Textbooks Cover Both FE and RE, and the Hausman Test Exists—So Why Do Most Empirical Studies Just Use Fixed Effects Models Directly?

1. Introduction: The Textbook Gives You Two Paths—Yet Empirical Researchers Almost Always Walk Only One

Open the panel data chapter of any intermediate econometrics textbook, and you are sure to encounter this standard narrative arc:

1. First, Pooled OLS—stack all observations together and run OLS, pretending that individual heterogeneity does not exist.
2. Next, Random Effects (RE)—acknowledge individual heterogeneity $u_i$, but assume it is uncorrelated with X. Estimate via GLS, which is more efficient than OLS.
3. Then, Fixed Effects (FE)—also acknowledge individual heterogeneity $u_i$, but allow it to be arbitrarily correlated with X. Estimate via the within transformation (or LSDV).
4. Finally, the Hausman test—compare the coefficient differences between FE and RE. If the difference is significant → reject RE (the assumption that $u_i$ is uncorrelated with X is overturned) → use FE. If the difference is not significant → fail to reject RE → use RE (because RE is more efficient).

The logic is rigorous and builds step by step. Then you open an empirical panel paper published in the AER, JPE, or QJE—and the authors simply run:

xtreg y x controls i.year, fe robust

No RE, no Hausman test, no discussion whatsoever of "why FE rather than RE." They do not even bother to mention RE in a footnote.

Why do textbooks devote an entire chapter to RE and the Hausman test, while the empirical frontier almost exclusively uses FE—and without conducting the test? This is not because frontier researchers are "lazy" or "do not understand the theory"—it is because, after graduating from the textbook, they have repeatedly learned one thing from real data: the core assumption of RE ($\text{Cov}(X_{it}, u_i) = 0$) almost never holds in observational micro data.

Core message: The practice of empirical research routinely using fixed effects models directly without a Hausman test rests on a fundamental asymmetry—FE is a consistent estimator under weaker assumptions than RE. FE only requires $\mathbb{E}[\varepsilon_{it} \mid X_{i1}, \dots, X_{iT}, u_i] = 0$ (the strict exogeneity condition), whereas RE additionally requires $\text{Cov}(X_{it}, u_i) = 0$—that is, unobservable individual heterogeneity must be uncorrelated with any X you choose. In observational micro data (e.g., labor, education, health, corporate finance), individual heterogeneity is almost naturally correlated with the key explanatory variables—more able people obtain more education, better-managed firms are more profitable and invest more, and health-conscious individuals exercise more and eat better—which is precisely the primary reason empirical researchers need panel data in the first place (to control for time-invariant confounders). Given this premise, the core assumption of RE is theoretically untenable. And the Hausman test—while econometrically elegant—has three major limitations in practice: it may lack power or be overly sensitive with small T and large N; it is a binary decision rule that cannot substitute for theoretical judgment; and in many cases, even when Hausman fails to reject RE, using the more conservative FE is considered the more prudent and defensible choice. The efficiency loss of FE (relative to RE) is limited in the typical structure of micro panels (large N, small T)—while its robustness gain is enormous. In short, the empirical frontier does not use RE because RE is wrong—but because on the real battlefield of observational data, FE almost never lets you down, whereas if RE's assumption fails, your entire paper rests on an inconsistent estimator. On an uncertain battlefield, choose the sturdier shield—this is the fundamental logic by which empirical researchers move from the textbook's symmetric narrative to asymmetric practice.

2. The Essential Difference Between FE and RE—Not Just "One Method" Versus "Another Method"

2.1 The Two Models Written Together

Let us return to the most basic setup. The standard form of the panel data model:

$Y_{it} = \beta_0 + \beta_1 X_{it} + u_i + \varepsilon_{it}$

where $u_i$ is the time-invariant heterogeneity of individual i (ability, culture, geographic location, institutional characteristics…), and $\varepsilon_{it}$ is the time-varying error.

This equation is the starting point of all divergence. FE and RE are not choices between "different models"—they are choices between different assumptions about the same $u_i$.

