SSRN Id559007 PDF

Event Study under Noisy Estimation Period
by Nihat Aktas, Eric de Bodt, Jean-Gabriel Cousin April 10, 2003
Abstract Since the contribution of Fama, Fisher, Jensen and Roll (1969), event studies have become an important reference tool for empirical research in finance. The original methodology has been improved in order to tackle numerous problems such as event-date uncertainty, event clustering, event-induced variance phenomena Somewhat surprisingly, the determination of the estimation period has attracted less interest. It remains most frequently routinely determined as a fixed window prior to the event announcement day, during which it is supposed that no other significant events have happened. In practice, in large sample studies, validation of this assumption on a case-bycase basis is out of reach, despite the fact that it is known to be violated for some specific corporate events. The case of merger and acquisitions, in particular the behavior of bidders who make repetitive acquisitions (and acquisition attempts), is a typical example. In this work we propose an adaptation of the basic methodology by taking explicitly into account the likely existence of firm-specific events during the estimation period. We first realize a standard specification and power analysis, following the Brown and Warner (1980, 1985) scheme. We then show that the proposed method changes significantly the inferences realized on a sample of around 580 merger and acquisition operations concerning the bidders abnormal returns.
Aktas Universit Catholique de Louvain 1 place des Doyens 1348 Louvain-la-Neuve BELGIUM 32-10-478-447 32-10-478-324 aktas@fin.ucl.ac.be de Bodt Universit de Lille 2-Esa 1 place Dliot - BP381 59020 Lille Cdex FRANCE 33-3-2090-7477 33-3-2090-7629 edebodt@hp-sc.univ-lille2.fr Cousin Universit de Lille 2-Esa 1 place Dliot - BP381 59020 Lille Cdex FRANCE 33-3-2090-7702 33-3-2090-7629 jgcousin@wanadoo.fr
Address Voice Fax e-mail
Acknowledgements We are grateful for constructive comments and suggestions from the participants of the SIFF 2002 (Rennes September) , AFFI 2002 (Paris December) meetings and Paris IX-Dauphine Research Seminar.
1. INTRODUCTION
Since Fama, Fisher, Jensen and Roll (1969), hereafter FFJR, event studies have become a standard empirical research methodology in finance. We will here take for granted that the FFJR approach is well known to the reader1. Applications have been so numerous that it would be impractical to try to list them exhaustively. Many suggestions have also been put forward to improve the basic FFJR scheme. For example Brown and Warner (1980, 1985) analyze the specification and power of several modifications of the FFJR scheme; Ball and Torous (1988) explicitly take into account the uncertainty about the event dates; Corrado (1989) introduces a robust test of significance; Boehmer et al. (1991) adapt the methodology in order to tackle the event-induced variance phenomenon; and Salinger (1992) systematically defines the correct abnormal returns standard errors, with and without event clustering, taking into account the correction factor due to the forecasting nature of the estimated abnormal returns during the event window. More recently, concerning the clustering phenomenon affecting the cross-sectional dependence of abnormal returns, Aktas et al. (2003) use the portfolio formation procedure when firms undergo the event on the same day (the event windows overlap perfectly) and advocate the use of a bootstrap method when the clustering is only partial. The estimation period, used to fit the parameters of the chosen return generating process, has been less extensively analyzed. It is most often defined as a period preceding the event, sufficiently long to estimate properly what should happen in the absence of an event. In studies using daily data, a window going from day -250 to day -30 relative to
See Chapter 4 of Campbell et al. (1997) for a review.
Event Study under Noisy Estimation Period, April 10
the event day is usually chosen, somewhat arbitrarily. A shorter period can then be excluded between the end of the estimation period and the beginning of the event period, if we wish to neutralize the impact of information leakages (or rumors) before the announcement. This mechanical choice of the estimation period is, however, not free of questions, particularly in the framework of large-sample empirical studies, which are currently more and more frequent (see e.g. Fuller et al. (2002), Mitchell and Stafford (2000)). When compiling the data for several hundreds (or thousands) of observations, it becomes impractical to analyze the estimation period, on a case-by-case basis, in order to be sure that it corresponds to a normal period, without any other significant firmspecific perturbing events. As we will show it, this generates a significant risk of bias to the analysis. Let us take the case of mergers and acquisitions (M&A) as a practical example. Imagine that the bidder has, during the months preceding a specific operation, realized other operations, as appears frequently to be the case2. The existence of such firm-specific
events in the estimation window will definitely affect the estimated parameters of the return generating process and, in this way, the estimated normal returns during the event window. Maybe more importantly, it will also impact significantly the variance of the return generating model residuals, which is frequently used to test the statistical significance of the observed abnormal returns around the event day3. Our results show that this reduces the power of the event study methodology (even when including the
2
Malatesta and Thompson (1985) have already pointed out that bidders frequently follow an acquisitions
program strategy, and suggested an adapted methodology to evaluate its global impact. The recent contribution of Fuller et al. (2002) emphasizes the importance of this point.
3
See Salinger (1992) for a systematic analysis of the ways to build a correct test of significance.
recent refinements proposed in the literature). In other words, it reduces its ability to detect abnormal returns when there are abnormal returns. This paper addresses the issue. We first suppose that, essentially for practical reasons, it is almost impossible to verify manually and systematically, on a case-by-case basis and for all the observations included in the sample, that there are non-firm-specific events perturbing the estimation period. We then propose an adaptation of the event study methodology in a way that automatically takes into account the potential presence of firm-specific perturbations in the estimation period. Our approach is essentially based on a combination of the well-established market model of Sharpe (1963) and the more recent Markov Switching Regression models (MSR), largely introduced and developed by Hamilton (1989, 1994), and significantly extended in Krolzig (1997). Our initial intuition is simple: the occurrence of firm-specific events has a significant impact on the firms return generating process, in particular on the variance of the generated returns. We will try to capture this perturbation by a switching regime model. We will then use the estimated parameters of the normal regime (which should correspond to the estimation window as initially defined by FFJR, a period without extraordinary events) as parameters to conduct the statistical analysis of abnormal returns during the event window. Our approach can be understood as a statistical filtering of the data, allowing us to neutralize perturbing events present in the estimation period, without requiring a manual case-by-case analysis. Another way to interpret our proposition is to see it as a better-specified return generating model, which takes into account the probability of firmspecific events occurrence, and therefore leads to statistical tests that are both better specified (this is a key point of our results) and more powerful. In fact, the way that we
propose to model the normal return is somewhat in line with Rolls (1987) intuition. According to Roll, the true return generating process seems to be better described by a mixture of two distributions: one corresponding to a state of information arrival, and the other to the normal return behavior.
The adopted approach in this paper is now classical in the field of event study methodology, and follows that of Brown and Warner (1980, 1985) or Boehmer et al. (1991). In a first step, using a large sample of firms (around 1,500 US quoted firms included in the S&P 1,500 index), we realize a specification and power analysis of several alternatives for the classical event study approach under a noisy estimation period. In particular, we show that when we shock the estimation period by introducing simulated abnormal returns, the results obtained using the standard approaches become badly specified. We also introduce a statistical test of difference of specification and power between the compared approaches. We show the importance of such a test before concluding as to the superiority of a specific methodological proposition. This point has been overlooked in almost all previous contributions in the field (e.g., Brown and Warner, 1980, 1985; Boehmer et al., 1991). We then show that our proposition seems generally to be more robust to such alterations. We finally illustrate, using a large dataset in the field of M&A, that taking into account the repetitive nature of the bidders behavior significantly modifies the inferences drawn from the bidders abnormal returns. In the light of the generally accepted result, that on average and for a global sample, bidders do not undergo significant abnormal returns, this is without doubt the main contribution of our paper.
