You copy every row of your data and rerun the regression. What happens to the t-statistic?

Statistics
Medium
A regression puzzle about standard errors and independent observations.
Published

November 16, 2026

NoteQuestion

You regress y on x by least squares with 30 observations and get a slope of 0.39 with a t-statistic of 1.6, from the standard formula that assumes independent errors. You then copy every row once, so the data set has 60 rows, and run the same regression. What happens to the slope and to its t-statistic?

Try it first, then open the solution.

The slope stays 0.39, and the t-statistic grows from 1.6 to 2.3. The copies add no information, yet the slope now looks significant at 5%.

The regression fits \(y_i = a + b\,x_i + e_i\) to the \(n\) rows, where \(e_i\) is the residual of row \(i\). The least-squares slope divides the covariance of \(x\) and \(y\) by the variance of \(x\), both averages over the rows:

\[\begin{aligned} b &= \frac{\operatorname{Cov}(x, y)}{\operatorname{Var}(x)}, \\ \operatorname{Cov}(x, y) &= \frac{1}{n} \sum_i (x_i - \bar x)(y_i - \bar y), \\ \operatorname{Var}(x) &= \frac{1}{n} \sum_i (x_i - \bar x)^2. \end{aligned} \tag{1}\]

Copying every row doubles each sum and doubles \(n\), so both averages, and the slope, stay the same. The standard error of the slope is

\[\begin{aligned} \operatorname{SE}(b) &= \frac{s}{\sqrt{n \operatorname{Var}(x)}}, \\ s^2 &= \frac{\sum_i e_i^2}{n - 2}, \end{aligned} \tag{2}\]

where \(s^2\) estimates the variance of the errors, the sum of squared residuals divided by \(n - 2\), and the 2 counts the two estimated coefficients, \(a\) and \(b\). Copying every row doubles \(n \operatorname{Var}(x)\) and \(\sum_i e_i^2\), while \(n - 2\) grows to \(2n - 2\), so

\[\operatorname{SE}_\text{copied} = \operatorname{SE} \times \sqrt{\frac{n - 2}{2n - 2}}. \tag{3}\]

The question gives three inputs:

Symbol Meaning Value
\(n\) number of rows before the copy 30
\(b\) slope 0.39
\(t\) t-statistic of the slope, \(b / \operatorname{SE}(b)\) 1.6

The slope stays the same, so dividing it by (3) gives

\[\begin{aligned} t_\text{copied} &= t \times \sqrt{\frac{2n - 2}{n - 2}} \\ &= 1.6 \times \sqrt{\frac{58}{28}} = 2.3. \end{aligned} \tag{4}\]

The factor \(\sqrt{58/28} = 1.44\) is close to \(\sqrt{2}\) because \(n\) doubled. Formula (2) assumes \(n\) independent errors, but a copied row repeats the error of its original, so the 60 rows hold only 30 independent errors. The same overcount appears with overlapping returns, such as twelve-month returns measured every month, where Newey–West standard errors with 11 lags are the usual fix.

Three wrong answers are common:

Answer The slip
both stay the same misses that \(n\) in (2) doubles, so the standard error falls
t doubles lets the standard error fall with \(n\), when (2) divides by \(\sqrt{n}\)
the slope changes misses that the averages in (1) stay the same when every row is copied
import numpy as np
import statsmodels.api as sm

rng = np.random.default_rng(264)
x = rng.normal(size=30)
y = 0.4 * x + rng.normal(size=30)

for copies in (1, 2):
    fit = sm.OLS(np.tile(y, copies), sm.add_constant(np.tile(x, copies))).fit()
    print(30 * copies, round(fit.params[1], 2), round(fit.bse[1], 3), round(fit.tvalues[1], 1))
30 0.39 0.239 1.6
60 0.39 0.166 2.3