Chapter 2.4 — When the Errors Misbehave: Heteroskedasticity, Clustering, HC/HAC

Maya has a clean-looking result and a problem she cannot name yet.

She is studying fair lending. She has pulled a dataset of N = 9{,}000 consumer loans — the kind of mid-prime installment loans a lot of her own friends’ families actually use — and she wants to know whether borrowers in a particular ZIP-code tier get charged higher interest rates after controlling for credit score, loan size, and income. She runs the regression her laptop makes easy. Out comes a coefficient on her tier variable of \hat\beta_1 = 0.42 percentage points, with a standard error of 0.07. The t-statistic is 0.42 / 0.07 = 6.0. Six. That is not “marginally significant”; that is a coefficient screaming through a megaphone. She writes it up.

Then her mentor asks one question: “Are those 9,000 loans really 9,000 independent pieces of evidence?”

They are not. The loans cluster by lender, and lenders set rates in correlated ways. They cluster by state, and states have correlated regulation and correlated local economies. Once Maya accounts for that, the same coefficient \hat\beta_1 = 0.42 comes back with a standard error of 0.19, and the t-statistic collapses from 6.0 to 2.2. Still significant — but the megaphone is now a normal speaking voice, and a referee could reasonably push back.

Here is the unsettling part, and it is the whole chapter in one line: her point estimate never moved. The 0.42 was right the whole time. What was wrong was the 0.07 — the number that tells you how much to trust the 0.42. In Ch 2.2 we proved that OLS is unbiased and, under the Gauss–Markov assumptions, the best linear unbiased estimator. None of that breaks here. What breaks is the standard error, and therefore every t-statistic, every p-value, every confidence interval you build on top of it.

The one-sentence result of this chapter: When the errors are heteroskedastic or correlated, \hat{\boldsymbol\beta} is still fine — but the classical standard-error formula lies, and you must replace it with a robust, clustered, or HAC formula that tells the truth.

We are going to take this apart carefully, because “your point estimate is fine, your t-stat lies” is one of the most important and most misunderstood facts in all of applied econometrics. We will start exactly where Ch 2.2 left us — with the classical variance formula — watch it break, and build the replacement.

1. Where we start: the classical variance and the assumption holding it up

Recall the setup from Ch 2.1. We write the model in matrix form,

\mathbf{y} = \mathbf{X}\boldsymbol\beta + \boldsymbol\varepsilon ,

where \mathbf{y} is the N\times 1 vector of outcomes (Maya’s interest rates), \mathbf{X} is the N\times K matrix of regressors (a column of ones for the intercept, plus tier, credit score, loan size, income), \boldsymbol\beta is the K\times 1 vector of coefficients we want, and \boldsymbol\varepsilon is the N\times 1 vector of errors — everything about each loan’s rate that our regressors do not explain. The OLS estimator is

\hat{\boldsymbol\beta} = (\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\mathbf{y} .

Now substitute \mathbf{y} = \mathbf{X}\boldsymbol\beta + \boldsymbol\varepsilon into that and simplify (the \mathbf{X}\boldsymbol\beta part collapses to \boldsymbol\beta, which is the algebra behind unbiasedness):

\hat{\boldsymbol\beta} = \boldsymbol\beta + (\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\boldsymbol\varepsilon .

This little equation is the engine room of the whole chapter. Read it slowly. The estimator equals the truth, \boldsymbol\beta, plus a sampling-error term (\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\boldsymbol\varepsilon that depends on which particular errors \boldsymbol\varepsilon the universe happened to deal this sample — exactly the “what if the draw had been different” question from Ch 1.3, now in matrix clothing. The whole variance of \hat{\boldsymbol\beta} comes from that second term.

Take its variance, treating \mathbf{X} as fixed (conditioning on \mathbf{X}, as in Ch 2.2):

\operatorname{Var}(\hat{\boldsymbol\beta}\mid \mathbf{X}) = (\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\, \underbrace{\operatorname{Var}(\boldsymbol\varepsilon \mid \mathbf{X})}_{=\;\boldsymbol\Omega}\, \mathbf{X}(\mathbf{X}'\mathbf{X})^{-1} .

I have given the middle object a name: \boldsymbol\Omega = \operatorname{Var}(\boldsymbol\varepsilon\mid\mathbf{X}), the N\times N variance–covariance matrix of the errors. Its diagonal entries are the variances of each error, \operatorname{Var}(\varepsilon_i); its off-diagonal entries are the covariances between errors of different observations, \operatorname{Cov}(\varepsilon_i,\varepsilon_j). Everything in this chapter is a story about what \boldsymbol\Omega really looks like.

In Ch 2.2 we made the homoskedasticity-and-no-correlation assumption:

\operatorname{Var}(\boldsymbol\varepsilon \mid \mathbf{X}) = \boldsymbol\Omega = \sigma^2 \mathbf{I}.

In plain language, that single equation makes two claims at once. First, homoskedasticity (“same scatter”): every error has the same variance \sigma^2, regardless of the regressors — a $2,000 loan and a $40,000 loan have rates that are equally hard to predict, in squared dollars. That puts the same \sigma^2 all down the diagonal. Second, no correlation across observations: the errors of any two different loans are uncorrelated, so every off-diagonal entry is zero. Together they make \boldsymbol\Omega a scalar times the identity matrix — a matrix with \sigma^2 on the diagonal and zeros everywhere else.

Plug \boldsymbol\Omega = \sigma^2\mathbf{I} into the sandwich and watch it collapse. The middle \sigma^2 is a scalar and slides out front; the \mathbf{I} vanishes; one \mathbf{X}'\mathbf{X} cancels against an inverse:

\operatorname{Var}(\hat{\boldsymbol\beta}\mid\mathbf{X}) = \sigma^2 (\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\mathbf{X}(\mathbf{X}'\mathbf{X})^{-1} = \sigma^2 (\mathbf{X}'\mathbf{X})^{-1} .

That is the classical variance formula from Ch 2.2. The standard error of any single coefficient \hat\beta_j is the square root of the j-th diagonal entry of \hat\sigma^2(\mathbf{X}'\mathbf{X})^{-1}, where \hat\sigma^2 = \frac{1}{N-K}\sum_i \hat\varepsilon_i^2 is the usual estimate of the error variance built from the residuals \hat\varepsilon_i. It is clean, it is fast, and it is what your software prints by default.

It is also only as true as the assumption \boldsymbol\Omega = \sigma^2\mathbf{I}. The rest of this chapter is what happens when that assumption is false — which, in finance data, is basically always.

2. Reveal the trick: the sandwich never went away

Here is the key idea, and it is liberating once you see it. We did not need \boldsymbol\Omega = \sigma^2\mathbf{I} to write down the variance. The honest formula is the full sandwich:

\boxed{\;\operatorname{Var}(\hat{\boldsymbol\beta}\mid\mathbf{X}) = (\mathbf{X}'\mathbf{X})^{-1}\,\mathbf{X}'\boldsymbol\Omega\,\mathbf{X}\,(\mathbf{X}'\mathbf{X})^{-1}\;}

The name is literal. The two (\mathbf{X}'\mathbf{X})^{-1} factors are the slices of bread; the \mathbf{X}'\boldsymbol\Omega\mathbf{X} in the middle is the filling. The classical formula is just the sandwich for the special diet \boldsymbol\Omega = \sigma^2\mathbf{I}. If the real \boldsymbol\Omega has different variances on the diagonal (heteroskedasticity) or nonzero off-diagonals (correlation), the bread is unchanged but the filling is different — and so the final standard errors are different.

This reframes the entire problem. We are never going to “fix” \hat{\boldsymbol\beta}; there is nothing wrong with it. The whole game is estimating the middle of the sandwich, \mathbf{X}'\boldsymbol\Omega\mathbf{X}, without pretending \boldsymbol\Omega = \sigma^2\mathbf{I}.

That sounds impossible at first. \boldsymbol\Omega is an N\times N matrix — with N = 9{,}000 loans that is 81 million entries, and we have only 9,000 data points. We cannot estimate 81 million numbers from 9,000 observations. This is the genuine difficulty, and the next three sections are three different clever answers to it:

White / HC (Section 3): We don’t need \boldsymbol\Omega itself, only the sandwiched combination \mathbf{X}'\boldsymbol\Omega\mathbf{X}, which is a small K\times K matrix. If errors are uncorrelated (diagonal \boldsymbol\Omega), we can estimate that small thing directly using each observation’s own squared residual.
Clustered (Section 4): If errors are correlated within groups but not across them, \boldsymbol\Omega is block-diagonal, and we estimate the middle using whole blocks.
HAC / Newey–West (Section 5): If errors are correlated across time in a way that fades with distance, we estimate the middle using nearby observations with decaying weights.

Same sandwich every time. Only the recipe for the filling changes.

2.1 A two-observation sandwich you can do by hand

Abstract matrices hide what is going on, so let us collapse the sandwich to the smallest case that still shows the mechanism: a single regressor (no intercept), so \mathbf{X} is just a column of numbers and the whole sandwich becomes ordinary arithmetic. Suppose we have only the simplest model y_i = \beta x_i + \varepsilon_i. Then \mathbf{X}'\mathbf{X} = \sum_i x_i^2 is a single number, and the sandwich for a diagonal \boldsymbol\Omega reads

\operatorname{Var}(\hat\beta) = \frac{\sum_i x_i^2\,\sigma_i^2}{\left(\sum_i x_i^2\right)^2}.

Now compare two worlds with the same regressors x = (1, 2, 3, 4), so \sum_i x_i^2 = 30.

World A (homoskedastic): every \sigma_i^2 = 1. The numerator is \sum_i x_i^2 \cdot 1 = 30, so \operatorname{Var}(\hat\beta) = 30 / 30^2 = 1/30 \approx 0.033. This is exactly the classical answer \sigma^2/\sum x_i^2 = 1/30.

