Parameter Recovery Study on IZI -- a Bayesian Analysis Tool for Emission Lines from H II Regions and Star-forming Galaxies

TL;DR

This study evaluates the accuracy of the Bayesian tool IZI in estimating emission line parameters from H II regions and star-forming galaxies, highlighting biases caused by model errors and providing guidelines for proper use.

Contribution

The paper systematically tests IZI's parameter recovery capabilities, identifies sources of bias, and offers practical recommendations and scripts for improved application.

Findings

IZI tends to underestimate or overestimate parameters by about 1-sigma when model errors are included.

Model error in the likelihood term causes biased parameter estimates, especially when it varies with parameters.

Recommendations are provided for users to perform parameter recovery tests tailored to their data.

Abstract

We present a series of parameter recovery test results of the Bayesian analysis tool IZI, which analyses emission lines from H II regions and star-forming galaxies and returns the estimates of the gas metallicity 12 + log (O/H), ionisation parameter log q, and nebular emission-line colour excess E(B-V). We created several mock datasets using IZI to represent a few different ideal or realistic datasets and performed parameter estimation on the mock data with IZI. We found that IZI underestimated or overestimated the parameters by approximately 1-$\sigma$ or greater when the model error was included, even when using all emission lines available in the model grids. We strongly recommend that IZI users run parameter recovery tests adjusted for their data before interpreting the IZI estimates. To encourage the appropriate use of IZI, we also share a script for parameter recovery tests. The cause of IZI's biased estimation is the substantial model error in the likelihood term, which varies with the model parameters. We thus note that any parameter estimation with a substantial, varying model error in the likelihood term could return biased estimates for the model parameters, such as in the case of NebulaBayes, another Bayesian analysis tool for photoionisation emission lines. We also note two issues relevant to setting the log q prior using the observed line ratio [S III]$\lambda$$\lambda$9068,9532/[S II]$\lambda$6717,6731 (the mismatch of line flux terms and violation of Bayes' theorem) and propose a way to avoid the issues.