MerQuR+Statistics Glossary

Statistics Glossary → Tobit Regression (Censored Regression)

Tobit Regression (Censored Regression)analysis

A regression model for outcomes that pile up at a lower or upper limit (e.g. many zeros), which treats values beyond the limit as censored when estimating the effects of the predictors.

Concept

Some measurements cannot go below or above a certain limit: people who spend nothing on a product are recorded as spending 0, and the true ability of those who get full marks on a test is cut off at the ceiling. A large share of the data then piles up at the limit. Ordinary linear regression (OLS) treats this pile-up as if it were ordinary values and typically pulls the slopes towards zero; the Tobit model was developed to remove this bias.

The model assumes that behind the observed value lies a latent variable y* that is not cut off by the limit: y* = β₀ + β₁x₁ + … + ε, ε ~ N(0, σ²). If y* falls inside the limit it is observed directly; if it falls beyond, all we know is that it is 'at or beyond the limit'. The likelihood uses the normal density for observations inside the limit and the probability of falling beyond the limit for censored ones, and the coefficients are estimated by maximum likelihood.

The coefficients are on the scale of the latent variable y*; to describe changes in the observed outcome, the marginal effects (on the observed mean, on the mean of those beyond the limit and on the probability of crossing it) should be read. The model assumes normally distributed errors with constant variance; unlike OLS, violating these assumptions can bias not only the standard errors but the coefficients themselves. It also assumes that the same factors drive both the decision to participate and the amount, in the same direction.

When to use it

  • When a continuous outcome clearly piles up at a limit (zero spending, zero hours worked, test scores hitting the ceiling) and values beyond the limit can be regarded as 'unobserved'.
  • If the zeros and the positive amounts come from different processes (e.g. the decision 'will they buy?' and the amount 'how much?' depend on different factors), two-part (hurdle) models or the Heckman selection model are more suitable.
  • If normality or constant variance is clearly violated, methods with weaker distributional assumptions, such as censored quantile regression (e.g. Powell's CLAD estimator), can be considered.

An everyday example

A supermarket chain wants to explain loyalty-card holders' monthly spending on organic products by income, household size and age. Of 1,000 customers, 600 never bought any organic product, so their spending shows as 0. A linear regression that takes these zeros as ordinary values understates the effect of income. The Tobit model treats the zero spenders as 'those whose interest in organic products falls below the threshold' and estimates how income raises both the chance of buying organic products and the amount spent by those who do.

Parameters in MerQur

Predictors (X) and Target (Y)

Concept: In Tobit the dependent variable is a continuous measure that piles up at a limit; the predictors are the variables that explain the latent tendency behind that measure.

Everyday example: In the supermarket example the target is monthly organic spending and the predictors are income, household size and age. Customers who spent 0 stay in the data; the model does not drop them but uses them as censored.

What it does: Selects the censored continuous dependent variable and the predictors.

Effect on the result: All outputs.

On the validation data: UCLA tobit, right censor 800: read coefficient 2.697939 (SE 0.618798), sigma 65.67672.

Left censor and Apply left censoring

Concept: Left censoring means a value cannot fall below a lower limit, so the true value of observations at or below that limit is unknown. The most common case is a pile-up at zero.

Everyday example: The 600 customers who bought no organic products show spending of 0; yet some of them are almost interested while others are not at all. Setting the left limit at 0 tells the model 'these people's tendency lies somewhere below 0, its exact value unknown'.

What it does: Sets the left censoring limit (observations with y <= L are left-censored; 0 = pile-up at zero). If the "Apply left censoring" box is not ticked, there is no left censoring (a right-censored-only model). If both sides are off, a validation warning is given.

Effect on the result: The number of censored observations and all estimates.

On the validation data: Affairs L = 0: 451 censored observations, rating coefficient -2.2850; L = -1: uncensored model (log-likelihood -1527.27). In UCLA, L = 500 gives a two-sided model.

→ Censoring (survival analysis)

Right censor and Apply right censoring

Concept: Right censoring means a value cannot exceed an upper limit, so for observations at or above that limit we only know that they are 'at least this much'.

Everyday example: If the store system records monthly spending in detail only up to 2,000 TL and stores anything above as '2,000 TL or more', those customers' true spending is unknown. Setting the right limit at 2,000 treats these records as cut off at the ceiling; both limits can be used together.

What it does: Sets the right censoring limit (observations with y >= the right limit are censored). While the box is ticked, 0 and negative values are also valid; when the value is changed, the box is ticked automatically.

Effect on the result: The number of censored observations and all estimates.

On the validation data: UCLA right limit 800: 17 censored observations, read 2.697939 (0.618798), sigma 65.67672; Affairs right limit 4: 80 censored, rating -2.2070.

→ Censoring (survival analysis)

Confidence Level (%90, %95, %99, custom)

Concept: Sets how 'sure' the interval estimates for the coefficients and for the error spread (σ) are; the higher the level, the wider the interval.

Everyday example: If each extra 1,000 TL of income is found to raise the latent spending tendency by 40 to 90 TL with 95% confidence, asking for 99% confidence makes this range somewhat wider.