World B (heteroskedastic): the scatter grows with x, say \sigma_i^2 = x_i^2, so the variances are (1, 4, 9, 16). The numerator is now \sum_i x_i^2 \cdot x_i^2 = 1 + 16 + 81 + 256 = 354, so \operatorname{Var}(\hat\beta) = 354 / 900 \approx 0.393 — almost twelve times larger. The classical formula, which would still report 1/30 \approx 0.033 here (it averages the variances and forgets that the big-variance points are also the high-x, high-leverage points), would understate the true standard error by a factor of \sqrt{0.393/0.033} \approx 3.5. A t-stat of 3.5 would really be a t-stat of 1. That is the entire heteroskedasticity problem in four data points, and notice it required no calculus and no asymptotics — just plugging the right \sigma_i^2 into the middle of the sandwich instead of pretending they are all equal. The rest of Section 3 is only about how to estimate those \sigma_i^2 when, unlike here, nobody hands them to you.

3. Heteroskedasticity: White / HC standard errors

3.1 What heteroskedasticity is, in a picture

Priya, working her insurance-claims data, has the cleanest possible example. She regresses the size of a wildfire claim on the insured value of the property. A $200,000 cabin can produce a claim anywhere from $0 (false alarm) to $200,000 (total loss) — a spread of $200,000. A $5 million estate can produce a claim anywhere from $0 to $5 million — a spread twenty-five times wider. The scatter of the errors grows with the regressor. That is heteroskedasticity (“different scatter”): \operatorname{Var}(\varepsilon_i) depends on i, so the diagonal of \boldsymbol\Omega is not a single repeated \sigma^2 but a list of different \sigma_i^2:

\boldsymbol\Omega = \operatorname{diag}(\sigma_1^2, \sigma_2^2, \dots, \sigma_N^2).

The off-diagonals are still zero — Priya’s claims are independent across unrelated properties — so this is a purely heteroskedastic, still-uncorrelated case. The classical formula \sigma^2(\mathbf{X}'\mathbf{X})^{-1} uses one average \hat\sigma^2 everywhere, which means it systematically misprices the uncertainty: too small where the true scatter is large, too large where it is small. There is no general rule for which way the overall t-stat goes — but it is wrong, and in finance it is usually wrong in the direction of overconfidence.

3.2 White’s idea (1980)

Halbert White’s insight was to stop chasing the impossible N\times N object and look at what we actually need. Write the middle of the sandwich for a diagonal \boldsymbol\Omega. Because the off-diagonals are zero, the matrix product \mathbf{X}'\boldsymbol\Omega\mathbf{X} becomes a simple sum over observations:

\mathbf{X}'\boldsymbol\Omega\mathbf{X} = \sum_{i=1}^{N} \sigma_i^2\,\mathbf{x}_i \mathbf{x}_i' ,

where \mathbf{x}_i is the K\times 1 row of regressors for observation i (written as a column). This is a K\times K matrix — for Maya, 5\times 5 — no matter how big N is. We do not need each \sigma_i^2 on its own; we need this weighted sum. And White’s trick is that each unknown \sigma_i^2 can be replaced by the single thing we observe for that point: its squared residual \hat\varepsilon_i^2. Any one \hat\varepsilon_i^2 is a terrible estimate of \sigma_i^2 — it’s based on one data point. But summed up across all N observations, the errors in those guesses average out, and the sum converges to the right thing. That gives the heteroskedasticity-consistent (HC) estimator:

\widehat{\operatorname{Var}}(\hat{\boldsymbol\beta})_{\text{HC0}} = (\mathbf{X}'\mathbf{X})^{-1}\left(\sum_{i=1}^{N}\hat\varepsilon_i^2\,\mathbf{x}_i\mathbf{x}_i'\right)(\mathbf{X}'\mathbf{X})^{-1}.

These are White standard errors, also called heteroskedasticity-robust or just robust standard errors. The “0” in HC0 is a clue that there are corrected versions HC1–HC3, which we get to next. Crucially, robust SEs are valid whether or not there is heteroskedasticity: if the errors happen to be homoskedastic, the robust estimator converges to the classical answer as N grows. In fact the small-sample correspondence is sharpest for HC1, which carries the same N/(N-K) degrees-of-freedom factor as the classical \hat\sigma^2: under homoskedasticity HC1 reproduces the classical SE almost exactly, while raw HC0 (lacking that factor) runs a touch smaller. Either way, robust SEs cost you almost nothing and protect you from a real failure — which is why “always report robust” is close to a default in modern finance work.

3.3 HC0, HC1, HC2, HC3 — the finite-sample corrections

White’s HC0 is consistent, meaning it gets the right answer as N\to\infty. But in finite samples it has a known flaw: OLS residuals are systematically too small. The fitting procedure bends the line toward each point, so \hat\varepsilon_i understates the true \varepsilon_i, which makes HC0 standard errors too small — overconfident again, in small samples. The fix is to inflate the squared residuals. The four common flavors differ only in how they inflate:

Flavor	Each \hat\varepsilon_i^2 is replaced by	What it corrects	Use when
HC0	\hat\varepsilon_i^2	nothing (raw White)	large N, you don’t care about small-sample bias
HC1	\dfrac{N}{N-K}\,\hat\varepsilon_i^2	a flat degrees-of-freedom inflation	the default in Stata; fine for moderate N
HC2	\dfrac{\hat\varepsilon_i^2}{1-h_{ii}}	leverage, observation by observation	small samples; principled choice
HC3	\dfrac{\hat\varepsilon_i^2}{(1-h_{ii})^2}	leverage, more aggressively	small N or high-leverage points; safest

The h_{ii} here are the diagonal entries of the hat matrix \mathbf{H} = \mathbf{X}(\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}' from Ch 2.1 — the leverage of observation i, a number between 0 and 1 measuring how much that point pulls the fitted line toward itself. A high-leverage point (an outlier in \mathbf{X}, like Priya’s one $5 million estate among cabins) has its residual shrunk the most by fitting, so HC2 and HC3 inflate it the most to compensate. HC1 ignores leverage and applies the same N/(N-K) factor to everyone.

The practical guidance is simple and worth memorizing. With a few hundred observations or more and no extreme leverage, HC1 (Stata’s default) is fine and HC0–HC3 barely differ. When N is small — say under 50 — or a handful of points have high leverage, the flavors diverge and HC3 is the conservative, recommended default (it is the most likely to keep you honest). HC3 is never badly anti-conservative, so if you are unsure, HC3 is the safe pick.

3.4 A number to make it concrete

Priya simulates 60 claims where the true error standard deviation grows linearly with insured value (textbook heteroskedasticity), with one $5M estate among $200K cabins to create leverage. On her coefficient of interest the four standard errors come out roughly:

Classical	HC0	HC1	HC2	HC3
0.085	0.121	0.127	0.140	0.169

The classical SE of 0.085 understates the true uncertainty by about half relative to HC3’s 0.169. If her coefficient were 0.30, the classical t-stat would read 0.30/0.085 = 3.5 (comfortably significant) while the HC3 t-stat reads 0.30/0.169 = 1.8 (not significant at the 5% level). Same coefficient. Same data. The only thing that changed is whether we told the truth about \boldsymbol\Omega — and it flipped the conclusion. Notice also how the four robust flavors fan out: HC0 at 0.121 and HC3 at 0.169 differ by a real margin precisely because that one $5M estate is a high-leverage point, and HC3’s (1-h_{ii})^2 denominator inflates its residual the hardest. In a sample with no leverage and a few hundred points, those four numbers would have agreed to the third decimal.

3.5 When it fails, what you actually see

Reveal-the-trick demands we say what the symptom looks like, so you can catch it in your own work rather than only in a textbook. Three tells, in increasing order of formality:

First, the residual plot. Plot the residuals \hat\varepsilon_i against a fitted value or against a regressor. Under homoskedasticity you see a structureless horizontal band of constant thickness. Under heteroskedasticity you see a fan or funnel — the band widens (or narrows) as you move along the axis. Priya’s wildfire residuals fan open toward expensive properties; Maya’s loan residuals often fan with loan size. This single plot catches most real cases and costs you one line of code.

Second, a formal test. The Breusch–Pagan test regresses the squared residuals on the regressors and asks whether they explain any of the variance (if scatter is constant, they should not); the White test does the same with squares and cross-products to catch more general patterns. A small p-value says “reject homoskedasticity.” But here is the modern attitude, and it is worth internalizing: you usually should not bother testing. Because robust SEs are valid whether or not heteroskedasticity is present, the cost of “just always using robust” is near zero, while the cost of testing, failing to reject by bad luck, and then trusting classical SEs is a published overconfident result. Test if a referee asks; otherwise default to robust.

Third, the SEs themselves disagree. If classical and robust standard errors are nearly identical, your errors are roughly homoskedastic and nothing was at stake. If they differ materially — as in Priya’s table — heteroskedasticity is real and the robust number is the one to report. The disagreement is the diagnostic.

4. Clustering: when errors hold hands within groups

4.1 The failure that White SEs do not fix

Robust SEs assume the off-diagonals of \boldsymbol\Omega are zero — errors uncorrelated across observations. In finance that assumption is heroic and usually false. Maya’s loans within the same lender share the lender’s pricing model, risk appetite, and quirks; their errors move together. Priya’s wildfire claims within the same state share that state’s fire season, building codes, and reinsurance market; their errors move together. Devon’s daily token returns within the same week share whatever sentiment shock hit crypto that week. These are within-cluster correlations, and robust SEs are blind to them because they only ever look at one observation at a time.

The stakes here are not academic. Go back to Maya’s fair-lending result from the opening. Her claim — that one borrower tier pays measurably more after controlling for credit — is the kind of finding that can support or sink a discrimination complaint, inform a regulator, or appear in a policy brief. If she reports a t-stat of 6.0 when the honest, clustered t-stat is 2.2, she has not just made a technical slip; she has overstated the strength of evidence on a question where overstating it has real consequences in both directions (a false alarm wastes enforcement attention; underpowered honesty lets real discrimination slide). The reason the naive number is so much too big is not that her 9,000 loans are fake data — they are real — but that they are not 9,000 independent pieces of evidence. Loans from the same lender largely repeat the same story. Counting them as independent is like polling one household nine times and reporting a sample of nine. The clustered standard error is the correction that stops you from doing that.