The next section of this paper is devoted to a short review of the classically proposed event-study approaches. Using the same notation, we introduce our approach and summarize some of the features of the MSR family of models. Section 3 is dedicated to the simulation work. We present the dataset used, the methodology followed to realize specification and power analyses, and our results. Section 4 presents the results obtained when applying our propositions to a real sample of European M&A operations. We report conclusions in Section 5.
2. EVENT STUDY METHODOLOGY
The seminal contribution of FFJR has been the starting point for an impressive diffusion of the event study methodology in empirical finance. Its constitutive steps are wellknown: determination of the event, determination of its announcement date, determination of the event and estimation windows, estimation of the return-generating model parameters with the estimation window dataset, computation of the abnormal returns (the residuals of the normal model during the event window), and if required, computation of the cumulative abnormal returns, averaging of the abnormal returns on the sample and construction of a statistical test of their significance. In this section, we focus on the choice of the return generating process and construction of the statistical test of significance, the two key points of concern in this work.
2.1. Return generating model
In the classical framework of event-study methodology, abnormal returns are defined as the forecast errors of a specific normal return-generating model. The classical Market Model (MM), introduced by Sharpe (1963), is the most frequently chosen model4:
R = + R +
jt j j mt
jt
(1)
where Rjt and Rmt are the return of asset j and the return of the market index at time t respectively. The residuals, jt, provide the estimation of the abnormal returns. Initially, the residuals were classically supposed to be identically and independently normally distributed (NIID). Numerous contributions have dealt with violations of these hypotheses. For example, Ruback (1982) suggests a way to cope with the existence of first-order auto-correlation in asset returns that is easy to implement. As our work focuses specifically on the NIID hypotheses, Section 2.2 is dedicated to it. The estimation of the MM parameters is realized by OLS.
2.2. Statistical test of significance In order to introduce the different approaches that will be submitted to shocks during the estimation period, we will use the same set of notations as Boehmer et al. (1991): 4
N: number of firms in the sample;
Brown and Warner (1980, 1985) have shown that the results of short-term event studies are not sensitive
to the choice of a specific return-generating process. These results are confirmed by Cowan and Sergeant (1996), who show that the use of Scholes and Williamss (1977) approach adds little. More recently, Aktas et al. (2003) find that there is no significant difference when comparing the results obtained with the constant mean model return, the market model or the Scholes and Williams approach, on a European sample of business combination announcements.
AjE: abnormal return of firm j at the event date; Ajt: abnormal return of firm j at date t; T: number of days during the estimation period; TE: number of days during the event period;
R m : average return of the market index during the estimation period;
S j : standard deviation of firm js abnormal returns during the estimation period;
SR jE : standardized abnormal returns of firm j at the event date, corrected in order

to take into account the forecasting nature of the estimated abnormal returns (see Boehmer et al. (1991) or Salinger (1992)). They are computed as:
SR jE = A jE
_ ^ ( R mE R m ) 2 1 + T S j 1 + _ T ( R mt R m ) 2 t =1
(2)
The Boehmer et al. (1991) standardized residuals test (BOEHMER)

As for each of the following tests, the authors test the null hypothesis of no crosssectional average (cumulative) abnormal returns around the event date. The BOEHMER test is similar in spirit to that of Patell (1976), but the authors propose the use of estimated cross-sectional variance of standardized abnormal returns instead of the theoretical one. This adaptation of the Patell (1976) approach allows account to be taken of the observed increase in both the time-series abnormal returns variance (securitys market model disturbance variance for Harrington and Shrider (2002)) and the cross-
sectional abnormal returns variance (effect size for Harrington and Shrider (2002)) around the event date (called by the authors the event-induced variance phenomenon, which is due to firm-specific reactions to the event announcement)5. The BOEHMER test takes the following form:
Z=
1 SR jE N j =1 1 N(N 1) SR (SR jE iE )2 j =1 i =1 N
N N
(3)
Boehmer et al. (1991) show that as defined, the test is well specified for a sample of firms selected from the NYSE-AMEX universe under the null hypothesis (no average abnormal returns), even if the cross-sectional variance increases during the event period. When there is no such increase, the BOEHMER test, moreover, is as powerful as the classical Patell (1976) standardized residuals test. Campbell and Wasley (1993) show the specification problem affecting the Patell (1976) test in case of event clustering (absence of event independence). The same limit affects, for the same reason, the BOEHMER test. The BOEHMER method has been so frequently used in empirical tests since its introduction that it has become the standard reference. We use it in the sequel as benchmark.
For a detailed analysis of the event-induced variance phenomenon, see Harrington and Shrider (2002). 8
The Corrado (1989) rank test (RANK)

Corrado (1989) introduces a test based on the ranks of abnormal returns, by this making any hypotheses concerning the abnormal returns distribution unnecessary (except that of independence among the observations). For each firm in the sample, the RANK test merges the estimation and event windows abnormal returns in a unique time series. Abnormal returns are then sorted and a rank is assigned to each day. Let Kjt be the rank assigned to firm js abnormal return on day t. By convention, rank one is assigned to the lowest observed abnormal returns. By construction, the mean rank is half the number of days in the constituted series of abnormal returns (the number of the days of the estimation period plus the number of days of the event period) plus one half. We denote it K . The RANK test then takes the following form:
1 N
(K
j =1
jE
K)
T=
S (K )
(4)
where the standard error, S(K), is :
S (K ) =
1 250
T +TE
t =1
1 N
(K
j =1
jt
K )) 2
(5)
As usually, the use of ranks neutralizes the impact of the form of the abnormal returns distribution (skewness, kurtosis, outliers,). Corrado (1989), Corrado and Zivney (1992) and Campbell and Wasley (1993) show using simulations that this test is generally well specified and robust. Campbell and Wasley (1993) also report that it seems to be robust
to the event-clustering phenomenon, when using the market model as return generating process. As the RANK test has also been classically used as a robust alternative to the BOEHMER, we will also use it in the sequel as a benchmark.
Other suggestions
Numerous other modifications of the initial FFJR methodology have been proposed. We briefly mention two of them because they could appear, at first sight, interesting alternatives to our propositions. Ball and Torous (1988) study the case of event date uncertainty. Using a maximum likelihood estimator, they simultaneously estimate, for each day of the event window, the abnormal returns, their variance and the probability of an event. Using simulations, the authors show that their approach is more powerful (more frequently detecting simulated abnormal returns) than the classical ones when the event date is uncertain. Conceptually, such an approach could be adapted to our problem (the potential presence of shocks during the estimation period). This approach seems attractive at first sight. While Ball and Torous (1988) estimate the probability of an event during the event window, we could, using the same approach, estimate the probability of an event during the estimation window. Days for which this probability is too high could be neutralized. In practice however, a thorough examination of the Ball and Torous approach reveals that the increase in the number of days in the studied period,6 and of the number of potential events,7 makes it computationally intractable.
6 7
The estimation period is classically far longer than the event period. We cannot exclude the presence of several firm-specific events during the estimation window.
