Selection Methods of Control Samples:
A Comparison of Two Matching Methodologies
Jong Hwan Yi
California State University, Los Angeles
School of Business and Economics
5151 State University Drive
Los Angeles, CA 90032
Tel: 323-343-2877
Fax: 323-343-2885
Email:
Selection Methods of Control Samples:
A Comparison of Two Matching Methodologies
Abstract
To
measure abnormal stock returns of a sample of firms in an event study, we often
use matching firm adjusted returns where returns of the control firms are
subtracted from the raw returns of the sample firms. In most financial studies, the control firms are selected by matching
industry and size (the I/S method).
That is, for each sample firm, a matching firm with the closest market
capitalization within the same 3-digit Standard Industry Classification (SIC)
code is selected. In this study, an
alternative control firm selection method based on earnings per share (the EPS
method) is compared to the traditional method.
The EPS method matches each sample firm with a control firm that has the
same EPS for a given fiscal year. While
the mean matching firm adjusted returns provided by the two methods are close
to the expected value of zero, the size of variances of the adjusted returns is
somewhat smaller for the I/S method, showing some superiority for the
traditional matching method based on industry and size.
Selection Methods
of Control Samples:
Introduction
To
measure abnormal stock returns of a sample of firms in an event study, we often
use matching firm adjusted returns. A
matching firm adjusted return is defined as the return of a sample firm minus
the return of its control firm over the same period. In most financial studies, these control firms are selected by
matching the industry and size of the sample firms (the I/S method). That is, for each sample firm, a matching
firm with the closest market capitalization within the same 3-digit Standard
Industry Classification (SIC) code is selected. Since two firms in the same industry are likely to be subject to
the same industry conditions, matching industry can isolate any
industry-specific factors that affect the stock returns of sample firms. Similarly, matching firm size attempts to
isolate any factors that can affect companies of certain size. An early empirical study of the size effect
is Banz (1981), who reports that small firms tend to have higher stock
performance than larger firms.
There are
numerous empirical studies that control for industry and size effects including
Ritter (1991) and Spiess and Affleck-Graves (1995). In these studies, matching firms are those that have the closest
firm size (usually proxied by the market capitalization) within the same
industry. Other studies, such as
Loughran and Ritter (1995) and Brav and Gompers (1995) control for the size of
the firm only.
This study tests an alternative
way of selecting matching firms for a sample of firms: matching based on
earnings per share (the EPS method).
While the conventional method of matching based on industry and size
tries to control for possible effects specific to certain industries and size,
matching based on earnings per share tries to control for the current level of
EPS in assessing abnormal returns of a stock relative to another. To the extent to which future stock returns
are correlated with current EPS, controlling for current EPS would isolate the
EPS effect that would contribute to any possible abnormal returns. Some of the studies that study the
relationship between earnings and stock returns include Basu (1983), Kim
(1997), and Chia, Czernkowski and Loftus (1997).
Obviously, in an event study
measuring abnormal stock returns, we would have more confidence in any
inferences that we might draw from a better methodology. The question of, which of the two matching
methods is more appropriate, seems to be largely an empirical one. Thus, we conduct an experiment using randomly
selected sample firms with two sets of matching firms based on the two methods
described above. The superior method
would be the one that produces a set of matching firms with closest returns to
those of the sample firms. One way to
test this is to measure the matching firm adjusted returns (the returns of
sample firms minus the returns of matching firms) over various holding
periods. Since the sample firms are
randomly selected from the general population, a superior method would be
defined as the one that yields matching firms with mean returns closest to zero
and the smallest variances.
The remaining sections are as
follows. In section 2, we describe the
selection process of sample firms and matching firms using the two
methodologies. Section 3 presents the
analysis and results. Section 4
summarizes and concludes the paper.
Selection
of sample firms and matching firms
In our experiment,
we first construct a set of sample firms by randomly selecting 500 firms from
both the NYSE and NASDAQ CRSP files. We
restrict our population to those that have publicly traded at least 5 years to
avoid any initial public offerings, which are known to have both short-run and
long-run abnormal returns. A summary of
abnormal short-run returns and long-run abnormal returns of initial public
offerings are reported in Smith (1986) and Ritter (1991), respectively. We also exclude firms with negative earnings
for any sample year because they may be under special circumstances, for
example, financial distress. Firms with
financial distress may be subject to some abnormal returns and may bias the
results of the study. For each of the 5
years during the 1987-1991 sample period, we randomly choose 100 firms. We study 5 sample years rather than a single
year for robustness of the test. For
the 500 firms in the sample, we construct two sets of matching firms using the
two different methods. The matching
firms are also selected from the same population as for the sample firms without
replacement.
