Tags: missing-data-imputation, data-science, statistics, mnar

Demystifying Heckman’s Two-Stage Model (Heckit) for MNAR Data

Heckman’s Two-Stage Model (commonly known as Heckit) is a Nobel Prize-winning statistical technique designed specifically to correct for sample selection bias.

In a Missing Not at Random (MNAR) scenario, data isn't missing by accident. Instead, individuals systematically "select" themselves out of being observed. For example, individuals with the highest wages might refuse to report them, or patients with the most severe side effects might drop out of a clinical trial.

If you run a standard Ordinary Least Squares (OLS) regression on only the available data, your coefficients will be deeply biased because you are ignoring the hidden mechanism governing why the data is missing. Heckit resolves this by treating missingness as an explicit, modelable choice.

---

The Mathematical Problem Heckit Solves

Imagine you want to model Wage based on Education. However, you can only observe wages for people who are actually employed.

1. The Primary Equation (Outcome): $Y = X\beta + u_1$ (Where $Y$ is Wage, $X$ is Education, and $u_1$ is the error term)

2. The Selection Equation (Missingness): $Z^* = W\gamma + u_2$ (Where $Z^*$ is a latent variable representing the propensity to work, $W$ represents factors influencing that choice like childcare or unearned income, and $u_2$ is the error term)

We only observe $Y$ if $Z^* > 0$. If $u_1$ and $u_2$ are correlated ($\rho \neq 0$), it means the unobserved factors that make someone likely to work (like high personal motivation) are the exact same unobserved factors that give them a higher wage. Standard regression fails here because the error term no longer averages out to zero.

---

How the Heckit Framework Works (The Two Stages)

Heckit elegantly bypasses this issue by breaking the problem into two distinct statistical steps:

Stage 1: The Selection Equation (Probit)

• The Process: You run a Probit regression on your entire dataset (both the complete cases and the missing cases). The binary dependent variable is whether the outcome data is observed ($1$) or missing ($0$).
• The Output: This stage estimates the probability that an individual sample will be observed based on their characteristics ($W$).
• The Inverse Mills Ratio (IMR): From this probit model, the algorithm calculates a value for every single data point called the Inverse Mills Ratio, usually denoted as $\lambda$ (lambda). The IMR represents the odds ratio of a sample being observed versus being missing given its observed features.

Stage 2: The Outcome Equation (OLS)

• The Process: You run your primary linear regression (OLS) only on the complete cases where data is observed.
• The Twist: You add the Inverse Mills Ratio ($\lambda$) computed from Stage 1 as an additional explanatory variable (a regressor) into the model.
• The Result: The model now looks like this: $Y = X\beta + \beta_\lambda \lambda + e$
• By including $\lambda$, you effectively soak up the variance caused by the selection bias. If the coefficient for $\lambda$ ($\beta_\lambda$) is statistically significant, it proves selection bias was present, and your newly adjusted $\beta$ coefficients are now mathematically unbiased.

---

🔑 Critical Requirement: The Exclusion Restriction

For Heckit to work reliably, you must satisfy a strict rule called the Exclusion Restriction.

Your Stage 1 model ($W$) must include at least one variable—known as an Instrumental Variable—that directly affects whether the data is missing, but has absolutely no direct effect on the final outcome ($Y$).

• Good Example (Wages): Number of young children at home. It heavily influences whether someone chooses to enter the workforce (Stage 1 missingness), but it does not directly dictate what an employer sets as a market hourly wage rate (Stage 2 outcome).
• Bad Example: Past work experience. It influences both whether you get a job and how much you get paid, failing the exclusion rule.

Note: Without a valid instrument, Heckit suffers from severe multicollinearity, making your final estimates highly unstable.

---

Summary Comparison: OLS vs. Heckit

Feature	Standard Regression (OLS)	Heckman Two-Stage (Heckit)
Data Scope	Uses complete cases only, discards missing rows.	Uses all rows for Stage 1; complete rows for Stage 2.
MNAR Capacity	Biased. Completely ignores the cause of missingness.	Unbiased. Explicitly models and penalizes selection bias.
Primary Risk	Misleading conclusions and skewed parameters.	Highly sensitive to choosing a bad instrumental variable.