Here is why ignoring clustering is so dangerous — and it is the single most expensive mistake in empirical finance, so it earns its own derivation. Suppose Maya has N loans split into G lenders (“clusters”) of equal size n = N/G, and within a lender the errors all share a common shock with within-cluster correlation \rho. There is a classic result (a special case of the Moulton factor) for how badly the naive standard error understates the truth. The variance of your estimate is inflated, relative to what the naive formula reports, by approximately

\text{true Var} \approx \text{naive Var} \times \big[\,1 + (n-1)\,\rho_x\,\rho_\varepsilon\,\big],

where n is the cluster size, \rho_\varepsilon is the within-cluster error correlation, and \rho_x is the within-cluster correlation of your regressor. Stare at the bracket. The damage scales with cluster size n: bigger clusters, worse understatement. With n = 200 loans per lender and even a mild \rho_x\rho_\varepsilon = 0.05, the bracket is 1 + 199\times0.05 \approx 11. Your true variance is eleven times the naive one — so your true standard error is \sqrt{11}\approx 3.3 times bigger, and your reported t-stat of 6.0 should really be about 6.0/3.3 \approx 1.8. That is exactly what happened to Maya in the opening, and now you can see it was not bad luck; it was arithmetic. Pouring in more observations within existing clusters does not buy you the independent evidence the naive formula thinks it does. What buys evidence is more clusters.

4.2 The cluster-robust sandwich

The fix is the same sandwich, with a middle built from whole clusters instead of single points. If errors are correlated within cluster g but uncorrelated across clusters, then \boldsymbol\Omega is block-diagonal — a big matrix with dense blocks down the diagonal (one block per cluster, capturing all the within-cluster covariances) and zeros everywhere else. The middle of the sandwich then becomes a sum over the G clusters:

\widehat{\operatorname{Var}}(\hat{\boldsymbol\beta})_{\text{cluster}} = (\mathbf{X}'\mathbf{X})^{-1}\left(\sum_{g=1}^{G}\mathbf{X}_g'\hat{\boldsymbol\varepsilon}_g\,\hat{\boldsymbol\varepsilon}_g'\mathbf{X}_g\right)(\mathbf{X}'\mathbf{X})^{-1},

where \mathbf{X}_g is the block of regressor rows for cluster g and \hat{\boldsymbol\varepsilon}_g is its vector of residuals. Compare this to White’s HC0 in §3.2. There the middle summed \hat\varepsilon_i^2\,\mathbf{x}_i\mathbf{x}_i' over single observations; here it sums \mathbf{X}_g'\hat{\boldsymbol\varepsilon}_g\hat{\boldsymbol\varepsilon}_g'\mathbf{X}_g over whole clusters, and the outer product \hat{\boldsymbol\varepsilon}_g\hat{\boldsymbol\varepsilon}_g' deliberately keeps the within-cluster covariances (the products \hat\varepsilon_{i}\hat\varepsilon_{j} for two loans i,j in the same lender) that White’s version threw away. White SEs are just cluster SEs where every cluster has size one. It is the same machine, zoomed out from points to groups.

The empirical-spec discipline from the Conventions now has teeth: when you write a spec you must name the clustering level, and that choice is an economic claim about where \boldsymbol\Omega’s off-diagonal blocks live. “Clustered by lender” asserts that two loans from the same lender have correlated errors and two loans from different lenders do not. If you believe the correlation actually lives at the state level (shared regulation, shared economy), you cluster by state, and the blocks get bigger.

4.3 The few-clusters caveat

There is a catch, and it bites hard in finance. The cluster-robust estimator is consistent as the number of clusters G \to \infty, not as N\to\infty. The asymptotics live in G, not in N. If Maya clusters by state, she has at most G = 50 clusters; if she clusters by some broad industry, maybe G = 12. With few clusters the cluster-robust SE is itself noisy and biased downward — overconfident again, the very disease we are treating.

The rough rule of thumb is that you want at least 30–50 clusters before you trust the cluster-robust normal-approximation t-test at face value. Below that, you reach for fixes: the small-sample t-distribution with G-1 degrees of freedom rather than the normal; or the wild cluster bootstrap (Cameron, Gelbach & Miller, 2008),¹ which resamples whole clusters and is the current standard remedy for the few-clusters problem.

5. HAC / Newey–West: correlation across time

5.1 The time-series version of the same disease

Sam trades. He regresses a strategy’s monthly return on a market-timing signal and gets a beautiful t-stat. But returns and residuals in financial time series are serially correlated (also called autocorrelated): this month’s error is correlated with last month’s, because slow-moving forces — a drifting risk premium, a multi-month liquidity squeeze, momentum itself — persist across periods. Now \boldsymbol\Omega has nonzero off-diagonals, but in a special pattern: the correlation is strongest between adjacent periods and fades as the time gap grows. \operatorname{Cov}(\varepsilon_t,\varepsilon_{t-1}) is large, \operatorname{Cov}(\varepsilon_t,\varepsilon_{t-12}) is near zero.

This is not clustering — there are no clean groups — and it is not pure heteroskedasticity. It is a third shape for \boldsymbol\Omega: a band of decaying correlations around the diagonal. Ignoring it, just like ignoring clustering, makes standard errors too small when the autocorrelation is positive (the usual case in finance), inflating t-stats.

5.2 Newey and West’s idea (1987)

Whitney Newey and Kenneth West built the heteroskedasticity-and-autocorrelation-consistent (HAC) estimator for exactly this. Same sandwich. The middle now includes not just each period’s own squared-residual term but cross-products of residuals from nearby periods, \hat\varepsilon_t\hat\varepsilon_{t-\ell}\,\mathbf{x}_t\mathbf{x}_{t-\ell}', for time gaps (“lags”) \ell = 1, 2, \dots, L. Because correlation fades with distance, Newey–West applies declining weights w_\ell = 1 - \tfrac{\ell}{L+1} (the Bartlett kernel): the lag-1 cross-products count almost fully, distant lags count for less, and beyond the chosen bandwidth L they count for nothing. The two declining-weight pieces are there both to honor the fading-correlation structure and, mathematically, to guarantee the resulting variance matrix is positive (a variance can’t be negative — without the weights this sum could misbehave).

The one knob you must set is the bandwidth L — how many lags to include. Too small and you miss real persistence; too large and you add noise. A common automatic choice is L \approx 4(T/100)^{2/9} (the Newey–West rule of thumb), where T is the number of time periods, and most software offers an automatic-bandwidth option. For Sam’s monthly data, setting L to a small multiple of 12 (a year or two of lags) is a sensible manual starting point. HAC SEs are the default for pure time-series regressions: predictive return regressions, factor-model time-series tests, anything indexed by t with one observation per period.

5.3 Why positive autocorrelation inflates t-stats, in one intuition

It is worth pinning down why serial correlation, like clustering, makes naive standard errors too small — because the mechanism is the same one, wearing a time-series costume. Recall the Moulton-style intuition from §4.1: when errors are positively correlated, each new observation carries less fresh information than the naive formula assumes, because part of it just echoes what you already saw. A time series with positive autocorrelation is exactly a chain of observations that echo their neighbors. If Sam has T = 120 monthly returns but each month’s residual is, say, 0.4-correlated with the previous month’s, his effective sample size — the number of genuinely independent monthly draws — is well below 120. The naive formula divides by the wrong (too large) effective T, so it reports a standard error that is too small and a t-stat that is too big. HAC repairs this by counting the lagged cross-products, which is arithmetically how it discovers that the 120 months are worth fewer than 120 independent observations.

The same logic explains the sign. Positive autocorrelation (the usual case in finance, where shocks persist) shrinks the effective sample and inflates t-stats — the dangerous direction. Negative autocorrelation (rarer, e.g. bid–ask bounce in high-frequency prices) would do the reverse, making naive t-stats too conservative. Either way, HAC gets it right because it estimates the actual band of off-diagonal covariances rather than assuming the band is empty.

6. Putting it together: the Petersen (2009) panel taxonomy

Now the hard case, and the one Maya and Priya actually live in: a panel, where you observe many firms (i) over many time periods (t) at once. A panel can have both diseases at once — errors correlated within a firm across time (a firm effect: a firm that is mispriced this year tends to be mispriced next year) and errors correlated across firms within the same period (a time effect: when the whole market drops, every firm’s residual drops together). Which standard error do you use?

This is the precise question Mitchell Petersen took up in a now-canonical 2009 paper, after noticing that finance papers were using a chaos of incompatible methods and getting standard errors that differed by factors of two or more on the same data. His contribution was not a new estimator so much as a clear taxonomy and a set of simulations showing which approach is right in which situation. The headline findings, which you should treat as practical defaults:

A persistent firm effect (the firm’s residuals correlated over time) is, in Petersen’s data, the dominant problem in corporate-finance panels, and it is the one most often ignored. The right tool is clustering by firm, which lets each firm’s errors be arbitrarily correlated across all its years — exactly the block-diagonal \boldsymbol\Omega of Section 4, with one block per firm spanning all its time periods. White SEs and Fama–MacBeth do not fix a firm effect and will leave you badly overconfident.
A time effect (a common shock hitting all firms in a period, like a market-wide crash) is better handled by clustering by time — one block per period, spanning all firms in it. Petersen notes that in many corporate panels a time effect is well absorbed by including time fixed effects (a dummy for each period, previewed by the demeaning idea in Ch 2.3); but if a residual common shock remains, clustering by time catches it.
Both effects present? Use two-way clustering (Cameron, Gelbach & Miller, 2011,² building on Petersen and Thompson), which clusters by firm and by time simultaneously, allowing both kinds of correlation in \boldsymbol\Omega at once. The mechanical recipe is elegant: compute the firm-clustered sandwich, add the time-clustered sandwich, and subtract the plain White (heteroskedasticity-only) sandwich so the overlap isn’t double-counted.