10
Nimalendran (1994) introduces an approach based on a return generating process that combines a Poisson process (in order to capture jumps in the distribution of returns) with a standard Brownian motion. The authors intuition is quite close to ours: the Poisson process allows dissociation of the normal behavior (the Brownian motion) from the days of information arrival. His findings suggest that, for an event with multiple announcements spread over a significant period of time, the proposed approach is more powerful than the classical ones. Again, a careful examination of the Nimalendran propositions reveals that it is not well-adapted for a short event period as the one on which we focus. Either the number of observations is too small to realize any estimation or the event window becomes too large to attribute the observed abnormal returns to a single event with confidence.
The Markov Switching Regression test (MSR)

The intuition on which the MSR test is based is quite straightforward: we anticipate that firm-specific events will change the return variance. It could be argued that it is in contradiction with the semi-strong form of efficiency hypothesis, under which the prices should adjust immediately to any public information announcement. It is in fact not the case if we take into account the uncertainty attached to firm-specific events. Let us take again the example of M&A. The announcement of a takeover bid does not guarantee its success. The initial announcement could be followed by the announcement of competitors bids, bid price revisions, target initiatives to block the operations, anti-trust authorities interventions All of this will generate a strong increase in the variance during the event period. Let us now imagine that such firm-specific events occur during
11
the chosen estimation period. Classical tests, such as the BOEHMER or the RANK tests, will in fact overestimate the residuals variance during the estimation period, leading to a downward bias (less powerful) in the test of significance during the event window. To deal with this bias, the MSR test that we introduce uses the Markov Switching Regression approach, largely introduced and developed by Hamilton (1989, 1994). We will suppose that the return generating process can be adequately modeled using a tworegime process8, one regime with normal variance and one regime with high variance (firm-specific event regime). In both regimes, the MM parameters are assumed to be the same. The return generating process is therefore the following:
R jt = j + j Rm,t + j ,1,t if St = 1 with j ,1,t following N (0, j ,1 ) R jt = j + j Rm,t + j ,2,t if St = 2 with j ,2,t following N (0, j ,2 ) and j ,2 > j ,1
(6)
where St is an indicator variable taking value 1 if we are in the low variance regime and 2 if we are in the high variance regime. In the sequel, we will refer to equation (6) as the two-state market model (TSMM). The proposed model is a direct and parsimonious extension of the classical MM. As the regime state variable St is not directly observable (recall that we assume the practical impossibility of verifying manually, on a case-bycase basis, the presence of firm-specific events during the estimation period), we have to
This hypothesis is supported by unreported results where, using the formal equivalence between ARMA
models and some families of Markov Switching Auto Regressive models developed in Krolzig (1997), we find that a two-regime model is an adequate representation of the return generating process, or three regimes in the presence of strong outliers.
12
specify its statistical properties. We rely for this on the Markov Switching Regression approach: St is described by a first-order Markov process (St depends only on St-1). The Markov chain is therefore defined by four transition probabilities:
P( S t = 1 S t 1 = 1) = p11 P( S t = 1 S t 1 = 2) = p21 P( S t = 2 S t 1 = 1) = p12 P( S t = 2 S t 1 = 2) = p22
(7)
where pij is the probability of regime i in t-1 and j in period t. All probabilities have to be positive and, for i equal to 1 and 2, pi1+pi2 must be equal to one. The model we propose is therefore based on the estimation of six parameters (, , 1, 2, p11 and p21) and, while much more flexible than the classical MM model (as we will stress in the next paragraph), it remains very parsimonious. The estimation of Markov Switching Regression models is fully presented in Hamilton (1994). It is based on a maximum likelihood approach, for which an efficient estimation algorithm has been developed9. The estimated probability of being in a specific state at a specific date is one of the interesting by-products of the advocated approach. It allows one, for some specific cases, to look for the reason why an increase in variance is observed. In other words, by looking at a period of high probability of a regime of high variance, it is possible to verify that these clusters of variance have indeed been generated by firm-specific events. Figure 1 presents a typical plot of the residuals and the estimated
Hamilton (1994) (p. 688-689) presents an Expectationminimization algorithm that proves to be
numerically very efficient.
13
regime probabilities10 obtained on a real M&A case (the acquisition of ICI Nylon activities by DuPont on the April 23, 1992). The analysis is realized on DuPont. The event is situated at day 235. The standard errors of regime 1 and regime 2 are respectively 0.99% and 1.6%. The event is clearly situated in regime 2, the one of high variance. The number of days respectively classified in regime 2 (1) are 77.8 (163.2) days. Interestingly, we see that the period between day 150 and day 210 is also classified in regime 2. We have checked the financial press to see whether firm-specific events could explain this highly volatile time period. Our investigation suggests that this is linked to the announcement of a deceptive result for the 4th quarter of 199111. Our MSR test is a straightforward adaptation of the BOEHMER test obtained by modifying the standardization procedure of the abnormal returns (equation 2). We now divide the estimated abnormal returns by their estimated standard deviation in regime 1 (low variance). We do not have to take into account, as in the BOEHMER test, the forecasting nature of the abnormal returns during the event window, the estimation of the TSMM being realized on the whole period (estimation window plus event window), as for the RANK test. Standardized residuals therefore take the following simple form:
SR jE = A jE / S j ,1
(8)
is the estimated standard deviation in the low variance regime. Contrary to where S j ,1
what might be expected at first sight, the use of the low-regime standard deviation does
10
The represented probabilities are the called smoothed probabilities, and are evaluated using the full set of
available observations.
11
On January 30, 1992, the Financial Times reports a loss for DuPont of US$ 240 millions.
14
not automatically produce a higher cross-sectional student t-statistic, as demonstrated in the homoscedastic case in Appendix 1. Note also that, for specification analysis, we use the average standard deviation of the two regimes with the probabilities of each regime as weight12. The use of this Markov Switching Regression model provides numerous interesting features. Beyond the above-quoted possibility of obtaining the estimated probability of being in a specific regime at a specific date, this specification is in line with Rolls (1987) intuition. In his Presidential Address to the American Finance Association, he clearly highlighted that the true return-generating process is better described by a mixture of two distributions, the first one corresponding to a state of information arrival and the other to the normal return behavior13. The specification has other attractive features. The Gaussian conditional distribution of returns (equation 6) could be misleading. While the model imposes a Gaussian assumption for the return distribution in each state, as shown in Hamilton (1994) and Krolzig (1997), it allows capturing the skewness and kurtosis in the unconditional distribution. The Markov Switching Regression framework also allows for conditional heteroscedasticity without imposing a specific form on the conditional
12
Specification analysis is done under the null hypothesis of no event. As we are working on real data and
we do not know whether there has been an event at the selected date, the weighted average standard deviation indeed represents its expected value.
13
It should however be emphasized that the author also suspects that a no-news day is characterized by a
mixed distribution, which is not taken into account here.
15
dependence of the variance (as in the (G)ARCH framework14). Another interesting feature of the approach is the explicit modeling of the market model disturbance variance effect pointed out by Harrington and Shrider (2002). Finally, the estimation process provides us simultaneously with the abnormal returns, the estimated variance in each regime, the probability of being in a specific regime at a specific date, and the estimated transition probabilities15.
