Print
E-MailKevin D. Chen is an Assistant Professor of Accounting at Duke University, John E. Core is the Nanyang Technological University Professor and a Professor of Accounting at MIT, and Wayne R. Guay is the Yagao Professor of Accounting at the University of Pennsylvania. This post is based on their SSRN working paper.
Although academics and practitioners generally view corporate governance as context-specific, existing measures of governance quality in the literature typically do not incorporate contextual information. That is, these measures assert the efficacy of specific governance mechanisms, often weighted equally, without considering that different types of firms might require different governance structures. While this practice undoubtedly stems from its convenience in operationalizing measures of governance quality, incorporating contextual factors would seem a natural next step in extending these measures and our understanding of the way governance influences corporate decision-making. How does context matter for corporate governance? To what extent is governance-relevant context observed or unobserved? Can observed contextual information be leveraged to improve measures of governance quality?
In our paper, Contextual Corporate Governance, we develop a novel prediction approach to investigate how context shapes firms’ governance choices. Drawing from prior theoretical and empirical literature, we begin by formulating a linear prediction model that incorporates observed context and use the model to examine how well observed contextual factors predict governance choices in out-of-sample data. The four observed contextual factors that we identify and include in the model are: the firm’s life cycle, nature of investments, operational complexity, and information environment. For example, theory suggests that the benefits and costs of anti-takeover provisions may change over a firm’s life cycle. Importantly, our prediction model can support (or undermine) the existence of a causal relation between observed context and governance choices without specifying an identification strategy, because prediction is a necessary (though not sufficient) condition for causality (Watts, 2014; Gow et al., 2023).
We first show that observed context predicts substantial out-of-sample variation in governance choices. Compared to a base model with no context (i.e., the unconditional model), the linear prediction model with observed context leads to a statistically significant increase in predictive accuracy for seven of the ten governance mechanisms that we examine. Averaged across all governance mechanisms, the linear prediction model with observed context improves predictive accuracy by 18%. Because prediction is a necessary condition for causality, our findings with out-of-sample predictive accuracy are consistent with a causal relation between the observed context and governance choices, confirming the commonly-held view that observed context matters for governance structure.
We next examine two potential improvements to the linear prediction model. First, observed context could have nonlinear effects on governance choices. For example, the information environment may play an especially vital role in young firms where there is relatively higher information asymmetry. To quantify the improvement from including nonlinearities, we build on the notion of model completeness developed in Fudenberg et al. (2022). Model completeness compares the prediction error of the linear prediction model to the lowest possible prediction error that is achievable given the observed contextual factors, which they term as irreducible error. Following Fudenberg et al. (2022), we estimate irreducible error using the predictions from the random-forest algorithm, a machine-learning method that aggregates the predictions from multiple decision trees. Averaged across all governance mechanisms, we show that the linear prediction model’s completeness is 0.32, indicating that of the total variation in governance choices that could possibly be predicted by the observed contextual factors (i.e., predictable variation), the linear prediction model predicts 32% with the remaining 68% captured by nonlinearities. This suggests that nonlinearities are highly important for understanding the role of context in influencing governance choices.
Second, there could be unobserved context (e.g., the CEO’s preferences) that matters for governance choices. While prior empirical studies commonly account for unobserved context through the use of firm fixed effects, this approach is not feasible in our prediction setting because there are firms in the out-of-sample data that do not belong in the training sample. To overcome this challenge, we use a machine-learning method, the k-modes clustering algorithm, to obtain information about unobserved context. This algorithm groups firms with similar governance structures together in clusters without using the observed contextual factors; therefore, the clusters that emerge from this algorithm provide additional information about unobserved context that can potentially enhance the prediction model (Chaturvedi et al., 2001). Averaged across all governance mechanisms, we show that including information from these clusters substantially improves on the prediction model with observed context and nonlinearities, increasing predictive accuracy by 27%. This suggests that unsupervised machine-learning algorithms like k-modes clustering can offer insights into unobserved context, enhancing our ability to explain out-of-sample variation in corporate governance choices.
One of the main challenges in corporate governance research is measuring a firm’s governance quality. Leveraging insights from the prediction analysis, we propose a new measure of governance quality, context consistent governance (CCG), which captures how close a firm’s actual governance structure is to the one predicted by its contextual factors. We find that CCG is positively associated with out-of-sample firm value and operating performance: a one standard deviation increase in CCG is associated with a 0.12 standard deviation increase in future firm value indicating that observed context can be useful for identifying well-governed firms. Comparing CCG to unconditional governance indices, we find that CCG has stronger associations with out-of-sample firm outcomes (with and without additional control variables), suggesting that incorporating contextual information can improve the measurement of governance quality.
Collectively, our results provide new evidence on the existence and strength of causal relations between observed contextual factors and corporate governance mechanisms using a prediction-based approach. These relations are important because they are often the focus of corporate governance theory and help us understand why firms make the governance choices they do. We also provide estimates of completeness for the linear model that has been the workhorse model in the empirical literature to examine the relations between firm characteristics and governance mechanisms. Our findings suggest that the linear model is highly incomplete and can be substantially improved by incorporating the nonlinear effects of contexts on governance choices. Further, we propose a new measure of governance quality and highlight the value of incorporating contextual information into the measurement of corporate governance.
