Static Panel Models With Unit Effects

Why Panel Models Separate Units From Time

Panel data record outcomes and covariates repeatedly for the same units: for example, the same hospitals over several years or the same workers across several quarters. Repeated observations make it possible to distinguish changes within a unit from persistent differences between units. That distinction matters when a unit characteristic affects the outcome and is also related to the covariates being studied.

A static panel model describes the outcome at a given time using covariates at that time, a component specific to the unit, and a time-varying disturbance. “Static” means the model does not include a lagged outcome as a regressor. It does not mean that covariates or disturbances cannot change over time.

Definition: For units $i=1,\ldots,N$ and observed periods $t=1,\ldots,T_i$, the linear static panel model with unit effects is $$ y_{it}=x_{it}'\beta+\alpha_i+u_{it}. $$ Here $y_{it}$ is the outcome for unit $i$ in period $t$; $x_{it}$ is a vector of observed covariates; $\beta$ is a vector of slopes common to the units and periods; $\alpha_i$ is a time-invariant unit effect; and $u_{it}$ is a time-varying disturbance. A common intercept, if needed, can be absorbed into the unit effects.

The unit effect $\alpha_i$ represents the unit’s persistent contribution to the outcome that is not already represented by the included covariates. It may combine observed or unobserved features that remain constant over the periods studied, such as a hospital’s stable location or a worker’s persistent attributes. The effect is not necessarily directly measured. The disturbance $u_{it}$, in contrast, represents influences on the outcome that vary across unit-period observations and are not otherwise included in the equation.

The equation is a model for a conditional mean only when its disturbance has conditional mean zero. In this setting, that condition is usually stated as strict exogeneity. Let $X_i=(x_{i1},\ldots,x_{iT_i})$ denote the unit’s entire observed covariate path.

Definition: Strict exogeneity requires $$ E[u_{it}\mid X_i,\alpha_i]=0 \qquad\text{for every observed period }t. $$ Thus, conditional on the unit effect and the full covariate path, the disturbance in any period has mean zero. Under this condition, $$ E[y_{it}\mid X_i,\alpha_i]=x_{it}'\beta+\alpha_i. $$

The condition concerns all periods’ covariates, not only the covariate observed at the same time as the disturbance. It permits $\alpha_i$ to be related to the covariate path: strict exogeneity conditions the disturbance on $\alpha_i$, but does not require the unit effect to be unrelated to $X_i$. It is also a conditional-mean restriction, not a claim that disturbances are independent over time, normally distributed, or equally variable.

What a Within-Unit Slope Means

Under the specified linear conditional mean, $\beta$ describes how the outcome’s conditional mean varies with the covariates for the same unit. For a single covariate $x_{it}$, a one-unit increase corresponds to a change of $\beta$ units in the conditional mean, holding the unit effect and other covariates fixed. For a vector of covariate changes, the corresponding mean change is their inner product with $\beta$.

This is a within-unit interpretation: it concerns differences in covariates and outcomes over time for a given unit, rather than comparisons between units that happen to have different covariate values. In the model, a direct comparison between two periods for the same unit illustrates why the unit effect is separated:

y_{it}-y_{is}=(x_{it}-x_{is})'\beta+(u_{it}-u_{is}).

The equality follows by subtracting the model for period $s$ from the model for period $t$. The same $\alpha_i$ appears in both equations, so $\alpha_i-\alpha_i=0$. Under strict exogeneity, the conditional expectation of the disturbance difference given $X_i$ and $\alpha_i$ is also zero. Therefore,

E[y_{it}-y_{is}\mid X_i,\alpha_i]=(x_{it}-x_{is})'\beta.

This identity explains the model’s within-unit interpretation; it is not, by itself, a claim that every realized outcome change equals its model-based conditional mean. The realized change also includes $u_{it}-u_{is}$. Nor does the conditional-mean interpretation alone make $\beta$ a causal effect. A causal interpretation needs an appropriate causal model and assumptions, including that time-varying confounding has been adequately addressed.

Worked Example: Interpreting a Change for One Unit

Suppose a hypothetical workforce model relates quarterly productivity $y_{it}$, measured in output units per shift, to training hours $x_{it}$, with $\beta=1.8$ output units per shift per training hour. For one worker, let $\alpha_i=5$. The worker completes 4 training hours in the first quarter and 7 in the second. Suppose the realized disturbances are $u_{i1}=0.5$ and $u_{i2}=-0.3$.

The model gives the first-quarter outcome: $$ y_{i1}=1.8(4)+5+0.5=12.7 $$ output units per shift. In the second quarter: $$ y_{i2}=1.8(7)+5-0.3=17.3 $$ output units per shift. The realized change is $17.3-12.7=4.6$ output units per shift.

The change in the model’s conditional mean for this 3-hour increase in training is $$ 1.8(7-4)=1.8(3)=5.4 $$ output units per shift. The realized change is 0.8 units smaller because the disturbance changed by $-0.3-0.5=-0.8$. Thus, $\beta$ describes the conditional-mean contrast, not a promise that every observed contrast equals 5.4.

Within-Unit Relationships and Pooled Associations

A pooled association compares observations across all units and periods without, by itself, distinguishing whether a difference comes from a change within a unit or a persistent difference between units. If units with large $\alpha_i$ also tend to have large or small values of a covariate, those persistent differences can contribute strongly to the pooled association. Consequently, a pooled association need not equal the common within-unit slope $\beta$.

The distinction is about what is being compared. A within-unit contrast holds the identity of the unit fixed and uses changes over time. A pooled association uses variation across all observations, including between-unit variation. Neither is automatically a causal effect; moreover, under this model they need not even have the same sign.

