Time-Invariant Regressors In Fixed-Effects Models

Why Time-Invariant Regressors Raise a Problem

A panel data set follows units—such as people, firms, or regions—over time. Some explanatory variables change from one observation to the next; others, such as a person's baseline characteristic or a firm's location, remain constant. A unit fixed-effects model can account for persistent differences across units, but this has a direct consequence: a regressor that is constant within each unit cannot be separated from that unit's fixed effect.

This is an identification issue, not evidence that the characteristic has no effect. The data do not distinguish a separate coefficient on the characteristic from the unrestricted unit effects. Interactions require a more careful check: a time-invariant characteristic can help define how an effect changes with a regressor that varies over time, because the product may itself vary within units.

Definition: In a panel with unit \(i\) observed at times \(t\), a regressor \(z_i\) is time-invariant if its value is the same at every observed time for unit \(i\). A unit fixed-effects model includes a separate intercept \(\alpha_i\) for each unit. A coefficient is identified when the observed design distinguishes its contribution from the contributions of the other model terms.

Within Transformation Removes the Main Effect

Consider the model with one time-varying regressor \(x_{it}\) and one time-invariant regressor \(z_i\):

y_{it}=\alpha_i+\beta x_{it}+\gamma z_i+u_{it}.

Here \(y_{it}\) is the outcome, \(\alpha_i\) is the unit effect, \(u_{it}\) is the disturbance, and \(\gamma\) is the proposed coefficient on \(z_i\). For each unit, subtract its time average from each variable. Write \(\bar y_i\), \(\bar x_i\), and \(\bar u_i\) for that unit's averages over its observed periods. Since \(z_i\) is constant, its average is \(z_i\). Thus the transformed equation is

y_{it}-\bar y_i =\beta(x_{it}-\bar x_i)+( \gamma z_i-\gamma z_i)+(u_{it}-\bar u_i) =\beta(x_{it}-\bar x_i)+(u_{it}-\bar u_i).

The unit effect disappears because it is constant within the unit, and the \(z_i\) term disappears for the same reason. The within-transformed data contain no separate variation with which to estimate \(\gamma\). By contrast, \(x_{it}\) can contribute to estimating \(\beta\) if it varies within units and is not collinear, after transformation, with the other included regressors. This is the same within-unit identification principle used in the fixed-effects estimator.

Worked Example: Demeaning a Time-Invariant Regressor

Suppose a hypothetical panel contains two firms observed for three years. Let \(z_i\) indicate whether a firm is in the North, and let \(x_{it}\) be annual training hours. The outcome is a hypothetical productivity score. For Firm A, \(z_A=1\) and the observations are \((x,y)=(2,10),(4,13),(6,12)\). For Firm B, \(z_B=0\) and the observations are \((1,8),(2,9),(3,11)\).

For Firm A, \(\bar x_A=(2+4+6)/3=4\) and \(\bar y_A=(10+13+12)/3=35/3\). Its demeaned \(z\) values are \(1-1=0\) in all three years. For Firm B, \(\bar x_B=(1+2+3)/3=2\), \(\bar y_B=(8+9+11)/3=28/3\), and its demeaned \(z\) values are also all zero because \(0-0=0\).

The within-transformed training values are \((-2,0,2)\) for Firm A and \((-1,0,1)\) for Firm B. There is within-firm variation in training, but none in the North indicator. A within regression can use the former variation to estimate a training slope, subject to the model's other regressors and identifying conditions. It cannot estimate a separate coefficient on North.

The Same Problem Appears as Rank Deficiency

The result is not a quirk of the within transformation. In a regression with unit indicators, the column for \(z_i\) is an exact linear combination of the unit-indicator columns. To see this, let \(D_i\) be the column that equals one for observations belonging to unit \(i\) and zero otherwise. On every row of the panel, exactly one unit indicator equals one. Therefore the column of repeated \(z_i\) values satisfies

z=\sum_{i=1}^{N}z_iD_i.

If the regression includes all unit indicators, the \(z\) column adds no new linearly independent column to the design matrix. If an intercept is included with a reference unit omitted, the same dependence holds: \(z\) can be written as a constant plus a combination of the included unit indicators. Either coding of unit fixed effects therefore leaves the main effect of \(z_i\) unidentified.

A matrix has full column rank only when none of its columns is an exact linear combination of the others. Here that condition fails. As a result, the least-squares coefficients on unit effects and on \(z_i\) are not uniquely determined. Software may omit \(z_i\), omit a redundant indicator, or report a singularity; that is a consequence of the specification, not a substantive estimate of zero.

Worked Example: A Linear Dependence in the Design Matrix

Consider three hypothetical units, with a binary characteristic \(z=(1,0,1)\), and suppose each is observed twice. The six rows are ordered by unit and then by time. The unit-indicator columns are

D_1=(1,1,0,0,0,0)',\quad D_2=(0,0,1,1,0,0)',\quad D_3=(0,0,0,0,1,1)'.

The repeated characteristic column is \(z=(1,1,0,0,1,1)'\). Direct addition gives

$$ z=1D_1+0D_2+1D_3=D_1+D_3. $$

Thus, in a regression containing \(D_1,D_2,D_3,z\), one column is redundant. For instance, adding any amount \(c\) to the coefficient on \(z\), subtracting \(c\) from the coefficient on \(D_1\), and subtracting \(c\) from the coefficient on \(D_3\) leaves every fitted value unchanged. The data can identify neither a unique coefficient on \(z\) nor unique unit-effect coefficients under this redundant parameterization.

