Latent Group Structures in Panel Data

Downloads

Each dataset is available as a labeled Stata .dta and its source file.

⇩ Download all data (ZIP)stata_codebook.do

Run stata_codebook.do in Stata once to attach long-form per-variable notes to the .dta files.

Load directly in code

Every file loads straight from GitHub (raw URLs). Swap the file name to load any dataset.

Stata

* Stata 14+ : `use` reads an https URL directly
global BASE "https://raw.githubusercontent.com/cmg777/starter-academic-v501/master/content/post/stata_panel_lasso_cluster/data/"
use "${BASE}saving.dta", clear
describe
notes

Python

!pip install -q pyreadstat
import pandas as pd
BASE = "https://raw.githubusercontent.com/cmg777/starter-academic-v501/master/content/post/stata_panel_lasso_cluster/data/"
df = pd.read_stata(BASE + "saving.dta")

# load every dataset at once
files = ["saving", "democracy"]
data = {f: pd.read_stata(BASE + f + ".dta") for f in files}

# pyreadstat (richest metadata) reads LOCAL files -> download first
import pyreadstat, urllib.request
urllib.request.urlretrieve(BASE + "saving.dta", "saving.dta")
df, meta = pyreadstat.read_dta("saving.dta")

Copy and paste this snippet in Google Colab app. https://colab.research.google.com/notebooks/empty.ipynb

R

# R : haven::read_dta auto-downloads an https URL
library(haven)
BASE <- "https://raw.githubusercontent.com/cmg777/starter-academic-v501/master/content/post/stata_panel_lasso_cluster/data/"
df <- read_dta(paste0(BASE, "saving.dta"))

Overview & sources

Companion data for a hands-on Stata tutorial that uses the Classifier-LASSO (C-LASSO) method of Su, Shi & Phillips (2016), via the classifylasso command (Huang, Wang & Zhou 2024), to discover latent group structures in panel data — subsets of units that share slope coefficients while groups differ. Two source datasets drive the post. saving is a balanced panel of 56 countries observed over 15 years (840 observations) on savings behavior, from the Su–Shi–Phillips (2016) replication. democracy is the Acemoglu, Naidu, Restrepo & Robinson (2019) democracy-and-growth panel of 98 countries from 1970–2010 (4,018 observations). The tutorial shows that the pooled democracy–growth effect of +1.055 masks a +2.151 effect in 57 countries and a −0.936 effect in 41 countries — a sign reversal exemplifying Simpson's paradox.

Data sources

Cite this data

Please cite this dataset as follows.

APA

Mendez, C. (2026). Identifying Latent Group Structures in Panel Data: The classifylasso Command in Stata [Data set]. https://carlos-mendez.org/post/stata_panel_lasso_cluster/

Su, L., Shi, Z., & Phillips, P. C. B. (2016). Identifying latent structures in panel data. Econometrica, 84(6), 2215–2264. https://doi.org/10.3982/ECTA12560 Acemoglu, D., Naidu, S., Restrepo, P., & Robinson, J. A. (2019). Democracy does cause growth. Journal of Political Economy, 127(1), 47–100. https://doi.org/10.1086/700936 Huang, W., Wang, Y., & Zhou, L. (2024). Identify latent group structures in panel data: The classifylasso command. Stata Journal, 24(1), 173–203. https://doi.org/10.1177/1536867X241233664

BibTeX

@misc{mendez2026statapanellassocluster,
  author       = {Mendez, Carlos},
  title        = {Identifying Latent Group Structures in Panel Data: The classifylasso Command in Stata},
  year         = {2026},
  howpublished = {\url{https://carlos-mendez.org/post/stata_panel_lasso_cluster/}},
  note         = {Data set}
}

@article{su2016identifying,
  author  = {Su, Liangjun and Shi, Zhentao and Phillips, Peter C. B.},
  title   = {Identifying Latent Structures in Panel Data},
  journal = {Econometrica},
  volume  = {84}, number = {6}, pages = {2215--2264}, year = {2016},
  doi     = {10.3982/ECTA12560}
}
@article{acemoglu2019democracy,
  author  = {Acemoglu, Daron and Naidu, Suresh and Restrepo, Pascual and Robinson, James A.},
  title   = {Democracy Does Cause Growth},
  journal = {Journal of Political Economy},
  volume  = {127}, number = {1}, pages = {47--100}, year = {2019},
  doi     = {10.1086/700936}
}
@article{huang2024classifylasso,
  author  = {Huang, Wenxin and Wang, Yuan and Zhou, Lin},
  title   = {Identify Latent Group Structures in Panel Data: The classifylasso Command},
  journal = {Stata Journal},
  volume  = {24}, number = {1}, pages = {173--203}, year = {2024},
  doi     = {10.1177/1536867X241233664}
}

Variable explorer search & filter all 21 variables

Type to filter by name or label, or use the chips to filter by type. Each row shows a mini distribution. Click a header to sort.

Cross-file variable index

Which file each variable appears in (● = present).

Construction & formulas

C-LASSO minimizes a penalized least-squares objective that shrinks each unit's slope vector toward one of K group centers, then re-estimates each group by plain OLS (postlasso) for valid inference.

• Objective: Q = (1/NT) Σ_i Σ_t (y_it − β_i'x_it)² + (λ/N) Σ_i Π_k ‖β_i − α_k‖ — SSR plus a product-form group penalty (the LASSO penalty λ_NT = c_λ·T^(−1/3)).
• Group membership: β_i = α_k if i ∈ G_k, for k = 1,…,K — each unit takes its group's slope vector.
• Number of groups: chosen to minimize an information criterion (IC); the complexity tuning parameter is ρ_NT = c_ρ·(NT)^(−1/2). K = 2 is selected in all specifications.
• Cluster-robust inference: standard errors are clustered by country (cluster(country)), adjusting for within-country serial correlation in the two-way fixed-effects (country + year) postlasso step.
• Dynamic bias correction: with a lagged dependent variable (lagsavings / ly1), the dynamic option applies the half-panel jackknife to remove Nickell bias before grouping.

The savings panel is pre-standardized (each variable demeaned and scaled to unit SD), so its coefficients are read in standard-deviation units. The democracy panel's lnPGDP and the GDP lags are log per-capita GDP × 100.

The datasets

Switch datasets with the tabs. Each shows the full variable dictionary plus a sortable statistics table with mini distributions and data coverage.

Expand all datasets expand to search (Ctrl/⌘+F) or print across all datasets

Known limitations & caveats

• Standardized savings variables. In saving.dta, savings, lagsavings, cpi, interest, and gdp are each standardized to mean 0, SD 1 — coefficients are in standard-deviation units, not raw economic magnitudes.
• Coded time index. In saving.dta, year is coded 1–15 (not calendar years); these 15 periods correspond to 1995–2010 in the source.
• Numeric country codes. Both panels identify units by numeric code (code in savings; country in democracy). The democracy file carries a name string; the savings file does not, so savings groups cannot be mapped to named countries.
• Scaled GDP variables. In democracy.dta, lnPGDP and its lags ly1–ly4 are log per-capita GDP multiplied by 100; lnTrade and lnMort are likewise scaled by 100.
• Associations, not proven causation. The group-specific C-LASSO estimates are conditional associations within the panel model; a causal reading of the democracy effect requires the same identifying assumptions as Acemoglu et al. (2019).
• Close IC in the democracy model. IC values across K = 1–5 span only 3.267–3.280; K = 2 is optimal but not dominant, so group assignments are sensitive to the tuning parameter ρ.