The first
set of firms is selected using the traditional method of matching by industry
and size (the I/S method). Our
selection process using the I/S method discussed here is similar to the one
used in Ritter (1991). For industry, we
use the Standard Industry Classification (SIC) code. The market capitalization of a firm, which proxies for the firm
size, is defined as the market price per share times the number of shares
outstanding. For each sample firm, we
select a matching firm that has the closest market capitalization within the
same 3-digit SIC code. For example, a
sample firm in 1990 is matched by a firm in the same 3-digit SIC code that has
the closest market capitalization at the end of the same year 1990.
From the
500 pairs of sample and matching firms, we drop the firms (either sample or
matching firms) that are delisted from CRSP return files before the three-year
anniversary (the longest holding period examined in this study) since the first
date of return calculation. This avoids
any selection bias that may result from including firms that would be merged,
bankrupt or liquidated. The final
numbers of observations are 394 pairs of firms for the I/S method.
Table 1 compares the market
capitalizations of the samples firms with those of the matching firms selected
by I/S method. The Market
capitalization is defined as the stock price at the first calendar day of a
year times the number of shares outstanding.
The mean (median) values are 150.9 (66.9) million dollars for the sample
firms and 133.2 (62.7) million dollars for the matching firms. The slightly greater numbers for the sample
firms are seen in every sample year except 1989. The difference, however, is not sizable and does not seem to
cause any significant bias in the results.
Table 1
Market capitalizations of
sample firms and matching firms based on industry and size
The
sample firms are randomly selected from CRSP tapes between 1987 and 1991. The matching firms are selected by using the
industry and size (I/S) method. Market
capitalization is the stock price at the first calendar day in a year times the
number of shares outstanding.
|
|
|
Sample Firms |
Matching Firms (I/S) |
|
Year |
Number of Pairs |
Mean (median) Market Cap. ($ millions) |
Mean (median) Market Cap. ($ millions) |
|
1987 |
81 |
145.5 (69.9) |
131.9 (54.8) |
|
1988 |
75 |
145.3 (60.4) |
123.4 (48.4) |
|
1989 |
88 |
96.8 (67.8) |
102.4 (74.2) |
|
1990 |
78 |
174.2 (70.2) |
167.2 (68.5) |
|
1991 |
72 |
160.8 (70.0) |
145.7 (69.1) |
|
Total |
394 |
150.9 (66.9) |
133.2 (62.7) |
We choose the second set of control
firms using the EPS matching method.
That is, for each sample firm, we select the matching firm that has the
same EPS for the same fiscal year.
Since there were many potential matching firms that had the same EPS for
a given sample firm, we randomly selected one with the same EPS. After excluding the firms that have been
delisted before three years, there were 354 pairs of sample and matching firms
with the same EPS. The sample is
smaller using the EPS method, compared to 394 for the I/S method, mainly
because a larger number of matching firms selected by this method was delisted
before three years. Another reason for
the different sample size is that in the I/S method, there were more cases
where both the sample firm and the matching firm were delisted before the three
years.
Table 2 shows the mean and median
EPS for the sample firms and the matching firms selected by using the EPS
method. Since each and every sample
firm was matched by a firm that had the same EPS, both the mean and the median
values are identical for both set of firms at 0.67 and 0.48 dollars,
respectively.
Table 2
Descriptive statistics of
sample firms and matching firms based on earnings per share
The
sample firms are randomly selected from CRSP tapes between 1987 and 1991. The matching firms are selected by using the
earnings per share (EPS) method.
|
|
|
Sample Firms |
Matching Firms (EPS) |
|
Year |
Number Of Pairs |
Mean (median) EPS ($) |
Mean (median) EPS ($) |
|
1987 |
78 |
0.67 (0.48) |
0.67 (0.48) |
|
1988 |
74 |
0.62 (0.45) |
0.62 (0.45) |
|
1989 |
75 |
0.64 (0.54) |
0.64 (0.54) |
|
1990 |
68 |
0.71 (0.47) |
0.71 (0.47) |
|
1991 |
59 |
0.67 (0.50) |
0.67 (0.50) |
|
Total |
354 |
0.67 (0.48) |
0.67 (0.48) |
Analysis
and Results
In order to compare the two
methods, we use a simplified version of the simulation event study used in
Brown and Warner (1980). Brown and
Warner were interested in finding the relative power of different event study
methodologies. Their methodology
involved taking a random sample of stocks and introducing artificial abnormal
performance.
