Hurdle Models Versus Tobit Models

Why the Source of Zero Expenditures Matters

Suppose a survey records each household’s monthly spending on a particular service. Many households report zero, while others report positive amounts. The same observed pattern can arise for different reasons. Some households may make a distinct decision about whether to use the service, followed by a separate decision about how much to spend. Alternatively, every household may have a latent desired expenditure, with values at or below zero recorded as zero. These explanations lead to different statistical models and different interpretations.

A hurdle model treats the zero-versus-positive outcome and the positive amount as two linked parts of the observed distribution. A Tobit model treats a zero as the censored observation of a latent outcome that is nonpositive. The key distinction is not whether the data contain zeros: both models can describe zeros. It is whether the zero probability and positive amounts are modeled separately or are both implied by one latent-outcome model.

Definition: In a hurdle model, an outcome is zero with probability $1-q(x)$ and positive with probability $q(x)$. Conditional on being positive, its amount follows a separately specified distribution supported on positive values. In a left-censored Tobit model with censoring point zero, the observed outcome is $Y=\max(0,Y^*)$, where the latent outcome $Y^*$ follows a regression model. The Tobit model’s probability of zero and distribution of positive outcomes are therefore linked by the latent regression and its error distribution.

The Hurdle Model: Separate Participation and Amount

Let $Y$ be expenditure and $x$ denote observed covariates. A hurdle model specifies the probability of a positive outcome, $q(x)=P(Y>0\mid x)$, and a conditional distribution for $Y$ given $Y>0$. Write the conditional positive mean as $m_+(x)=E[Y\mid Y>0,x]$. The model does not, by definition, require the covariates or parameters determining $q(x)$ to be the same as those determining the positive amounts.

P(Y=0\mid x)=1-q(x), \qquad f_Y(y\mid x)=q(x)g_+(y\mid x),\quad y>0, \qquad E[Y\mid x]=q(x)m_+(x).

Here $g_+$ is a density on positive amounts, and its integral over positive values is one. For example, a researcher might use a binary-response model for whether a household spends anything and a lognormal or gamma distribution for expenditure among spenders. The separate positive-outcome distribution is not required to be any particular one; it should be chosen to represent the positive data and research question.

The equation for the unconditional mean follows by conditioning on whether expenditure is positive: zero expenditure contributes nothing, while positive expenditure occurs with probability $q(x)$ and has conditional mean $m_+(x)$. Thus a covariate can change expected spending by changing participation, the amount among participants, or both. This decomposition is descriptive of the model and does not by itself establish a causal effect.

Worked Example: A Hurdle Model for Monthly Service Spending

Consider hypothetical data on 200 households’ monthly spending on a service. Of these, 110 households spend $0 and 90 spend a positive amount. Suppose the mean among the 90 positive spenders is $120 per month. Calculate the observed participation share and overall sample mean, then interpret them.

The observed positive share is $90/200=0.45$, and the observed zero share is $110/200=0.55$. The positive-spending mean is $120 per month. The corresponding overall sample mean is

\bar{Y} = \frac{110(0)+90(120)}{200} = \frac{10{,}800}{200} = \$54.00\text{ per household per month}.

Equivalently, the sample mean is the positive share times the mean among positive spenders: $0.45\times \$120=\$54.00$. In a hurdle-model interpretation, the two quantities answer distinct questions: 45% of these households spend anything, and positive spenders average $120 per month. This calculation is a description of the hypothetical sample, not evidence that the same values hold in a wider population.

The Tobit Model: Censoring a Latent Outcome

In the standard left-censored Tobit model at zero, the latent outcome and observed outcome are

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

The assumption that the error is conditionally normal with mean zero and common variance $\sigma^2$ is part of this standard specification. It implies that the probability of observing a positive amount and the distribution of the amount among positive observations both come from the same latent regression. In particular, $Y=0$ means the latent value was at or below zero; it does not mean that a household’s latent desired expenditure was exactly zero.

