# F2-dev-01 — Two-way FE slope with firm-clustered uncertainty

# Question

Estimate β in y_it = a_i + g_t + β x_it + u_it (firm and year fixed effects), with standard errors that allow arbitrary within-firm dependence over time. The data are simulated (synthetic).

# Data and exclusions

data/firm_panel.csv: 1200 rows read. 100 firms (IDs 1001–1100) × 12 years (2008–2019), one row per firm–year. I checked for missing values, duplicated firm–year keys, imbalance, and firms with no within-firm variation in x. None of these problems turned up, so no rows were excluded (data_issues = []). All 1200 observations are used.

# Method

• Baseline: pooled OLS of y on x with an intercept.
• Two-way FE: OLS of y on x plus firm and year dummies (LSDV), which is numerically identical to the two-way within estimator in a balanced panel. Residual df = 1200 − 1 − 99 − 11 − 1 = 1088.
• Conventional SE: the homoskedastic OLS SE from the LSDV fit (df = 1088).
• Clustered SE: the CR1 sandwich estimator clustered by firm_id (100 clusters). It uses the statsmodels small-sample factor G/(G−1)·(N−1)/(N−K), where K counts all dummies (like Stata areg/reg … i.firm, vce(cluster firm)).
• 95% CI: β̂ ± t_{0.975, G−1} · se_cluster, with G−1 = 99 df (t = 1.984).
• No randomness is involved (random_seed = null).

# Results

The pooled slope is about twice the FE slope. This means x is correlated with the firm and/or year effects, so the naïve estimate is biased upward relative to the within-firm, within-year association.

# Uncertainty

The clustered SE is about 1.44 times the conventional SE. This suggests positive within-firm dependence in the residuals and/or x over time, which the conventional SE ignores. The result changes little under other small-sample conventions (see sensitivity in results.json):

• clustered SE with no finite-sample correction: 0.0446;
• clustered SE with K excluding the absorbed fixed effects (Stata xtreg, fe style): 0.0448;
• normal-critical-value CI: [0.5122, 0.6967].

I report the most conservative variant (0.0471) as se_cluster.

# Limitations and open issues

• β is a within-firm, within-year association. It is not a causal effect unless x is strictly exogenous given the fixed effects.
• Inference relies on asymptotics in the number of firms. G = 100 is moderate, so it should be adequate.
• The choice of finite-sample correction for the clustered SE is a convention. The alternatives differ by about 5%.
• These are synthetic data. The conclusions apply only to this dataset.