Ch. 2: Self-Selection Models in Corporate Finance 79
12.2. Matching and long-run performance:Cheng (2003),Li and Zhao (2006)
A vast literature on market efficiency examines the long-run stock return after events
such as IPOs, SEOs, share repurchases, listing changes, etc. The semi-strong version of
the efficient markets hypothesis predicts that long-run returns should be zero on aver-
age. However, several papers report empirical evidence against the efficiency hypothesis
(Fama, 1998). In most studies, post-event buy-and-hold returns are systematically pos-
itive or negative relative to benchmarks over periods of three to five years.Chapter 1
(Kothari and Warner, 2007) offers an overview of this literature. We focus on applica-
tions of matching models to assess long-run performance.
To test whether abnormal returns are zero or not, one needs a model of benchmark
returns. As discussed inChapter 1, the standard approach, is to match an event firm
with a non-event firm on between two and four characteristics that include size, book-
to-market, past returns, and perhaps industry. This method runs into difficulties when
the number of dimensions becomes large and the calipers become fine, when it becomes
difficult to generate matching firms. Propensity score (PS) based matching methods
reviewed in Section4.3.2are potentially useful alternatives in this scenario. Two recent
papers,Cheng (2003)andLi and Zhao (2006)use PS methods to reexamine the long-
term performance of stock returns after SEOs. Both papers find that while characteristic-
by-characteristic matching results in significant long-term abnormal returns after SEOs,
abnormal returns are insignificant if one uses propensity score based matching methods
instead.
Cheng (2003)studies SEOs offered between 1970 and 1997 for which necessary
COMPUSTAT data are available on firm characteristics. She finds significant buy-and-
hold abnormal returns of between−6% and−14% over three to five years in the full
sample and various sub-samples when matches are constructed on size, industry and
book-to-market. She then uses three logit models, one for each decade, to predict the
probability of issuance. Several firm characteristics such as size, book-to-market, in-
dustry, R&D, exchange, as well as 11-month past returns predict the issuance decision.
Cheng matches each issuer with a non-issuer in the SEO year with a similar propen-
sity score (i.e., predicted probability). She finds little evidence of significant abnormal
returns except for one sub-sample in the 1970s.
Li and Zhao undertake an exercise similar to that inCheng (2003)for issuers from
1986 to 1997. They show that characteristic-by-characteristic matching produces inade-
quate matches between issuers and non-issuers in terms of average size.^30 They estimate
propensity scores with size, book-to-market, and past returns in three quarters prior to
issuance, one model per year, and add interaction terms for better predictions and delete
firms as necessary to have a common support. In their final sample, conventional match-
ing gives average three-year buy-and-hold abnormal returns of−16%, but this drops to
an insignificant−4% with PS matching.
(^30) Medians are not reported, so it is hard to assess the role of outliers.