Count Data Models

Use this page as a cautious guide to count-style outcomes and their modeling needs, while keeping the current product support boundary explicit.

What this method is

This page is a concept guide for outcomes that behave like counts rather than ordinary continuous measurements.

Count outcomes often have features that make plain OLS interpretations uncomfortable. They may be bounded below by zero, may cluster around small integers, and may show variance patterns that do not behave like a simple continuous response.

The current repo does not confirm one dedicated end-to-end count-model branch, so this page should be read as guidance rather than as a confirmed implementation manual.

When to use or avoid

Use this page when:

  • your outcome is a count of events, occurrences, or items
  • zero values matter substantively
  • the response distribution is clearly not behaving like an unrestricted continuous variable

Avoid reading this page as proof that the product already exposes every count-data method directly.

  • Do not assume Poisson or negative-binomial modeling is available now just because the page exists.
  • Do not use this page to claim that a dedicated zero-inflation workflow is already implemented.

Required inputs

Conceptually, count models usually need:

  • a response variable that represents counts
  • predictors or grouping variables suitable for the scientific question
  • enough context to judge whether overdispersion or excess zeros may matter

The current public product contract does not yet confirm a dedicated count-model input surface.

Outputs

No dedicated count-model output schema was confirmed in this pass.

The safest current framing is:

  • this page helps readers understand when count-style modeling may be more appropriate than a simple linear reading
  • dedicated public outputs for Poisson, negative-binomial, or zero-inflated models are not yet confirmed here

No dedicated count-model figure family was confirmed in this pass.

Interpretation notes

  • Count data often call for a different modeling mindset than continuous outcomes.
  • Poisson, negative-binomial, and zero-inflated approaches are useful concepts to understand here, even when the current product boundary is still incomplete.
  • The current repo evidence is stronger for describing what is not yet confirmed than for documenting a stable count-model feature path.

Limits

  • The current GLMM public route confirms gaussian and binomial, but not a dedicated count family such as Poisson.
  • No end-to-end count-model route, output contract, or tests were confirmed in this pass.
  • This page should stay conceptual until stronger implementation evidence exists.

TODO (Phase02+)

  • Confirm dedicated count-model support before making method-level implementation claims.