Write $\mu=x'\beta$ and $a=\mu/\sigma$. The Tobit model implies $P(Y>0\mid x)=\Phi(a)$ and $P(Y=0\mid x)=\Phi(-a)$, where $\Phi$ is the standard normal cumulative distribution function. For $y>0$, the observed density is the latent normal density at $y$; conditional on being positive, that density is divided by $\Phi(a)$. Thus the positive distribution is not freely chosen independently of the zero probability.

The observed-outcome mean is not simply the latent mean $\mu$. Negative latent values are replaced by zero, rather than appearing as negative observed expenditures. The standard Tobit model implies

E[Y\mid x] = \mu\Phi(a)+\sigma\phi(a) = P(Y>0\mid x)\,E[Y\mid Y>0,x],

where $\phi$ is the standard normal density. The first expression gives the unconditional mean of the censored outcome; the second expresses the same mean as positive probability times conditional positive mean.

Worked Example: Tobit Predictions for Monthly Spending

For a hypothetical household profile, suppose a Tobit model has latent mean $\mu=\$80$ per month and error standard deviation $\sigma=\$50$ per month. Calculate the probability of positive spending, the mean among positive spenders, and the overall expected observed spending. Assume the standard Tobit model above.

First, $a=80/50=1.60$. Standard normal values are $\Phi(1.60)\approx0.9452$ and $\phi(1.60)\approx0.1109$. Therefore, the model-implied probability of positive spending is about 0.9452, and the probability of a zero is $\Phi(-1.60)\approx0.0548$.

The conditional mean among positive spenders is

E[Y\mid Y>0,x] = \mu+\sigma\frac{\phi(a)}{\Phi(a)} \approx \$80+\$50\left(\frac{0.1109}{0.9452}\right) \approx \$85.87\text{ per month}.

The overall expected observed expenditure is then

E[Y\mid x] \approx 0.9452\times \$85.87 \approx \$81.16\text{ per month}.

The unrounded Tobit mean formula gives the same result to two decimal places: $80(0.9452)+50(0.1109)\approx81.16$. The model predicts a positive amount for most households with this profile, and its expected observed expenditure includes the small probability of a zero. The latent mean of $80 is not the observed-outcome mean.

Why the Models Impose Different Restrictions

A short derivation makes the Tobit restriction clear. For $Y^*=\mu+\sigma Z$, where $Z\sim N(0,1)$, the observed outcome is zero when $Z\leq-a$ and equals $\mu+\sigma Z$ when $Z>-a$. Hence

E[Y\mid x] = \int_{-a}^{\infty}(\mu+\sigma z)\phi(z)\,dz = \mu\Phi(a)+\sigma\phi(a).

For the last equality, $\int_{-a}^{\infty}\phi(z)\,dz=\Phi(a)$, by symmetry of the standard normal distribution, and $\int_{-a}^{\infty}z\phi(z)\,dz=\phi(-a)=\phi(a)$, since the derivative of $-\phi(z)$ is $z\phi(z)$. Dividing the mean by $P(Y>0\mid x)=\Phi(a)$ yields the conditional positive mean in the example. The same $a$, $\mu$, and $\sigma$ determine both the participation probability and the positive-amount distribution.

A hurdle model makes no such latent-normal link by default. Its probability of positive spending can be low while its positive amounts are large, or high while positive amounts are modest. Whether particular covariates appear in both parts is a modeling choice. In Tobit, by contrast, a covariate that shifts $\mu$ changes both the probability of crossing zero and the distribution of positive amounts according to the model’s shared latent index.

Feature	Hurdle model	Left-censored Tobit model
Meaning of zero	A separate zero outcome in the participation part	A latent outcome at or below the censoring point
Positive probability	Specified by $q(x)$	Implied by the latent regression and error distribution
Positive amounts	Have a separately specified positive distribution	Follow the latent distribution conditional on exceeding zero
Link between the two parts	Not imposed unless the researcher chooses it	Imposed by the shared latent model and its assumptions
Typical substantive question	How do factors relate to spending at all versus the amount among spenders?	How does a latent outcome, censored at zero, map to observed spending?