3. SPECIFICATION AND POWER TESTS
The investigation of the specification and power of the MSR test follows the procedure introduced in Brown and Warner (1980, 1985) and classically used since then (see e.g. Boehmer et al. (1991), Corrado (1989), Corrado and Zivney (1992), Cowan (1992), Cowan and Sergeant (1996)). In contrast with Monte-Carlo simulation, in the framework of which data are generated using a theoretical return-generating process specification, the authors build their data samples from real returns computed from the CRSP16 database. They randomly pick firms in the NYSE/AMEX/NASDAQ universe. For each firm, they randomly determine an event date and generate a simulated event by
14
Hilliard and Savickas (2000) propose an event study methodology that is built on the GARCH
framework. Aktas et al. (2001) is another example of the use of GARCH models to estimate the variance of the abnormal returns.
15
Following the advice of Professor Hamilton, all estimations presented in this paper have been realized
under the Ox econometric software, using the Krolzig MSVAR package, freely downloadable at http://www.nuff.ox.ac.uk/Users/Doornik/index.html.
16
Center for Research in Security Prices (Chicago University).
16
injecting pertubations into the real price series. Proceeding in this way avoids any assumption concerning the return generating process and resembles as closely as possible the situation faced when using event studies on a real dataset.
3.1. Data
Our firms universe is composed of the S&P 1500 index. The S&P 1500 Index is a combination of the S&P 500, the S&P MidCap 400 and the S&P SmallCap 600 indexes. As a result, it provides broad representation of the U.S. Stock Market, accounting for 87% of the U.S. stock market value. The total market value of the index at December 31, 2002 was US$ 9,135 billions. The mean (median) company market value was US$ 6,090 millions (US$ 1,239 million). This comprised 1,028 (68.5%) companies quoted on the NYSE, 462 (30.8%) on the NASDAQ and 10 (0.6%) on the AMEX. Table 1 presents the repartition of companies by sector. The three largest sectors in the index were the financials (20.10%), information and technology (14.2 %) and health care (14.6%). For all the firms included in this index, we use daily prices from January 1, 1990 to March 1, 2002. We use as market portfolio the Standard & Poors 500 index. The data were obtained from Datastream, accessed at the Universit de Lille 2.
3.2. Sample generation

All firms and event dates are randomly chosen with replacement. For each simulation, we build 250 samples of 50 firms. The estimation window length is 200 days and the event date is situated at day 250. To be included in a sample, the firm must have at least 50 available prices during the estimation window, and no missing prices in the 30 days
17
around the event date (from day 230 to day 260). Panel A of Table 2 presents some descriptive statistics on the MM and TSMM residuals. Both models produce centered residuals. The TSMM residuals exhibit a somewhat higher level of kurtosis. Panel B of Table 2 presents cross-sectional properties of sample-wide mean abnormal returns on the event day. As expected, increasing the portfolio size leads to an abnormal return distribution closer to the Gaussian distribution.
3.3. Event generation

The aim of our simulation work is to study the specification and power of the MSR test, as compared to the BOEHMER and RANK tests, when shocking the estimation window. The mean simulated abnormal returns injected into the estimation window will be 0%, +/1%, +/- 2%, and +/- 4%. We generate both positive and negative abnormal returns in order to take into account the unknown nature of the events likely to affect the estimation window. These shocks will be stochastic. As in Boehmer et al. (1991), the variance of the
2 Gaussian distribution will have the form k 2 j , where j is the estimated variance from
firm j during the initial estimation window, and k takes values 1 and 2, which allow doubling and tripling of the variance of the abnormal returns. The number and nature of events during the estimation window are determined in a twostep process. First, a random drawing from a Poisson distribution with a mean of 2 is realized. Assume that this number corresponds to *. It will represent the number of events during the estimation window. Then, we make a random draw in a uniform process for each day of the estimation window. If the probability obtained for day t is less than (*/200), we generate a corporate event at that date. The length (in number of days)
18
of this event is again randomly drawn from a Poisson distribution (but with a mean of 4 this time). Such an approach allows the generation of a random event during the estimation window17. Figure 2 presents a typical result obtained using this procedure. The plain line is the initial returns time-series and the dotted one is the returns time-series obtained after event generation. Four events are generated at dates T=37, 50, 68 and 124. The initial estimation-period standard error is 1.17% and becomes, after event generation, 1.31%.
We generate abnormal returns at the event day as in Brown and Warner (1980, 1985). A constant is added to the observed day 0 return for each security. The abnormal performances simulated are 0% for specification analysis, and + 0.5% and + 1% for the power analysis18. The new return of day 0 for firm i is denoted Ri,0 . To produce stochastic abnormal returns (the event-induced variance), each securitys day 0 return, Ri,0 , is transformed to double (k=1) or triple (k=2) its variance. For k=1, the following transformation is realized:
R'i,0 = Ri,0 +(Ri, X Ri)
(9)
where R'i,0 is the transformed return, Ri, X is the firm i return randomly selected from the
estimation period, and Ri is the securitys average return in the estimation period. Note
17
The way we simulate the shocks during the estimation period is in line with Nimalendran (1994), where
the author uses an approach based on a return generating process that combines a Poisson process, to capture the arrival of information, with a standard Brownian motion.
18
In unreported results, we show that beyond a simulated abnormal performance of 1%, all three compared
methods (BOEHMER, RANK and MSR) are highly powerful.
19
that this transformation does not affect the mean of the event-day abnormal return day. For k=2, we add randomly two mean-adjusted returns selected from the estimation window. This procedure allows us to simulate event-induced variance without any parametric hypothesis on the distribution of the return generating process19.
3.4. Results Our simulations are too numerous to be analyzed on a case-by-case basis. The details of the specification analysis are given in Appendix 2 and for the power analysis in Appendix 3. These appendices are organized as follow: Appendix 2 Panel A: specification analysis without simulated event during the estimation period. Appendix 2 Panel B: specification analysis with simulated event during the estimation period (with average simulated abnormal return of 0%, +/-1%, +/-2%, +/-4%). Appendix 3 Panel A: power analysis without simulated event during the estimation period. Appendix 3 Panel B: power analysis with simulated event during the estimation period (with average simulated abnormal return of 0%, +/-1%, +/-2%, +/-4%).
To interpret all these results we propose the following regression analysis: the dependent variables will be either the specification error in percentage terms (absolute value of the difference between the chosen confidence level (5%) and the obtained proportion of
19
It should be noted that this corresponds to a proportional error structure (or proportional heteroscedasticity) as defined in Harrington and Shrider (2002). 20
samples with significant abnormal returns) or the power (percentage of samples where the generated abnormal returns are correctly detected). As our dependent variables are percentages, we have to take into account the inherent heteroscedastic nature of our regression models in order to build correct inferences. We follow the procedure presented in Appendix 4. It leads us to use as dependent variable either the logistic transform of the specification error or the logistic transform of the power, in order to produce adequate statistical tests. The independent variables represent the conditions of the simulation. We use the following: DBOEHMER: dummy variable equal to 1 for BOEHMER test; DRANK: dummy variable equal to 1 for RANK test; DMSR: dummy variable equal to 1 for MSR test; VAREvent: multiplicative coefficient applied to the variance of the abnormal returns generated at the event date (k coefficient introduced in Section 3.3); AREst: average abnormal return generated during the estimation window; VAREst: multiplicative coefficient applied to the variance of the abnormal returns generated during the estimation window (k coefficient introduced in Section 3.2); SAREst: dummy variable equal to 1 if AREst is negative; RT: dummy variable equal to 1 if the specification is evaluated in the right tail.
