Conflict between cluster-based signaling variable and clustered standard errors in insider trading study

External question — Cross Validated Stack Exchange Author: Alfred Ljungkvist Original post: https://stats.stackexchange.com/questions/675857 License: CC BY-SA 4.0 — https://creativecommons.org/licenses/by-sa/4.0/ Adaptation: HTML converted to plain text; contact email addresses removed. I am currently working on a finance thesis examining insider trading on Nasdaq Stockholm (2020–2026), utilizing a dataset of over 24,000 transactions. Study Design : I use a pooled OLS approach at the transaction level to predict abnormal returns (relative to the OMXSGI index) over 3, 6, 12, and 24-month horizons. The independent variables include: Cluster Signal : A dummy variable indicating if >3 unique insiders in the same firm traded in the same direction within a rolling 6-month window (Net Purchase Ratio > |0.95|). Position : Dummy for CEO/CFO vs. Board members. Transaction Size : Relative to market cap. Controls : Firm size (ln Market Cap) and sector dummies. The Econometric Challenge : Since the dataset contains multiple observations per firm over time, the residuals are likely correlated within firms. To address this, I have run two specifications: one with HC3 robust standard errors and one with standard errors clustered at the firm level. The specific problem I am facing is a conceptual and statistical conflict regarding the Cluster Signal. The logic behind this variable is to explicitly capture the informative value of multiple insiders acting on the same private information—essentially measuring dependent behavior within the firm. However, by clustering standard errors at the firm level, I am correcting for within-firm correlation. Results: When moving from HC3 to firm-clustered standard errors, the Cluster Signal loses significance at shorter horizons (3–6 months) but remains significant at 12–24 months. My Questions : Is there an inherent conflict in using firm-clustered standard errors when the main variable of interest is designed to capture a "cluster" effect within that same unit (the firm)? Does clustering the errors risk "absorbing" the very correlation the signal is intended to measure? In this specific context, which estimation choice is considered more robust? Is it academically defensible to report both specifications (HC3 vs. Clustered) as a form of sensitivity/robustness analysis, or does the firm-level clustering always supersede HC3 in a transaction-level panel like this?I would appreciate any insights on the best path forward for interpreting these conflicting results. Best regards