Carto Gwr

Geographically Weighted Regression (GWR)

Builds CARTO Workflows that model spatially varying relationships between a dependent variable and one or more independent variables using GWR. Unlike global regression (one set of coefficients for the entire study area), GWR produces local coefficients per spatial unit, revealing how relationships change across space. Example: "bedrooms add $50k to price in downtown but only $20k in suburbs."

Prerequisites: Load carto-create-workflow for the development process, JSON structure, and validation commands — it covers both access paths (the MCP server's workflow tools such as create_workflow, validate_workflow, and run_workflow when attached; the carto workflows CLI otherwise; routing signals in carto-basics/references/access-paths.md).

Instructions

A GWR workflow follows this pipeline:

Source Data -> (Filter) -> Spatial Indexing (H3/Quadbin) -> Aggregation (dependent + independent vars per cell) -> GWR -> Save

Step 1: Load Source Data

Use native.gettablebyname. The input table must contain at least one numeric dependent variable and one or more numeric independent (predictor) variables.

Success: Node outputs a table with the necessary numeric columns.

Step 2: Filter (if needed)

Use native.wheresimplified or native.where to narrow the dataset (e.g. remove nulls from key columns, filter by category or date range).

Success: Output contains only rows with valid, non-null values for the dependent and all independent variables.

Step 3: Spatial Indexing

If the data is not already indexed, convert point geometries to spatial index cells:

• native.h3frompoint for H3
• native.quadbinfromgeopoint for Quadbin

If the data already contains an H3 or Quadbin column (common for pre-aggregated datasets), skip this step.

Resolution guidance:

Success: Every row has a spatial index column (e.g. h3).

Step 4: Aggregate per Cell

Use native.groupby to produce one row per cell with aggregated values for the dependent and all independent variables:

• Group by: the spatial index column (h3)
• Aggregation: price,avg,bedrooms,avg,bathrooms,avg (adapt to the actual columns)

The dependent variable should be aggregated with avg or sum depending on what makes sense. Independent variables are typically averaged.

Success: Output has exactly one row per unique cell, with numeric columns for the target and all predictors.

Step 5: Run GWR

Use native.gwr with:

Kernel options: gaussian (recommended -- smooth distance decay), uniform, triangular, quadratic, quartic.

K-ring size: Controls the neighborhood radius.

• Too small (1-2): noisy, unstable coefficients.
• Too large (5+): over-smoothed, approaches global regression.
• Start with 3 as a balanced default.

Success: Output contains per-cell columns: index, intercept, one coefficient column per independent variable, r_squared, and residual. (See the Provider casing note in Gotchas — Snowflake surfaces these UPPERCASE.)

Step 6: Save

Use native.saveastable to persist results. The spatial index column is directly visualizable in CARTO Builder -- style the map by coefficient columns to create coefficient maps showing spatial variation.

Success: Validated workflow that can be uploaded via create_workflow (MCP) or carto workflows create (CLI).

Output Columns

Gotchas

• Provider casing & SQL dialect. This skill documents columns in lowercase (BigQuery / Databricks / Postgres / Redshift convention). On Snowflake, unquoted identifiers surface UPPERCASE — reference H3, INDEX, PRICE, R_SQUARED, INTERCEPT, etc. in expressions. See carto-create-workflow/references/providers/<provider>.md for casing rules and SQL dialect equivalents.
• The GWR component requires the Analytics Toolbox. Offline validation (validate_workflow over MCP, or carto workflows validate) can't resolve the AT location — confirm it with the CLI-only carto workflows verify-remote --connection <conn> before running.
• The dependent variable must be continuous and numeric. Categorical targets need a different approach (e.g. classification).
• Cells with null values in ANY variable (dependent or independent) will be excluded from the model. Pre-filter or impute nulls before running GWR.
• Multicollinearity between independent variables degrades results. If two predictors are highly correlated (e.g. bedrooms and total_rooms), drop one or combine them. Check correlation before including multiple similar variables.
• r_squared per cell indicates local model fit. Very low values across many cells suggest important predictors are missing from the model.
• The features_columns input is an array of column names (e.g. ["bedrooms", "bathrooms"]), not a comma-separated string.
• The output column is named index, not the original spatial index column name. If joining back to original data, rename it with native.renamecolumn.
• Sparse data at high resolutions leads to unreliable coefficients. Ensure enough cells have data for all variables before choosing a high resolution.

Reference Templates

Workflow template (available in CARTO Workspace): "Applying Geographical Weighted Regression (GWR) to model the local spatial relationships in your data"

Example use case: Analyzing Airbnb ratings in Los Angeles -- models overall_rating vs value_review, cleanliness, location, enriched with Data Observatory sociodemographics. Uses H3 res 7, kring 3, Gaussian kernel.

Common Variations