Statistical Methods
Record the Reference Distribution and Method in an Anderson–Darling Check
Define the task
An Anderson–Darling check compares data with a specified distribution, with particular sensitivity to tail discrepancies. State the distribution family, how parameters are fitted and what observations are eligible. A statistic does not have a universal reference cutoff across families or fitting procedures. Preserve fitting diagnostics, especially when numerical estimation is difficult. An unsuccessful fit should not be reported as a successful goodness-of-fit test with an assumed reference value.
Carry out the check
Current SciPy documentation offers an explicit interpolated p-value method or a configured Monte Carlo method. It describes older critical-value outputs as deprecated, so record the installed version and method rather than assuming a result object always contains the same fields. For simulation-based calibration, preserve the random generator, number of simulations and fitting scheme. Do not reuse a normal-family cutoff for a Weibull fit. A goodness-of-fit result evaluates a chosen model against available data; it does not establish that the model is scientifically true.
Worked hypothetical example
In a hypothetical simulation check, 49 of 999 simulated statistics are at least as extreme as the observed statistic. A plus-one tail estimate is (49+1)/(999+1)=0.05, with resolution 1/1000. This illustrates Monte Carlo resolution, not an actual software result or a universal Anderson–Darling p-value. Verify the implemented tail and simulation method before reproducing this convention. Keep the observed statistic and fitting details with the uncertainty introduced by finite simulation.
Method checked 18 November 2026 against Primary method documentation: anderson.