# Residual-level HD diagnostics on NANOGrav 15-yr and EPTA DR2: which nulls survive real PTA sampling (pre-registered replication)

**URL:** <https://ask.igwn.org/t/residual-level-hd-diagnostics-on-nanograv-15-yr-and-epta-dr2-which-nulls-survive-real-pta-sampling-pre-registered-replication/1592>\
**Category:** Data Analysis\
**Created:** [September 26, 2026, 11:41am UTC](https://ask.igwn.org/t/residual-level-hd-diagnostics-on-nanograv-15-yr-and-epta-dr2-which-nulls-survive-real-pta-sampling-pre-registered-replication/1592 "2026-09-26T11:41:31Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![Kretski](https://yyz1.discourse-cdn.com/flex031/user_avatar/ask.igwn.org/kretski/32/614_2.png) [@Kretski](https://ask.igwn.org/u/Kretski)\
**Post date:** [September 26, 2026, 11:41am UTC](https://ask.igwn.org/t/residual-level-hd-diagnostics-on-nanograv-15-yr-and-epta-dr2-which-nulls-survive-real-pta-sampling-pre-registered-replication/1592/1 "2026-09-26T11:41:31Z")

</div>

I’ve published a small methods note and code on residual-level cross-pulsar statistics for PTAs, tested on public NANOGrav 15-yr and EPTA DR2 data.

Main finding is a negative one: an omnibus cross-correlation statistic that looked calibrated on idealised simulations (false-positive rate 0.065) gives 0.69 on real NANOGrav sampling and 1.00 on EPTA. Temporal-shift nulls also fail for monopole/angular tests. Only sky-scramble nulls stayed calibrated in all three data sets. Ablations trace the failure to variance that changes at the same dates for many pulsars (backend upgrades, campaign cadence) and to red-noise-dominated series.

A sky-scramble HD test orthogonalised against monopole + the full 6-dim dipole subspace was then pre-registered (frozen, tagged code; calibration gate before any real-data run) and applied to EPTA DR2: p = 0.0135 (DR2full, primary), p = 0.0065 (DR2new, secondary). This recovers the known PTA HD signal with a simple method — it is not a new detection.

Preprint: [Residual-level Hellings–Downs diagnostics on NANOGrav 15-yr and EPTA DR2: calibration failures and a pre-registered replication | Zenodo](https://doi.org/10.5281/zenodo.22976009)  
Code, pre-registration, amendments, logs: [GitHub - Kretski/residual-first-framework: Residual-level Hellings–Downs diagnostics for PTAs: calibration failures and a pre-registered EPTA replication · GitHub](https://github.com/Kretski/residual-first-framework)

Feedback on the calibration analysis especially welcome — e.g. whether others have seen the shared-heteroscedasticity problem with shift-based nulls.

(The code and text were developed with AI assistance, Claude by Anthropic; all runs and decisions are mine and the full commit history is public.)

I’m aware that the PTA collaborations use the noise-marginalised optimal statistic, sky scrambles and joint HD/monopole/dipole fits — this is a much simpler residual-level analogue. The part I think may be useful to others is the calibration record: which nulls fail on real sampling, and why.