Explain Statistical Output with Large Language Models • statlingua

The statlingua R package is designed to help bridge the gap between complex statistical outputs and clear, human-readable explanations. By leveraging the power of Large Language Models (LLMs), statlingua helps you effortlessly translate the dense jargon of statistical models—coefficients, p-values, model fit indices, and more—into straightforward, context-aware natural language.

Whether you’re a student grappling with new statistical concepts, a researcher needing to communicate findings to a broader audience, or a data scientist looking to quickly draft reports, statlingua makes your statistical journey smoother and more accessible.

Why statlingua?

Statistical models are powerful, but their outputs can be intimidating. statlingua empowers you to:

Democratize Understanding: Make complex analyses accessible to individuals with varying levels of statistical expertise.
Enhance Learning & Education: Students can gain a deeper intuition for model outputs, connecting theory to practical application. Use it as an interactive learning aid to demystify statistical concepts.
Foster Interdisciplinary Collaboration: Researchers from diverse fields can more easily interpret and discuss analytical results, leading to richer insights.
Streamline Reporting & Consulting: Quickly generate initial drafts of interpretations for reports and presentations, saving time and ensuring clarity for clients or stakeholders.
Drive Data-Informed Decisions: Business professionals can better grasp statistical findings, enabling more confident data-driven decision-making without needing to become statistical experts themselves.
Accelerate Prototyping & Exploration: Rapidly understand model summaries during iterative data exploration, allowing for faster assessment and refinement of analyses.

By providing clear and contextualized explanations, statlingua helps you focus on the implications of your findings rather than getting bogged down in technical minutiae.

Supported Models

As of now, statlingua explicitly supports a variety of common statistical models in R, including:

Objects of class "htest" (e.g., from t.test(), prop.test()).
Linear models (lm()) and Generalized Linear Models (glm()).
Linear and Generalized Linear Mixed-Effects Models from packages nlme (lme()) and lme4 (lmer(), glmer()).
Generalized Additive Models (gam() from package mgcv).
Survival Regression Models (survreg(), coxph() from package survival).
Proportional Odds Logistic Regression (polr() from package MASS).
Decision Trees (rpart() from package rpart).
…and more, with a robust default method for other model types!

Installation

statlingua is not yet on CRAN, but you can install the development version from GitHub:

if (!requireNamespace("remotes")) {
  install.packages("remotes")
}
remotes::install_github("bgreenwell/statlingua")

You’ll also need to install the ellmer package, which you can obtain from CRAN:

install.packages("ellmer")  # >= 0.2.0

API Key Setup & ellmer Dependency

statlingua doesn’t directly handle API keys or LLM communication. It acts as a sophisticated prompt engineering toolkit that prepares inputs and then passes them to ellmer. The ellmer package is responsible for interfacing with various LLM providers (e.g., OpenAI, Google AI Studio, Anthropic).

Please refer to the ellmer package documentation for detailed instructions on:

Setting up API keys (usually as environment variables like OPENAI_API_KEY, GEMINI_API_KEY, etc.).
Specifying different LLM models and providers.
Other configuration and model parameter options.

Once ellmer is installed and has access to an LLM provider, statlingua will seamlessly leverage that connection.

Quick Example: Explaining a Linear Model

# Ensure you have an appropriate API key set up first!
# Sys.setenv(GEMINI_API_KEY = "<YOUR_API_KEY_HERE>") 

library(statlingua)

# Fit a polynomial regression model
fm_cars <- lm(dist ~ poly(speed, degree = 2), data = cars)
summary(fm_cars)

# Define some context (highly recommended!)
cars_context <- "
This model analyzes the 'cars' dataset from the 1920s. Variables include:
  * 'dist' - The distance (in feet) taken to stop.
  * 'speed' - The speed of the car (in mph).
We want to understand how speed affects stopping distance in the model.
"

# Establish connection to an LLM provider (in this case, Google Gemini)
client <- ellmer::chat_google_gemini(echo = "none")  # defaults to gemini-2.0-flash

# Get an explanation
explain(
  fm_cars,                 # model for LLM to interpret/explain
  client = client,         # connection to LLM provider
  context = cars_context,  # additional context for LLM to consider
  audience = "student",    # target audience
  verbosity = "detailed",  # level of detail
  style = "markdown"       # output style
)

# Ask a follow-up question
client$chat(
  "How can I construct confidence intervals for each coefficient in the model?"
)

For more examples, including output, see the introductory vignette.

Extending statlingua to Support New Models

One of statlingua’s core strengths is its extensibility. You can add or customize support for new statistical model types by crafting specific prompt components. The system prompt sent to the LLM is dynamically assembled from several markdown files located in the inst/prompts/ directory of the package.