The decision rule Petersen leaves you with is the empirical-spec discipline made concrete: figure out where your residuals are correlated — across time within a firm, across firms within a period, or both — and cluster on that dimension (or those dimensions). It is an economic judgment about \boldsymbol\Omega, not a software default, and the same dataset can demand different clustering depending on what you put on the left-hand side. (Ch 5.4 is a full reader’s guide to this paper; here we want the operational rule.)

How do you tell which effect dominates without guessing? Petersen offers a practical diagnostic that you can run yourself. Cluster by firm and look at the standard error; then cluster by time and look again. Whichever clustering moves the standard error more away from the naive White number is pointing at the bigger source of correlation. If clustering by firm roughly doubles your SE but clustering by time barely changes it, you have a firm effect and little else — firm-clustering is enough. If both move it a lot, you have both effects and want two-way clustering. This is exactly the experiment nb2.4 walks you through, and it turns the “which cluster?” question from a matter of taste into something you can read off your own output.

A subtlety worth flagging now and developing in Weeks 3–4: fixed effects and clustering are not substitutes — they do different jobs. A firm fixed effect (a dummy for each firm, Ch 2.3’s demeaning) removes the firm’s average level from both sides; it handles a firm-specific mean that never changes. Clustering by firm handles correlation in the leftover residuals over time — the part a fixed effect cannot touch because it is about co-movement, not means. You often want both at once: firm fixed effects to absorb permanent firm differences and clustering by firm to honor the within-firm residual correlation that remains. Conflating the two (“I have firm fixed effects, so I don’t need to cluster by firm”) is a common and costly error, and it is the kind of thing the empirical-spec line in your write-up forces you to get explicit about.

A worked taste of the stakes, in the spirit of Petersen’s own tables. Maya assembles a panel of default outcomes for many firms over many years and regresses default on a leverage measure. The coefficient is the same in every column; only the SE flavor changes. The illustrative numbers below show the shape of what happens — the SE inflates as you allow more correlation into \boldsymbol\Omega, dragging the t-stat across the significance line. (For real, reproducible magnitudes built from a simulated panel, see nb2.4: on its 120-firm \times 20-year panel the identical \hat\beta=0.30 carries a classical t of 5.70 that collapses to a firm-clustered t of 2.19 — the same story, with numbers you can re-run. The smaller §7 toy panel just below shows the same collapse at its own scale.)

SE flavor	Std. error (illustrative)	t-stat	Honest?
Classical (OLS)	0.018	5.0	No — assumes \sigma^2\mathbf{I}
White (HC1)	0.021	4.3	No — ignores within-firm correlation
Clustered by firm	0.044	2.0	Yes, if the firm effect dominates
Two-way (firm + year)	0.052	1.7	Yes, if both effects are present

These figures are illustrative, not the output of a committed script; the live numbers are in nb2.4. The qualitative point is exact, though: a classical t-stat around 5 and a two-way-clustered t-stat around 1.7 are computed from identical point estimates on identical data. One says “publish”; the other says “not yet.” The difference is entirely a story about \boldsymbol\Omega, and choosing the wrong one is how real papers get retracted.

7. The code

Every flavor in this chapter is one argument in statsmodels. The point of seeing them side by side is to feel how much the t-stat moves while \hat\beta sits perfectly still.

import numpy as np
import pandas as pd
import statsmodels.formula.api as smf

# --- A toy clustered panel: firms over years, with a firm-level error component ---
rng = np.random.default_rng(8)
n_firms, n_years = 40, 15
firm = np.repeat(np.arange(n_firms), n_years)
year = np.tile(np.arange(n_years), n_firms)

# the regressor itself has a persistent firm-level piece (correlated within firm) --
# this is what makes ignoring clustering so costly (the Moulton logic of Section 4.1):
x_firm = np.repeat(rng.normal(size=n_firms), n_years)
x = x_firm + 0.5 * rng.normal(size=n_firms * n_years)      # the key regressor
firm_shock = np.repeat(rng.normal(size=n_firms), n_years)  # persistent firm effect
year_shock = np.tile(rng.normal(size=n_years), n_firms)    # common time shock
eps = 1.5 * firm_shock + year_shock + rng.normal(size=n_firms * n_years)
y = 0.30 * x + eps                                         # TRUE slope is 0.30

df = pd.DataFrame({"y": y, "x": x, "firm": firm, "year": year})
fit = smf.ols("y ~ x", data=df).fit()                  # one OLS fit, reused below
xi = list(fit.params.index).index("x")                 # column position of the x coef

def se(res, label):
    # robust results return plain numpy arrays, so index by position, not name
    b = np.asarray(res.params)[xi]
    s = np.asarray(res.bse)[xi]
    print(f"{label:<22} beta={b:.3f}  se={s:.3f}  t={b/s:.2f}")

se(fit,                                          "classical")
se(fit.get_robustcov_results("HC1"),             "robust HC1")
se(fit.get_robustcov_results("HC3"),             "robust HC3")
se(fit.get_robustcov_results("cluster", groups=df["firm"]),  "cluster by firm")
se(fit.get_robustcov_results("cluster",
       groups=df[["firm", "year"]]),             "two-way (firm+year)")
# HAC / Newey-West (for a pure time series; shown for completeness):
se(fit.get_robustcov_results("HAC", maxlags=4),  "Newey-West HAC(4)")

Run it and you will see the coefficient on x glued near 0.27 in every line (close to the true 0.30) while the standard error roughly doubles from classical (about 0.073) to firm-clustered (about 0.144), because we deliberately built a strong, persistent firm effect into both the regressor and the error. That doubling is the chapter in one screen of output: the estimate is fine; the t-stat was lying — it falls from about 3.7 to about 1.9, crossing the significance line without the point estimate budging. Change firm_shock to zero (and drop the firm piece of x) and the gap closes — robust and classical agree when the errors really are well behaved, which is the reassurance that you lose nothing by defaulting to the honest formula.

8. The mental checklist

When you sit down with a regression and ask “which standard error?”, walk the decision in this order — it is just “what does \boldsymbol\Omega look like?” in plain questions:

Could the error variance differ across observations? In finance, essentially always yes. So never use classical SEs as your final answer — use at least robust (HC1, or HC3 if N is small or leverage is high).
Are observations grouped, with correlated errors inside a group (firms, lenders, states, schools)? Then cluster on that group — and check you have enough clusters (aim for \gtrsim 30–50; otherwise use the wild cluster bootstrap).
Is it a pure time series with serial correlation? Use HAC / Newey–West, and set the bandwidth to cover the persistence you expect.
Is it a panel (many units × many periods)? Diagnose the correlation structure à la Petersen (2009): firm effect → cluster by firm; time effect → time FEs or cluster by time; both → two-way cluster.

The recurring lesson, and the reason this chapter sits between the Gauss–Markov guarantees of Ch 2.2 and the bias problems of Ch 2.5: getting \hat\beta right and getting its uncertainty right are two separate jobs. Ch 2.5 is about when \hat\beta itself is biased — a different and in some ways more dangerous failure. This chapter was entirely about the second job. A biased estimator points at the wrong answer; a wrong standard error makes you wrongly confident about whatever answer you have. Both will get a finding rejected. Only one of them will get it rejected after it’s been published.

Your Turn

Open nb2.4 — SE flavors on a clustered panel. You will build a panel with a tunable firm effect and a tunable time effect, then watch the five standard-error flavors (classical, HC1, HC3, firm-clustered, two-way) move as you turn the knobs — including the unsettling experiment of shrinking the number of clusters until the cluster-robust SE stops being trustworthy. The goal is to see, with your own simulated data, the gap between the t-stat that lies and the t-stat that tells the truth.

Check questions.

Maya runs a regression and gets \hat\beta_1 = 0.50 with a classical SE of 0.10 (t = 5.0). After clustering by lender the SE rises to 0.25 (t = 2.0). A classmate says “so the clustering shrank your coefficient.” In one or two sentences, explain exactly what is wrong with that statement and what actually changed.
Priya has a panel of 8,000 insurance-claim records but only 9 states. She clusters her standard errors by state. Why should she be nervous about the resulting t-statistics, and name one thing she could do about it.
Sam regresses monthly strategy returns on a signal and reports classical standard errors. His residuals are positively autocorrelated from month to month. Is his reported t-statistic likely too big or too small, and which standard-error flavor should he switch to? Explain why in terms of what the off-diagonal entries of \boldsymbol\Omega look like in his case.

Cameron, A. C., Gelbach, J. B., & Miller, D. L. (2008). Bootstrap-Based Improvements for Inference with Clustered Errors. Review of Economics and Statistics, 90(3), 414–427. The lab in nb2.4 has you watch a cluster-robust SE get unreliable as you shrink G, which is the kind of thing you only really believe once you have made it happen on your own screen.↩︎
Cameron, A. C., Gelbach, J. B., & Miller, D. L. (2011). Robust Inference With Multiway Clustering. Journal of Business & Economic Statistics, 29(2), 238–249.↩︎