To compare the two methods, we
construct two series of matching firm adjusted holding period returns based on
dividend and split-adjusted stock prices.
The matching adjusted holding period returns are defined as the holding
period return of a sample firm minus the holding period return of its matching
firm. The four holding periods we
examine are 6 month, one year, two years and three years beginning the first
trading day of each calendar year.
Those are
ARET6 = Rsample,6 –
Rmatch,6
ARET12 = Rsample,12
– Rmatch,12
ARET24 = Rsample,24
– Rmatch,24
ARET36 = Rsample,36
– Rmatch,36
where Rsample,j is the j-month holding period returns
of a sample firm
Rmatch,j is the j-month holding period
returns of its matching firm
and ARETj is the j-month matching firm adjusted return.
Because
this experiment involves a random selection of seasoned stocks, we would expect
to find no abnormal performance of our sample firms (zero mean adjusted
returns) if our matching schemes are appropriate. Our objective is to find the method that provides a better fit to
the returns of our sample firms, that is we are searching for the method whose
adjusted return series have a mean of zero and a smaller variance of the mean.
Table
3 presents the first comparison of the two methods. For holding periods of 6, 12, 24, and 36 months, we report the
mean and median matching firm adjusted returns using of the two methods. While the mean returns provided by the two
methods are close to the expected value of zero, it appears that for each of
the four holding periods the I/S method produces mean returns closer to zero in
absolute value. For the I/S method, the
mean returns for the four periods are 0.0048, 0.0042, -0.0143 and –0.0003,
respectively. The mean returns for the
EPS method are 0.0154, 0.0111, -0.0279 and –0.0096. In terms of the median adjusted returns, each of the methods
yields lower absolute values in two of the four holding periods. In any case, all the mean and median returns
are not significantly different from zero, and thus neither method can be
chosen as the superior one based on them.
Table 3
Comparison of mean (median)
matching firm adjusted returns and their variances of the two matching methods:
industry and size versus earnings per share
This
table reports the mean (median) matching firm adjusted returns and their
variances of a s ample of seasoned firms using two different methods of
selecting the control sample. The
industry and size method selects a matching firm by choosing the firm with the
closest market capitalization among the firms that share the same 3-digit SIC
code. The earnings per share method
selects a matching firm based on the EPS of a fiscal year. ARET6, ARET12, ARET
24 and ARET36 are 6-month, one-year, two-year and three-year
matching firm adjusted holding period returns, respectively. F-statistic is obtained by dividing the
larger variance by the smaller variance.
This assumes that the two samples are independent.
|
Abnormal Returns |
Industry and size (n=394) |
Earnings per share (n=354) |
F-statistic |
||
|
|
Mean (median) |
Variance of mean |
Mean (median) |
Variance of mean |
|
|
ARET6 |
0.0048 (0.0084) |
0.07456 |
0.0154 (0.0134) |
0.05849 |
1.27 |
|
ARET12 |
0.0042 (0.0091) |
0.07022 |
0.0111 (0.0069) |
0.04823 |
1.46* |
|
ARET24 |
-0.0143 (0.0018) |
0.05688 |
-0.0279 (-0.0133) |
0.06678 |
1.17 |
|
ARET36 |
-0.0003 (-0.0048) |
0.04130 |
-0.0096 (-0.0001) |
0.05220 |
1.26 |
* The
two variances are significantly different from each other at the 5% level.
We now turn to the variances of
the mean adjusted returns presented in Table 3. An examination of the variances shows that for two of the four
periods (6 and 12 month), the EPS method yields smaller variance. However, as is the case with the mean
returns, the variances are close to each other. To formally test the difference of variances, F-statistics can be
computed by dividing the larger variance by the lower variance. This test shows that, except for the
12-month period in which the EPS method has a smaller variance, there is no
significant difference between the variances of the mean returns. Moreover, since the returns produced by the
two methods are not likely to be independent from each other, the F-test may be
biased. Therefore, the overall evidence
presented in Table 3 does not suggest that either matching method is superior.
Another
way to compare the predictive ability of the two matching methods is to examine
the performance of the two methods at the individual firm level. Comparing the performance of the two methods
for each of our sample firms, we should prefer the method that more frequently
produces smaller (in absolute value) adjusted returns. Again, this criterion is based on our
expectation for the adjusted returns to be zero.