The main function explain() uses S3 dispatch. When explain(my_model_object, ...) is called, R looks for a method like explain.class_of_my_model_object(). If not found, explain.default() is used.

Prompt Directory Structure

The prompts are organized as follows within inst/prompts/:

common/: Contains base prompts applicable to all models.
- role_base.md: Defines the fundamental role of the LLM.
- caution.md: A general cautionary note appended to explanations.
audience/: Markdown files for different target audiences (e.g., novice.md, researcher.md). The filename (e.g., “novice”) matches the audience argument in explain().
verbosity/: Markdown files for different verbosity levels (e.g., brief.md, detailed.md). The filename matches the verbosity argument.
style/: Markdown files defining the output format (e.g., markdown.md, json.md). The filename matches the style argument.
models/<model_class_name>/: Directory for model-specific prompts. <model_class_name> should correspond to the R class of the statistical object (e.g., “lm”, “glm”, “htest”).
- instructions.md: The primary instructions for explaining this specific model type. This tells the LLM what to look for in the model output, how to interpret it, and what assumptions to discuss.
- role_specific.md (Optional): Additional role details specific to this model type, augmenting common/role_base.md.

Example: Adding Support for `vglm` from the `VGAM` package

Let’s imagine you want to add dedicated support for vglm (Vector Generalized Linear Models) objects from the VGAM package.

Create New Prompt Files: You would create a new directory inst/prompts/models/vglm/. Inside this directory, you’d add:
- inst/prompts/models/vglm/instructions.md: This file will contain the detailed instructions for the LLM on how to interpret vglm objects. You’d detail what aspects of summary(vglm_object) are important, how to discuss coefficients (potentially for multiple linear predictors), link functions, model fit statistics specific to vglm, and relevant assumptions.
```
    You are explaining a **Vector Generalized Linear Model (VGLM)** (from `VGAM::vglm()`).

**Core Concepts & Purpose:**
VGLMs are highly flexible, extending GLMs to handle multiple linear predictors and a wider array of distributions and link functions, including multivariate responses.
Identify the **Family** (e.g., multinomial, cumulative) and **Link functions**.

**Interpretation:**
* **Coefficients:** Explain for each linear predictor. Pay attention to link functions (e.g., log odds, log relative risk). Clearly state reference categories.
* **Model Fit:** Discuss deviance, AIC, etc.
* **Assumptions:** Mention relevant assumptions.
```
- inst/prompts/models/vglm/role_specific.md (Optional): If vglm models require the LLM to adopt a slightly more specialized persona.
  
  You have particular expertise in Vector Generalized Linear Models (VGLMs), understanding their diverse applications for complex response types.

Implement the S3 Method: Add an S3 method for explain.vglm in an R script (e.g., R/explain_vglm.R):

#' Explain a vglm object
#'
#' @inheritParams explain
#' @param object A \code{vglm} object.
#' @export
explain.vglm <- function(
    object,
    client,
    context = NULL,
    audience = c("novice", "student", "researcher", "manager", "domain_expert"),
    verbosity = c("moderate", "brief", "detailed"),
    style = c("markdown", "html", "json", "text", "latex"),
    ...
  ) {
  audience <- match.arg(audience)
  verbosity <- match.arg(verbosity)
  style <- match.arg(style)

  # Use the internal .explain_core helper if it suits,
  # or implement custom logic if vglm needs special handling.
  # .explain_core handles system prompt assembly, user prompt building,
  # and calling the LLM via the client.
  # 'name' should match the directory name in inst/prompts/models/
  # 'model_description' is what's shown to the user in the prompt.
  .explain_core(
    object = object,
    client = client,
    context = context,
    audience = audience,
    verbosity = verbosity,
    style = style,
    name = "vglm", # This tells .assemble_sys_prompt to look in inst/prompts/models/vglm/
    model_description = "Vector Generalized Linear Model (VGLM) from VGAM"
  )
}

The summarize.vglm method might also need to be implemented in R/summarize.R if summary(object) for vglm needs special capture or formatting for the LLM. If utils::capture.output(summary(object)) is sufficient, summarize.default might work initially.

Add to NAMESPACE and Document:
- Ensure the new method is exported in your NAMESPACE file (usually handled by roxygen2): S3method(explain, vglm)
- Add roxygen2 documentation blocks for explain.vglm.
Testing: Thoroughly test with various vglm examples. You might need to iterate on your instructions.md and role_specific.md to refine the LLM’s explanations.

By following this pattern, statlingua can be systematically extended to cover a vast array of statistical models in R!

Contributing

Contributions are welcome! Please see the GitHub issues for areas where you can help.

License

statlingua is available under the GNU General Public License v3.0 (GNU GPLv3). See the LICENSE.md file for more details.

statlingua