# Chapter 2.4 — When the Errors Misbehave: Heteroskedasticity, Clustering, HC/HAC Maya has a clean-looking result and a problem she cannot name yet. She is studying fair lending. She has pulled a dataset of $N = 9{,}000$ consumer loans — the kind of mid-prime installment loans a lot of her own friends' families actually use — and she wants to know whether borrowers in a particular ZIP-code tier get charged higher interest rates after controlling for credit score, loan size, and income. She runs the regression her laptop makes easy. Out comes a coefficient on her tier variable of $\hat\beta_1 = 0.42$ percentage points, with a standard error of $0.07$. The t-statistic is $0.42 / 0.07 = 6.0$. Six. That is not "marginally significant"; that is a coefficient screaming through a megaphone. She writes it up. Then her mentor asks one question: *"Are those 9,000 loans really 9,000 independent pieces of evidence?"* They are not. The loans cluster by lender, and lenders set rates in correlated ways. They cluster by state, and states have correlated regulation and correlated local economies. Once Maya accounts for that, the *same* coefficient $\hat\beta_1 = 0.42$ comes back with a standard error of $0.19$, and the t-statistic collapses from $6.0$ to $2.2$. Still significant — but the megaphone is now a normal speaking voice, and a referee could reasonably push back. Here is the unsettling part, and it is the whole chapter in one line: **her point estimate never moved.** The $0.42$ was right the whole time. What was wrong was the $0.07$ — the number that tells you how much to *trust* the $0.42$. In Ch 2.2 we proved that OLS is unbiased and, under the Gauss–Markov assumptions, the best linear unbiased estimator. None of that breaks here. What breaks is the *standard error*, and therefore every t-statistic, every p-value, every confidence interval you build on top of it. > **The one-sentence result of this chapter:** When the errors are heteroskedastic or > correlated, $\hat{\boldsymbol\beta}$ is still fine — but the classical standard-error formula > lies, and you must replace it with a robust, clustered, or HAC formula that tells the truth. We are going to take this apart carefully, because "your point estimate is fine, your t-stat lies" is one of the most important and most misunderstood facts in all of applied econometrics. We will start exactly where Ch 2.2 left us — with the classical variance formula — watch it break, and build the replacement. --- ## 1. Where we start: the classical variance and the assumption holding it up Recall the setup from Ch 2.1. We write the model in matrix form, $$ \mathbf{y} = \mathbf{X}\boldsymbol\beta + \boldsymbol\varepsilon , $$ where $\mathbf{y}$ is the $N\times 1$ vector of outcomes (Maya's interest rates), $\mathbf{X}$ is the $N\times K$ matrix of regressors (a column of ones for the intercept, plus tier, credit score, loan size, income), $\boldsymbol\beta$ is the $K\times 1$ vector of coefficients we want, and $\boldsymbol\varepsilon$ is the $N\times 1$ vector of errors — everything about each loan's rate that our regressors do not explain. The OLS estimator is $$ \hat{\boldsymbol\beta} = (\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\mathbf{y} . $$ Now substitute $\mathbf{y} = \mathbf{X}\boldsymbol\beta + \boldsymbol\varepsilon$ into that and simplify (the $\mathbf{X}\boldsymbol\beta$ part collapses to $\boldsymbol\beta$, which is the algebra behind unbiasedness): $$ \hat{\boldsymbol\beta} = \boldsymbol\beta + (\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\boldsymbol\varepsilon . $$ This little equation is the engine room of the whole chapter. Read it slowly. The estimator equals the truth, $\boldsymbol\beta$, plus a *sampling-error term* $(\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\boldsymbol\varepsilon$ that depends on which particular errors $\boldsymbol\varepsilon$ the universe happened to deal this sample — exactly the "what if the draw had been different" question from Ch 1.3, now in matrix clothing. The whole variance of $\hat{\boldsymbol\beta}$ comes from that second term. Take its variance, treating $\mathbf{X}$ as fixed (conditioning on $\mathbf{X}$, as in Ch 2.2): $$ \operatorname{Var}(\hat{\boldsymbol\beta}\mid \mathbf{X}) = (\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\, \underbrace{\operatorname{Var}(\boldsymbol\varepsilon \mid \mathbf{X})}_{=\;\boldsymbol\Omega}\, \mathbf{X}(\mathbf{X}'\mathbf{X})^{-1} . $$ I have given the middle object a name: $\boldsymbol\Omega = \operatorname{Var}(\boldsymbol\varepsilon\mid\mathbf{X})$, the $N\times N$ **variance–covariance matrix of the errors.** Its diagonal entries are the variances of each error, $\operatorname{Var}(\varepsilon_i)$; its off-diagonal entries are the covariances between errors of different observations, $\operatorname{Cov}(\varepsilon_i,\varepsilon_j)$. *Everything* in this chapter is a story about what $\boldsymbol\Omega$ really looks like. In Ch 2.2 we made the **homoskedasticity-and-no-correlation** assumption: $$ \operatorname{Var}(\boldsymbol\varepsilon \mid \mathbf{X}) = \boldsymbol\Omega = \sigma^2 \mathbf{I}. $$ In plain language, that single equation makes two claims at once. First, **homoskedasticity** ("same scatter"): every error has the *same* variance $\sigma^2$, regardless of the regressors — a \$2,000 loan and a \$40,000 loan have rates that are equally hard to predict, in squared dollars. That puts the same $\sigma^2$ all down the diagonal. Second, **no correlation across observations**: the errors of any two different loans are uncorrelated, so every off-diagonal entry is zero. Together they make $\boldsymbol\Omega$ a scalar times the identity matrix — a matrix with $\sigma^2$ on the diagonal and zeros everywhere else. Plug $\boldsymbol\Omega = \sigma^2\mathbf{I}$ into the sandwich and watch it collapse. The middle $\sigma^2$ is a scalar and slides out front; the $\mathbf{I}$ vanishes; one $\mathbf{X}'\mathbf{X}$ cancels against an inverse: $$ \operatorname{Var}(\hat{\boldsymbol\beta}\mid\mathbf{X}) = \sigma^2 (\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\mathbf{X}(\mathbf{X}'\mathbf{X})^{-1} = \sigma^2 (\mathbf{X}'\mathbf{X})^{-1} . $$ That is the **classical variance formula** from Ch 2.2. The standard error of any single coefficient $\hat\beta_j$ is the square root of the $j$-th diagonal entry of $\hat\sigma^2(\mathbf{X}'\mathbf{X})^{-1}$, where $\hat\sigma^2 = \frac{1}{N-K}\sum_i \hat\varepsilon_i^2$ is the usual estimate of the error variance built from the residuals $\hat\varepsilon_i$. It is clean, it is fast, and it is what your software prints by default. It is also only as true as the assumption $\boldsymbol\Omega = \sigma^2\mathbf{I}$. The rest of this chapter is what happens when that assumption is false — which, in finance data, is basically always. --- ## 2. Reveal the trick: the sandwich never went away Here is the key idea, and it is liberating once you see it. We did **not** need $\boldsymbol\Omega = \sigma^2\mathbf{I}$ to write down the variance. The honest formula is the full **sandwich**: $$ \boxed{\;\operatorname{Var}(\hat{\boldsymbol\beta}\mid\mathbf{X}) = (\mathbf{X}'\mathbf{X})^{-1}\,\mathbf{X}'\boldsymbol\Omega\,\mathbf{X}\,(\mathbf{X}'\mathbf{X})^{-1}\;} $$ The name is literal. The two $(\mathbf{X}'\mathbf{X})^{-1}$ factors are the slices of bread; the $\mathbf{X}'\boldsymbol\Omega\mathbf{X}$ in the middle is the filling. The classical formula is just the sandwich for the special diet $\boldsymbol\Omega = \sigma^2\mathbf{I}$. If the real $\boldsymbol\Omega$ has *different* variances on the diagonal (heteroskedasticity) or *nonzero* off-diagonals (correlation), the bread is unchanged but the filling is different — and so the final standard errors are different. This reframes the entire problem. We are never going to "fix" $\hat{\boldsymbol\beta}$; there is nothing wrong with it. The whole game is **estimating the middle of the sandwich, $\mathbf{X}'\boldsymbol\Omega\mathbf{X}$, without pretending $\boldsymbol\Omega = \sigma^2\mathbf{I}$.** That sounds impossible at first. $\boldsymbol\Omega$ is an $N\times N$ matrix — with $N = 9{,}000$ loans that is 81 million entries, and we have only 9,000 data points. We cannot estimate 81 million numbers from 9,000 observations. This is the genuine difficulty, and the next three sections are three different clever answers to it: - **White / HC (Section 3):** We don't need $\boldsymbol\Omega$ itself, only the sandwiched combination $\mathbf{X}'\boldsymbol\Omega\mathbf{X}$, which is a small $K\times K$ matrix. If errors are *uncorrelated* (diagonal $\boldsymbol\Omega$), we can estimate that small thing directly using each observation's own squared residual. - **Clustered (Section 4):** If errors are correlated *within groups but not across them*, $\boldsymbol\Omega$ is block-diagonal, and we estimate the middle using whole blocks. - **HAC / Newey–West (Section 5):** If errors are correlated *across time* in a way that fades with distance, we estimate the middle using nearby observations with decaying weights. Same sandwich every time. Only the recipe for the filling changes. ### 2.1 A two-observation sandwich you can do by hand Abstract matrices hide what is going on, so let us collapse the sandwich to the smallest case that still shows the mechanism: a *single* regressor (no intercept), so $\mathbf{X}$ is just a column of numbers and the whole sandwich becomes ordinary arithmetic. Suppose we have only the simplest model $y_i = \beta x_i + \varepsilon_i$. Then $\mathbf{X}'\mathbf{X} = \sum_i x_i^2$ is a single number, and the sandwich for a diagonal $\boldsymbol\Omega$ reads $$ \operatorname{Var}(\hat\beta) = \frac{\sum_i x_i^2\,\sigma_i^2}{\left(\sum_i x_i^2\right)^2}. $$ Now compare two worlds with the same regressors $x = (1, 2, 3, 4)$, so $\sum_i x_i^2 = 30$. *World A (homoskedastic):* every $\sigma_i^2 = 1$. The numerator is $\sum_i x_i^2 \cdot 1 = 30$, so $\operatorname{Var}(\hat\beta) = 30 / 30^2 = 1/30 \approx 0.033$. This is exactly the classical answer $\sigma^2/\sum x_i^2 = 1/30$. *World B (heteroskedastic):* the scatter grows with $x$, say $\sigma_i^2 = x_i^2$, so the variances are $(1, 4, 9, 16)$. The numerator is now $\sum_i x_i^2 \cdot x_i^2 = 1 + 16 + 81 + 256 = 354$, so $\operatorname{Var}(\hat\beta) = 354 / 900 \approx 0.393$ — almost **twelve times larger.