Organizational cultural strength as the negative cross-entropy of mindshare: a measure based on descriptive text

Organizational cultural strength, the degree to which organizational members think cohesively, is a crucial factor influencing organizational structure, performance, diversity, and inclusion. Strong cultures can foster self-organization and resilience, but they can also lead to groupthink and hinder innovation or diversity. Despite its theoretical importance, measuring cultural strength reliably and scalably has been challenging. Prior research has struggled with defining and operationalizing the concept, measuring differences across cultures while accounting for idiosyncratic elements, and accessing sufficient data for robust comparisons. This paper aims to address these challenges by proposing a novel measure and applying it to a large-scale dataset.

Existing literature on organizational cultural strength reveals a long-standing fascination with the concept, but also a lack of clarity in its definition and measurement. Early work focused on defining strong cultures as homogenous, thick, stable, intense, cohesive, or displaying consensus and intensity. More recent work attempted mathematical formalizations, defining strong cultures as those with high homogeneity in members' beliefs. Empirical research has utilized survey data, but these studies have been limited by small sample sizes and reliance on pre-defined cultural dimensions (etic approach), potentially neglecting idiosyncratic elements (emic approach). The authors identify three primary limitations in the existing literature: (1) lack of a clear conceptual bridge between the intuitive notion of a strong culture and its operationalization; (2) challenges in reliably measuring differences across organizational cultures while being sensitive to organization-specific cultural elements; and (3) limited availability of large-scale data for quantitative comparisons.

This study proposes a novel measure of organizational cultural strength as the negative average cross-entropy of individual mindshare distributions over firm-specific cultural attributes. The mindshare distributions are estimated from employee-generated text reviews on Glassdoor.com using Latent Dirichlet Allocation (LDA), a topic modeling technique. The LDA algorithm extracts firm-specific cultural attributes discussed in employee reviews, summarizing each review as a probability distribution representing the individual's mindshare over these attributes. The study uses data from 2,900,436 reviews across 94,868 US firms from 2008-2020. To mitigate potential response biases, the authors focus on firms with high Glassdoor coverage (at least 25%) and employ several robustness checks, including examination of excess kurtosis in satisfaction ratings to assess polarization. The authors primarily analyze the "pros" section of Glassdoor reviews, given its presumed higher relevance to actual organizational attributes, but also analyze the "cons" section for comparative purposes. The number of topics in LDA was optimized based on coherence maximization, ensuring greater semantic distance among the identified topics. Bivariate correlations and OLS regressions are used to validate the cultural strength measure against theoretically relevant covariates, while controlling for variables like firm revenues, Glassdoor coverage, and the proportion of current employees providing reviews. Firm industry, state, and type fixed effects are included in the regressions. Robustness checks, including bootstrapping and re-analysis using a discretized cultural strength variable, were conducted.

The study finds that the proposed measure of cultural strength correlates with theoretically expected covariates. Smaller firms (negative correlation), older firms (positive correlation), and geographically concentrated firms (negative correlation) exhibit stronger cultures. Consistent with socialization theory, reviewers' average tenure shows a positive correlation with cultural strength. Exploratory analyses reveal intriguing associations. Firms with greater gender imbalance (positive correlation) and fewer Glassdoor occupational codes (negative correlation) display stronger cultures. These bivariate findings are largely supported by multivariate regression analyses, which further control for firm-level variables and include fixed effects for industry, state, and firm type. The study finds that variations in cultural strength are larger between than within industries, geographies, and firm typologies. Analyses of the "cons" section of reviews show generally consistent results, though with some notable differences: older firms and firms with longer reviewer tenure display weaker cultures in the "cons" analysis. The analyses of satisfaction data showed that no evidence of polarization was found.

The findings support the validity and reliability of the proposed measure of organizational cultural strength. The consistent relationships between the measure and theoretically predicted covariates demonstrate that it captures meaningful organizational-level properties and is not solely driven by biases related to the reviewer sample. The exploratory findings regarding gender imbalance and role differentiation raise important questions about the potential interplay between cultural strength and demographic composition/organizational structure. These correlations, however, do not necessarily imply causality and warrant further investigation. The greater variation in cultural strength between rather than within industry, geographic, and firm typologies indicates a need for careful consideration when selecting exemplars of strong cultures.

This paper contributes to the literature by offering a clear, operationalizable definition of organizational cultural strength as the negative average cross-entropy of mindshare distributions. The proposed measure leverages readily available large-scale textual data, enabling researchers to study cultural strength in a comprehensive manner. Future research could explore different types of organizational cultures, investigate the causal mechanisms behind observed relationships, and apply the measure to various organizational contexts, including mergers and acquisitions, strategic alliances, and responses to organizational disruptions.

The primary limitation of the study stems from the use of Glassdoor reviews, which are not a random sample of organizational employees. While efforts were made to mitigate response bias, the results may not fully represent the cultures of the sampled companies. The correlational nature of the findings also limits causal inference; future research is needed to test the causal mechanisms linking cultural strength to gender imbalance and role differentiation. The large sample size also leads to a significant reduction in observations when fixed effects are added to the model.