2.2 Random Effects (RE): Assuming $u_i$ and X Are "Strangers"

The core assumption of RE is:

$\text{Cov}(X_{it}, u_i) = 0 \quad \text{for all } t$

In plain language: in every period, each X you choose (education, income, investment…) has no linear correlation whatsoever with the unobservable yet ever-present individual heterogeneity $u_i$ (ability, motivation, innate endowment). $u_i$ is a purely "random disturbance"—it induces correlation across observations of the same individual, but it is a "stranger" to your explanatory variables of interest.

Under this assumption, you can use generalized least squares (GLS)—applying a "quasi-demeaning" transformation to the data (subtracting a fraction of the individual mean rather than the whole, with the fraction depending on the relative magnitudes of $\sigma_u^2$ and $\sigma_\varepsilon^2$), then running OLS. GLS uses both between variation and within variation to estimate β—it is more "data-efficient" than FE, which uses only within variation. Therefore, when the RE assumption holds, RE is more efficient than FE (smaller standard errors).

When might the RE assumption plausibly hold?

• Experimental data (randomly assigned treatment—treatment status is uncorrelated with subjects' individual characteristics).
• Certain hierarchical models where $u_i$ is interpreted as "random sampling error."
• When you genuinely believe that all confounders have been explicitly included as control variables (for further discussion, see Section 5.3 below).

2.3 Fixed Effects (FE): Allowing $u_i$ and X to Be "Old Friends"—Correlated However You Like

FE makes no assumption whatsoever about the relationship between $u_i$ and X. $\text{Cov}(X_{it}, u_i)$ can take any value—positive, negative, large, small—FE handles them all.

FE eliminates $u_i$ through the within transformation:

$\tilde{Y}_{it} = Y_{it} - \bar{Y}_i = \beta_1 (X_{it} - \bar{X}_i) + (\varepsilon_{it} - \bar{\varepsilon}_i)$

$u_i$ is a constant—so $\bar{u}_i = u_i$, and $u_i - \bar{u}_i = 0$. $u_i$ is completely eliminated. No matter how strongly it correlates with X—it no longer exists. You do not need to know what it is, you do not need to measure it, and you do not need to assume anything about its relationship with anything else.

The cost of FE is: between variation (differences in X across individuals) is discarded. If the variation in X is mainly between individuals (e.g., years of education barely changes after entering the workforce), FE may make it difficult to estimate β precisely. But the cost buys you robustness—your estimator is consistent and unbiased under weaker conditions than RE.

2.4 A One-Sentence Summary

3. Why Does the RE Assumption Almost Never Hold in Observational Data?

3.1 Individual Heterogeneity Is Not "Random Noise"—It Is the Result of Choices

In observational data, individuals are not randomly assigned to different values of X. People choose their own education levels, firms choose their own investment scales, and cities choose their own governance models. These choices are endogenous—they are systematically correlated with individuals' unobservable characteristics ($u_i$).

Three classic examples:

• Returns to education: You study "the effect of years of education on income." $u_i$ includes "innate ability." More able people—do better in school, are more likely to continue studying (more years of education), and earn more even with the same years of education. $\text{Cov}(\text{Education}_{it}, u_i) > 0$—the RE assumption is shattered.

• Corporate investment: You study "the effect of the debt ratio on corporate investment." $u_i$ includes "management quality." Better-managed firms—may be better at controlling their debt ratios (choosing lower leverage) while also having higher investment levels (because they have better projects). $\text{Cov}(\text{Debt}_{it}, u_i) \neq 0$—the RE assumption is shattered once again.

• Health behavior: You study "the effect of exercise on BMI." $u_i$ includes "health awareness." Health-conscious people—exercise more and are also more disciplined in other dietary and lifestyle habits (lower BMI). $\text{Cov}(\text{Exercise}_{it}, u_i) \neq 0$—the RE assumption fails yet again.

This is not the exception—it is the norm. In observational data across the social sciences, the assumption that $\text{Cov}(X_{it}, u_i) = 0$ is not "occasionally violated"—it is "almost never true." And this is precisely the primary reason you need panel data in the first place: if you could estimate β unbiasedly with cross-sectional data (i.e., if all confounders could be measured and controlled), you would not need panel data. The core value of panel data is to allow you to control for those unobservable, time-invariant confounders ($u_i$). But if $u_i$ were observable—you would have put it in the regression long ago. $u_i$ is $u_i$—rather than a control variable called $X_k$—precisely because you cannot measure it. And a factor you cannot measure that affects Y—in virtually all observational studies, is almost always correlated with X.

