duckreg : very fast out-of-memory regressions with duckdb

python package to run stratified/saturated regressions out-of-memory with duckdb. The package is a wrapper around the duckdb package and provides a simple interface to run regressions on very large datasets that do not fit in memory by reducing the data to a set of summary statistics and runs weighted least squares with frequency weights. Robust standard errors are computed from sufficient statistics, while clustered standard errors are computed using the cluster bootstrap.

See examples in notebooks/introduction.ipynb.

• install

pip install duckreg

• dev install (preferably in a venv) with

(uv) pip install git+https://github.com/apoorvalal/duckreg.git

or git clone this repository and install in editable mode.

---

Currently supports the following regression specifications:

1. DuckRegression: general linear regression, which compresses the data to y averages stratified by all unique values of the x variables
2. DuckMundlak: One- or Two-Way Mundlak regression, which compresses the data to the following RHS and avoids the need to incorporate unit (and time FEs)

$$ y \sim 1, w, \bar{w}_{i, .}, \bar{w}_{., t} $$

3. DuckDoubleDemeaning: Double demeaning regression, which compresses the data to y averages by all values of $w$ after demeaning. This also eliminates unit and time FEs

$$ y \sim (W_{it} - \bar{w}_{i, .} - \bar{w}_{., t} + \bar{w}_{., .}) $$

4. DuckMundlakEventStudy: Two-way mundlak with dynamic treatment effects. This incorporates treatment-cohort FEs ($\psi_i$), time-period FEs ($\gamma_t$) and dynamic treatment effects $\tau_k$ given by cohort X time interactions.

$$ y \sim \psi_i + \gamma_t + \sum_{k=1}^{T} \tau_{k} D_i 1(t = k) $$

All the above regressions are run in compressed fashion with duckdb.

---

references:

methods:

• Arkhangelsky and Imbens (2023)
• Wooldridge 2021
• Wong et al 2021

libraries:

• Grant McDermott's duckdb lecture