RougeLM

“Equations from District 7. Linear models for people the algorithm forgot.”

RougeLM is an R package providing datasets and utilities for teaching linear models in the context of the BetaBit StatPunk universe — a cyberpunk dystopia set in WaszKrak megapolis, 2047, where every citizen is reduced to a number by the LifeCalc algorithm.

The package is a companion to the book “Equations from District 7: A Practical Guide to Linear Models”.

Installation

# From GitHub (development version)
# install.packages("remotes")
remotes::install_github("BetaAndBit/RougeLM")

Datasets

Dataset	Description	n	Variables
`lifecalc`	LifeCalc social scoring — multicollinearity & regularisation	5,000	25

`lifecalc`

A simulated dataset of 5,000 WaszKrak residents scored by the fictional LifeCalc algorithm. The outcome variable SocialScore is generated by a sparse true model:

SocialScore ≈ 0.40 × DistrictScore
             − 16   × prior_flag
             − 3.5  × gender
             + 0.18 × EmploymentScore
             + 0.12 × EducationScore
             + 0.10 × MedicalScore
             + ...
             + ε

The remaining 17 variables are correlated proxies organised into four clusters — education, employment, health, behaviour — that introduce structured multicollinearity. The dataset is designed for:

Demonstrating VIF and the cost of multicollinearity
Comparing AIC and BIC for model selection
Showing how LASSO recovers the sparse true model
Contrasting LASSO and Ridge regularisation

Usage

library(RougeLM)
library(glmnet)

data(lifecalc)

# 1. Check correlation clusters
lifecalc_cor_clusters(lifecalc)

# 2. Full OLS — inflated standard errors from multicollinearity
model_full <- lm(SocialScore ~ ., data = lifecalc)
summary(model_full)

# 3. VIF diagnostic
lifecalc_vif(lifecalc)

# 4. LASSO — recover the sparse true model
X  <- model.matrix(SocialScore ~ ., data = lifecalc)[, -1]
y  <- lifecalc$SocialScore
cv <- cv.glmnet(X, y, alpha = 1, nfolds = 10)

# Which variables survive?
coef(cv, s = "lambda.min")

# 5. Ridge — all variables retained, coefficients shrunk
cv_ridge <- cv.glmnet(X, y, alpha = 0, nfolds = 10)
coef(cv_ridge, s = "lambda.min")

Variable clusters

DistrictScore  ──────────────────────────────────────────────── SocialScore
prior_flag     ──────────────────────────────────────────────── (dominant)
               │
               ├── Education: EducationScore, LiteracyScore,
               │              CuriosityScore, VerificationScore
               │
               ├── Employment: EmploymentScore, NetworkScore,
               │               ConsumptionScore, MobilityScore
               │
               ├── Health: MedicalScore, SleepScore, RecoveryScore,
               │           NutritionScore, StressIndex,
               │           GeneticRiskScore, ChronicLoadScore
               │
               └── Behaviour: ComplianceScore, NarrativeScore,
                              RoutineScore, AttentionScore,
                              DisplacementScore

Functions

Function	Description
`generate_lifecalc(n, seed)`	Regenerate the dataset with custom size and seed
`lifecalc_cor_clusters(data)`	Tidy correlation summary by cluster
`lifecalc_vif(data)`	VIF table for the full OLS model

The BetaBit universe

Beta and Bit are statistical Robin Hoods in a world where data is power.

“We’re poor but we have something most data scientists don’t.” “What?” “The ability to look in a mirror without disgust.”

The lifecalc dataset recreates the data they spent a year collecting — one subscore at a time — before Ghost gave them the full model and they finally understood what they had been looking at.

prior_flag coefficient: −16. The most important feature in the model. Set before you were born. Non-reversible.