3.2 The RE Assumption Is Mathematically "Zero Covariance"—Economically "People Do Not Make Choices Based on Their Own Characteristics"

The RE assumption that $\text{Cov}(X_{it}, u_i) = 0$ is equivalent to assuming: when people make choices that affect X, they completely ignore (or are completely unaffected by) their own time-invariant characteristics. If $u_i$ is "innate ability," this assumption says "when people decide whether to attend college, ability plays no role in their consideration"—which is obviously false.

This is RE's fundamental dilemma: $u_i$ needs to be in the model precisely because it affects both Y and X. If it did not affect X—then you could simply leave it in the error term without omitted variable bias, and you would not need panel data structure at all. The RE assumption is precisely "the reason panel data is needed does not exist"—a paradox.

4. The Hausman Test—Theoretically Elegant, with Three Major Limitations in Practice

4.1 The Principle of the Hausman Test

The core idea of the Hausman (1978) test is remarkably subtle:

• FE is consistent regardless of whether $\text{Cov}(X_{it}, u_i)$ is zero.
• RE is consistent and more efficient when $\text{Cov}(X_{it}, u_i) = 0$, but inconsistent when $\text{Cov}(X_{it}, u_i) \neq 0$.

Therefore—if the FE and RE estimates are statistically significantly different → the consistency condition for RE does not hold → use FE. If the difference is not significant → fail to reject RE's consistency condition → use RE (because it is more efficient).

The test statistic:

$H = (\hat{\beta}_{FE} - \hat{\beta}_{RE})' [\text{Var}(\hat{\beta}_{FE}) - \text{Var}(\hat{\beta}_{RE})]^{-1} (\hat{\beta}_{FE} - \hat{\beta}_{RE}) \sim \chi^2(K)$

xtreg y x1 x2, fe
estimates store FE
xtreg y x1 x2, re
estimates store RE
hausman FE RE

p < 0.05 → reject RE (the FE-RE difference exceeds sampling error) → use FE.

4.2 Limitation One: The Hausman Test Is a Binary Decision—But It Cannot Substitute for Theoretical Judgment

A statistical test tells you "whether the difference is significant"—but it does not tell you "which choice is more reasonable in your research context." Consider two scenarios:

• Scenario A: You have a good instrumental variable or natural experiment, and $u_i$ is theoretically unlikely to be correlated with X. Hausman p = 0.06—marginally insignificant. Should you use RE? If theory strongly supports RE, perhaps—but most researchers would still choose FE, because they know the Hausman test lacks power with small T.

• Scenario B: You are using observational data, and X is clearly endogenous (e.g., education and ability). Hausman p = 0.12—"insignificant." But do you genuinely believe ability and education are uncorrelated? You do not need a test to tell you—you already know from economic theory and common sense that they are correlated. In this case, even if Hausman tells you "you should use RE," using RE would still be dishonest—because you already know RE's premise does not hold.

The Hausman test is a useful diagnostic tool—but it should not be the sole decision-maker you rely on. It answers "whether the data contain sufficient evidence to reject RE"—not "whether RE has theoretical credibility in your empirical context." When the answers to these two questions point in different directions—trust theory and common sense, not a test that may lack power.

4.3 Limitation Two: With Large N, the Hausman Test Almost Always Rejects RE

As N → ∞, both the FE and RE estimates converge—but to different limits (if $\text{Cov}(X_{it}, u_i) \neq 0$, RE converges to a biased value). In large-N panels, standard errors become very small—even tiny, economically insignificant FE-RE differences can be statistically significant.

The result: In large-N micro panels, the Hausman test almost always rejects RE—not because the FE and RE estimates differ enormously, but because N is so large that any subtle difference is detected as "significant." In this case, the Hausman test provides no additional information—it merely repeats your prior belief ("in observational data, $u_i$ is certainly correlated with X"), but cloaked in the garb of statistical significance.

This raises a question: if you already know Hausman will almost certainly reject RE—why report it at all? This is also one reason many empirical researchers no longer treat the Hausman test as a core diagnostic tool.

4.4 Limitation Three: With Small T, the Hausman Test May Lack Power

In the opposite direction: when T is small (T = 3–5), both FE and RE estimates have substantial sampling variation—larger standard errors. Even if RE is indeed inconsistent, the Hausman test may fail to detect the FE-RE difference—because the noise is too large. In this case, Hausman p > 0.05 may not be because the RE assumption holds—but because the test lacks sufficient information to make a judgment.

