Essays on Corporate Culture

Abstract

In this dissertation, I examine important questions in the measurement of organizational culture using computational textual analysis of corporate earnings conference calls. The research topics include contextual measurement of corporate integrity using a multi-agent LLM architecture in the first essay and measuring organizational culture through the Competing Values Framework using the same methodology in the second essay. Both of my essays are co-authored with Professor Shantanu Dutta with first essay currently under a revise-and-resubmit decision at Information Systems Frontier. The first essay addresses a fundamental measurement challenge in the corporate culture literature: dictionary-based approaches to measuring corporate integrity from corporate disclosures count every occurrence of culture-related keywords as evidence of the underlying construct, regardless of the context in which those words appear. We develop and validate a multi-agent large language model (LLM) pipeline, based on a blackboard architecture with eleven specialized agents, to measure four components of corporate integrity (i.e., Accountability, Compliance, Moral Integrity, and Honesty) from the Q&A sections of earnings conference call transcripts. Our contextual approach reveals that fewer than 30% of the segments containing integrity-related dictionary terms are meaningfully related to the integrity construct, demonstrating that traditional dictionary methods substantially overcount meaningful cultural discourse. The multi-agent classification system achieves 95.67% accuracy and a Cohen’s kappa of 0.913 against human annotations, with five independent stability runs producing an inter-run mean kappa of 0.945. To validate the resulting measure empirically, we examine its association with established indicators of corporate governance quality and reporting integrity. We find that firms with greater board independence and higher female board representation engage in more integrity-related discourse, consistent with prior research linking these governance characteristics to stronger ethical oversight. We also find that firms with higher contextual integrity scores are more likely to disclose non-fraudulent restatements, consistent with the view that integrity-oriented cultures promote transparent correction of reporting problems rather than increasing the incidence of misreporting. Furthermore, higher integrity scores are associated with lower discretionary accruals, higher dividend payouts, and greater access to trade credit, in line with theoretical predictions that integrity culture constrains opportunistic reporting, enhances credibility with capital providers, and reduces counterparty risk. Across these analyses, the dictionary-based measure exhibits systematically weaker and, in several cases, statistically insignificant associations, underscoring that contextual interpretation is not merely methodological refinement but a necessary step for recovering economically meaningful integrity signals from corporate discourse. The second essay implements the methodological innovations of the first essay to the Competing Values Framework (CVF), the most widely adopted theoretical model of organizational culture. Despite its theoretical maturity, empirical measurement of the CVF has relied predominantly on dictionary-based word counts from 10-K filings, an approach that is both context-insensitive and confined to heavily curated corporate documents. We adapt the multi-agent blackboard architecture to classify earnings call Q&A discourse into the four CVF cultural orientations (i.e., Clan, Adhocracy, Hierarchy, and Market) using over 183,000 transcripts spanning 2008 to 2024. Our classification identifies approximately 20% of total segments as containing genuine cultural content, with Market discourse predominating, followed by Hierarchy, Adhocracy, and Clan, a distribution that reflects the externally oriented communicative demands of the earnings call setting. We conduct the first systematic construct-validation exercise that pairs each CVF dimension with the specific firm outcomes its own theoretical logic predicts, and we benchmark our contextual scores against the widely used dictionary-based measures on the same validation outcomes. Our contextual CVF scores produce theoretically consistent and statistically significant associations across all four dimensions (i.e., Clan with labor intensity, Adhocracy with innovation, Market with profitability, and Hierarchy with regulatory discipline) whereas the dictionary-based scores yield inconsistent and often economically negligible results. These findings demonstrate that all four dimensions of the CVF carry distinct economic content when measured with sufficient contextual precision, and that the apparent irrelevance of certain CVF dimensions in prior work likely reflects measurement noise rather than a genuine absence of cultural influence on firm behavior.