Bayesian Regression Analysis: population vs phoneme inventory size

The assocaite between the two variables: population size vs phoneme inventory size has been critically discussed in

Steven Moran, Daniel McCloy, and Richard Wright. Revisiting the population vs phoneme inventory correlation. Language, 88(4):877–893, 2012

This regression analysis employs a Bayesian approached implemented with R and Stan（an R interface to Stan in rethinking package).

Bayesian data analysis: count all the ways data can happen, accoding to assumptions. Assumptions with more ways that are consistent with data are more plausible.

Analysis

The analysis fits data to the following models:

• a simple linear regression with log(population size) as independent variable and log(sound inventory size) as dependent variable

• a hierarchical model with language families as random effect

• a hierarchical model with continents as random effect

• a hierarchical model with both language families and continents as random effect

Then the analysis are then repeated with a Poisson regression instead of a linear regression, with untransformed number of phonemes as dependent variable.

• Posterior distributions are estimated using Stan.

• A proper prior is determined through predictive checks.

• Models are compared via WAIC.

Code

• View the code

• View on Kaggle

• View Jupyter notebook