This is the most dangerous combination: small T + large N micro panels. Large N may push Hausman toward rejection, small T may push it toward non-rejection—the two pull against each other, and the final p-value is highly dependent on the specific structure of the data. In this situation, relying on Hausman's binary output—rather than a set of theoretical analyses—is not robust.

5. Why Can Empirical Research Use FE Directly Without Testing?—Five Legitimate Reasons

5.1 Reason One: FE Is Always the "Safe Vote"—Asymmetric Consistency

This is the most fundamental logic. Consider a decision table:

FE is consistent in every scenario (under strict exogeneity). RE is consistent in only one scenario—but that scenario almost never exists in observational data. Between two options, one guarantees you survive only in a fairy tale, the other guarantees you survive both in the fairy tale and on the battlefield—which should you choose on the battlefield?

The asymmetric cost:

• Choosing FE when RE would have been valid → you lose some efficiency—slightly larger standard errors, slightly wider confidence intervals, but the coefficient is correct in expectation. With small T, this efficiency loss is actually limited.
• Choosing RE when FE should have been used → you lose consistency—your coefficient is asymptotically wrong. Your entire paper rests on an inconsistent estimator.

In an academic culture that prizes "better safe than sorry"—FE is the dominant strategy.

5.2 Reason Two: In Micro Panels, T Is Small—FE's Efficiency Loss Is Not That Large

RE is more efficient than FE—that is a fact. But the magnitude of this "efficiency advantage" depends on the specific structure of the data. RE's efficiency gain over FE depends on $\sigma_u^2 / \sigma_\varepsilon^2$ and T. The conclusion is: when T is small, RE's efficiency advantage is limited.

Intuition: When T is small, each individual's "within information" is scarce, and within variation itself is imprecise. But RE's main advantage lies in also using between variation. Yet when T is small—especially when T is far smaller than N—the between variation contains a large amount of N information, and the portion discarded by FE can be partially compensated by the large cross-sectional sample size. Furthermore, when T is very small (e.g., T = 3), within variation itself is also small—meaning that even the FE estimator using "only within variation" will not be much worse than the RE estimator using all information, because in extremely short panels, the information content of the time dimension is inherently limited.

The cost of efficiency loss is empirically acceptable: standard errors go from 0.08 to 0.10—your conclusion may shift from "significant at the 0.1% level" to "significant at the 1% level," but rarely from "significant" to "insignificant." By contrast, the cost of inconsistency is catastrophic.

5.3 Reason Three: FE Has a Cleaner Causal Interpretation

FE uses only within variation—that is, "changes in the same individual over time." This identification source has a natural causal intuition: after fixing all time-invariant characteristics of the individual (observed and unobserved), how do changes in X affect changes in Y?

In contrast, RE mixes within variation and between variation (cross-sectional comparisons across individuals). And the variation in "why individual A's X is higher than individual B's X" is highly endogenous—it is far more likely to be contaminated by $u_i$ than time-dimension variation. In the framework of causal inference, "changes within an individual" are closer to a clean quasi-experiment than "differences between individuals." FE automatically focuses on the more credible portion of variation—this is not merely an econometric choice but also a research design stance.

This is also why difference-in-differences (DID), event studies, and panel IV are all built on the FE framework—not the RE framework. The causal inference turn in contemporary empirical research naturally favors the logic of FE.

5.4 Reason Four: FE + Clustered Standard Errors Is the Standard Configuration in Panel Empirics—Reviewers Do Not Require You to Explain It

If you are using observational micro panel data and include unit fixed effects and year fixed effects in your regression—reviewers will almost never ask "why did you not consider random effects?" This is not because reviewers forgot or are unaware of RE's existence—but because a consensus has formed in the field: FE is the benchmark specification for panel empirics, and RE is an exception that requires special justification.

What you need to explain in your paper is not "why FE"—but "why RE." If you choose RE, you need to argue in the main text: why is $\text{Cov}(X_{it}, u_i) = 0$ credible in your research context? This typically requires experimental data, a natural experiment, or a special design (such as hierarchical models where $u_i$ is modeled as group-level random effects).

