Cross Validated
2026-11-04 13:59 UTC
By Efey
AI-113-20261104-social-media-ee327991
Why does a short stationary bootstrap underestimate uncertainty for a contrast of correlated AR(1) series?
I am studying the finite-sample behavior of a one-sided test for a contrast of four bounded, correlated time series. The following example is entirely synthetic ; it contains no application data. For four stationary zero-mean series, the estimand is $$ \theta=\frac12[(\mu_1-\mu_2)+(\mu_3-\mu_4)]=0. $$ I test $H_0:\theta\leq 0$ versus $H_1:\theta>0$ at $\alpha=0.05$ . Each series has 200 observations. A stationary Gaussian AR(1) common factor and four AR(1) idiosyncratic factors generate correlated signs, with $\rho=0$ or $0.90$ . The common-factor mixing coefficient is $0.75$ , and observations are transformed into bounded values $\operatorname{sign}(Z_t)(2,1,2,1)$ . By symmetry, each population mean, and therefore the contrast, is exactly zero. All four series are resampled with the same circular stationary-bootstrap date indices , preserving contemporaneous dependence. I use an expected block length of 8 observations, 500 bootstrap resamples per dataset, and the bootstrap standard deviation of the contrast as the standard error in a one-sided normal test. For comparison only, I estimate a Monte Carlo reference standard deviation from the same 400 independently generated datasets used to compute the rejection rates . This is an in-sample diagnostic reference, not an independently estimated oracle. It is not available for real data. Reproducible Python/NumPy code import numpy as np DAYS, TRIALS, RESAMPLES, BLOCK = 200, 400, 500, 8 CRITICAL = 1.6448536269514722 def sample_dat…
I am studying the finite-sample behavior of a one-sided test for a contrast of four bounded, correlated time series. The following example is entirely synthetic ; it contains no application data. For four stationary zero-mean series, the estimand is $$ \theta=\frac12[(\mu_1-\mu_2)+(\mu_3-\mu_4)]=0. $$ I test $H_0:\theta\leq 0$ versus $H_1:\theta>0$ at $\alpha=0.05$ . Each series has 200 observations. A stationary Gaussian AR(1) common factor and four AR(1) idiosyncratic factors generate correlated signs, with $\rho=0$ or $0.90$ . The common-factor mixing coefficient is $0.75$ , and observations are transformed into bounded values $\operatorname{sign}(Z_t)(2,1,2,1)$ . By symmetry, each population mean, and therefore the contrast, is exactly zero. All four series are resampled with the same circular stationary-bootstrap date indices , preserving contemporaneous dependence. I use an expected block length of 8 observations, 500 bootstrap resamples per dataset, and the bootstrap standard deviation of the contrast as the standard error in a one-sided normal test. For comparison only, I estimate a Monte Carlo reference standard deviation from the same 400 independently generated datasets used to compute the rejection rates . This is an in-sample diagnostic reference, not an independently estimated oracle. It is not available for real data. Reproducible Python/NumPy code import numpy as np DAYS, TRIALS, RESAMPLES, BLOCK = 200, 400, 500, 8 CRITICAL = 1.6448536269514722 def sample_dat…
Full article content could not be extracted automatically. Read the original below.
Source:
Cross Validated
· stats.stackexchange.com