scpc-python 0.1.0b4

SCPC inference in Python

███████  ██████ ██████   ██████       ██████  ██    ██ ████████ ██   ██  ██████  ███    ██ 
██      ██      ██   ██ ██            ██   ██  ██  ██     ██    ██   ██ ██    ██ ████   ██ 
███████ ██      ██████  ██      █████ ██████    ████      ██    ███████ ██    ██ ██ ██  ██ 
     ██ ██      ██      ██            ██         ██       ██    ██   ██ ██    ██ ██  ██ ██ 
███████  ██████ ██       ██████       ██         ██       ██    ██   ██  ██████  ██   ████ 
                                                                                           

scpc-python: Spatial Correlation-Robust Inference

scpc-python provides spatial correlation-robust inference for regression coefficients following Müller and Watson (2022, 2023).

Installation

You can install scpc-python from PyPI (coming soon) with:

uv venv --python 3.11
uv pip install scpc-python

or from GitHub with:

uv venv --python 3.11
uv pip install git+https://github.com/DGoettlich/scpc-python.git

Basic Usage

This example starts from the transformed branch of the Becker, Boll, and Voth (2026) workflow. In the full workflow, you would first use spurtest to decide whether to stay in levels or transform. Here we assume the transformed branch and move directly to the regression plus scpc step.

# assumes spur-python is installed
from spur import load_chetty_data, spurtransform
from scpc import scpc
import statsmodels.formula.api as smf

data = load_chetty_data()

data = data[~data["state"].isin(["AK", "HI"])][
    ["am", "gini", "fracblack", "lat", "lon"]
]
data = data.dropna(subset=["am", "gini", "fracblack", "lat", "lon"]).copy()

transformed = spurtransform(
    "am ~ gini + fracblack",
    data,
    lon="lon",
    lat="lat",
    transformation="lbmgls",
    prefix="h_",
)

fit = smf.ols("h_am ~ h_gini + h_fracblack", data=transformed).fit()

result = scpc(
    fit,
    data=transformed,
    lon="lon",
    lat="lat",
    cvs=True,
)

scpc() returns an SCPCResult object:

• result.scpcstats: the main inference table with coefficient estimates, standard errors, t statistics, p values, and 95% interval endpoints
• result.scpccvs: optional stored critical values at 32%, 10%, 5%, and 1%
• result.avc: the average pairwise correlation bound used in the analysis
• result.c0: the kernel scale implied by avc
• result.cv: the unconditional 5% critical value
• result.q: the number of retained non-constant spatial principal components

Key Arguments

The most important scpc() arguments in the workflow above are:

• model: the fitted regression model
• data: the data frame
• lon, lat: the geodesic coordinate column names
• coords_euclidean: use this instead of lon / lat when coordinates are Euclidean rather than geographic
• cluster: optional clustering column
• ncoef: how many coefficients to report
• avc: upper bound on the average pairwise correlation
• uncond: whether to skip the conditional adjustment
• cvs: whether to store the extra critical values