This also explains why the Hausman test is increasingly rare in empirical papers: in most micro panels, the outcome of the Hausman test is entirely predictable (rejecting RE). A predictable outcome is not information—so reporting it becomes an empty ritual. Reviewers do not ask you about Hausman—they only ask "what is your identification strategy? Which confounders does your FE specification control for?"

5.5 Reason Five: In Real Research, the Identification Strategy Determines the Econometric Method—Not the Test Statistic

In an empirical paper, the choice of econometric method does not begin with "should I run FE or RE?"—it begins with "what causal effect do I want to estimate? What identification assumptions are needed? Which confounders can I control, and which can I not?" The FE model in panel data is already chosen at the research design stage—because the researcher needs it to control for time-invariant individual heterogeneity.

The core question of the Hausman test—"should I use RE or FE?"—presupposes that the researcher has no strong prior judgment on modeling strategy. But good empirical researchers are precisely the opposite—they should know before even seeing the data: in my research context, is $u_i$ independent of X? If the answer is no (and it almost always is)—FE is the only choice. The Hausman test merely confirms ex post a fact you already knew.

6. When Should You Actually Consider RE?—Not All Panels Must Be FE

Having said so much in favor of FE—RE is not a useless antique. In the following three scenarios, RE deserves serious consideration.

6.1 Scenario One: You Have a Key Time-Invariant Variable—And It Is Your Object of Study

This is FE's biggest pain point: FE cannot estimate the coefficient of a time-invariant variable. If you study "the effect of gender on wages"—gender does not vary over time in a panel, and FE's within transformation will eliminate it (gender − individual mean(gender) = 0). If you want to estimate coefficients for variables such as gender, race, birth year, or distance from birthplace—you can only use RE or pooled OLS (and accept their assumptions).

Workaround: The Hausman-Taylor estimator or the Mundlak-Chamberlain device (see Section 6.3 below)—these allow you to estimate coefficients on time-invariant variables while relaxing some of RE's assumptions.

6.2 Scenario Two: Experimental or Quasi-Experimental Data—Treatment Is Indeed Randomly Assigned

If your data come from a randomized controlled trial (RCT) or a carefully designed natural experiment—where treatment status is indeed randomly assigned and uncorrelated with any individual characteristics (including $u_i$)—then RE is not only feasible but more efficient. RE uses more information, produces narrower confidence intervals, and makes your experimental conclusions more powerful.

But even in this case, reviewers may still expect you to show FE results as a benchmark—because FE is the more conservative choice, and in a randomized experiment, FE and RE should give very similar coefficients (which is corroborating evidence that randomization succeeded).

6.3 Scenario Three: You Flexibly Model the Relationship Between $u_i$ and X—The Mundlak-Chamberlain Device

There is a "third path"—neither pure FE (discarding all between variation) nor pure RE (assuming $u_i$ is completely uncorrelated with X). Mundlak (1978) and Chamberlain (1984) proposed an intermediate strategy: explicitly adding individual means as additional control variables in the RE model.

$Y_{it} = \beta_0 + \beta_1 X_{it} + \gamma \bar{X}_i + u_i + \varepsilon_{it}$

where $\bar{X}_i$ is the time average of X within each individual. By adding $\bar{X}_i$, you explicitly model the correlation between $u_i$ and X (because $\bar{X}_i$ absorbs the part of $u_i$ that is linearly related to X).

If $\gamma = 0$ (i.e., the coefficient on $\bar{X}_i$ is insignificant) → no evidence that $u_i$ is correlated with X → RE is fine. If $\gamma \neq 0$ → $\bar{X}_i$ has already controlled for the correlation → the corrected RE is equivalent to FE (for coefficients on time-varying variables).

* Mundlak approach
bysort id: egen x1_mean = mean(x1)
bysort id: egen x2_mean = mean(x2)
xtreg y x1 x2 x1_mean x2_mean, re
* Test whether x1_mean x2_mean are jointly significant
test x1_mean x2_mean

The Mundlak-Chamberlain device is far less common in practice than directly using FE—but it offers a more refined way of thinking: rather than choosing between FE and RE, you can obtain a more flexible estimator by explicitly modeling the relationship between $u_i$ and X within RE. Many contemporary advanced panel models (such as CRE—Correlated Random Effects) are built on this idea.

