Decomposing Tobit Marginal Effects

Why Decompose a Tobit Effect?

In a left-censored Tobit model, a covariate can affect an observed outcome in two ways: it can change the probability that the outcome is positive, and it can change the outcome’s mean among cases with positive values. Those channels combine to determine the change in the unconditional mean, which includes both zeros and positive outcomes.

The distinction matters when interpreting coefficients. A coefficient in the latent equation is not, by itself, the marginal effect on the observed outcome. The marginal effect depends on the covariate values through the probability of being above the censoring point. This tutorial derives the three relevant effects and shows how the two channels add up.

The Tobit Model and Its Three Means

Suppose the latent outcome \(Y^*\) follows a left-censored Tobit model with censoring point zero:

Y^* = x'\beta+\varepsilon,\qquad \varepsilon\mid x\sim N(0,\sigma^2),\qquad Y=\max(0,Y^*),

Here \(x\) is a vector of covariates, \(\beta\) is the vector of latent-index coefficients, and \(\sigma>0\) is constant. The normality and constant-variance assumptions are part of this model. Define the index \(\eta=x'\beta\), its standardized version \(z=\eta/\sigma\), and let \(\Phi\) and \(\phi\) denote the standard normal cumulative distribution function and density.

Definition: The inverse Mills ratio for a positive outcome is \(\lambda(z)=\phi(z)/\Phi(z)\). Under the stated Tobit model, the probability of a positive observed outcome is \(P(Y>0\mid x)=\Phi(z)\); the conditional mean among positive outcomes is \(E(Y\mid Y>0,x)=\eta+\sigma\lambda(z)\); and the unconditional observed-outcome mean is \(E(Y\mid x)=\eta\Phi(z)+\sigma\phi(z)\).

The conditional mean is specifically the mean of \(Y\) among cases with \(Y>0\). Since positive observed outcomes equal their latent values, it is also \(E(Y^*\mid Y^*>0,x)\). It is not the unconditional mean of \(Y^*\), which is \(\eta\), nor is it the unconditional mean of the censored observed outcome.

Why the Conditional Mean Has This Form

Conditional on \(x\), \(Y>0\) exactly when \(Y^*>0\), or equivalently when \(\varepsilon>-\eta\). Standardizing the error gives the event \(Z>-z\), where \(Z=\varepsilon/\sigma\sim N(0,1)\). Its probability is \(\Phi(z)\), by symmetry of the standard normal distribution.

To calculate the conditional error mean, use \(\phi'(u)=-u\phi(u)\). Integrating over the truncated tail gives:

E(Z\mid Z>-z) =\frac{\int_{-z}^{\infty}u\phi(u)\,du}{\Phi(z)} =\frac{\phi(-z)}{\Phi(z)} =\frac{\phi(z)}{\Phi(z)} =\lambda(z).

Therefore \(E(\varepsilon\mid Y>0,x)=\sigma\lambda(z)\). Adding the index \(\eta\) gives \(E(Y\mid Y>0,x)=\eta+\sigma\lambda(z)\). For the unconditional observed mean, the censored cases contribute zero, so \(E(Y\mid x)=P(Y>0\mid x)E(Y\mid Y>0,x)\). Multiplying out and using \(\Phi(z)\lambda(z)=\phi(z)\) yields \(\eta\Phi(z)+\sigma\phi(z)\).

Worked Example: Calculating All Three Quantities

Suppose a hypothetical observation has \(\eta=1\) outcome unit and \(\sigma=2\) outcome units. Then \(z=1/2=0.5\). The standard normal values are \(\Phi(0.5)\approx0.691462\) and \(\phi(0.5)\approx0.352065\), giving \(\lambda(0.5)\approx0.509160\).

Probability of a positive outcome. \(P(Y>0\mid x)=\Phi(0.5)\approx0.691462\), or about 69.15%.

Conditional mean among positive outcomes. \(E(Y\mid Y>0,x)=1+2(0.509160)=2.01832\) outcome units.

Unconditional observed mean. Using the probability times the positive-outcome mean gives approximately \(0.691462(2.01832)=1.39559\) outcome units. The direct formula gives \(1(0.691462)+2(0.352065)=1.39559\) outcome units, the same result up to rounding.

