Why bossanova?

Confusion arises not from a disagreement about core beliefs, but from the absence of a shared language for discussing tools.

Motivation¶

Bossa Nova emerged in 1950s Rio de Janeiro as a fusion of samba and jazz—a synthesis of Brazilian and North American musical traditions that became something entirely its own. When critics dismissed it as Americanized imitation, composer Antônio Carlos Jobim countered remarked how this synthesis had become “the biggest influence on American music in the last thirty years.”

In a similar spirit bossanova aims to bridge two traditionally distinct statistical cultures Breiman, 2001 Shmueli, 2010:

1. Explanation modeling culture assumes data comes from a stochastic model. The goal is to estimate parameters and test hypotheses. “Is the effect significant?”

2. Prediction modeling culture treats the data-generating mechanism as unknown. The goal is prediction. “How accurately can we forecast new observations?”

Most experimental social science training grew from the first culture, while much modern machine learning grew from the latter Yarkoni & Westfall, 2017. And the tools we use reflect this divide: R grew from statisticians asking “what are the effects?”; scikit-learn grew from computer scientists asking “what’s the test accuracy?”

But the division is artificial. Researchers need both. A neuroscientist fitting a mixed-effects model wants to know whether an effect is “real” and whether it generalizes beyond the current sample. A data scientist building a classifier wants predictive accuracy and some understanding of why the model works. The practical consequence is tooling fragmentation. A typical analysis might require (in Python):

• One package for mixed models (e.g. pymer4 Jolly, 2018)

• Another for machine-learning style modeling (sklearn)

• A third for time-series and marginal effects (statsmodels)

• etc

Each package has its own conventions, data structures, and mental models adding to the cognitive overhead and distracting from the goal: what is my model telling me?

A Cultural Bridge¶

After fitting any model, you can ask three fundamentally different questions:

The third question—“What would happen if...?”—is where the cultures meet. This is the domain of marginal effects Searle et al., 1980Lenth, 2016Arel-Bundock et al., 2024: the model’s implications under controlled changes.

“What if?” questions have deep roots. Judea Pearl’s Ladder of Causation Pearl & Mackenzie, 2018 distinguishes three levels of reasoning:

1. Association: “What do I expect to see?” — Observational data, correlations

2. Intervention: “What if I do X?” — The do(X=x) operator Pearl, 1995Pearl, 2009

3. Counterfactuals: “What if I had done X differently?” — Individual-level potential outcomes Rubin, 1974

Traditional regression lives on Rung 1—it describes patterns in data. Causal inference Hernán & Robins, 2020 asks how to climb higher: when can we interpret associations as interventions?

Marginal effects occupy a productive middle ground. Computationally, they answer: “What does the model predict under counterfactual covariate values?” This resembles interventional reasoning but remains a Rung 1 calculation—a summary of the fitted model, not a causal claim. Whether the resemblance is causally meaningful depends entirely on assumptions the model cannot verify: no unmeasured confounding, correct functional form, no selection bias. These require domain knowledge to evaluate Shmueli, 2010.

The value of marginal effects isn’t that they automatically provide causal answers. It’s that they phrase questions in substantively meaningful terms. Instead of “the coefficient is 0.3,” you ask “what’s the predicted difference between treatment and control, averaging over the covariate distribution?” or “what would happen if the treatment was twice as strong?” This reframing connects statistical output to the counterfactual questions researchers actually care about Harrell, 2015.

A Bicycle for the Mind¶

Richard McElreath describes statistical models as golems—powerful constructs that do exactly what you tell them, whether or not that’s what you actually wanted McElreath, 2020. A golem has no intent; it follows instructions. If your instructions are wrong, the golem will faithfully execute them. But as engineers we can help users wrangle their golem a bit more easily.

Take the simple Wilkinson formula syntax familiar to R users Wilkinson & Rogers, 1973Chambers & Hastie, 1992: y ~ x1 * x2. While simple to understand for beginners its strength lies in its declarative specification of structure, separate from the how of linear algebra and statistical inference. bossanova extends this idea with a handful of grammar-like methods Wickham, 2010 that encompass the full combinatorial space of all the ways you might use a model as a tool:

1. Formulas: declare what your model is

2. Methods: declare what question you want to ask

• simulate(): given a formula, generate data under certain assumptions

• fit(): given a formula and data, estimate parameters and quantify fit

• predict(): given new data, quantify generalization

• explore(): given a fitted model, ask questions as marginal effects

3. Inference: declare how you want to answer it

• infer(): is chained after any of the methods above to quantify uncertainty around estimation and generalization

The goal isn’t to hide complexity behind “smart” defaults that guess your intent, but make the golem’s behavior legible to help you interpret what you’re asking for. Statistical jargon creates unnecessary confusion. The terms “fixed” and “random” effects have at least five conflicting definitions across statistics, econometrics, and machine learning Gelman, 2005Gelman & Hill, 2007. bossanova follows Gelman’s recommendation: we use varying for effects that differ across groups, avoiding terminology that means different things to different communities.

Conclusion¶

Bossa Nova roughly means “new trend” or “new wave” in Portuguese. The musical genre emerged from a desire to create something fresh by synthesizing existing traditions—not rejecting them, but building on them with a different sensibility. bossanova aspires to the same: a synthesis of statistical cultures, explanation and prediction, that becomes a tool for exploration.