dynobo / lmdiag Goto Github PK

Python Library providing Diagnostic Plots for Linear Regression

License: MIT License

Python 100.00%

lmdiag's Introduction

lmdiag

Python Library providing Diagnostic Plots for Linear Regression Models. (Like plot.lm in R.)

I built this, because I missed the diagnostics plots of R for a university project. There are some substitutions in Python for individual charts, but they are spread over different libraries and sometimes don't show the exact same. My implementation tries to copycat the R-plots, but I didn't reimplement the R-code: The charts are just based on available documentation.

Installation

pip install lmdiag

Usage

lmdiag generates plots for fitted linear regression models from statsmodels, linearmodels and scikit-learn.

You can find some usage examples in this jupyter notebook.

Example

import numpy as np
import statsmodels.api as sm
import lmdiag

# Fit model with random sample data
np.random.seed(20)
X = np.random.normal(size=30, loc=20, scale=3)
y = 5 + 5 * X + np.random.normal(size=30)
X = sm.add_constant(predictor)  # intercept required by statsmodels
lm = sm.OLS(y, X).fit()

# Plot lmdiag facet chart
lmdiag.style.use(style="black_and_red")  # Mimic R's plot.lm style
fig = lmdiag.plot(lm)
fig.show()

Methods

Draw matrix of all plots:

lmdiag.plot(lm)
Draw individual plots:

lmdiag.resid_fit(lm)

lmdiag.q_q(lm)

lmdiag.scale_loc(lm)

lmdiag.resid_lev(lm)
Print description to aid plot interpretation:

lmdiag.help() (for all plots)

lmdiag.help('<method name>') (for individual plot)

Increase performance

Plotting models fitted on large datasets might be slow. There are some things you can try to speed it up:

1. Tune LOWESS-parameters

The red smoothing lines are calculated using the "Locally Weighted Scatterplot Smoothing" algorithm, which can be quite expensive. Try a lower value for lowess_it and a higher value for lowess_delta to gain speed at the cost of accuracy:

lmdiag.plot(lm, lowess_it=1, lowess_delta=0.02)
# Defaults are: lowess_it=2, lowess_delta=0.005

(For details about those parameters, see statsmodels docs.)

2. Change matplotlib backend

Try a different matplotlib backend. Especially static backends like AGG or Cairo should be faster, e.g.:

import matplotlib
matplotlib.use('agg')

Setup development environment

python -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'
pre-commit install

Certification

lmdiag's People

Contributors

Stargazers

Watchers

Forkers

kschan923 t-hdd lystahi jboverio

lmdiag's Issues

Python 2 compatibility

@dynobo I want to start off by saying that I love this package. Inferential statistics is super easy in R thanks to things like plot.lm, so lmdiag really helps bridge the gap between Python and R! I was hunting down an easy way to do something similar to plot.lm in Python without having to manually build a bunch of the plots, so I was very happy to find this package.

So I found one case in info.py where you used some Python 3-specific string formatting that causes this package to not work for Python 2. If you're only planning on having this be a Python 3 package, I completely respect that, though it would be nice if that were the case to add a note of that in the README. If you are interested in making this package work for Python 2, here is a super quick fix I made in a fork of this package that fixes the issue for Python 2 (note: I haven't done extensive testing based on this fix, I just know it lets the package be imported without raising an error).

Example:

df = sm.datasets.get_rdataset("ames", "openintro").data
res = smf.ols("np.log10(price) ~ Q('Overall.Qual') + np.log(area)", df).fit()

lmdiag

%%time
lmdiag.plot(res)

CPU times: user 15.1 s, sys: 215 ms, total: 15.3 s
Wall time: 16.1 s

LinearRegDiagnostic

%%time
LinearRegDiagnostic(res)()

CPU times: user 2.17 s, sys: 125 ms, total: 2.29 s
Wall time: 2.17 s