For the comparison described
above, we would need the same sample size for the two methods. Thus we use the 354 sample firms with
matching firms selected by both the I/S and the EPS methods. All these firms have complete return data
for three years examined. For each
method, we compute the adjusted returns over the holding periods of 6, 12, 24
and 36 months. Then, for each firm at
each holding period, we observe which of the two methods produces a lower (in
absolute value) adjusted return. The
method that more frequently produces lower adjusted returns would be the
superior method. If the two methods are
equally good ways to choose the control firms, we would expect that each of the
two methods would have smaller abnormal returns for 50% of the firms.
The results are presented in Table
4. The results show that for each of
the four holding periods, the I/S method has lower adjusted returns more than
50% of the time. The percentage ranges
from a low of 56.2% (36 month holding period) to a high of 64.4% (24 month
holding period). If we assume that the
returns produced by the two methods are independent from each other, a simple
binomial test can be used to see if these percentages are significantly
different from the 50%. The test shows
that the difference is significant at the 5% level for each holding
period. Although to the extent of any
cross-sectional correlation in the security returns, the tests may be biased,
the overall results suggest that the traditional matching method based on
industry and size is superior.
Table 4
Comparison of two matching
methods at individual firm level
This
table reports the frequency for which the industry and size matching method
produces a smaller adjusted return (in absolute value) than the EPS matching
method. Adj (I/S) is the absolute value
of an individual firm’s adjusted return when matching on industry and size is
used. Adj (EPS) is the absolute value
of an individual firm’s adjusted return when matching on earnings per share is
used. A total of 354 sample firms are
used in the experiment.
|
Holding Period |
Number of firms in the sample |
Number of firms with Adj (I/S) < Adj (EPS) |
% of firms with Adj (I/S) < Adj (EPS) |
|
6-month |
354 |
219 |
61.9* |
|
One-year |
354 |
206 |
58.2* |
|
Two-year |
354 |
228 |
64.4* |
|
Three-year |
354 |
199 |
56.2* |
* Significantly
higher than 50% by binomial test at 5% level.
Summary
and Conclusion
In this study, we run an
experiment to compare two selection methods of control sample. To measure abnormal returns of a sample of
stocks in an event study, we often use matching firm adjusted returns where
returns of the control firms are subtracted from the raw returns of the sample
firms. In most financial studies, the
control firms are selected by matching industry and size (the I/S method). An alternative matching method based on
earnings per share (the EPS method) is compared to the traditional method. The EPS method matches each sample firm with
a control firm that has the same EPS for the same fiscal year. While the mean matching firm adjusted
returns provided by the two methods are close to the expected value of zero,
the size of variances of adjusted returns is somewhat smaller for the I/S
method. Also, at the individual firm
level, the I/S method produced a larger number of matching firms whose holding
period returns were closer to the returns of the sample firms, showing some superiority
for the traditional matching method based on industry and size.
References
Banz, Rolf, 1981, The Relationship
between Return and Market Value of Common Stocks, Journal of Financial
Economics 9, 3-18.
Basu, Sanjoy, 1983, The
Relationship between Earnings Yield, Market Value, and Return for NYSE Common
Stocks: Further Evidence, Journal of Financial Economics 12, 129-156.
Brav, Alon and Paul Gompers, 1997,
Myth or Reality? The Long-Run
Underperformance of Initial Public Offerings: Evidence from Venture and
Nonventure Capital-backed Companies, Journal of Finance 52, 1791-1821.
Brown, Stephen and Jerold Warner,
1980, Measuring Security Price Performance, Journal of Finance 8,
205-258.
Chia Y, R. Czernkowski and J.
Loftus, 1997, The Association of Aggregate and Disaggregatd Earnings with
Annual Stock Returns, Accounting and Finance 37, 111-128.
Kim, Dongcheol, 1977, A
Reexamination of Firm Size, Book-to-Market, and Earnings Price in the
Cross-Section of Expected Stock Returns, Journal of Financial and
Quantitative Analysis 32, 463-489.
Loughran, Tim and Jay Ritter,
1995, The New Issues Puzzle, Journal of Finance 50, 23-51.
Ritter, Jay, 1991, The Long-Run
Performance of Initial Public Offerings, Journal of Finance 46, 3-27.
Smith, Clifford Jr., 1986,
Investment Banking and the Capital Acquisition Process, Journal of Financial
Economics 15, 3-30.
Spiess, K and J. Affleck-Graves,
1995, The Long-Run Performance following Seasoned Equity Offerings, Journal
of Financial Economics 38, 243-267.