The subtle statistics of the distance ladder: On the distance prior and selection effects

Technical Summary: The Subtle Statistics of the Distance Ladder: On the Distance Prior and Selection Effects

Problem Statement
Modern cosmology faces the "Hubble tension," a significant discrepancy ($\gtrsim 5\sigma$) between the Hubble constant ($H_0$) inferred from the local distance ladder and that derived from the Cosmic Microwave Background (CMB) under the $\Lambda$CDM model. While systematic errors in distance indicators are often scrutinized, the statistical methodology underlying distance-ladder analyses—specifically the choice of priors on latent distance parameters and the modeling of selection effects—is rarely considered a significant contributor to this tension. This paper argues that inappropriate statistical modeling in these areas introduces appreciable biases in inferred $H_0$.

Methodology
The authors employ a combination of analytic derivations, mock data simulations, and re-analyses of real-world datasets to investigate two distinct statistical issues:

1. The Distance Prior Effect: Distances are latent parameters inferred from observables (e.g., apparent magnitude and redshift). The authors demonstrate that because objects are intrinsically uniformly distributed in three-dimensional volume, the correct prior for distance $r$ must scale as $\pi(r) \propto r^2$ (the "volume prior"). Many current analyses implicitly assume a flat prior on the distance modulus $\mu$ (where $\mu \propto \ln r$), which corresponds to $\pi(r) \propto 1/r$.

   • Analytic Approach: The authors derive the shift in the maximum a posteriori (MAP) estimate of $H_0$ resulting from different prior exponents $k$ (where $\pi(r) \propto r^k$) in a simplified Bayesian forward model. They show that for a volume-limited sample with negligible redshift uncertainty, the bias in $\ln H_0$ is proportional to the difference in the prior exponent $\Delta k$.
   • Simulation: Markov Chain Monte Carlo (MCMC) simulations using mock data confirm that a uniform-in-volume prior ($k=2$) yields an unbiased posterior for $H_0$, whereas a uniform-in-$\mu$ prior ($k=-1$) produces a significant bias toward lower distances and higher $H_0$.

2. Selection Effects: The authors analyze how selection criteria (volume-limited, redshift-limited, and magnitude-limited) interact with the distance prior within a Bayesian framework.

   • They derive the necessary correction factors for the posterior distribution based on the probability of an object entering the sample.
   • They compare these Bayesian forward-modeling results with frequentist $\chi^2$ estimators, which treat distances as deterministic functions of redshift and lack an explicit prior mechanism.

Key Contributions and Results

• Quantification of Prior Bias: The study quantifies that using a flat prior on distance modulus (common in Bayesian analyses) instead of a volume prior biases $H_0$ high. For a sample with magnitude uncertainty $\sigma_m = 0.1$ mag and $N=2000$ galaxies, the shift between a flat-$\mu$ prior and a volume prior corresponds to a $6.2\sigma$ bias.
• Interplay with Selection:
  • Redshift Selection: In the limit of negligible redshift uncertainty ($\sigma_z \to 0$), a redshift selection factor ($H_0^{3N}$) exactly cancels the volume prior effect if a flat-$\mu$ prior is used. This creates a "fortuitous cancellation" where an unphysical prior yields an unbiased result, but only under specific conditions.
  • Magnitude Selection: For magnitude-limited samples, the selection factor does not cancel the incorrect prior, leading to significant bias unless the volume prior is correctly applied.
  • Frequentist $\chi^2$: The authors show that frequentist methods are unbiased for volume-limited samples only if $\sigma_z \to 0$. When $\sigma_z > 0$, linear error propagation introduces biases that depend on the sign of the correction term.
• Case Study I: CosmicFlows-4 (CF4):
  • Re-analyzing the Tully-Fisher subset of CF4, the authors find that the standard analysis (equivalent to a flat-$\mu$ prior) yields $H_0 \approx 75.1$ km s$^{-1}$ Mpc$^{-1}$.
  • Imposing the correct volume prior ($k=2$) shifts the result to $H_0 \approx 66.9$ km s$^{-1}$ Mpc$^{-1}$, a dramatic $\approx 55\sigma$ shift.
  • The authors note that this shift is illustrative; the true value likely lies between the volume-limited and redshift-selected limits depending on the actual, complex selection function of the CF4 compilation, which is not fully homogeneous.
• Case Study II: SH0ES:
  • Applying the volume prior correction to the SH0ES pipeline (which uses a frequentist approach with simulation-based bias corrections) suggests a shift of $\approx 1.7\sigma$ toward lower $H_0$ (from $73.04$ to $\approx 71.3$ km s$^{-1}$ Mpc$^{-1}$).
  • The authors argue that the SH0ES pipeline's existing simulation-based bias corrections likely already account for these effects, implying the nominal SH0ES result may already be unbiased, provided the simulations accurately model both the intrinsic source distribution and selection cuts.

Significance and Claims
The paper claims that the statistical methodology of distance-ladder studies is a critical, often overlooked component of the Hubble tension. The primary significance lies in demonstrating that:

1. Unbiased Inference Requires Physical Priors: In Bayesian analyses, assuming a flat prior on distance modulus is physically incorrect for a volume-distributed universe and leads to biased $H_0$ estimates unless specific, often unrealistic, selection conditions are met.
2. Selection Modeling is Non-Trivial: Selection effects do not merely "correct" the volume prior; they introduce distinct factors into the posterior. A principled approach requires a generative model that accounts for both the intrinsic volume distribution and the specific selection function.
3. Current Limitations: Many state-of-the-art analyses (both frequentist and Bayesian) neglect the volume prior or treat selection effects inadequately. While simulation-based corrections (as used in SH0ES) can mitigate these biases, they rely on the accuracy of the underlying simulation models.
4. Future Requirements: To achieve precise constraints, future surveys must prioritize samples with known, homogeneous selection criteria. This allows for the implementation of principled Bayesian forward models or accurate post-hoc simulation-based corrections.

The authors conclude that without accurate modeling of both distances (via the correct volume prior) and selection effects, inferred cosmological parameters like $H_0$ and peculiar velocities may suffer from systematic biases that scale with sample size, potentially exacerbating or obscuring the Hubble tension.