** The classical formula, which would still report $1/30 \approx 0.033$ here (it averages the variances and forgets that the big-variance points are also the high-$x$, high-leverage points), would understate the true standard error by a factor of $\sqrt{0.393/0.033} \approx 3.5$. A t-stat of $3.5$ would *really* be a t-stat of $1$. That is the entire heteroskedasticity problem in four data points, and notice it required no calculus and no asymptotics — just plugging the right $\sigma_i^2$ into the middle of the sandwich instead of pretending they are all equal. The rest of Section 3 is only about how to *estimate* those $\sigma_i^2$ when, unlike here, nobody hands them to you. --- ## 3. Heteroskedasticity: White / HC standard errors ### 3.1 What heteroskedasticity is, in a picture Priya, working her insurance-claims data, has the cleanest possible example. She regresses the size of a wildfire claim on the insured value of the property. A \$200,000 cabin can produce a claim anywhere from \$0 (false alarm) to \$200,000 (total loss) — a spread of \$200,000. A \$5 million estate can produce a claim anywhere from \$0 to \$5 million — a spread twenty-five times wider. The *scatter of the errors grows with the regressor.* That is **heteroskedasticity** ("different scatter"): $\operatorname{Var}(\varepsilon_i)$ depends on $i$, so the diagonal of $\boldsymbol\Omega$ is not a single repeated $\sigma^2$ but a list of different $\sigma_i^2$: $$ \boldsymbol\Omega = \operatorname{diag}(\sigma_1^2, \sigma_2^2, \dots, \sigma_N^2). $$ The off-diagonals are still zero — Priya's claims are independent across unrelated properties — so this is a *purely* heteroskedastic, still-uncorrelated case. The classical formula $\sigma^2(\mathbf{X}'\mathbf{X})^{-1}$ uses one average $\hat\sigma^2$ everywhere, which means it systematically misprices the uncertainty: too small where the true scatter is large, too large where it is small. There is no general rule for which way the *overall* t-stat goes — but it is wrong, and in finance it is usually wrong in the direction of overconfidence. ### 3.2 White's idea (1980) Halbert White's insight was to stop chasing the impossible $N\times N$ object and look at what we actually need. Write the middle of the sandwich for a diagonal $\boldsymbol\Omega$. Because the off-diagonals are zero, the matrix product $\mathbf{X}'\boldsymbol\Omega\mathbf{X}$ becomes a simple sum over observations: $$ \mathbf{X}'\boldsymbol\Omega\mathbf{X} = \sum_{i=1}^{N} \sigma_i^2\,\mathbf{x}_i \mathbf{x}_i' , $$ where $\mathbf{x}_i$ is the $K\times 1$ row of regressors for observation $i$ (written as a column). This is a $K\times K$ matrix — for Maya, $5\times 5$ — no matter how big $N$ is. We do *not* need each $\sigma_i^2$ on its own; we need this weighted sum. And White's trick is that each unknown $\sigma_i^2$ can be replaced by the single thing we observe for that point: its **squared residual** $\hat\varepsilon_i^2$. Any one $\hat\varepsilon_i^2$ is a terrible estimate of $\sigma_i^2$ — it's based on one data point. But summed up across all $N$ observations, the errors in those guesses average out, and the *sum* converges to the right thing. That gives the **heteroskedasticity-consistent (HC) estimator**: $$ \widehat{\operatorname{Var}}(\hat{\boldsymbol\beta})_{\text{HC0}} = (\mathbf{X}'\mathbf{X})^{-1}\left(\sum_{i=1}^{N}\hat\varepsilon_i^2\,\mathbf{x}_i\mathbf{x}_i'\right)(\mathbf{X}'\mathbf{X})^{-1}. $$ These are **White standard errors**, also called **heteroskedasticity-robust** or just **robust** standard errors. The "0" in HC0 is a clue that there are corrected versions HC1–HC3, which we get to next. Crucially, robust SEs are valid whether or not there is heteroskedasticity: if the errors *happen* to be homoskedastic, the robust estimator converges to the classical answer as $N$ grows. In fact the small-sample correspondence is sharpest for **HC1**, which carries the same $N/(N-K)$ degrees-of-freedom factor as the classical $\hat\sigma^2$: under homoskedasticity HC1 reproduces the classical SE almost exactly, while raw HC0 (lacking that factor) runs a touch *smaller*. Either way, robust SEs cost you almost nothing and protect you from a real failure — which is why "always report robust" is close to a default in modern finance work. ### 3.3 HC0, HC1, HC2, HC3 — the finite-sample corrections White's HC0 is *consistent*, meaning it gets the right answer as $N\to\infty$. But in finite samples it has a known flaw: OLS residuals are systematically **too small.** The fitting procedure bends the line toward each point, so $\hat\varepsilon_i$ understates the true $\varepsilon_i$, which makes HC0 standard errors too *small* — overconfident again, in small samples. The fix is to inflate the squared residuals. The four common flavors differ only in *how* they inflate: | Flavor | Each $\hat\varepsilon_i^2$ is replaced by | What it corrects | Use when | |--------|-------------------------------------------|------------------|----------| | **HC0** | $\hat\varepsilon_i^2$ | nothing (raw White) | large $N$, you don't care about small-sample bias | | **HC1** | $\dfrac{N}{N-K}\,\hat\varepsilon_i^2$ | a flat degrees-of-freedom inflation | the default in Stata; fine for moderate $N$ | | **HC2** | $\dfrac{\hat\varepsilon_i^2}{1-h_{ii}}$ | leverage, observation by observation | small samples; principled choice | | **HC3** | $\dfrac{\hat\varepsilon_i^2}{(1-h_{ii})^2}$ | leverage, more aggressively | small $N$ or high-leverage points; safest | The $h_{ii}$ here are the diagonal entries of the **hat matrix** $\mathbf{H} = \mathbf{X}(\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'$ from Ch 2.1 — the **leverage** of observation $i$, a number between 0 and 1 measuring how much that point pulls the fitted line toward itself. A high-leverage point (an outlier in $\mathbf{X}$, like Priya's one \$5 million estate among cabins) has its residual shrunk the most by fitting, so HC2 and HC3 inflate it the most to compensate. HC1 ignores leverage and applies the same $N/(N-K)$ factor to everyone. The practical guidance is simple and worth memorizing. With a few hundred observations or more and no extreme leverage, HC1 (Stata's default) is fine and HC0–HC3 barely differ. When $N$ is small — say under 50 — or a handful of points have high leverage, the flavors diverge and HC3 is the conservative, recommended default (it is the most likely to keep you honest). HC3 is never badly anti-conservative, so if you are unsure, HC3 is the safe pick. ### 3.4 A number to make it concrete Priya simulates 60 claims where the true error standard deviation grows linearly with insured value (textbook heteroskedasticity), with one \$5M estate among \$200K cabins to create leverage. On her coefficient of interest the four standard errors come out roughly: | Classical | HC0 | HC1 | HC2 | HC3 | |-----------|-----|-----|-----|-----| | 0.085 | 0.121 | 0.127 | 0.140 | 0.169 | The classical SE of $0.085$ understates the true uncertainty by about half relative to HC3's $0.169$. If her coefficient were $0.30$, the classical t-stat would read $0.30/0.085 = 3.5$ (comfortably significant) while the HC3 t-stat reads $0.30/0.169 = 1.8$ (not significant at the 5% level). Same coefficient. Same data. The only thing that changed is whether we told the truth about $\boldsymbol\Omega$ — and it flipped the conclusion. Notice also how the four robust flavors fan out: HC0 at $0.121$ and HC3 at $0.169$ differ by a real margin precisely *because* that one \$5M estate is a high-leverage point, and HC3's $(1-h_{ii})^2$ denominator inflates its residual the hardest. In a sample with no leverage and a few hundred points, those four numbers would have agreed to the third decimal. ### 3.5 When it fails, what you actually see Reveal-the-trick demands we say what the *symptom* looks like, so you can catch it in your own work rather than only in a textbook. Three tells, in increasing order of formality: First, **the residual plot.** Plot the residuals $\hat\varepsilon_i$ against a fitted value or against a regressor. Under homoskedasticity you see a structureless horizontal band of constant thickness. Under heteroskedasticity you see a *fan* or *funnel* — the band widens (or narrows) as you move along the axis. Priya's wildfire residuals fan open toward expensive properties; Maya's loan residuals often fan with loan size. This single plot catches most real cases and costs you one line of code. Second, **a formal test.** The Breusch–Pagan test regresses the squared residuals on the regressors and asks whether they explain any of the variance (if scatter is constant, they should not); the White test does the same with squares and cross-products to catch more general patterns. A small p-value says "reject homoskedasticity." But here is the modern attitude, and it is worth internalizing: **you usually should not bother testing.** Because robust SEs are valid whether or not heteroskedasticity is present, the cost of "just always using robust" is near zero, while the cost of testing, failing to reject by bad luck, and then trusting classical SEs is a published overconfident result. Test if a referee asks; otherwise default to robust. Third, **the SEs themselves disagree.** If classical and robust standard errors are nearly identical, your errors are roughly homoskedastic and nothing was at stake. If they differ materially — as in Priya's table — heteroskedasticity is real and the robust number is the one to report. The disagreement *is* the diagnostic. --- ## 4. Clustering: when errors hold hands within groups ### 4.1 The failure that White SEs do not fix Robust SEs assume the off-diagonals of $\boldsymbol\Omega$ are zero — errors uncorrelated across observations. In finance that assumption is heroic and usually false. Maya's loans within the same lender share the lender's pricing model, risk appetite, and quirks; their errors move together. Priya's wildfire claims within the same state share that state's fire season, building codes, and reinsurance market; their errors move together. Devon's daily token returns within the same week share whatever sentiment shock hit crypto that week. These are **within-cluster correlations**, and robust SEs are blind to them because they only ever look at one observation at a time. The stakes here are not academic. Go back to Maya's fair-lending result from the opening. Her claim — that one borrower tier pays measurably more after controlling for credit — is the kind of finding that can support or sink a discrimination complaint, inform a regulator, or appear in a policy brief. If she reports a t-stat of $6.0$ when the honest, clustered t-stat is $2.2$, she has not just made a technical slip; she has overstated the strength of evidence on a question where overstating it has real consequences in both directions (a false alarm wastes enforcement attention; underpowered honesty lets real discrimination slide). The reason the naive number is *so* much too big is not that her 9,000 loans are fake data — they are real — but that they are not 9,000 *independent* pieces of evidence. Loans from the same lender largely repeat the