Interpreting 2SLS Coefficients in Linear Models

What a 2SLS Coefficient Measures

A 2SLS coefficient is interpreted first through the model in which it appears: it describes the change in the modeled outcome associated with a one-unit change in a regressor, holding the other included regressors fixed. When that regressor is an endogenous treatment, 2SLS uses instruments to estimate its coefficient. Whether the estimate supports a causal interpretation depends on the model and the instrument assumptions, not on the coefficient’s units alone.

Consider an outcome \(Y_i\), treatment \(D_i\), observed covariates \(X_i\), and excluded instrument(s) \(Z_i\) for unit \(i\). For a single endogenous treatment, a linear structural equation is

Here, \(\tau\) is the structural treatment coefficient, \(\gamma\) is the vector of covariate coefficients, and \(u_i\) contains other determinants of the outcome. The units of \(\tau\) are outcome units per treatment unit. If \(Y_i\) is a test score in points and \(D_i\) is hours of a program, \(\tau\) is points per program hour. If \(D_i\) is an indicator coded zero or one, \(\tau\) is the modeled outcome difference between treatment and control.

When the Coefficient Has a Causal Meaning

The units of the coefficient follow from the model regardless of whether the instrument is valid. A causal interpretation needs more. In this linear setting, the central instrument conditions are that the excluded instruments are relevant for the endogenous treatment, satisfy the exclusion restriction, and are exogenous in the structural equation, conditional on the included covariates. In a conditional moment formulation, exogeneity requires \(E[Z_i u_i\mid X_i]=0\); the instrument must also provide enough independent variation to identify the treatment coefficient after accounting for \(X_i\). These assumptions are substantive and cannot be verified just by reading a 2SLS output table.

The constant-effect specification says that the treatment effect does not vary with the included covariates or across units. In potential-outcome terms, its core implication is that changing treatment from \(d\) to \(d+1\) changes the outcome by the same amount \(\tau\) for each unit, conditional on the model. If effects vary, a single 2SLS coefficient need not equal the effect for every person or the average effect for the whole population; its interpretation can depend on the source of instrument-induced treatment variation. Thus, “one additional unit of treatment changes the outcome by \(\hat{\tau}\)” is a causal statement only when the assumptions justify that target.

The identifying logic can be stated directly. Let \(R_i\) collect the structural regressors, including \(D_i\) and the included covariates, and let \(Q_i\) collect the instruments, including the included covariates as their own instruments. If \(Y_i=R_i'\beta+u_i\) and the instrument moment condition holds, then

If \(E[Q_iR_i']\) has sufficient rank to identify \(\beta\), these population moments determine the structural coefficients. In particular, the moment condition explains why 2SLS targets the structural coefficient rather than the ordinary least-squares association. It does not make the coefficient causal if the instrument assumptions or the structural model are wrong.

Covariates Do Not Change the Treatment Coefficient’s Units

Including covariates changes what is held fixed in the model, but it does not change the basic units of the treatment coefficient. In \(Y_i=\alpha+\tau D_i+X_i'\gamma+u_i\), the coefficient \(\tau\) remains outcome units per treatment unit. Each \(\gamma_j\) is the modeled outcome change per unit of its covariate, holding the other regressors fixed. The causal interpretation of a covariate coefficient is a separate question; its presence as an adjustment variable does not by itself make that coefficient causal.

Interpretation also depends on how variables are measured and coded. If treatment is recorded in tens of hours rather than hours, a one-unit increase means ten hours. If an outcome is measured in thousands of dollars, the coefficient is in thousands of dollars per treatment unit. Always translate the model’s actual coding back into the context before reporting the estimate.

Interactions Make the Treatment Effect Depend on a Covariate

A constant treatment coefficient is not appropriate when the model allows the treatment effect to vary with a measured characteristic. For a covariate \(W_i\), consider

The modeled change in outcome for a one-unit increase in \(D_i\), at a specified value \(W_i=w\), is the derivative with respect to treatment:

Therefore, \(\tau\) is the treatment effect at \(W_i=0\), and \(\delta\) is the change in the treatment effect for a one-unit increase in \(W_i\). If \(W_i\) is binary, \(\tau\) is the treatment effect in the reference group \(W_i=0\), while \(\tau+\delta\) is the effect in the group \(W_i=1\). In an IV model, both \(D_i\) and \(D_iW_i\) are endogenous regressors and need instruments; a common instrument set includes \(Z_i\) and the interaction \(Z_iW_i\), along with the included exogenous regressors. Identification and validity must hold for the model with these regressors and instruments.

Centering and Reporting Conditional Effects

When \(W_i\) is continuous, \(W_i=0\) may be outside the observed or substantively meaningful range. In that case the main treatment coefficient can be hard to interpret even though the model is mathematically well defined. Centering \(W_i\) at a meaningful value \(c\), by defining \(W_i^*=W_i-c\), makes the treatment main effect refer to \(W_i=c\). The fitted treatment effect at a given original value \(w\) remains \(\tau+\delta w\); under the centered model, it is represented by the re-expressed coefficients at \(w-c\).

For example, if \(W_i\) is age and the model is centered at age 40, the treatment main effect is the modeled treatment effect at age 40, not at age zero. Centering changes the reference value and the main-effect coefficient, but not the fitted values or the treatment effect at any fixed age. The interaction coefficient retains the interpretation of how much the treatment effect changes per additional year of age.

When an interaction is present, report the treatment effect at meaningful covariate values rather than interpreting \(\hat{\tau}\) alone as the effect for everyone. Uncertainty for a sum such as \(\hat{\tau}+w\hat{\delta}\) must account for the uncertainty and covariance of both estimates; the point estimate alone is not a confidence interval. The interpretation of the sum is still in outcome units per treatment unit.

Diagnostics, Alternatives, and Common Pitfalls

Before interpreting a 2SLS coefficient, check the actual equation and coding. Confirm which variable is the outcome, what a one-unit treatment change means, and whether the model includes an interaction. With an interaction, identify the reference value for the covariate and calculate the relevant sum. Check that reported units match any rescaling. Finally, assess whether the instrument assumptions and identification conditions are defensible; a plausible-looking coefficient is not evidence that they hold.

A common mistake is to read the main treatment coefficient in an interaction model as the treatment effect for every unit. In the regional example, that would report 240 dollars per month for rural participants. The correction is to substitute \(W_i=1\): \(240+90=330\) dollars per month. A second mistake is to say that a valid instrument has been established because the 2SLS coefficient has sensible units. Units are a matter of model interpretation; instrument validity concerns the assumptions linking \(Z_i\) to the structural error \(u_i\).

OLS and 2SLS coefficients can share the same units and model form, but they generally answer different questions when treatment is endogenous. OLS describes a conditional linear association under its own assumptions; 2SLS uses instrument-induced variation and requires valid instruments for a structural causal interpretation. A first-stage coefficient describes how the instrument predicts treatment, not the effect of treatment on the outcome. In a simple just-identified setting, a reduced-form outcome change and a first-stage treatment change can help explain the source of the IV estimate, but neither should be mistaken for the structural treatment coefficient itself.

A single linear coefficient is also limited when effects vary across units or the functional form is misspecified. Under heterogeneity, the coefficient need not represent a universal effect; under misspecification, its simple per-unit interpretation may not describe the actual outcome relationship. Treat the constant-effect reading as a model-based claim and state the target population and assumptions clearly.