Two comments are worth noting: coefficients of regressions presented in Tables 3 to 5 are coefficients of classical linear regression models. As these models are built with all dummy variables (e.g., DBOEMER, DRANK and DMSR) but without constant (to avoid the
21
classical perfect colinearity problem), the coefficients are sub-sample averages of either specification error (Table 3) or power (Tables 4 and 5). inferences are realized using the two-step procedures described in Appendix 4, in order to take into account the specific nature of the dependent variable. Moreover, they represent tests of differences of mean between the BOEHMER test (our benchmark) and the RANK or MSR tests. In Table 3, column one, for example, the average specification error for the BOEHMER test is 2.35%, 3.79% for the RANK test (and the difference is statistically significant at the 1% level) and 1.51% for the MSR test (and the difference with the BOEHMER rate is significant at the 5% level).
Specification analysis Results for specification analysis are summarized in Table 3. The main results are the following: regression 1 compares BOEHMER, RANK and MSR tests on the whole set of simulations. MSR is the best-specified test (1.5% average specification error) and the difference with the BOEHMER is significant at the 5% level. The RANK test is the worst specified (3.79% average specification error and the difference with the BOEHMER is significant at 1%). regression 2 focuses on the effect of the generated abnormal returns variance at the event date (event-induced variance). We get the same result except that the difference between the MSR test and BOEHMER is not significant any more.
22
regression 3 focuses on the effect of the increasing shocks during the estimation window, the main point of our analysis. The specification errors clearly increase for the three tests. MSR remains the best-specified test. The difference of the specification error with the BOEHMER approach is significant at 1%.
regression 4 focuses on the effect of the generated abnormal returns variance during the estimation window (variance of AREst). A first interesting result is that the impact of the increase in the variance of AREst is far less dramatic on the specification error than the one of the) AREst level. MSR remains the bestspecified test (with a difference compared to BOEHMER test significant at the 10% level).
regression 5 presents the effect of the sign of the generated average abnormal returns during the estimation period. Our main conclusions remain valid (the difference of the means between the MSR and BOEHMER tests being of no significance).
regression 6 tests whether the specification results are asymmetric. The least asymmetric test seems to be the RANK test, but no differences of means are significant.
The main conclusion is that the MSR test clearly dominates the BOEHMER and RANK tests with respect to the specification, particularly when the estimation period is shocked. As in Cowan and Sergeant (1996), we find that the RANK test is seriously misspecified.
23
Power analysis Results for power analysis are presented in Tables 4 and 5. Table 4 summarizes results for low event-date generated abnormal returns (0.5%) and Table 5 for medium ones (1%)20.
The main results in Table 4 (0.5% generated abnormal returns at the event day) are the following: regression 1 compares RANK and MSR tests to BOEHMER, on the set of simulations with no event-induced variance. MSR is the most powerful test in these circumstances. The difference of power with the BOEHMER test is significant at the 5% level. regression 2 compares RANK and MSR tests to BOEHMER on the whole set of simulations (with and without event-induced variance). Regression 3 focuses on the effect of the generated abnormal returns variance at the event date. Regression 4 focuses on the effect of the increasing shocks during the estimation window, the main point of our analysis. Regression 5 focuses on the effect of the generated abnormal returns variance during the estimation window (variance of AREst). From these four regressions it emerges that the RANK test is the most powerful in each case. It is a direct consequence of its bad specification. As the RANK test detects events too frequently when there are none, it also detects them more frequently when they are present. The MSR test reveals itself to be more powerful than the BOEHMER in each case. The differences of power rate are,
20
We focus on low or moderate shocks as Harrington and Shrider (2002). It is indeed for these levels that power is an issue in event studies tests. 24
however, not significant. This result is particularly interesting and casts some doubts on the results published up to now in the field of event study methodologies. It can not be concluded from a set of simulations that one proposed test dominates another. An adequate analysis of the statistical significance of the observed difference of power rate must be conducted in order to validate the results. regression 6 presents the effect of the sign of the generated average abnormal returns during the estimation period. No clear-cut result emerges. Table 5 confirms these conclusions for the case of 1% generated abnormal returns.
Results of Section 3 allow us to conclude that the RANK test is generally badly specified. The BOEHMER test becomes badly specified under a shocked estimation window, while the MSR is far more robust to this phenomenon. As for power, the MSR test dominates the BOEHMER test but the differences are not statistically significant. The RANK test appears to be the most powerful one, but this is a straightforward consequence of its misspecification.
4. THE CASE OF MERGERS AND ACQUISITIONS
As mentioned previously, merger and acquisition announcements, and more particularly the analysis of the bidder stock price performance, are well suited to test the proposed Markov Switching event study approach. This section is devoted to the application of the
25
proposed methodology to a real sample of merger and acquisition announcements. In the sequel we first present the considered M&A sample, and then the results.
4.1. Data Our data was obtained from statistics relating to the activities of the DGC (Directorate General for Competition of the European Commission). Table 6 provides summary information on proposed combinations that had been notified to the DGC since the inception of the EC regulation n 4064/89, up to December 200021. It usually takes quite some time after the intervention, for the EC to file an official report on its web site. Consequently, we were obliged to restrict our analysis to notifications from 1990 through 2000 inclusive; later cases were mostly incomplete. The total number of notified combinations during this period was 1573 (see Panel A). Many proposed business combinations involved small or closely-held firms with no readily available market price information, so they could not be included in this study. To be included in our sample, at least one of the subject firms must be quoted on a national stock exchange; 874 of the 1573 notified operations, involving 1535 different firms, satisfy this requirement. Among these 1535 different firms 582 are bidders and 486 are targets, the remainder being firms involved in joint venture operations. Panel B of Table 6 presents a breakdown of our sample operations by year of notification.
21
Regulation EC n 4064/89 set up the notification obligation for European area business combinations. A
detailed presentation of the European regulation and notification criteria can be found in Aktas et al. (2001).
26
Market data were obtained from Datastream, accessed at the Universit de Lille 2. For announcement dates, two separate sources were checked: the SDC Database edited by Thomson Financial and, depending on the country, the financial press (Les Echos, Financial Times, Wall Street Journal, etc.). The SDC Database and the financial press have also been used to collect supplementary information such as the size of the deal, the means of payment, the type of combination, the presence of rumors in the months preceding the combination, etc.
Because the firms involved were traded on various national exchanges, it was necessary to collect local market information about each exchange and to select a market index, which will be employed in the usual way to construct abnormal returns. The countries involved, the stock market indexes selected, and the local currencies are listed in Table 7.
Lastly, two numbers deserve to be quoted. The average number of acquisitions by bidder in our sample is 3.52. This by no means represents the total average number of acquisition attempts by the bidders in our sample, but only the ones that have been notified to the DGC. All acquisition attempts that need not be notified to the European authorities are excluded from it. Also excluded from this number are all other significant corporate events that could have shocked the estimation period. The second important figure is the number of times an acquisition attempt has been announced within the estimation period of another attempt. It amounts to 113 in our sample. In other words, we are sure that for almost one case out of five, the estimation period includes another M&A announcement!