same story. Counting them as independent is like polling one household nine times and reporting a sample of nine. The clustered standard error is the correction that stops you from doing that. Here is *why ignoring clustering is so dangerous* — and it is the single most expensive mistake in empirical finance, so it earns its own derivation. Suppose Maya has $N$ loans split into $G$ lenders ("clusters") of equal size $n = N/G$, and within a lender the errors all share a common shock with within-cluster correlation $\rho$. There is a classic result (a special case of the Moulton factor) for how badly the naive standard error understates the truth. The variance of your estimate is inflated, relative to what the naive formula reports, by approximately $$ \text{true Var} \approx \text{naive Var} \times \big[\,1 + (n-1)\,\rho_x\,\rho_\varepsilon\,\big], $$ where $n$ is the cluster size, $\rho_\varepsilon$ is the within-cluster error correlation, and $\rho_x$ is the within-cluster correlation of your regressor. Stare at the bracket. The damage scales with **cluster size $n$**: bigger clusters, worse understatement. With $n = 200$ loans per lender and even a mild $\rho_x\rho_\varepsilon = 0.05$, the bracket is $1 + 199\times0.05 \approx 11$. Your true variance is *eleven times* the naive one — so your true standard error is $\sqrt{11}\approx 3.3$ times bigger, and your reported t-stat of $6.0$ should really be about $6.0/3.3 \approx 1.8$. That is *exactly* what happened to Maya in the opening, and now you can see it was not bad luck; it was arithmetic. Pouring in more observations *within existing clusters* does not buy you the independent evidence the naive formula thinks it does. What buys evidence is more **clusters.** ### 4.2 The cluster-robust sandwich The fix is the same sandwich, with a middle built from whole clusters instead of single points. If errors are correlated *within* cluster $g$ but uncorrelated *across* clusters, then $\boldsymbol\Omega$ is **block-diagonal** — a big matrix with dense blocks down the diagonal (one block per cluster, capturing all the within-cluster covariances) and zeros everywhere else. The middle of the sandwich then becomes a sum over the $G$ clusters: $$ \widehat{\operatorname{Var}}(\hat{\boldsymbol\beta})_{\text{cluster}} = (\mathbf{X}'\mathbf{X})^{-1}\left(\sum_{g=1}^{G}\mathbf{X}_g'\hat{\boldsymbol\varepsilon}_g\,\hat{\boldsymbol\varepsilon}_g'\mathbf{X}_g\right)(\mathbf{X}'\mathbf{X})^{-1}, $$ where $\mathbf{X}_g$ is the block of regressor rows for cluster $g$ and $\hat{\boldsymbol\varepsilon}_g$ is its vector of residuals. Compare this to White's HC0 in §3.2. There the middle summed $\hat\varepsilon_i^2\,\mathbf{x}_i\mathbf{x}_i'$ over single *observations*; here it sums $\mathbf{X}_g'\hat{\boldsymbol\varepsilon}_g\hat{\boldsymbol\varepsilon}_g'\mathbf{X}_g$ over whole *clusters*, and the outer product $\hat{\boldsymbol\varepsilon}_g\hat{\boldsymbol\varepsilon}_g'$ deliberately keeps the within-cluster covariances (the products $\hat\varepsilon_{i}\hat\varepsilon_{j}$ for two loans $i,j$ in the same lender) that White's version threw away. White SEs are just cluster SEs where every cluster has size one. It is the *same machine*, zoomed out from points to groups. The empirical-spec discipline from the Conventions now has teeth: when you write a spec you must **name the clustering level**, and that choice is an economic claim about where $\boldsymbol\Omega$'s off-diagonal blocks live. "Clustered by lender" asserts that two loans from the same lender have correlated errors and two loans from different lenders do not. If you believe the correlation actually lives at the state level (shared regulation, shared economy), you cluster by state, and the blocks get bigger. ### 4.3 The few-clusters caveat There is a catch, and it bites hard in finance. The cluster-robust estimator is consistent as the **number of clusters $G \to \infty$**, not as $N\to\infty$. The asymptotics live in $G$, not in $N$. If Maya clusters by state, she has at most $G = 50$ clusters; if she clusters by some broad industry, maybe $G = 12$. With few clusters the cluster-robust SE is itself noisy and *biased downward* — overconfident again, the very disease we are treating. The rough rule of thumb is that you want **at least 30–50 clusters** before you trust the cluster-robust normal-approximation t-test at face value. Below that, you reach for fixes: the small-sample $t$-distribution with $G-1$ degrees of freedom rather than the normal; or the **wild cluster bootstrap** (Cameron, Gelbach & Miller, 2008),[^cgm2008] which resamples whole clusters and is the current standard remedy for the few-clusters problem. [^cgm2008]: Cameron, A. C., Gelbach, J. B., & Miller, D. L. (2008). Bootstrap-Based Improvements for Inference with Clustered Errors. *Review of Economics and Statistics*, 90(3), 414–427. The lab in nb2.4 has you watch a cluster-robust SE get unreliable as you shrink $G$, which is the kind of thing you only really believe once you have made it happen on your own screen. [^cgm2011]: Cameron, A. C., Gelbach, J. B., & Miller, D. L. (2011). Robust Inference With Multiway Clustering. *Journal of Business & Economic Statistics*, 29(2), 238–249. --- ## 5. HAC / Newey–West: correlation across time ### 5.1 The time-series version of the same disease Sam trades. He regresses a strategy's monthly return on a market-timing signal and gets a beautiful t-stat. But returns and residuals in financial time series are **serially correlated** (also called **autocorrelated**): this month's error is correlated with last month's, because slow-moving forces — a drifting risk premium, a multi-month liquidity squeeze, momentum itself — persist across periods. Now $\boldsymbol\Omega$ has nonzero off-diagonals, but in a special pattern: the correlation is strongest between *adjacent* periods and **fades as the time gap grows.** $\operatorname{Cov}(\varepsilon_t,\varepsilon_{t-1})$ is large, $\operatorname{Cov}(\varepsilon_t,\varepsilon_{t-12})$ is near zero. This is not clustering — there are no clean groups — and it is not pure heteroskedasticity. It is a third shape for $\boldsymbol\Omega$: a band of decaying correlations around the diagonal. Ignoring it, just like ignoring clustering, makes standard errors too small when the autocorrelation is positive (the usual case in finance), inflating t-stats. ### 5.2 Newey and West's idea (1987) Whitney Newey and Kenneth West built the **heteroskedasticity-and-autocorrelation-consistent (HAC)** estimator for exactly this. Same sandwich. The middle now includes not just each period's own squared-residual term but **cross-products of residuals from nearby periods**, $\hat\varepsilon_t\hat\varepsilon_{t-\ell}\,\mathbf{x}_t\mathbf{x}_{t-\ell}'$, for time gaps ("lags") $\ell = 1, 2, \dots, L$. Because correlation fades with distance, Newey–West applies **declining weights** $w_\ell = 1 - \tfrac{\ell}{L+1}$ (the Bartlett kernel): the lag-1 cross-products count almost fully, distant lags count for less, and beyond the chosen **bandwidth** $L$ they count for nothing. The two declining-weight pieces are there both to honor the fading-correlation structure and, mathematically, to guarantee the resulting variance matrix is positive (a variance can't be negative — without the weights this sum could misbehave). The one knob you must set is the bandwidth $L$ — how many lags to include. Too small and you miss real persistence; too large and you add noise. A common automatic choice is $L \approx 4(T/100)^{2/9}$ (the Newey–West rule of thumb), where $T$ is the number of time periods, and most software offers an automatic-bandwidth option. For Sam's monthly data, setting $L$ to a small multiple of 12 (a year or two of lags) is a sensible manual starting point. HAC SEs are the default for pure time-series regressions: predictive return regressions, factor-model time-series tests, anything indexed by $t$ with one observation per period. ### 5.3 Why positive autocorrelation inflates t-stats, in one intuition It is worth pinning down *why* serial correlation, like clustering, makes naive standard errors too small — because the mechanism is the same one, wearing a time-series costume. Recall the Moulton-style intuition from §4.1: when errors are positively correlated, each new observation carries *less* fresh information than the naive formula assumes, because part of it just echoes what you already saw. A time series with positive autocorrelation is exactly a chain of observations that echo their neighbors. If Sam has $T = 120$ monthly returns but each month's residual is, say, 0.4-correlated with the previous month's, his *effective* sample size — the number of genuinely independent monthly draws — is well below 120. The naive formula divides by the wrong (too large) effective $T$, so it reports a standard error that is too small and a t-stat that is too big. HAC repairs this by counting the lagged cross-products, which is arithmetically how it discovers that the 120 months are worth fewer than 120 independent observations. The same logic explains the sign. *Positive* autocorrelation (the usual case in finance, where shocks persist) shrinks the effective sample and inflates t-stats — the dangerous direction. *Negative* autocorrelation (rarer, e.g. bid–ask bounce in high-frequency prices) would do the reverse, making naive t-stats too conservative. Either way, HAC gets it right because it estimates the actual band of off-diagonal covariances rather than assuming the band is empty. --- ## 6. Putting it together: the Petersen (2009) panel taxonomy Now the hard case, and the one Maya and Priya actually live in: a **panel**, where you observe many firms ($i$) over many time periods ($t$) at once. A panel can have *both* diseases at once — errors correlated within a firm across time (a **firm effect**: a firm that is mispriced this year tends to be mispriced next year) *and* errors correlated across firms within the same period (a **time effect**: when the whole market drops, every firm's residual drops together). Which standard error do you use? This is the precise question Mitchell Petersen took up in a now-canonical 2009 paper, after noticing that finance papers were using a chaos of incompatible methods and getting standard errors that differed by factors of two or more *on the same data*. His contribution was not a new estimator so much as a clear **taxonomy and a set of simulations** showing which approach is right in which situation. The headline findings, which you should treat as practical defaults: - **A persistent firm effect** (the firm's residuals correlated over time) is, in Petersen's data, the dominant problem in corporate-finance panels, and it is the one most often ignored. The right tool is **clustering by firm**, which lets each firm's errors be arbitrarily correlated across all its years — exactly the block-diagonal $\boldsymbol\Omega$ of Section 4, with one block per firm spanning all its time periods. White SEs and Fama–MacBeth do *not* fix a firm effect and will leave you badly overconfident. - **A time effect** (a common shock hitting all firms in a period, like a market-wide crash) is better handled by **clustering by time** — one block per period, spanning all firms in it. Petersen notes that in many corporate panels a time effect is well absorbed by including **time fixed effects** (a dummy for each period, previewed by the demeaning idea in Ch 2.3); but if a *residual* common shock remains, clustering by time catches it. - **Both effects present?