Marginal Effects for a Continuous Covariate

For a continuous covariate \(x_j\), let \(\beta_j\) be its coefficient in the index. A marginal effect is a derivative with respect to \(x_j\), holding the other covariates fixed. Since \(\partial\eta/\partial x_j=\beta_j\) and \(\partial z/\partial x_j=\beta_j/\sigma\), differentiating the positive-outcome probability gives:

\frac{\partial P(Y>0\mid x)}{\partial x_j} =\frac{\beta_j}{\sigma}\phi(z).

The derivative of the inverse Mills ratio is \(\lambda'(z)=-\lambda(z)\{z+\lambda(z)\}\). Differentiating the conditional mean therefore gives:

\frac{\partial E(Y\mid Y>0,x)}{\partial x_j} =\beta_j\left[1-\lambda(z)\{z+\lambda(z)\}\right].

This is the change in the mean outcome among positive cases, not the change in the unconditional mean. The unconditional observed-outcome effect follows by differentiating its mean formula:

\begin{aligned} \frac{\partial E(Y\mid x)}{\partial x_j} &=\beta_j\left[\Phi(z)+z\phi(z)-z\phi(z)\right]\\ &=\beta_j\Phi(z). \end{aligned}

The cancellation uses \(\eta/\sigma=z\) and the derivative \(\phi'(z)=-z\phi(z)\). The resulting unconditional effect is the latent-index coefficient multiplied by the probability of a positive outcome.

Decomposing the Unconditional Effect

Write \(P=\Phi(z)\) for the positive-outcome probability and \(m=\eta+\sigma\lambda(z)\) for the conditional mean among positive outcomes. The unconditional mean is \(Pm\). The product rule separates its derivative into a probability channel and a conditional-mean channel:

\begin{aligned} \frac{\partial E(Y\mid x)}{\partial x_j} &=m\frac{\partial P}{\partial x_j} +P\frac{\partial m}{\partial x_j}\\ &=\underbrace{\beta_j\phi(z)\{z+\lambda(z)\}}_{\text{probability channel}} +\underbrace{\beta_j\left[\Phi(z)-\phi(z)\{z+\lambda(z)\}\right]}_{\text{conditional-mean channel}}\\ &=\beta_j\Phi(z). \end{aligned}

The probability channel is the change in the chance of a positive outcome, multiplied by the current conditional mean. The conditional-mean channel is the change in the mean among positive outcomes, weighted by the current probability of being positive. These are local effects at the specified covariate values; both generally vary with \(z\).

Worked Example: Continuous-Covariate Decomposition

Consider a hypothetical Tobit model with \(\eta=1\), \(\sigma=2\), and \(\beta_j=0.8\) outcome units per unit of \(x_j\). Thus \(z=0.5\), \(\Phi(z)\approx0.691462\), \(\phi(z)\approx0.352065\), and \(\lambda(z)\approx0.509160\). From the previous calculation, the positive-outcome conditional mean is about \(2.01832\) outcome units.

Probability effect. \(\partial P/\partial x_j=(0.8/2)(0.352065)=0.140826\). A one-unit local increase in \(x_j\) is associated with an increase of about 0.1408 in the probability of a positive outcome at this covariate value.

Conditional-mean effect. Here \(z+\lambda(z)=0.5+0.509160=1.009160\), so \[ \frac{\partial m}{\partial x_j} =0.8[1-(0.509160)(1.009160)] \approx0.38894 \] outcome units per unit of \(x_j\).

Probability channel. Using the displayed values, \(m(\partial P/\partial x_j)\approx(2.01832)(0.140826)=0.28423\) outcome units per unit of \(x_j\).

Conditional-mean channel and total. \(P(\partial m/\partial x_j)\approx(0.691462)(0.38894)=0.26894\) outcome units per unit of \(x_j\). The two displayed channel contributions sum to approximately \(0.55317\). Directly, \(\beta_j\Phi(z)=(0.8)(0.691462)=0.55317\) outcome units per unit of \(x_j\). The small difference, if any, in the last digit is due to rounding.