27
4.2. Results
Table 8 displays the results for the bidders and compares the proposed MSR method to the considered benchmark (BOEHMER). Among the 582 sample bidders, 514 converge with the MSR method (circa 88% convergence). The significance of the SCAR (Standardized Cumulative Abnormal Return) over the period that goes from day -3 to day +2 relative to the announcement day increases with the MSR approach. The resultant tstat is 1.95 (p-value: 5.1%) for the MSR approach, while being 1.67 (p-value: 9.5%) with the BOEHMER test. This corresponds to a 17% increase in the significance of the statistic. To a large extent, this result confirms our initial intuition in the sense that firmspecific events during the estimation period influence the significance of the abnormal return of the bidders when we use standard tests (RANK or BOEHMER). The bidder result obtained with the MSR method contrasts with the prior findings in the literature (see, for example, the literature review of Jensen and Ruback (1983)). Prior studies have documented that the average abnormal gain to bidder companies, around the announcement day of the M&A operation, is of low significance or none at all. Taking into account the probability of a shocked estimation window with the MSR test, the observed abnormal returns become much more significant. The average number of acquisition attempts by bidders (3.52 in our sample) and the number of cases for which we are sure that there has been such an attempt during the estimation window (113 out of 582 cases in our sample) gives a strong foundation to the result. This leads us to think that previous published results understate the real value impact of M&A for bidders.
28
Are our results trivial? In other words, does the MSR test automatically generate a more significant cross-sectional student test? Table 9 displays the results for the targets. Among the 486 sample targets, 436 converge with the MSR method (circa 89% convergence). The significance of the two tests (BOEHMER and MSR) is comparable; the t-stats are 10.70 and 10.88 for the BOEHMER and MSR respectively. These results are consistent with Appendix 1. The increase in the statistical significance is not an automatic by-product of the MSR test.
29
5. CONCLUSION
The recent contribution of Fuller et al. (2002) stresses one important dimension of bidders behavior in the field of mergers and acquisitions: they are often repetitive acquirers. This has already been pointed out by Malatesta and Thompson (1985), who propose an adapted methodology to evaluate the wealth effects of acquisitions programs. Such repetitive acquisitions (attempts) potentially create significant shocks during the estimation window used in the classical event study methodology. Could it be that such behavior would have some influence on inferences drawn concerning the wealth effect of M&A for bidders? In other words, do potential corporate events during the estimation window affect statistical inferences drawn in the classical event study framework?
We first study the specification and power of three alternative cross-sectional tests of abnormal returns built upon the seminal contribution of Fama et al. (1969). The first one is the Boehmer et al. (1991) proposition which is known to tackle the event-induced variance phenomenon. The second one is the Corrado rank test (1989), which is robust to departures from the Gaussian abnormal returns hypothesis. We introduce a third one, in the spirit of the Roll (1987) results. The author shows, in his presidential address to the American Finance Association, that the return generating process could better be described as a mixture of Gaussian distributions, describing a two-state return generating process. The first one would correspond to a no-news regime, and the second one to an information-arrival regime. Our test builds on this result to propose a two-regime market model generating process, with state dependent variance and stochastic regime transition
30
model (described by a Markov chain). The main results of our simulation work are that the Corrado (1989) rank is generally misspecified (as has already been pointed out in previous contributions) and that the Boehmer et al. (1991) test becomes misspecified as shocks in the estimation window appear. Our test reveals itself to be robust to this problem. While it also reveals itself to be more powerful than the Boehmer et al. (1991) test, we illustrate that the difference of power rates is not statistically significant. Incidentally (? By the way is too conversational), our methodology stresses the importance of constructing a clean statistical test of significance when comparing several methodological alternatives in the field of event studies. A higher power or a better specification on a set of simulations is not sufficient to draw conclusions concerning the superiority of the proposed methodology without testing for the statistical significance of the results. Such a test must take into account the heteroscedastic nature of proportion data. We finally apply the proposed methodology to a sample of real M&A operations. Our sample includes 582 bidders and 486 targets involved in operations notified to the European Commission during the period 1990-2000. One major result emerges from this analysis. Bidders abnormal returns, while only marginally significant using the classical Boehmer et al. (1991) methodology, become significant at the 5% level when taking into account the probability of a disturbed estimation window.
31
REFERENCES Aktas, N., E. de Bodt, M. Levasseur and A. Schmitt, 2001, The emerging role of the European Commission in merger and acquisition monitoring: The Boeing / Mc Donnell Douglas case, European Financial Management 7, 447-480. Aktas N., E. de Bodt and R. Roll, 2003, Market response to European regulation of business combinations, University of California at Los Angeles, Working Paper. Ball, C. and W. Torous, 1988, Investigating security-price performance in the presence of event-date uncertainty, Journal of Financial Economics 22, 123-154. Boehmer, E., J. Musumeci and A. Poulsen, 1991, Event-study methodology under conditions of event induced variance, Journal of Financial Economics 30, 253-272. Brown, S. and J. Warner, 1980, Measuring security price performance, Journal of Financial Economics, 8, 205-258. Brown, S. and J. Warner, 1985, Using daily stock returns: The case of event studies, Journal of Financial Economics 14, 3-31. Campbell, J., A. Lo and A. MacKinlay, 1997, The Econometrics of Financial Markets", Princeton University Press. Campbell, C. and C. Wasley, 1993, Measuring security price performance using daily NASDAQ return, Journal of Financial Economics 33, 73-92. Corrado, C.J., 1989, A nonparametric test for abnormal security price performance in event studies, Journal of Financial Economics 23, 385-395. Corrado, C.J. and T.L. Zivney, 1992, The specification and power of the sign test in event study hypothesis tests using daily stock returns, Journal of Financial and Quantitative Analysis 27, 465-478. Cowan A., 1992, Nonparametric event study tests, Review of Quantitative Finance and Accounting 2, 343-358. Cowan, A. and A. Sergeant, 1996, Trading frequency and event study test specification, Journal of Banking and Finance 20, 1731-1757. Fama, E., L. Fisher, M. Jensen and R. Roll, 1969, The adjustment of stock prices to new information, International Economic Review 10, 1-21. Fuller K., J. Netter and M. Stegemoller, 2002, What do returns to acquiring firms tell us? Evidence from firms that make many acquisitions, Journal of Finance, August 2002. Greene W., 2002, Econometric Analysis, Fourth edition, Prentice Hall. Hamilton J., 1989, A New Approach to the Economic Analysis of Nonstationary Time Series and The Business Cycle. Econometrica 57, 357-84. Hamilton J., 1994, Time series analysis, Princeton University Press. Harrington S.E., D. Shrider, 2002, All Events Induce Variance: Analyzing Abnormal Returns When Effects Vary Across Firms, SSRN Working Paper. Hilliard J. and R. Savickas, 2000, On stochastic volatility and more powerful parametric tests of event effects on unsystematic returns. FEN Working Paper. Jensen, M. and R. Ruback, 1983, The market for corporate control, Journal of Financial Economics 11, 5-50. Krolzig H.-M., 1997, Markov-Switching Vector Autoregressions, University of Oxford, UK Lecture Notes in Economics and Mathematical Systems, Vol. 454. Springer, XIV.
