Selection Effects in Periodic X-ray Data from Maximizing Detection Statistics

TL;DR

Selecting data to maximize pulsation detection statistics can introduce biases in parameter estimates and significance levels, affecting the interpretation of NICER pulsar observations.

Contribution

This paper demonstrates that the data selection procedure used in pulsation searches can bias results and alter significance estimates, highlighting the need for careful analysis.

Findings

Biases in mean count rate and pulsation amplitude due to data selection.

Altered null-distribution of the H-test affects significance estimates.

Estimated biases can impact neutron star radius constraints.

Abstract

The Neutron Star Interior Composition Explorer (NICER) records data of exceptional quality on the energy-dependent X-ray pulse profile of pulsars. However, in searching for evidence of pulsations, Guillot et al. (2019) introduce a procedure to select an ordered subset of the data that maximizes a detection statistic (the H-test). I show that this procedure can degrade subsequent analyses using an idealized model with a stationary expected count rates from both noise and signal. Specifically, the data-selection procedure biases the inferred mean count rate to be too low, biases the inferred pulsation amplitude to be too high, and that the size of these biases scales strongly with the amount of data that is rejected and the true signal amplitude. The procedure also alters the null-distribution of the H-test rendering nominal detection significance estimates overly optimistic. While the idealized model does not capture all the complexities of real NICER data, it suggests that these biases could be important for NICER's observations of J0740+6620 and other faint pulsars (observations of J0030+0451 are likely less affected). I estimate that these selection effects may introduce a bias of $\mathcal{O}(10\%)$ on average in the inferred modulation depth of lightcurves like J0740+6620's, and may be as large as $\mathcal{O}(50\%)$ for fainter pulsars. However, the change for a single dataset like J0740+6620 is expected to be a shift between $-5\%$ and $+20\%$. This could imply that the lower limit on J0740+6620's radius is slightly larger than it should be, although preliminary investigations suggest the change in the radius constraints are $\mathcal{O}(1\%)$ with real J0740+6620 data.