient, but it can differ from the likelihood-ratio test in finite samples, especially when estimates are far from the restrictions.

When a Two-Part Alternative Is Not Nested

Many two-part specifications do not use a probit participation equation and a truncated normal positive-outcome distribution. For instance, an analyst might use a logit model for positivity and a lognormal distribution for positive values. That model generally does not contain Tobit as a special case. Its parameter space and distributional assumptions differ, so there is no set of ordinary equality restrictions that turns it into the Tobit model.

In that situation, subtracting maximized log likelihoods and comparing the result with a chi-squared distribution is not justified by the nested-model likelihood-ratio result. A fit comparison can still be useful, but it answers a different question. For models fitted to the same observations and the same observed outcome, information criteria such as AIC can summarize a trade-off between fit and parameter count. Predictive performance on held-out data can address predictive goals. Neither comparison is automatically a test of Tobit restrictions.

Practical Interpretation and Common Pitfalls

A rejection of the Tobit restrictions indicates that the specific common latent process is inadequate relative to the nested alternative. It does not identify a unique explanation. The problem could lie in the participation equation, the positive-outcome distribution, the shared regressors, or other aspects of the assumed error distribution. Conversely, failure to reject does not prove the Tobit model; it means the test did not find sufficient evidence against the restrictions in the available sample.

Common mistake: “The two-part logit-lognormal model has a higher log likelihood, so the likelihood-ratio test rejects Tobit.” A higher maximized likelihood alone is expected when comparing a genuinely nested, less restricted model, but the standard chi-squared likelihood-ratio reference distribution requires nesting and regularity conditions. A logit-lognormal alternative is generally non-nested. The correction is to use a nested alternative with explicit restrictions for a likelihood-ratio or Wald test, or to report an appropriate non-nested fit or prediction comparison without labeling it a nested restriction test.