7. Common Misconceptions

7.1 Misconception One: "Hausman Test Insignificant → Should Use RE"

An insignificant Hausman test means the data do not contain sufficient evidence to distinguish between FE and RE—which could be because (a) the RE assumption indeed holds, or (b) T is too small, both FE and RE standard errors are too large, and the test lacks power. In observational data, the probability of (b) is far higher than (a). Using RE when Hausman is insignificant—you may not be making a legitimate efficiency gain, but rather an incorrect specification choice that you lack the power to detect.

7.2 Misconception Two: "RE Is More Efficient When Its Assumption Holds → I Should Test First and Then Decide"

This logic is theoretically correct. But the problem is: in observational data, "the RE assumption holds" is such a fragile premise—it requires that your X is uncorrelated with all unobserved, time-invariant individual characteristics. How confident are you in making this claim? If you are only 50% confident—then you have a 50% probability that the entire study rests on an inconsistent estimator. Gambling consistency risk for efficiency gains—this is not a good bet.

7.3 Misconception Three: "FE Controls for All Omitted Variables—So with FE, You Do Not Need Control Variables"

FE only controls for time-invariant omitted variables. Any time-varying omitted variable—if it is correlated with both X and Y—remains in the error term and will still bias your β. FE is not the end point of causal inference—it is merely a tool that controls for one specific dimension (time-invariant confounders). You still need to carefully choose time-varying control variables and, where possible, seek sources of exogenous variation (instrumental variables, DID, regression discontinuity).

7.4 Misconception Four: "My Panel Has Large T—So RE's Efficiency Advantage Is Large, and I Should Use RE"

When T is large, RE's efficiency advantage is indeed more pronounced. But large T also means you have more data to estimate each individual's $u_i$—and these estimates of $u_i$ can be used to "see" how large $\text{Cov}(X, u_i)$ is. Moreover, when T is large, the time dimension is long enough that nonstationarity issues may begin to emerge—FE's within transformation is more robust to nonstationarity than RE's quasi-differencing. In long panels, FE is not less reasonable—it may be more reasonable, but you simultaneously need to worry about time series issues (panel unit roots, panel cointegration—see the "Panel Data—Stationarity" article in this book).

8. Summary

Empirical research directly using fixed effects models without a Hausman test—five core reasons:

1. Asymmetric consistency. FE is consistent both when the RE assumption holds and when it fails—RE is consistent in only one case. And that "one case" ($u_i$ uncorrelated with X) almost never holds in observational micro data—because individual heterogeneity is precisely the reason X is chosen.

2. The cost of efficiency loss is limited—the cost of inconsistency is catastrophic. In typical micro panels with large N and small T, FE's efficiency loss relative to RE is modest—standard errors go from 0.08 to 0.10. The cost of inconsistency, by contrast, is that the entire study rests on a wrong estimator. Trading efficiency for safety—this is a judgment you can make without Hausman.

3. FE has a cleaner causal interpretation. FE identifies β using only "changes within the same individual over time"—in the language of causal inference, this is a more credible source of variation than "differences between individuals." The causal inference turn in contemporary empirical research (DID, event studies, panel IV) naturally favors FE.

4. FE is the industry benchmark for panel empirics. Reviewers do not ask "why FE"—they ask "why RE." FE is the default; RE is an exception requiring special justification. In most empirical papers, the Hausman test is merely a ritual with a known answer—which is why it is increasingly rarely reported.

5. Identification strategy comes first; econometric method follows. Good empirical research does not begin with "which model should I run"—it begins with "how do I identify the causal effect of X on Y? What confounders do I need to control?" If the answer is "I need to control for time-invariant individual heterogeneity"—then FE is the natural implementation of that strategy. No test is needed—because the premise of testing is that you have no theoretical prior between the two options. But you do have one for your research context.

