Non-Gaussian Error Distributions for Limited Outcomes

Why the Error Distribution Matters

A censored outcome does not reveal the latent value for every observation. For an observation at the censoring point, the data tell us that the latent outcome lies in a region, not exactly where it lies. The likelihood therefore uses a probability for censored observations and a density for uncensored observations. Both depend on the assumed error distribution.

The familiar Tobit specification uses Gaussian errors, but the likelihood construction does not require normality. A distribution with heavier tails can assign different probabilities to extreme latent values; a skewed distribution can assign different probabilities on the two sides of its center. Those differences can affect the fitted regression coefficients and the predicted probability of censoring. Changing the error distribution is a substantive modeling choice, not merely a change in notation.

For this parameterization, assume the errors are independent across observations conditional on the regressors, and that the chosen \(F\) correctly describes their common standardized distribution. The errors have mean zero and variance one, so \(\mu_i\) is the latent conditional mean and \(\sigma\) is the error standard deviation. The censoring rule and threshold must also be correctly recorded. The likelihood is conditional on the observed regressors; it does not by itself establish a causal interpretation for \(\beta\).

Building the Likelihood With a General Distribution

If \(Y_i=c_i\), the observation indicates \(Y_i^*\leq c_i\). Its probability is \(P(Y_i^*\leq c_i\mid x_i)=F(z_i)\). If \(Y_i>c_i\), the latent outcome is observed exactly. Transforming its density from the standardized error scale to the outcome scale gives \(\frac{1}{\sigma}f((y_i-\mu_i)/\sigma)\). The factor \(1/\sigma\) is the change-of-scale term; omitting it gives an incorrect density for \(Y_i\).

Let \(d_i=1\) when \(Y_i>c_i\), and \(d_i=0\) when \(Y_i=c_i\). For a sample of independent observations, the likelihood and log-likelihood are

The likelihood is maximized over the regression coefficients and scale, using the specified \(F\) and \(f\). The general structure stays the same when the error distribution changes; the CDF and density, and therefore the numerical likelihood and its maximizer, change. For a right-censored observation, the corresponding probability is an upper-tail probability \(1-F(z_i)\). For two-limit censoring, observations at each limit use the probability of the appropriate tail, while observations between the limits use a density.

Standardizing Alternative Error Distributions

Comparisons are easiest when each candidate error has mean zero and variance one. Then the same \(\sigma\) measures standard deviation across models, and \(z_i\) has the same standardized interpretation. A standard logistic random variable \(V\), with CDF \(\Lambda(v)=1/(1+\exp(-v))\), has variance \(\pi^2/3\). Thus \(u=V/a\), where \(a=\pi/\sqrt{3}\), has unit variance and CDF and density

This logistic error is symmetric, like a Gaussian error, but has heavier tails. A useful skewed alternative is the standardized Gumbel maximum distribution. If \(G\) has CDF \(\exp\{-\exp(-g)\}\), its mean is Euler’s constant \(\gamma\approx0.577216\) and its standard deviation is \(s_G=\pi/\sqrt{6}\approx1.28255\). Setting \(u=(G-\gamma)/s_G\) gives

The standardized Gumbel error has mean zero and variance one, but it is asymmetric: its left and right tails have different shapes. In particular, its standardized CDF at zero is about 0.5704, rather than 0.5. This is not a calculation error; zero is its mean, not its median.

Tail Shape and Skewness in Censoring Probabilities

For a fixed regression index and scale, a censored observation contributes \(F(z_i)\). Differences between candidate CDFs therefore directly change the likelihood contribution. Tail comparisons are especially important when the threshold is far from the latent mean, because then the censoring event may correspond to a tail probability. Skewness matters even at thresholds that are not especially extreme: for a skewed distribution, equal positive and negative standardized distances need not produce symmetric probabilities.

How Distributional Choice Changes Fitted Effects

A coefficient in \(\mu_i=x_i'\beta\) describes a change in the latent conditional mean, holding other regressors fixed. It is not automatically the change in the observed conditional mean or in the probability of being censored. For a left-censored outcome, the censoring probability is

For a continuous regressor \(x_{ij}\), assuming \(c_i\) and \(\sigma\) do not change with that regressor, differentiating gives

Thus the same latent-index coefficient can imply different changes in censoring probability under different error densities. In estimation, the distribution affects more than this probability formula: it determines the likelihood contributions from both censored and uncensored observations. Consequently, maximizing different likelihoods on the same data can produce different estimates of \(\beta\) and \(\sigma\). That difference should be described as model sensitivity, not as evidence that one distribution is correct simply because its fitted coefficient is larger.

Likelihood Construction: A Short Justification

The likelihood formula follows from the observation rule, rather than from a special Gaussian result. For an uncensored outcome, write \(z=(y_i-\mu_i)/\sigma\). Since \(Y_i^*=\mu_i+\sigma u_i\), the CDF of the latent outcome is \(F((y-\mu_i)/\sigma)\). Differentiating with respect to \(y\) gives its density:

For a censored outcome, the observed value is \(c_i\) whenever \(Y_i^*\leq c_i\). The probability of that event is

These are mutually exclusive observation types: a record contributes either the probability of falling in the censored region or the density at its observed uncensored value. Conditional independence lets us multiply those contributions across observations. This justifies the likelihood for any continuous error distribution with the stated CDF and density, provided the model assumptions hold.

Practical Interpretation and Common Pitfalls

Alternative distributions are useful when there is a substantive reason to expect asymmetry or tail behavior unlike the Gaussian model. They are not a mechanical remedy for every mismatch between model and data. A distribution with the desired shape must still be appropriately standardized, and the likelihood must reflect the actual censoring rule. In particular, using a density at the censoring point in place of a probability discards the information correctly represented by the censoring event.

• Pitfall: treating the latent error’s variance as one without standardizing it. A raw logistic error has variance \(\pi^2/3\), not one. Use the standardized form, or explicitly account for that variance in the scale parameter.
• Pitfall: using a density contribution for a censored record. A censored record tells us the latent outcome lies in a region. Its likelihood contribution is the probability of that region, such as \(F(z_i)\), not \(f(z_i)\).
• Pitfall: assuming zero means a 50% censoring probability. That is true when the threshold equals the median under a symmetric error distribution. A skewed distribution’s mean and median need not coincide; the standardized Gumbel example has \(F_{\mathrm{G}}(0)\approx0.5704\).
• Pitfall: reading a coefficient as an observed-outcome effect. The coefficient shifts the latent mean. Effects on censoring probabilities and observed outcomes depend on the distribution, scale, and censoring threshold.