Measuring corporate diversification effect sizes from observational data using Bayesian causal approaches
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
This dissertation investigates three different, but related questions where the traditional models in the finance literature are unable to provide a concrete answer. Firm diversification is one of the highly controversial issues in financial literature which has been extensively studied. However, mixed findings in the literature reveal the inefficiency of previously applied methods. In this dissertation, novel customized causal models are applied to investigate the different effects of diversification strategy and provide efficient insights for strategy setting. The goal is to address the limitations of existing methods for financial applications by designing bespoke Bayesian statistical models, applying those models to real-world data, and extending them with customized models of causal inference. ☐ The first application focuses on investigating the causal relationship between firm diversification and bankruptcy risk. Specifically, the causal effect of diversification on bankruptcy risk for US public firms using the Bayesian g-formula is examined. This study confirms the presence of a causal relationship between firm diversification and bankruptcy risk and discover results for US-based firms that are consistent with coinsurance theory; namely, diversification strategy reduces bankruptcy risk. However, the causal relationship is weaker than implied by most previous correlation literature, suggesting that the value of firm diversification has been overstated in related research and that causal techniques are useful in teasing out the true magnitude of an effect. ☐ The second part is motivated by investigating the joint causal effects of corporate diversification on firm risk and performance for US public firms. The correlation between the outcome variables is considered and a joint statistical model is designed to investigate the causal effects on firm risk and performance. Also, to examine the potential heterogeneity in causal effects and avoid relying solely on making statements about the average causal effect, the existing Bayesian g-formula is extended to provide estimates of quantile treatment effect. Combining causal models and Bayesian multilevel models with quantile analysis allows us to quantify and visualize the considerable heterogeneous effects of corporate diversification on firms’ performance and risk. In addition, the determinants of the heterogeneous effects are visualized to help understanding why the causal effect is different across the outcome distributions. Given the heterogeneity, this work shows how to estimate firm-level effects of diversification. The probability of any potential causal effect direction is provided for both the entire population as well as for single firms. These probabilities coupled with individual and population estimation of causal effect provide more comprehensive insights for the decision makers. ☐ The third part of this dissertation provides a methodology for extending causal effect estimation. The aim of the third part is to improve the predictive power of Inverse probability weighting (IPW); which is one the most popular and widely used causal models and relies on parametric logistic regression model to estimate propensity scores. However, the assumptions required by IPW for regression models may not hold for many datasets. To mitigate these limitations, Bayesian Additive Regression Trees (BART) is applied as a machine learning alternative for logistic regression within the IPW framework. BART can extract the non-linear and interaction among variables to propose more accurate propensity score estimations. In addition, a Bayesian approach is replaced with the frequentist one in original model to estimate the effect size in the pseudo-population, which provides the posterior predictive distribution of the causal effect. The new IPW model performance is evaluated in terms of bias and RMSE. The results suggest that the new IPW model improves both propensity score and ATE estimations. The causal effect of diversification strategy on US public firms’ risk-adjusted performance is examined by the new IPW model. ☐ In summary, this dissertation focuses on proposing more comprehensive causal analysis for decision makers in the financial area. As opposed to most of the related studies, this study motivates decision makers for strategical changes by providing more reliable causal inferences instead of making decisions using biased associations. The three applications of this study all differ slightly in focus but contribute to a well-rounded study of designing advanced causal models. The first contribution is application driven, but the second and third parts are methodologically focused as well as being application oriented. Taken together, this work proposes several flexible causal models to solve real financial problems by providing more insights of strategy’s effects.
Description
Keywords
Bankruptcy Risk, Bayesian Inference, Causal Inference, Corporate Diversification, Firm performance and risk, Machine learning