Features

Overview

All pymer4 models operate on long-format pandas dataframes. These dataframes should contain columns for a dependent variable, independent variable(s), and optionally a column for a group/cluster identifiers.

Currently, pymer4 contains 3 different model classes:

  • Lm for ordinary-least-squares and weighted-least-squares regression optionally with robust standard errors

  • Lmer for multi-level models estimated using glmer() in R.

  • Lm2 for two-stage ordinary-least-squares in which a separate Lm model is fit to every group/cluster and inference is performed on the coefficients across all groups/clusters. This is also known as the “summary statistics approach” and is an alternative to multi-level models estimated using Lmer, which implicitly allow for both random-intercepts and random-slopes but shares no information across each groups/clusters to help during estimation.

Standard regression models

Lm models which are equivalent to lm() in R with the following additional features:

  • Automatic inclusion of confidence intervals in model output

  • Optional empirically bootstrapped 95% confidence intervals

  • Cluster-robust, heteroscedasticity-robust or auto-correlation-robust, ‘sandwich estimators’ for standard errors (note: these are not the same as auto-regressive models)

  • Weighted-least-squares models (experimental)

  • Permutation tests on model parameters

Multi-level models

Lmer models which are equivalent to glmer() in R with the following additional features:

  • Automatic inclusion of p-values in model output using lmerTest

  • Automatic inclusion of confidence intervals in model output

  • Automatic conversion and calculation of odds-ratios and probabilities for logit models

  • Easy access to group/cluster fixed and random effects as pandas dataframes

  • Random effects plotting using seaborn

  • Easy post-hoc tests with multiple-comparisons correction via emmeans

  • Easy model predictions on new data

  • Easy generation of new data from a fitted model

  • Optional permuted p-value computation via within cluster permutation testing (experimental)

  • note that Lmer’s usage of coef, fixef, and ranef differs a bit from R:

  • coef = summary(model) in R, i.e. “top level” estimates, i.e. the summary output of the model that can be used to make predictions on new datasets and on which inference (i.e. p-values) are computed

  • fixef = coef(model) in R, i.e. “group/cluster” level fixed effects, conceptually similar to coefficients obtained from running a seperate Lm (lm in R) for each group/cluster

  • ranef = ranef(model) in R, i.e. “group/cluster” level random effects, deviance of each cluster with respect to “top level” estimates

Other Features

  • Highly customizable functions for simulating data useful for standard regression models and multi-level models

  • Convenience methods for plotting model estimates, including random-effects terms in multi-level models

  • Statistics functions for effect-size computation, permutations of various 1 and 2 sample tests, bootstrapping of various 1 and 2 sample tests, and two-one-sided equivalence tests