Worked Example: A Negative Pooled Association and Positive Within-Unit Slopes

Consider hypothetical observations on two units and two periods. Let the model be $y_{it}=2x_{it}+\alpha_i+u_{it}$, with $u_{it}=0$ for all four observations, $\alpha_A=100$, and $\alpha_B=0$.

Unit	Period	$x_{it}$	$y_{it}$
A	1	1	102
A	2	2	104
B	1	10	20
B	2	11	22

Within unit A, $x$ rises by 1 and $y$ rises by $104-102=2$. Within unit B, $x$ also rises by 1 and $y$ rises by $22-20=2$. Both within-unit slopes are $2$, matching the model’s $\beta$.

Across all four observations, however, the pooled means are $\bar{x}=6$ and $\bar{y}=62$. The pooled linear association slope, calculated as the sum of cross-products divided by the sum of squared deviations in $x$, is $$ \frac{\sum (x_{it}-\bar{x})(y_{it}-\bar{y})}{\sum (x_{it}-\bar{x})^2} = \frac{(-5)(40)+(-4)(42)+(4)(-42)+(5)(-40)} {(-5)^2+(-4)^2+4^2+5^2} = \frac{-736}{82} \approx -8.976. $$

The pooled association is negative even though the within-unit slopes are positive. Unit A has both a much larger unit effect and lower $x$ values than unit B; its persistent outcome difference dominates the pooled comparison. The pooled calculation summarizes these four observations, but it does not recover the within-unit slope in this example.

Why Strict Exogeneity Is a Full-Path Condition

It is tempting to check only whether the disturbance in a period is conditionally mean zero given the covariate in that same period. Strict exogeneity is stronger: the disturbance must have conditional mean zero given covariates from every observed period. This rules out, for example, a current disturbance that helps determine a future covariate. Such feedback can make the full covariate path informative about an earlier disturbance.

Worked Example: A Future Covariate Reveals an Earlier Disturbance

Consider a hypothetical two-period panel with $\alpha_i=0$. Let $x_{i1}=0$, let $u_{i1}$ take values $-1$ and $1$ with equal probability, and set $x_{i2}=u_{i1}$. Let $u_{i2}$ be independent of $u_{i1}$ and have mean zero. Use the outcome equation $y_{it}=3x_{it}+u_{it}$.

In period 1, the contemporaneous covariate is constant, so $E[u_{i1}\mid x_{i1}]=E[u_{i1}]=0$. In period 2, independence and the zero mean of $u_{i2}$ give $E[u_{i2}\mid x_{i2}]=0$. These contemporaneous checks hold.

But the full covariate path includes $x_{i2}=u_{i1}$. Therefore, $$ E[u_{i1}\mid x_{i1},x_{i2},\alpha_i]=x_{i2}, $$ which equals either $-1$ or $1$, not zero. Strict exogeneity fails for period 1. The failure occurs because the earlier disturbance determines a future covariate, even though the contemporaneous checks are satisfied.

Model Conditions and Interpretation Limits

The model’s interpretation relies on more than writing down an equation. The following conditions clarify what the slopes represent and what the model does not guarantee.

Stable unit effects: $\alpha_i$ is constant over the observation window. A unit characteristic that changes over time cannot be represented by this time-invariant component alone.
Common linear slopes: the same $\beta$ applies to the units and periods in the model. If slopes differ by unit or period, one common coefficient may not describe the relevant contrasts.
Strict exogeneity: conditional on $\alpha_i$ and the entire covariate path, each disturbance has mean zero. Contemporaneous mean-zero conditions alone are insufficient.
Relevant within-unit variation: a within-unit slope requires covariate variation over time. To distinguish multiple slopes, the within-unit changes must also contain enough independent variation to separate their contributions.

Strict exogeneity does not rule out every source of misspecification. For example, if an omitted time-varying factor affects both a covariate and the outcome, it may be part of $u_{it}$ and violate the condition. It also does not establish that a change in a covariate was randomly assigned. A causal reading requires additional assumptions appropriate to the application.

Common Mistakes

Calling every panel slope a between-unit relationship. The model’s $\beta$ is interpreted through covariate changes within a unit under its assumptions; a pooled association also reflects comparisons between units.
Checking only contemporaneous exogeneity. The wrong reasoning is that $E[u_{it}\mid x_{it}]=0$ is enough. Strict exogeneity conditions on the full covariate path, so feedback between an earlier disturbance and a later covariate can violate it.
Equating a realized change with its conditional mean. A unit’s observed outcome change includes the change in its disturbances. The slope describes the model-based conditional-mean contrast.
Interpreting the unit effect as necessarily observed or unrelated to covariates. It can represent unobserved persistent features and can be associated with the covariate path.
Claiming causation from the equation alone. A within-unit interpretation is not automatically a causal one; time-varying confounding and other causal assumptions still matter.

Key takeaway: A static panel model separates a common covariate slope, a persistent unit effect, and a time-varying disturbance. Under strict exogeneity, the slope describes a within-unit conditional-mean relationship. A pooled association mixes within-unit changes with between-unit differences and can tell a different story.

Check Your Understanding

Use the model and assumptions above to answer these questions.

In $y_{it}=x_{it}'\beta+\alpha_i+u_{it}$, which component represents a persistent unit-specific contribution, and which represents time-varying unexplained influences?
Why does strict exogeneity condition on the entire covariate path rather than only $x_{it}$?
If one covariate increases by 2 units within a unit and its slope is 1.5, what is the model-based conditional-mean change, holding other covariates fixed?
How can a pooled association have the opposite sign from the within-unit slopes?
Does strict exogeneity by itself guarantee a causal interpretation of $\beta$? Explain briefly.