Worked Comparison: Contributions for a Zero and a Positive Amount

Worked Example: Comparing Model-Implied Likelihood Contributions

For illustration, consider one hypothetical household reporting $0 and another reporting $100 in monthly expenditure. Use a hurdle specification with positive probability $q=0.45$ and an exponential positive-amount distribution with mean $120. Compare the observations’ likelihood contributions with those from the Tobit specification $\mu=\$80$, $\sigma=\$50$. These specified parameter values illustrate what each model implies; they are not fitted to a common dataset.

Under the hurdle model, the zero observation has probability $1-q=0.55$. For a positive amount $y$, the exponential density is $g_+(y)=(1/120)\exp(-y/120)$, measured per dollar. The $100 observation therefore contributes

qg_+(100) = 0.45\left(\frac{1}{120}e^{-100/120}\right) \approx 0.001630\text{ per dollar}.

The contribution is the probability of crossing the hurdle multiplied by the density of the positive amount. The zero observation and the positive observation use different parts of the model.

For the Tobit specification, the zero probability is $\Phi(-1.60)\approx0.0548$. For a positive observed amount of $100, the contribution is the latent normal density evaluated at that amount:

\frac{1}{50}\phi\left(\frac{100-80}{50}\right) = \frac{\phi(0.40)}{50} \approx 0.007365\text{ per dollar}.

The Tobit model’s positive density is tied to the same latent normal distribution that gives its zero probability. The contributions show how the models assign probability and density under the stated assumptions; comparing two individual contributions does not establish which model better fits a dataset. That requires considering the full data and a suitable model-comparison strategy.

Choosing the Model to Match the Question

The choice should follow the outcome-generating account and the target of interpretation, rather than the mere presence of many zeros. A hurdle model is often a natural specification when any expenditure represents a distinct participation or purchasing decision and positive spending has its own distribution. It allows the determinants of spending anything to differ from the determinants of the amount conditional on spending.

A Tobit model is appropriate when it is substantively plausible that a continuous latent outcome exists for every observation and values at or below the censoring threshold are recorded at the threshold. It is more restrictive: the latent conditional distribution and the censoring rule jointly determine zeros and positive amounts. Those restrictions can be useful if credible, but they should not be adopted simply because a Tobit model can produce zeros.

For expenditure, a recorded zero might mean no purchase, no use, a decision not to participate, or a value censored by a reporting or measurement rule. These interpretations are not interchangeable. The researcher should clarify what the observation process means and what quantity the analysis aims to describe before selecting a model. As with other regression models, a covariate association in either part is not automatically a causal effect.

Common Mistake

Pitfall: “There are many zero expenditures, so a Tobit model is automatically appropriate.” This reasoning confuses the observed distribution with the mechanism producing it. Both a hurdle model and a Tobit model can generate zero observations, but only Tobit interprets them as latent outcomes at or below a censoring point and links their probability to positive amounts through a specified latent distribution. Identify the meaning of zero and assess whether that shared latent-distribution restriction is credible.

A second error is to interpret the Tobit latent mean as the average observed expenditure. The latent mean can include negative values that are recorded as zero. The observed mean instead uses the censoring rule, as in $E[Y\mid x]=\mu\Phi(\mu/\sigma)+\sigma\phi(\mu/\sigma)$. In a hurdle model, the corresponding unconditional mean is $q(x)m_+(x)$.

Key takeaway: Hurdle and Tobit models answer different questions. Hurdle models specify the probability of any positive outcome and the distribution of positive amounts as separate components. Tobit models derive both from one censored latent outcome. The appropriate model depends on the substantive meaning of zeros and the credibility of the restrictions linking zeros to positive values.

Check Your Understanding

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

In a hurdle model, what does $q(x)$ represent, and how does it combine with the conditional positive mean to give the unconditional mean?
In a Tobit model censored at zero, what latent-outcome event produces an observed zero?
A hurdle model predicts a positive-spending probability of 0.30 and a mean of $200 per month among positive spenders. What is its implied unconditional mean?
Why can a Tobit model not generally choose its zero probability and positive-amount distribution independently?
What substantive feature of the data-generating process would make a Tobit interpretation more credible than a hurdle interpretation?