Measurement errors and scaling relations in astrophysics: a review

This review article considers some of the most common methods used in astronomy for regressing one quantity against another in order to estimate the model parameters or to predict an observationally expensive quantity using trends between object values. These methods have to tackle some of the awkward features prevalent in astronomical data, namely heteroscedastic (point-dependent) errors, intrinsic scatter, non-ignorable data collection and selection effects, data structure and non-uniform population (often called Malmquist bias), non-Gaussian data, outliers and mixtures of regressions. We outline how least square fits, weighted least squares methods, Maximum Likelihood, survival analysis, and Bayesian methods have been applied in the astrophysics literature when one or more of these features is present. In particular we concentrate on errors-in-variables regression and we advocate Bayesian techniques.

💡 Research Summary

This review article provides a comprehensive synthesis of regression techniques that have been adopted in astrophysics to relate one observable quantity to another, especially when the data exhibit the peculiarities typical of astronomical surveys. The authors begin by enumerating the challenges that distinguish astronomical datasets from those encountered in many other fields: heteroscedastic measurement uncertainties that vary from object to object, intrinsic astrophysical scatter that cannot be attributed to observational error, non‑ignorable selection effects (including Malmquist bias), non‑Gaussian error distributions, censored or truncated observations (upper limits), outliers, and cases where multiple physical populations are mixed together. These complications render naïve ordinary least‑squares (OLS) regression inadequate, as OLS assumes homoscedastic, Gaussian errors and no intrinsic scatter.

The first methodological section revisits OLS and weighted least‑squares (WLS). While WLS can accommodate point‑dependent variances by assigning each datum a weight equal to the inverse of its variance, the authors stress that this approach still relies on accurate knowledge of the variance and on the assumption that the errors are independent and normally distributed. In many astronomical contexts, the variance itself is estimated from the data, leading to potential bias.

The review then turns to errors‑in‑variables (EIV) regression, where uncertainties affect both the independent and dependent variables. By formulating a latent “true” value for each object and expressing the observed quantities as the sum of the latent value and a measurement error term, the problem becomes amenable to maximum‑likelihood estimation (MLE). The authors derive the likelihood function for a linear EIV model, discuss extensions to non‑linear (e.g., log‑log) relations, and outline numerical strategies such as the Expectation‑Maximization (EM) algorithm and Newton‑Raphson optimization. They also address the incorporation of intrinsic scatter as an additional variance component, thereby separating observational noise from genuine astrophysical dispersion.

Censored data, common when a source falls below a detection threshold, are handled through survival analysis techniques. The review explains how the Kaplan‑Meier estimator can be used to reconstruct the underlying distribution in the presence of upper limits, and how Cox proportional‑hazards models or Tobit regression can be employed when a parametric form for the relationship is assumed. By explicitly modeling the detection process, these methods avoid the systematic over‑estimation of slopes that would arise from discarding non‑detections.

The centerpiece of the article is a detailed exposition of Bayesian regression. The authors argue that Bayesian hierarchical models provide a natural and unified framework for simultaneously treating heteroscedastic measurement errors, intrinsic scatter, and selection biases. Priors encode external astrophysical knowledge (e.g., theoretical scaling‑law expectations) and can be chosen to be weakly informative when such knowledge is limited. Posterior inference is performed via Markov Chain Monte Carlo (MCMC) sampling, with practical examples implemented in Stan, PyMC3, and the emcee package. The Bayesian EIV formulation treats the true latent variables as additional parameters, allowing the model to propagate uncertainty from the measurement stage all the way to the final scaling‑relation coefficients. Moreover, the hierarchical structure can include a population‑level model for the distribution of true values, thereby correcting for Malmquist bias in a principled way. Sensitivity analyses demonstrate how different prior choices affect the posterior, and diagnostic tools (e.g., posterior predictive checks) are recommended to assess model adequacy.

To illustrate the practical impact of each technique, the review surveys several astrophysical case studies: the Tully‑Fisher relation for spiral galaxies, the mass‑luminosity relation for galaxy clusters, the period‑luminosity relation for Cepheid variables, and the X‑ray temperature‑luminosity scaling for galaxy clusters. In each case, the authors compare results obtained with OLS/WLS, MLE‑based EIV, survival‑analysis‑adjusted fits, and full Bayesian hierarchical models. The consensus is that while point estimates (e.g., slopes) are often similar across methods, the uncertainty quantification and bias correction differ markedly. Bayesian models consistently yield broader, more realistic credible intervals that incorporate both measurement error and intrinsic scatter, and they explicitly correct for selection effects that would otherwise skew the inferred scaling laws.

In the concluding section, the authors summarize the trade‑offs: OLS/WLS are computationally cheap but inadequate for complex error structures; MLE/EIV offers a solid frequentist alternative but can be numerically demanding and sensitive to variance estimates; survival analysis is indispensable when dealing with censored data; Bayesian hierarchical modeling is the most flexible and statistically rigorous, at the cost of higher computational load and the need for careful prior specification. The review suggests future directions such as hybrid approaches that combine Bayesian priors with MLE optimization, and the integration of machine‑learning techniques (e.g., Gaussian processes) to capture non‑linearities while still respecting the full error model. Overall, the paper serves as a valuable guide for astronomers seeking to extract reliable scaling relations from data plagued by the myriad imperfections inherent to observational astrophysics.