OLS with Fixed vs Random \(X\): What Actually Changes?

Background

In regression courses, you will eventually hear the phrase: “OLS works whether \(X\) is fixed or random.” That statement is correct, but dangerously compressed.

The distinction between fixed and random regressors is not about how you compute \(\hat\beta\). The algebra is identical. The difference is in what is random, what we condition on, and how we interpret sampling statements.

The goal of this note is to make that distinction precise, and to clarify what does—and does not—depend on treating \(X\) as fixed.

Notation

Consider the well-known linear model

\[Y = X\beta + \varepsilon,\]

where:

• \(Y \in \mathbb{R}^n\),
• \(X \in \mathbb{R}^{n \times p}\) with full column rank,
• \(\beta \in \mathbb{R}^p\),
• \(\varepsilon \in \mathbb{R}^n\) with \(\mathbb{E}[\varepsilon \mid X] = 0\) and \(\operatorname{Var}(\varepsilon \mid X) = \sigma^2 I\).

The standard OLS estimator is

\[\hat\beta = (X^\top X)^{-1} X^\top Y.\]

The key question is: are we conditioning on \(X\), or is \(X\) itself a random object in the data-generating process?

A Closer Look

Let’s take a closer look at the two cases.

Fixed \(X\): Classical Linear Model

In the classical setup, \(X\) is treated as fixed (non-stochastic). Then, all randomness comes from the error term \(\varepsilon\).

Conditional on \(X\),

\[\mathbb{E}[\hat\beta \mid X] = \beta,\] \[\operatorname{Var}(\hat\beta \mid X) = \sigma^2 (X^\top X)^{-1}.\]

Inference is therefore conditional inference. Confidence intervals and \(t\)-tests are statements about the distribution of \(\hat\beta\) given this specific design matrix, \(X\).

This framework is natural in designed experiments, where \(X\) is literally chosen by the researcher.

Random \(X\): Econometric View

In most observational settings, \(X\) is random. We observe i.i.d. draws \((X_i, Y_i)\) from an unknown joint distribution, \(F_{X,Y}\). Under standard regularity conditions, the same OLS estimator, \(\hat\beta\), satisfies

\[\hat\beta \;\xrightarrow{p}\; \beta \quad \text{if} \quad \mathbb{E}[X_i \varepsilon_i] = 0.\]

The asymptotic variance becomes

\[ \operatorname{Avar}(\hat\beta)=\left( \mathbb{E}[X_i X_i^\top] \right)^{-1}\mathbb{E}[X_i X_i^\top \varepsilon_i^2]\left( \mathbb{E}[X_i X_i^\top] \right)^{-1}.\]

Under homoskedasticity, this simplifies to

\[\sigma^2 \left( \mathbb{E}[X_i X_i^\top] \right)^{-1}.\]

The algebra mirrors the fixed-\(X\) case, but the interpretation changes: we are no longer conditioning on a specific realization of \(X\); we are averaging over its distribution.

What Actually Changes?

Three things matter.

First, the object of inference. With fixed \(X\), inference is conditional on the design. With random \(X\), inference is about repeated sampling of \(F_{X,Y}\).

Second, exogeneity assumptions. In the fixed-\(X\) model, we require \(\mathbb{E}[\varepsilon \mid X] = 0\). In the random-\(X\) case, we need the same condition, but it now constrains the joint distribution: it says that once we know the regressors, there is no systematic remaining signal in the error term. Violations become statements about endogeneity, meaning \(X\) is statistically related to omitted factors inside \(\varepsilon\).

Third, robustness. Heteroskedasticity-robust standard errors are naturally derived in the random-\(X\) framework, where the conditional variance may depend on \(X_i\). In other words, different parts of the regressor distribution can come with different noise levels, so inference has to account for that variation rather than rely on a single common variance.

What does not change is the formula for \(\hat\beta\). Nor does unbiasedness depend on \(X\) being fixed; it depends on the conditional mean-zero assumption.

Bottom Line

• The OLS estimator is algebraically identical whether \(X\) is fixed or random.
• Fixed-\(X\) inference is conditional; random-\(X\) inference averages over the joint distribution.
• Consistency hinges on \(\mathbb{E}[X_i \varepsilon_i] = 0\), not on whether \(X\) is stochastic.
• Robust variance formulas arise naturally once \(X\) is treated as random.

Where to Learn More

For a classical treatment, see Greene’s Econometric Analysis, which clearly distinguishes fixed and stochastic regressors. Wooldridge’s Econometric Analysis of Cross Section and Panel Data provides a modern random-\(X\) perspective with emphasis on exogeneity conditions and robust inference.

References

Greene, W. H. (2018). Econometric Analysis.

Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data.