With equal Gaussian log errors, 95% MD intervals ignoring correlation had 26.7% coverage in ds008695 b≤900 fits without an intercept after linear log reference interpolation

We tested the low b diffusion MRI design in OpenNeuro ds008695 version 1.0.0. The same numerical design occurs in 22 raw AP scan records. We used linear interpolation of b=0 log references and a tensor fit with no intercept. Raw log errors were independent Gaussian values with equal variance. The complete mean diffusivity standard error was 5.7446 times the error from the exact covariance diagonal alone. The corresponding nominal 95% interval had exact coverage of 26.7035%. In 100,000 raw noise draws, this interval had 26.638% coverage. The full covariance interval had 95.044% coverage. Generalized least squares reduced the modeled standard error by 10.2704%. It gave the same tensor estimate as a fit of the original signals with a piecewise linear log baseline. We also tested reference noise ratios, a b>=200 subset, gradient norms, and the separate high b design. The package preserves code, all small source files, provenance, matrices, simulation results, a fitting tool, and offline reproduction instructions. All 58 result files were reproduced byte for byte. All 136 source files passed a second download hash check. This is a fixed design and model result. No measured image signal or released derivative pipeline was tested.

diffusion MRI · mean diffusivity · diffusion tensor imaging · reference noise · covariance · confidence intervals · generalized least squares · Gaussian log errors