Binary Covariates Require a Finite Decomposition

For a binary covariate, a derivative is not the relevant contrast. Instead, compare the model predictions at \(x_j=0\) and \(x_j=1\). If \(P_0,m_0\) and \(P_1,m_1\) are the corresponding probabilities and conditional means, the change in the unconditional mean has an exact decomposition:

$$ P_1m_1-P_0m_0 =(P_1-P_0)m_0+P_1(m_1-m_0). $$

This version assigns the probability change the baseline conditional mean \(m_0\), and weights the conditional-mean change by the probability in state 1. Other algebraic decompositions are possible, but the total difference is the same. A finite contrast should not be described as a derivative or as the effect of a small continuous change.

Worked Example: Decomposing a Binary-Covariate Contrast

Suppose a hypothetical binary covariate changes the index from \(\eta_0=0\) to \(\eta_1=1\), with \(\sigma=1\). At state 0, \(z_0=0\), so \(P_0=0.5\), \(m_0=\lambda(0)\approx0.797885\), and the unconditional mean is \(P_0m_0\approx0.398942\) outcome units.

At state 1, \(z_1=1\), \(\Phi(1)\approx0.841345\), and \(\phi(1)\approx0.241971\). Thus \(\lambda(1)\approx0.287600\), \(P_1\approx0.841345\), and \(m_1=1+0.287600=1.287600\) outcome units. The unconditional mean is \(P_1m_1=1(0.841345)+0.241971\approx1.083315\) outcome units.

The total difference is \(1.083315-0.398942=0.684373\) outcome units. For the decomposition, \(P_1-P_0\approx0.341345\) and \(m_1-m_0\approx0.489715\). The probability channel is \((0.341345)(0.797885)\approx0.272354\) outcome units. The conditional-mean channel is \((0.841345)(0.489715)\approx0.412019\) outcome units. Their sum, \(0.272354+0.412019=0.684373\), agrees with the total contrast up to rounding.

Interpretation and a Common Pitfall

The three quantities answer different questions. The derivative of \(\Phi(z)\) describes how a covariate changes the probability of avoiding censoring. The derivative of \(\eta+\sigma\lambda(z)\) describes how it changes the mean among positive outcomes. The derivative \(\beta_j\Phi(z)\) describes how it changes the mean of the observed outcome, counting censored zeros as zero. Each effect is in its stated units: probability change per covariate unit for the first, and outcome units per covariate unit for the mean effects.

Pitfall: “The Tobit coefficient is \(0.8\), so one more unit of the covariate raises the observed outcome by \(0.8\) units.” This treats the latent-index coefficient as the observed-outcome marginal effect. The correction is that the unconditional marginal effect is \(\beta_j\Phi(z)\), while the probability and positive-mean channels each have their own formulas.
Pitfall: Calling the conditional mean among positive outcomes the unconditional outcome mean. The correction is that the unconditional mean equals the positive-outcome probability times the conditional mean, because censored observations contribute zero.
Pitfall: Applying a derivative formula to a binary covariate and calling it the 0-to-1 effect. The correction is to compute predictions in both states and compare them using a finite difference.

These formulas describe model-based associations conditional on the covariates and Tobit assumptions. A causal interpretation requires additional design or identification assumptions; the decomposition alone does not establish causation. In applied reporting, it is useful to state the covariate values at which continuous marginal effects are evaluated, or to specify the two states compared for a discrete covariate.

Key takeaway: Under the left-censored normal Tobit model, the unconditional marginal effect is \(\beta_j\Phi(z)\). It decomposes into a probability channel and a conditional-mean channel, whose covariate-specific contributions sum to that total.

Check Your Understanding

Use the model definitions and decomposition above to answer these questions.

Why is the conditional mean among positive outcomes not the same as the unconditional observed-outcome mean?
For a continuous covariate, what is the marginal effect on the probability of a positive outcome?
Explain in words what \(m(\partial P/\partial x_j)\) represents in the product-rule decomposition.
If \(z=0\), \(\sigma=1\), and the latent-index coefficient for a continuous covariate is \(0.5\), what is the unconditional marginal effect on the observed mean?
Why should a binary covariate’s 0-to-1 contrast be calculated as a finite difference rather than by interpreting a derivative?