When an Interaction Can Be Estimated

Suppose the model also contains a variable \(w_{it}\) that changes over time, and consider an interaction:

y_{it}=\alpha_i+\beta w_{it}+\gamma z_i+\delta(z_iw_{it})+u_{it}.

The main effect \(\gamma z_i\) is still absorbed. But the interaction \(z_iw_{it}\) may vary within a unit because \(w_{it}\) varies. After subtracting its unit mean, the interaction becomes \(z_i(w_{it}-\bar w_i)\), which need not be zero. Its coefficient \(\delta\) is estimable if this transformed interaction is not an exact linear combination of the other transformed regressors. The presence of an interaction does not automatically guarantee that condition; it must be checked in the particular design.

For a binary \(z_i\), the interaction permits the slope on \(w_{it}\) to differ between the two groups. The coefficient \(\beta\) is the slope for the \(z_i=0\) group, and \(\delta\) is the difference between the slopes for the \(z_i=1\) and \(z_i=0\) groups. This interpretation concerns a slope difference, not a separately estimated level difference between groups. If the interaction is with a common time variable, it can similarly represent different time patterns across groups, provided the pattern is not collinear with other included time-varying terms.

Worked Example: Estimating a Group-Specific Time Pattern

Consider four hypothetical units observed at \(t=0\) and \(t=1\). Let \(z_i=0\) for Units 1 and 2 and \(z_i=1\) for Units 3 and 4. Their outcomes are \(10,12\); \(12,14\); \(20,25\); and \(22,26\), respectively, with each pair listed in time order. Fit the fixed-effects specification

y_{it}=\alpha_i+\theta t+\delta(z_it)+u_{it}.

The main effect of \(z_i\) is omitted because it is absorbed by the unit effects. Demeaning within each unit gives \(\tilde t=t-\bar t_i=(-0.5,0.5)\) for every unit. For Units 1 and 2, the demeaned outcomes are \((-1,1)\) in each case. For Unit 3 they are \((-2.5,2.5)\), and for Unit 4 they are \((-2,2)\). The demeaned interaction \(\widetilde{z_it}=z_i(t-\bar t_i)\) is zero for Units 1 and 2 and equals \((-0.5,0.5)\) for Units 3 and 4.

Let \(q_{it}=\tilde t\) and \(r_{it}=\widetilde{z_it}\). The sums for the two-regressor least-squares calculation are

\sum q_{it}^2=2,\quad \sum r_{it}^2=1,\quad \sum q_{it}r_{it}=1,\quad \sum q_{it}\tilde y_{it}=6.5,\quad \sum r_{it}\tilde y_{it}=4.5.

For example, \(\sum q_{it}\tilde y_{it}=1+1+2.5+2=6.5\), where each unit contributes half its outcome change. The normal equations are therefore

\begin{pmatrix}2&1\\1&1\end{pmatrix} \begin{pmatrix}\hat\theta\\\hat\delta\end{pmatrix} = \begin{pmatrix}6.5\\4.5\end{pmatrix}.

The determinant is \(2(1)-1(1)=1\), so the transformed regressors are not perfectly collinear. Solving gives \(\hat\theta=2\) and \(\hat\delta=2.5\). In this hypothetical example, the estimated time pattern for the \(z_i=0\) group is 2 outcome-score units per time period, and the estimated difference in time patterns for the \(z_i=1\) group is 2.5 units per period. This calculation illustrates algebraic estimability; a causal interpretation would require additional assumptions about the disturbance and the design.

A Practical Identification Check

A reliable way to assess a proposed term is to apply the within transformation to it and inspect what remains. A main effect \(z_i\) becomes zero. An interaction \(z_iw_{it}\) becomes \(z_i(w_{it}-\bar w_i)\), which may be nonzero. Then check whether the transformed term is linearly independent of the other transformed regressors. If it is not, its coefficient cannot be uniquely estimated even though it varies within some units.

This also explains why “the characteristic varies across units” is not enough to estimate its main effect in a fixed-effects model. Fixed effects use comparisons within units, not the between-unit differences that distinguish one constant characteristic value from another. A variable can be substantively important and still lack an independently identified coefficient under this specification.

Key takeaway: Unit fixed effects absorb every regressor that is constant within units. A main effect of a time-invariant characteristic is therefore unidentified, not estimated as zero. An interaction with a time-varying variable can remain estimable when it has within-unit variation and is not collinear with the other included terms.

Common Mistakes

Interpreting an omitted main effect as no effect. Omission reflects exact collinearity with unit effects; it does not establish that the characteristic has no substantive relationship with the outcome.
Assuming every interaction is identified. A product is not automatically estimable. After removing unit means, it must retain variation not explained exactly by other included regressors.
Calling the interaction coefficient a group-level difference. In a group-by-time-varying-regressor interaction, the coefficient describes a difference in slopes, not a time-invariant outcome-level gap.
Believing different software output means different identification. Software may drop different redundant columns, but the underlying rank deficiency remains.

Check Your Understanding

Use the within transformation and the design-matrix argument to answer these questions.

In a unit fixed-effects model, what happens to a regressor \(z_i\) that is constant over time for every unit?
Why does including unit indicators make a separate main-effect coefficient on \(z_i\) unidentified?
If \(z_i\) is binary and \(w_{it}\) varies over time, what does the coefficient on \(z_iw_{it}\) describe in a model that also includes \(w_{it}\)?
Give one reason an interaction involving a time-invariant characteristic might still fail to be identified.
Does failure to estimate a separate fixed-effects coefficient on \(z_i\) show that \(z_i\) has no relationship with the outcome? Explain.