With 42 supernovae in 37 galaxies, how do we know SH0ES results is robust?

Question

SH0ES is a collaboration that's anchoring the cosmic distance ladder by matching Cepheid data with Type 1a supernovae. Both of these are standard candles, so if we know the distance to a particular Cepheid and Type 1a supernova in the same galaxy, we can infer the distance to other galaxies where only one standard candle is visible.

My question is about this paragraph in the following Science article:

Source

Adam Riess of Johns Hopkins University and the team he leads, known as SH0ES, have turned this method into a fine art, deriving $H_0$ by measuring hundreds of Cepheids in 37 nearby galaxies that are host to 42 type Ia supernovae and then extending measurements out to 277 more distant supernovae. In 2022, using the Hubble Space Telescope, the team put $H_0$ at $73 km/s/Mpc$ with a 1.4% precision.

With only 42 supernovae in 37 galaxies, it's inevitable that some galaxies have only one supernova. This makes it seem like the supernova cannot be used as a standard candle - even if its intrinsic brightness is known, there could be some systematic effect that blocks some of its light. For example, a Type 1 a supernova on the other side of the Milky Way would necessarily be obscured by the galaxy center (as viewed by us). Some similar dust-like obscuring effect would make the galaxy distance no longer well-estimated. It seems like we need several supernovae in the same galaxy, all of which give the same distance estimate, before we can be confident in the result.

How do we know the SH0ES results are robust?

Answer 1

Lets read the actual publication, Adam G. Riess et al (2022) ApJL 934 L7 to understand what the authors did to account for these possible biases. Notably the authors of the paper write

The SH0ES program has been designed to improve upon past determinations of H0 by (1) extending the range of Cepheid observations with ACS and WFC3 to reach the hosts of a large sample of "ideal" SNe Ia, free from the preceding problems; (2) using near-infrared (NIR) observations of all Cepheids in SN Ia hosts with NICMOS and WFC3 to reduce the systematic uncertainty associated with the reddening laws for Cepheids and their hosts and the Cepheid metallicity dependence; and (3) calibrating Cepheids with new, geometric distances tied directly with HST to the Cepheids in SN Ia hosts to nullify zero-point uncertainties.

and further (emphasis mine)

A number of systematic differences in the first-generation calibration of SNe Ia by Sandage et al. (2006) were quantified by Riess et al. (2005, Table 16). These differences, totaling about 20%, arose from several effects which were amplified by small sample statistics: problematic SN Ia data such as photographic photometry, highly extinguished objects, and poorly sampled light curves;

Thus the main point of this study is to try to un-bias the individual observations from any reddening or extinction by taking into account the spectra or at least observations in several well-calibrated colour filters of both the Cepheids as well as the supernovae. The effects of extinction are wavelength-dependent, so that these situations are directly detectably in spectral data and can subsequently be accounted for in an analysis. The whole section §3.4 describes the de-reddening of the used Cepheids with extensive references itself. The whole chapter §4 deals with other constraints and possible biases, including the description of how the SN data themselves are treated in chapter §4.8 by comparing data from different sources and especially different wavelength for the same SN.

An improvement in the current analysis of calibrator SNe Ia over R16 is our use of multiple SN light-curve data sets for most calibrators, 77 sets in all for 42 SNe Ia, a mean of ∼2 independent sets per SN, reducing measurement errors (not intrinsic scatter, which is covariant among multiple samples of the same SN) by a mean factor of 1.4

All this said, one single observation point does not drive the results, but $H_0$ of course is a fit through the redshift-distance data. Thus the overall result does not (strongly) depend on the error made with a single of these measurements. And the robustness of this fit has been confirmed by simulating many results using a Monte-Carlo method (chapter §5) and in chapter §6 they look at the influence on the results using (partially) different analysis or selection criteria.

So all-in-all it's my impression that this is a well-written paper which details both the data sampling and data treatment in excellent detail with good justification and discussion of the impact of the choices made - so quite worth the read.