** Use **two-way clustering** (Cameron, Gelbach & Miller, 2011,[^cgm2011] building on Petersen and Thompson), which clusters by firm *and* by time simultaneously, allowing both kinds of correlation in $\boldsymbol\Omega$ at once. The mechanical recipe is elegant: compute the firm-clustered sandwich, add the time-clustered sandwich, and subtract the plain White (heteroskedasticity-only) sandwich so the overlap isn't double-counted. The decision rule Petersen leaves you with is the empirical-spec discipline made concrete: **figure out where your residuals are correlated — across time within a firm, across firms within a period, or both — and cluster on that dimension (or those dimensions).** It is an economic judgment about $\boldsymbol\Omega$, not a software default, and the same dataset can demand different clustering depending on what you put on the left-hand side. (Ch 5.4 is a full reader's guide to this paper; here we want the operational rule.) How do you tell *which* effect dominates without guessing? Petersen offers a practical diagnostic that you can run yourself. Cluster by firm and look at the standard error; then cluster by time and look again. Whichever clustering moves the standard error *more* away from the naive White number is pointing at the bigger source of correlation. If clustering by firm roughly doubles your SE but clustering by time barely changes it, you have a firm effect and little else — firm-clustering is enough. If both move it a lot, you have both effects and want two-way clustering. This is exactly the experiment nb2.4 walks you through, and it turns the "which cluster?" question from a matter of taste into something you can read off your own output. A subtlety worth flagging now and developing in Weeks 3–4: **fixed effects and clustering are not substitutes — they do different jobs.** A firm fixed effect (a dummy for each firm, Ch 2.3's demeaning) removes the firm's *average* level from both sides; it handles a firm-specific mean that never changes. Clustering by firm handles *correlation in the leftover residuals* over time — the part a fixed effect cannot touch because it is about co-movement, not means. You often want both at once: firm fixed effects to absorb permanent firm differences *and* clustering by firm to honor the within-firm residual correlation that remains. Conflating the two ("I have firm fixed effects, so I don't need to cluster by firm") is a common and costly error, and it is the kind of thing the empirical-spec line in your write-up forces you to get explicit about. A worked taste of the stakes, in the spirit of Petersen's own tables. Maya assembles a panel of default outcomes for many firms over many years and regresses default on a leverage measure. The coefficient is the same in every column; only the SE flavor changes. The illustrative numbers below show the *shape* of what happens — the SE inflates as you allow more correlation into $\boldsymbol\Omega$, dragging the t-stat across the significance line. (For real, reproducible magnitudes built from a simulated panel, see nb2.4: on its 120-firm $\times$ 20-year panel the identical $\hat\beta=0.30$ carries a classical t of $5.70$ that collapses to a firm-clustered t of $2.19$ — the same story, with numbers you can re-run. The smaller §7 toy panel just below shows the same collapse at its own scale.) | SE flavor | Std. error (illustrative) | t-stat | Honest? | |-----------|-----------|--------|---------| | Classical (OLS) | 0.018 | 5.0 | No — assumes $\sigma^2\mathbf{I}$ | | White (HC1) | 0.021 | 4.3 | No — ignores within-firm correlation | | Clustered by firm | 0.044 | 2.0 | Yes, if the firm effect dominates | | Two-way (firm + year) | 0.052 | 1.7 | Yes, if both effects are present | These figures are illustrative, not the output of a committed script; the live numbers are in nb2.4. The qualitative point is exact, though: a classical t-stat around $5$ and a two-way-clustered t-stat around $1.7$ are computed from *identical* point estimates on *identical* data. One says "publish"; the other says "not yet." The difference is entirely a story about $\boldsymbol\Omega$, and choosing the wrong one is how real papers get retracted. --- ## 7. The code Every flavor in this chapter is one argument in `statsmodels`. The point of seeing them side by side is to feel how much the t-stat moves while $\hat\beta$ sits perfectly still. ```python import numpy as np import pandas as pd import statsmodels.formula.api as smf # --- A toy clustered panel: firms over years, with a firm-level error component --- rng = np.random.default_rng(8) n_firms, n_years = 40, 15 firm = np.repeat(np.arange(n_firms), n_years) year = np.tile(np.arange(n_years), n_firms) # the regressor itself has a persistent firm-level piece (correlated within firm) -- # this is what makes ignoring clustering so costly (the Moulton logic of Section 4.1): x_firm = np.repeat(rng.normal(size=n_firms), n_years) x = x_firm + 0.5 * rng.normal(size=n_firms * n_years) # the key regressor firm_shock = np.repeat(rng.normal(size=n_firms), n_years) # persistent firm effect year_shock = np.tile(rng.normal(size=n_years), n_firms) # common time shock eps = 1.5 * firm_shock + year_shock + rng.normal(size=n_firms * n_years) y = 0.30 * x + eps # TRUE slope is 0.30 df = pd.DataFrame({"y": y, "x": x, "firm": firm, "year": year}) fit = smf.ols("y ~ x", data=df).fit() # one OLS fit, reused below xi = list(fit.params.index).index("x") # column position of the x coef def se(res, label): # robust results return plain numpy arrays, so index by position, not name b = np.asarray(res.params)[xi] s = np.asarray(res.bse)[xi] print(f"{label:<22} beta={b:.3f} se={s:.3f} t={b/s:.2f}") se(fit, "classical") se(fit.get_robustcov_results("HC1"), "robust HC1") se(fit.get_robustcov_results("HC3"), "robust HC3") se(fit.get_robustcov_results("cluster", groups=df["firm"]), "cluster by firm") se(fit.get_robustcov_results("cluster", groups=df[["firm", "year"]]), "two-way (firm+year)") # HAC / Newey-West (for a pure time series; shown for completeness): se(fit.get_robustcov_results("HAC", maxlags=4), "Newey-West HAC(4)") ``` Run it and you will see the coefficient on `x` glued near $0.27$ in every line (close to the true $0.30$) while the standard error roughly *doubles* from classical (about $0.073$) to firm-clustered (about $0.144$), because we deliberately built a strong, persistent firm effect into both the regressor and the error. That doubling is the chapter in one screen of output: **the estimate is fine; the t-stat was lying** — it falls from about $3.7$ to about $1.9$, crossing the significance line without the point estimate budging. Change `firm_shock` to zero (and drop the firm piece of `x`) and the gap closes — robust and classical agree when the errors really are well behaved, which is the reassurance that you lose nothing by defaulting to the honest formula. --- ## 8. The mental checklist When you sit down with a regression and ask "which standard error?", walk the decision in this order — it is just "what does $\boldsymbol\Omega$ look like?" in plain questions: 1. **Could the error variance differ across observations?** In finance, essentially always yes. So never use classical SEs as your final answer — use at least **robust (HC1, or HC3 if $N$ is small or leverage is high)**. 2. **Are observations grouped, with correlated errors inside a group** (firms, lenders, states, schools)? Then **cluster** on that group — and check you have enough clusters (aim for $\gtrsim 30$–$50$; otherwise use the wild cluster bootstrap). 3. **Is it a pure time series with serial correlation?** Use **HAC / Newey–West**, and set the bandwidth to cover the persistence you expect. 4. **Is it a panel (many units × many periods)?** Diagnose the correlation structure à la **Petersen (2009)**: firm effect → cluster by firm; time effect → time FEs or cluster by time; both → two-way cluster. The recurring lesson, and the reason this chapter sits between the Gauss–Markov guarantees of Ch 2.2 and the bias problems of Ch 2.5: **getting $\hat\beta$ right and getting its uncertainty right are two separate jobs.** Ch 2.5 is about when $\hat\beta$ itself is biased — a different and in some ways more dangerous failure. This chapter was entirely about the second job. A biased estimator points at the wrong answer; a wrong standard error makes you wrongly *confident* about whatever answer you have. Both will get a finding rejected. Only one of them will get it rejected *after* it's been published. --- ## Your Turn Open **nb2.4 — SE flavors on a clustered panel.** You will build a panel with a tunable firm effect and a tunable time effect, then watch the five standard-error flavors (classical, HC1, HC3, firm-clustered, two-way) move as you turn the knobs — including the unsettling experiment of shrinking the number of clusters until the cluster-robust SE stops being trustworthy. The goal is to *see*, with your own simulated data, the gap between the t-stat that lies and the t-stat that tells the truth. **Check questions.** 1. Maya runs a regression and gets $\hat\beta_1 = 0.50$ with a classical SE of $0.10$ (t = 5.0). After clustering by lender the SE rises to $0.25$ (t = 2.0). A classmate says "so the clustering shrank your coefficient." In one or two sentences, explain exactly what is wrong with that statement and what actually changed. 2. Priya has a panel of 8,000 insurance-claim records but only **9 states**. She clusters her standard errors by state. Why should she be nervous about the resulting t-statistics, and name one thing she could do about it. 3. Sam regresses monthly strategy returns on a signal and reports classical standard errors. His residuals are positively autocorrelated from month to month. Is his reported t-statistic likely too big or too small, and which standard-error flavor should he switch to? Explain why in terms of what the off-diagonal entries of $\boldsymbol\Omega$ look like in his case.