One-sentence conclusion:

"The textbook gives you two panel models—symmetrically, fairly, each with its own pros and cons. But what the textbook does not tell you is that in the real world of observational data, the assumptions of these two models are not 'each with its own applicable scenario'—rather, one assumption (FE: arbitrary correlation allowed) almost always holds, while the other (RE: zero correlation required) almost never holds. This asymmetry plays out repeatedly in panel empirics across all social sciences—people choose their years of education, firms choose their investment scales, cities choose their policies—each choice is living evidence of the correlation between $u_i$ and X. Given this premise, FE is not 'the more prudent choice'—it is the only consistent choice. And the Hausman test, in 99% of micro panels, merely uses a p-value to confirm what you already know: $\text{Cov}(X_{it}, u_i) \neq 0$. So empirical researchers do not report Hausman—not because they do not understand it, but because it is not telling them anything they do not already know. FE is the shield, RE is the spear—on the battlefield of observational data, you do not know where the enemy (the correlation between $u_i$ and X) is or how large it is—you raise the shield first, and only then consider thrusting the spear. And most of the time, raising the shield well is already winning the battle."

9. Presentation Suggestions for Bilibili/WeChat Official Account

• Bilibili video: Use "shield vs. spear" as the core visual metaphor. Opening split screen: left side reads "FE = Shield," a sturdy large shield—no matter how strong the correlation between $u_i$ and X, the shield can block it ($u_i$ is eliminated in the within transformation). Right side reads "RE = Spear," a sharper spear (smaller standard errors, more efficient estimation), but the spear can only hit in one scenario—when $u_i$ and X are uncorrelated. Voiceover: "The textbook teaches you two weapons. In exams, you use the Hausman test to decide which to use. But on the real battlefield of observational data—you almost never know whether $u_i$ and X are correlated. Under this uncertainty, do you raise the shield first or thrust the spear first?" Act One "Why RE almost never holds": three scenarios in parallel—education (high ability → more schooling + higher wages), corporate (good management → lower leverage + higher investment), health (health awareness → more exercise + lower BMI). In all three scenarios, an invisible shadow ($u_i$) simultaneously connects X and Y—arrows pointing in the same direction. Annotation: "$u_i$ is not random noise—it is the reason people make choices. If $u_i$ did not affect X, you would not need panel data." Act Two "How FE eliminates $u_i$": animation—three time points for individual i ($Y_{i1}, Y_{i2}, Y_{i3}$), each subtracting the individual mean $\bar{Y}_i$, with $u_i$ disappearing as it is subtracted (red X mark). Annotation: "Within transformation = using each individual as their own benchmark. $u_i$ is a constant—completely eliminated." Act Three "The dilemma of the Hausman test": with large N (N = 10000), Hausman's p-value is always < 0.05—a row of p-values flashes across the screen, all in red. With small T (T = 3), Hausman's p-value may be > 0.05—but only because the noise is too large, not because the RE assumption holds. Annotation: "With large N, Hausman almost always rejects RE—you know it will reject. With small T, Hausman may fail to reject—but you should not trust it. Either way, you do not need a p-value to make the decision for you."

• WeChat Official Account: Make the FE vs. RE core assumption comparison table the central infographic. Present the asymmetric decision matrix (cost of choosing FE vs. cost of choosing RE) as a two-column comparison chart. Present Hausman's three limitations (always rejects with large N, lacks power with small T, cannot substitute for theoretical judgment) as three warning cards. Present the Mundlak-Chamberlain third path (adding $\bar{X}_i$ in RE) as a code card. Present the four common misconceptions as correction cards. Present the "identification strategy → econometric method" logic chain of causal inference as a flowchart.

• Recommended titles:

  • Main title: "Textbooks Cover RE and Hausman, but Empirics Only Use FE—Is That Justified?"
  • Alternative title: "Fixed Effects vs. Random Effects—Why the Choice in Real Research Is Simpler Than the Textbook Suggests"
  • New media title: "The Textbook Tells You to Run the Hausman Test—but 99% of Empirical Papers Directly Use FE. Whom Should You Trust?"

• Key quotes:

  "The choice between FE and RE is not 'each with its own applicable scenario'—rather, FE is consistent in every scenario, while RE is consistent in only one. And that scenario—where $u_i$ and X are completely uncorrelated—requires not a statistical test in observational data, but a fairy tale."

  "The Hausman test answers 'whether the data contain sufficient evidence to reject RE'—not 'whether RE has theoretical credibility in your research context.' In observational data, the answer to the second question has always been 'no'—so you do not need the answer to the first question."

  "The cost of choosing FE is efficiency—slightly larger standard errors. The cost of choosing RE is inconsistency—your coefficients are asymptotically wrong. In an academic paper, you can