An empirical test of the various theoretical arguments for the size factor in the Fama-French three-factor framework

The Fama-French (1992, 1993) three-factor model provides a succinct description of cross-sectional variations in stock returns. The Fama-French three-factor model not only has been widely accepted in journal-quality empirical asset pricing research, but also has been incorporated into introductory finance textbooks. However, lacking any rigorous theoretical foundation, the Fama-French three-factor "model" is essentially driven by empirical regularities. To justify the pricing of size in multi-factor asset pricing, this dissertation focuses on the various theoretical arguments to explain the size factor within the Fama-French framework, and empirically tests which hypothesis best fits the underlying data generating process.; The size factor (or effect) in this dissertation is used in both the factor loading and firm characteristic contexts. Assuming rational asset pricing, I focus on testing the default risk factor hypothesis versus the illiquidity/information costs story as the competing explanations for the size factor. In my empirical study, empirical variables proxying for these two competing hypotheses are retrieved or constructed using market and accounting data from CRSP and COMPUSTAT databases respectively. Then, using both the Fama-MacBeth time series and panel data approaches, I conduct three sets of empirical regression analyses at both industry and overall market level for the period of 1980--1999.; In summary, my results offer supporting evidence for both the default risk factor hypothesis and the illiquidity/information story when they are evaluated alone, but favor the default risk theory when they are examined side by side. Further, among the default risk factors, market-based default risk factor proxies are found to provide relatively more information than accounting-based default risk factor proxies. Finally, my results do not seem to be driven by the survivorship bias or other econometric issues. They are also robust to model selection tests such as log likelihood ratio tests and the Vuong (1989) test.