West by Midwest: Package Maintenance, Transition and Handover

May 14, 2026

Emmeans Background

• For estimating marginal means from a linear modeling object
• Legacy package: first released November 5, 2017
  
• replacing lsmeans’ (first released August 15, 2012)
  
• methods based on Searle, Speed, and Milliken (1980) Population marginal means in the linear model: An alternative to least squares means, The American Statistician 34(4), 216-221 doi:10.1080/00031305.1980.10483031.

West by Midwest: Where We Are

West by Midwest: Where We’ve Lived

Russ’s Computing Background

• Skeletran, Fortran IV
• Algorithms for noncentral \(F\), beta, and \(t\)
• Power and sample size
  • Pascal
  • Xlisp-stat
  • Java
• Other R packages: rsm, lsmeans, estimability, vigindex

Origins of ‘lsmeans’

• Preliminary thoughts about how to obtain LSmeans focused on interpretation of model coefficients
  • What to do with covariates?
  • Doesn’t help that SAS Type III estimable functions do this wrong (IMHO)
• Salary equity analysis for administration
  • Focus on what model predicts
  • Duh! Those are the linear functions of coefficients that I need

Differing Application Perspectives

• LSMs use equal weighting
• Makes sense to experimenters
• Others who deal with observational data, not so much
• Alernative weighting: Proportional, Outer, Flat
• Causal effects: G computation (counterfactuals)
• Other packages or platforms: Stata, effects, marginaleffects, margins

emmeans (2017–)

• Re-structure of lsmeans with improved architecture and class structure, customizability
• Expanding to generalized linear models, link functions and response transformations
• Version 2.0.3 has 45 source files (17,500 lines)
  • 9,700 lines of R code
  • 7,800 lines of comments or embedded documentation
• 243 functions (79 public), 16 vignettes
• 121 reverse depends/imports/suggests/enhances

Getting into R package development

• My first R package was rsm
• I had existing code that worked, and used package.skeleton() function (still available) to bundle it up
  • But this required writing separate Rd files, etc.
• Now the way to go is to use RStudio (or Positron) – free download – and create a new project
• Use roxygen2 package to incorporate documentation, imports, etc.
• Wickham, H and Bryan, J (2023) R Packages (2d ed), O’Reilly Media, Inc.

Observations on package development

• Writing code is fun
• Writing documentation & vignettes is less fun, and time-consuming. But really important
• Your stuff is there for others to see – and they will look

My pet peeves with some other people’s packages

• Cramped,runtogethercode. Use whitespace
• Over-abstraction. Use meaningful (but short) variable names
• Re-using the same variable or object name in a string of examples

Bug reports

• Same peeves, also unnecessary elaboration

Transition experience

• Julia …
  • encouraged me to add pkgdown site (July ’24)
  • set up GitHub action to update site automatically
  • started monitoring the Issues page
  • began answering some issues
  • major overhaul graphics theme
  • took over as official maintainer (August ’25, version 2.0.0)

Transitioning = Stress of change + Benefits

• ‘pkgdown’ as mentioned before
• Different work habits & availability
• Repo management – working in branches
• Coding style (e.g., I like “=” rather than “<-”)
  • but similar styles with whitespace and indentation
• Joint responsibility for decisions (e.g., no “stars”)

Programming Background

• Originally trained in SAS
  
• Started using R in 2010 when I was unable to access all the predictions from PROC PLS (SAS v9.3)
  
• R in those days was hard (no tidyverse, ggplotting, or RStudio), but it had enough functionality and flexibility to make it worth the effort.
  
• Learned C, python, bash shell scripting. Programming is satisfying!
• I pay attention to developements in the R ecosystem and read source code frequently.

Programming Philosophy

• I have an interest in clean code: functions with a clear (single) purpose, good error catching and informative messages. Don’t go overboard on functionality (a function is not a multi-tool)
• Style: code needs to breathe! We aren’t bit packing code and sending it to the moon, so use whitespace. Indent code (to indicate function nesting) because it’s easier to read and understand
• Use existing libraries for package enhancement (e.g. ‘[pbkrtest(https://CRAN.R-project.org/package=pbkrtest )]’ for Kenward-Rogers degrees of freedom), but try to limit software dependencies that are required for yours to run.

emmeans::cld()

Daily CRAN Downloads

More Usage/Popularity Stats

• Downloads: 6,000 daily downloads
• Reverse dependencies: 207
• Code Mentions on GitHub: 21,600
• Citations: ?????
• GitHub Stars: 414
• Contributors: 17
• Issues filed: 532

There is Not Much Guidance on Library Handoff

How Did It Start? (Julia)

1. I filed an issue in 2024 (to make ‘pkgdown’ website) and assisted with implementing it, realized Russ was 77 years old, and inquired about package succession.
2. I did not contribute much for the next year; I answered issues, and looked more into package functionality.
3. I became more involved starting last August with some encouragement from Russ (“Could you address this…?”)

Challenges: Breaking Dependencies

Internal or external

• Some issues are complex. The problem might be straightforward to describe, but what it is causing it is not easy to resolve (example.
  
• Fixing one issue may introduce new problems in the package (example)
  
• Many packages depend on emmeans, so every change has to be weighed with that in mind. What if I break something?
  
• We have a script to check for breaking dependencies. But be careful, if something is broken downstream, that will be our problem.

Collaboration

Russ is #14, I am #2

Collaboration

• We are collaborating two levels: package maintenance and everyday git issues
  
• Provide transition tools: guides to package structure, how to prepare for CRAN
  
• Clarify expectations: who is addressing what issue and how?
  
• Need clear two-way communication (we email each other frequently)
  
• Make changes as pull requests (PR’s) and request code review
  
• Pave the way to easy “wins”
  
• Patience, kindness, and encouragement will go a long ways
  

How to Get Involved

The open source software universe needs you!

• Some authors struggle to keep up with demand (see the 200+ issues currently open in glmmTMB)
• Popular packages are abandoned regularly when the original author no longer has time. Any package ever on CRAN is available in a public archive.
• There are many opportunities; ROpenSci runs a ‘Help Wanted’ page with requests for help in solving issues or looking for new maintainers

Why SAS Type III Anova is wrong

(when covariates are involved)

library(emmeans)
fiber.add = lm(strength ~ diameter + machine, data = fiber) # why have this? 
fiber.int = lm(strength ~ diameter * machine, data = fiber)

emmeans(fiber.add, "machine") |> contrast("consec") |> test(joint = TRUE)

 df1 df2 F.ratio p.value
   2  11   2.611  0.1181

emmeans(fiber.int, "machine") |> contrast("consec") |> test(joint = TRUE)

 df1 df2 F.ratio p.value
   2   9   2.814  0.1124

# What SAS type III ANOVA does
emmeans(fiber.int, "machine", at = list(diameter = 0)) |> 
    contrast("consec") |> test(joint = TRUE)

 df1 df2 F.ratio p.value
   2   9   0.475  0.6367

SAS compares intercepts, not LSmeans