32
Malatesta, P. and R. Thompson, 1985, Partially Anticipated Events: A Model of Stock Price Reactions with an Application to Corporate Acquisitions, Journal of Financial Economics 14, 237-250. Mitchell, M. and E. Stafford, 2000, Managerial decisions and long-term stock price performance, Journal of Business 73, 287-330. Nimalendran M., 1994, Estimating the effects of information surprises and trading on stock returns using a mixed jump-diffusion model, Review of Financial Studies 7, 451-473. Patell, J., 1976, Corporate forecasts of earnings per share and stock price behavior: Empirical tests,, Journal of Accounting Research 14, 246-276. Roll, R., 1987, R-Squared, Journal of Finance 43, 541-566. Ruback, R., 1982, The effect of discretionary price control decisions on equity values, Journal of Financial Economics 10, 83-105. Salinger, M., 1992, Standard errors in event studies, Journal of Financial and Quantitative Analysis 27, 39-53. Scholes, M. and J. Williams, 1977, Estimating betas from nonsynchronous data, Journal of Financial Economics 5, 309-328. Sharpe, W., 1963, A simplified model for portfolio analysis, Management Science 9, 277-293.
33
Figure 1
A typical regime decomposition using the two-state market model (TSMM) Figure 1 presents the typical plot of the residuals and the estimated regime probabilities obtained on a real M&A case (the acquisition of ICI Nylon activities by DuPont on April 23,1992). The analysis is realized on DuPont. The event is situated at day 235. The standard errors of regime 1 and regime 2 are respectively 0.99% and 1.6%. The event is clearly situated in regime 2, the one of high variance. 77.8 (163.2) days are classified in regime 2 (1). .
MSH(2)-ARX(0), 1 - 241
RI Mean(RI)
0.050 0.025 0.000 -0.025 0 1.0
20 40 60 Probabilities of Regime 1
filtered pred icted smoothed
80
100
120
140
160
180
200
220
240
0.5
0 1.0
20 40 60 Probabilities of Regime 2
filtered pred icted smoothed
80
100
120
140
160
180
200
220
240
0.5
20
40
60
80
100
120
140
160
180
200
220
240
34
Figure 2
Event generation during the estimation window Figure 2 presents a typical result obtained using the procedure described in Section 3.3. The plain line is the initial returns times series and the dotted one is the returns time series obtained after event generations. Four events (average of 1% and k=2) are generated at dates T=37, 50, 68 and 124. The initial estimation period standard error is 1.17% and becomes, after event generation, 1.31%.
6,0%
4,0%
2,0%
0,0%
-2,0%
-4,0%
-6,0% 1 11 21 31 41 51 61 71 81 91 101 111 121 131 141 151 161 171 181 191
35
Table 1
Table 1 presents the S&P 1500 Sectors breakdown at December 31, 2002. Financials, HealthCare and Information Technology are the more heavily weighted ones. The total market value of the index at December 31, 2002 is US$ 9,135 billion. The mean (median) company market value is US$ 6,090 million (US $ 1,239 million). The numbers of quoted companies are: 1,028 (68.5%) companies quoted on the NYSE, 462 (30.8%) on the NASDAQ and 10 (0.6%) on the AMEX.
Global Industry Classification Sectors (December 31, 2002) Number of Cos Consumer Discretionary Consumer Staples Energy Financials Health Care Industrials Informationn Technology Materials Telecommunication Services Utilities 276 75 76 200 146 253 266 100 18 90
% of Market Cap 13.80% 9.00% 6.10% 20.10% 14.60% 12.00% 14.20% 3.10% 3.80% 3.30%
36
Table 2
Panel A Descriptive statistics on the market model (MM) and two-state market model (TSMM) residuals. Reported numbers are the sample mean of 10,000 estimates. Both models produce centered residuals. The TSMM residuals exhibit a somewhat higher level of kurtosis. Performance measure Market model Market model - MSR Mean 0.0000 0.0009 Standard deviation 0.0249 0.0251 Skewness 0.2802 0.2871 Kurtosis 6.0834 6.5259
Panel B - Cross-sectional properties of sample-wide mean abnormal returns on the event day. Each statistic is based on 250 values of the average abnormal returns, one for each sample. For a given sample, the average abnormal return is the arithmetic average of the abnormal returns of the individual securities in the sample. Results are reported for sample sizes of 10, 25 and 50 securities. As expected, increasing the portfolio size leads to an abnormal-return distribution closer to the Gaussian distribution. Performance measure Market model Portfolio size 50 25 10 50 25 10 Standard deviation 0.0037 0.0056 0.0084 0.0037 0.0055 0.0083
Mean -0.0006 -0.0007 -0.0008 0.0004 0.0002 0.0001
Skewness 0.0750 0.3180 -0.0211 0.0877 0.2975 -0.0174
Kurtosis 3.0620 3.1593 3.6008 3.1844 3.0762 3.5474
Market model - MSR
37
Table 3 Specification analysis

The dependent variable is the absolute value of the specification (5%-reject proportion). We use the following independent variables: DBOEHMER: dummy variable equal to one for BOEHMER test; DRANK: dummy variable equal to one for RANK test; DMSR: dummy variable equal to one for MSR test; VAREvent: multiplicative coefficient applied to variance of abnormal returns generated at the event date (k coefficient introduced in Section 3.3); AREst: average abnormal return generated during the estimation window; VAREst: multiplicative coefficient applied to variance of abnormal returns generated during the estimation window (k coefficient introduced in Section 3.3); SAREst: dummy variable equal to one if AREst is negative; RT: dummy variable equal to one if specification is evaluated in the right tail. Coefficients are obtained using classical OLS estimation. As these models are built with all dummy variables (e.g., DBOEHMER, DRANK and DMSR) but without constant (to avoid the classical perfect colinearity problem), the coefficients are sub-sample averages of specification error. Inferences are realized using the two-step procedure described in Appendix 4 in order to take into account the specific nature of the dependent variable. Moreover, they represent tests of differences of mean between the BOEHMER test (our benchmark) and RANK or MSR tests.
Specification Error | 5%- H0's Reject | weighted OLSQ : Dependent Variable Logit (ESPEC) 1 2 3 4 5 Coef. Coef. Coef. Coef. Coef. -4.14 -3.71 *** -4.49 ** -3.61 -2.87 *** -3.98 -259.60 -225.88 -307.94 *** -3.50 -3.07 ** -3.93 * -4.27 -3.76 ** -4.31 6 Coef.
DBOEHMER DRANK DMSR DBOEHMER*VAREvent DRANK*VAREvent DMSR*VAREvent DBOEHMER*AREst DRANK*AREst DMSR*AREst DBOEHMER*VAREst DRANK*VAREst DMSR*VAREst DBOEHMER*SAREst (1 if AREst is negative) DRANK*SAREst DMSR*SAREst DBOEHMER*RT (1 if right tail) DRANK*RT DMSR*RT
-4.08 -4.46 -4.02 270 270 270 270
Number of Observations 270 270 * difference relative to Boehmer mean significant at 10% level ** difference relative to Boehmer mean significant at 5% level *** difference relative to Boehmer mean significant at 1% level
38
Table 4 Power analysis (0.5%)

The dependent variable is the power rate for 0.5% generated abnormal returns. We use the following independent variables: DBOEHMER: dummy variable equal to one for BOEHMER test; DRANK: dummy variable equal to one for RANK test; DMSR: dummy variable equal to one for MSR test ; VAREvent: multiplicative coefficient applied to variance of abnormal returns generated at the event date (k coefficient introduced in Section 3.3); AREst: average abnormal return generated during the estimation window; VAREst: multiplicative coefficient applied to variance of abnormal returns generated during the estimation window (k coefficient introduced in Section 3.3); SAREst: dummy variable equal to one if AREst is negative. Coefficients are obtained using classical OLS estimation. As these models are built with all dummy variables (e.g., DBOEHMER, DRANK and DMSR) but without constant (to avoid the classical perfect colinearity problem), the coefficients are sub-sample averages of power rates. Inferences are realized using the two-step procedure described in Appendix 4 in order to take into account the specific nature of the dependent variable. Moreover, they represent tests of differences of mean between the BOEHMER test (our benchmark) and RANK or MSR tests.