What it does: Sets the Wald confidence interval (z) for the coefficients and the sigma interval.

Effect on the result: Intervals only.

On the validation data: UCLA read 95% [1.4851, 3.9108] (with the z distribution). The sigma interval is a Wald on the log scale [59.20, 72.87].

→ Confidence level and confidence interval

Marginal effects (McDonald-Moffitt)

Concept: Splits the effect on the latent variable into parts that are meaningful for the observed outcome: effects on observed mean spending, on the mean of those who spend, and on the probability of crossing the limit (spending at all). The effect on the observed mean equals the coefficient times the probability of crossing the limit.

Everyday example: Let the income coefficient be 60 TL per 1,000 TL for the latent tendency. If 40% of customers buy organic products, the observed spending of an average customer rises by about 60 x 0.40 = 24 TL. Part of this rise comes from new customers starting to buy organic products and part from existing buyers spending more.

What it does: At the mean of X, gives the effects on the latent outcome y*, the observed E[y], E[y | uncensored] and P(uncensored) with the delta-method SE.

Effect on the result: An additional table only.

On the validation data: Affairs yearsmarried: dE[y*]/dx 0.5541; dE[y]/dx 0.1295 (= b x P(uncensored) 0.2338); dE[y | uncensored]/dx 0.1298; dP(uncensored)/dx 0.0206.

→ Average marginal effect (AME)

Left limit column and Right limit column

Concept: The censoring limit does not have to be the same for everyone; each observation can have its own lower or upper limit. In that case a column giving the limit for each row is used instead of a constant value.

Everyday example: If some branches' systems cap spending records at 2,000 TL and others at 3,000 TL, each customer's row carries the upper limit of their own branch. The model then judges each record against its own ceiling.

What it does: Observation-specific censoring limit (in place of the constant value); an empty cell means there is no limit for that observation.

Effect on the result: When the limit columns equal the constant limit, the result does not change (only the label and note).

On the validation data: With columns of constant 0 and 4, the result equals the constant cut. There is no comparison dataset with limits that vary by observation.

→ Censoring (survival analysis)

Robust (sandwich) SE

Concept: Sandwich standard errors give a more reliable estimate of uncertainty when the assumption about the error spread does not fully hold. In Tobit, however, such violations can bias the coefficients themselves, so robust SEs alone do not fix the problem.

Everyday example: If high-income customers' spending is far more scattered than that of low-income customers, robust SEs show the uncertainty of the income coefficient more realistically. Still, if this difference in spread is large, the coefficient itself should be questioned.

What it does: Gives sandwich (HC0) standard errors.

Effect on the result: SE and intervals only.

On the validation data: Affairs (L = 0) sandwich SE: yearsmarried 0.1372, rating 0.3935.

→ Robust (sandwich) standard errors (HC0-HC3)

Assumption diagnostics (CM)

Concept: Conditional moment tests check the two assumptions Tobit relies on: normally distributed errors and constant variance. If p < α, the assumption may be violated and the coefficients may be biased.

Everyday example: If the normality test is significant, spending among organic buyers may be right-skewed, for instance because of a few very heavy spenders. If the heteroskedasticity test is significant, the spread of spending changes with income or household size; alternatives such as a two-part model should then be considered.

What it does: Gives the conditional moment normality (df 2) and heteroskedasticity (df = number of X) tests.

Effect on the result: An additional table only.

On the validation data: Affairs normality chi-square 4.311, p 0.116; heteroskedasticity 4.986, p 0.418 (by OPG regression).

→ Normality assumption

Example on real data

Data: Fair 1978 Affairs, R AER (limit columns added)

Variables: target affairs (yearly count, piled up at zero), predictors age, yearsmarried, religiousness, occupation, rating

  • Question: How is the number of extramarital affairs (affairs) related to age, years married, religiousness, occupation and marital rating? Since 451 of 601 people reported 0, the dependent variable is left-censored at 0; OLS would be biased, so Tobit is used.
  • Model test: LR χ²(5) = 78.32, p < .001 (null model LL = -744.74); the predictors are jointly significant.
  • Result: Years married increases the latent propensity (b = 0.554, z = 4.12, p < .001); age (b = -0.179, p = .023), religiousness (b = -1.686, p < .001) and marital rating (b = -2.285, z = -5.60, p < .001) decrease it. Occupation is not significant (b = 0.326, p = .200).
  • Model fit: LL = -705.58, AIC = 1425.15, BIC = 1455.94; σ = 8.247 (95% CI 7.23-9.41), the residual spread of the latent variable is large.
  • Limitation: The coefficients are on the scale of the latent (uncensored) variable y*; marginal effects should be computed for the change in the observed number of affairs. Tobit assumes normality and equal variance.
  • APA: Tobit regression (left censoring at 0, 451/601 censored), LR χ²(5) = 78.32, p < .001; rating b = -2.28, z = -5.60, p < .001.

Related concepts

Censoring (survival analysis)Normality assumptionStandard errorAverage marginal effect (AME)Robust (sandwich) standard errors (HC0-HC3)HeteroscedasticityTruncationMaximum likelihood (ML) estimation

The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.