Study of average rejection rates for abnormal return of +0.5% weighted OLSQ : Dependent Variable Logit (Rejection rates) 1 2 3 4 5 6 Coef. Coef. Coef. Coef. Coef. Coef. DBOEHMER -0.37 -1.07 DRANK -0.20 -0.70 *** DMSR -0.11 ** -0.89 DBOEHMER*VAREvent -0.93 DRANK*VAREvent -0.60 *** DMSR*VAREvent -0.84 DBOEHMER*AREst -38.87 DRANK*AREst -25.08 ** DMSR*AREst -31.39 DBOEHMER*VAREst -0.68 DRANK*VAREst -0.43 *** DMSR*VAREst -0.56 DBOEHMER*SAREst (1 if AREst is negative) -0.73 DRANK*SAREst -0.43 * DMSR*SAREst -0.82 Number of Observations 45 135 * difference relative to Boehmer mean significant at 10% level ** difference relative to Boehmer mean significant at 5% level *** difference relative to Boehmer mean significant at 1% level 135 135 135 135
39
Table 5 Power analysis (1 %)

The dependent variable is the power rate for 1% generated abnormal returns. We use the following independent variables: DBOEHMER: dummy variable equal to one for BOEHMER test; DRANK: dummy variable equal to one for RANK test; DMSR: dummy variable equal to one for MSR test; VAREvent: multiplicative coefficient applied to variance of abnormal returns generated at the event date (k coefficient introduced in Section 3.3); AREst: average abnormal return generated during the estimation window; VAREst: multiplicative coefficient applied to variance of abnormal returns generated during the estimation window (k coefficient introduced in Section 3.3); SAREst: dummy variable equal to one if AREst is negative. Coefficients are obtained using classical OLS estimation. As these models are built with all dummy variables (e.g., DBOEMER, DRANK and DMSR) but without constant (to avoid the classical perfect colinearity problem), the coefficients are sub-sample averages of power rates. Inferences are realized using the two-step procedure described in Appendix 4 in order to take into account the specific nature of the dependent variable. Moreover, they represent tests of differences of mean between the BOEHMER test (our benchmark) and RANK or MSR tests.
Study of average rejection rates for abnormal return of +1% weighted OLSQ : Dependent Variable Logit (Rejection rates) 1 2 3 4 5 6 Coef. Coef. Coef. Coef. Coef. Coef. DBOEHMER 2.26 0.92 DRANK 3.10 *** 1.53 ** DMSR 2.56 * 1.08 DBOEHMER*VAREvent 0.08 DRANK*VAREvent 0.39 *** DMSR*VAREvent 0.13 DBOEHMER*AREst 32.85 DRANK*AREst 58.34 * DMSR*AREst 39.75 DBOEHMER*VAREst 0.57 DRANK*VAREst 0.98 ** DMSR*VAREst 0.67 DBOEHMER*SAREst (1 if AREst is negative) 1.34 DRANK*SAREst 1.85 DMSR*SAREst 1.23 Number of Observations 45 135 * difference relative to Boehmer mean significant at 10% level ** difference relative to Boehmer mean significant at 5% level *** difference relative to Boehmer mean significant at 1% level 135 135 135 135
40
Table 6
Table 6 Panel A presents summary statistics on combinations that have been notified to the European Community since the inception of the regulation in 1990 through the latest month in our data sample (December 2000). Panel B presents a breakdown by year of notification for our sample operations. Panel A. Number of cases notifying the EC by year Year N 1990 12 1991 63 1992 60 1993 58 1994 95 1995 110 1996 131 1997 172 1998 235 1999 292 2000 345 Total 1573
Panel B. Sample operations by year Year N 1990 12 1991 44 1992 37 1993 40 1994 53 1995 66 1996 59 1997 86 1998 105 1999 150 2000 222 Total 874
41
Table 7
Table 7 reports the home country of our sample firms and the corresponding local market index as well as the currency. Country Australia Austria Belgium Bermuda Canada Denmark Finland France Germany Greece Hong Kong Ireland Italy Japan Luxembourg Netherlands Norway Portugal Singapore South Africa Spain Sweden Switzerland UK USA Total N 5 8 24 2 21 11 24 221 267 2 1 2 74 35 1 88 11 3 1 7 26 61 57 250 334 1535 Index S&P ASX 200 Weiner Boerse Index Brussels all Shares MSCI World Price Index Toronto 300 Copenhagen SE HEX CAC40 DAX Kurs Price Index DJ Euro Stoxx Price Index Hang Seng Ireland SE Milan Comit NIKKEI 225 Luxembourg SE 13 CBS All Share Oslo SE General DJ Euro Stoxx Price Index Singapore DBS 50 Price Index JSE Industrial Madrid SE General Affrsvarlden weighted all shares Swiss Market Index FTSE 100 S&P 500 Currency22 Dollar Schilling* Franc* Dollar Dollar Danish Krone Finish Markka* Franc* Mark* Euro Dollar Punt* Lira* Yen Franc* Guilder* Norwegian Krone Euro Dollar Rand Peseta* Swedish Krona Franc Pound Dollar
22
Since January 1, 1999, euroland countries indicated by an asterisk have maintained fixed exchange rates
with the euro (and hence with each other).
42
Table 8
Table 8 presents the Standardized Abnormal Return (SAR) and Standardized Cumulative Abnormal Return (SCAR) for Bidders (among the 582 sample bidders, 514 converge with the MSR method (circa 88% of convergence)) around the initial announcement date. BSR is the Boehmer et al. test (1991) and MSR is the Markov Switching Regression test (see Section 2.2).
BSR SAR SE t-stat SCAR (-3;+2) Var(SCAR) t-stat
-3 -0.0038 0.0534 -0.0713 0.9485 0.5690 1.6669
-2 -0.0468 0.0580 -0.8070
-1 0.2387 0.0894 2.6705
0 0.2084 0.1051 1.9834
1 0.0478 0.0558 0.8569
2 0.5042 0.5436 0.9276
MSR SAR SE t-stat SCAR (-3;+2) Var(SCAR) t-stat
-3 0.0578 0.0811 0.7120 1.2619 0.6471 1.9502
-2 -0.0265 0.0846 -0.3133
-1 0.4030 0.1487 2.7097
0 0.1907 0.1694 1.1257
1 0.1067 0.0872 1.2238
2 0.5301 0.5886 0.9005
43
Table 9
Table 9 presents the Standardized Abnormal Return (SAR) and Standardized Cumulative Abnormal Return (SCAR) for Targets (among the 486 sample targets, 436 converge with the MSR method (circa 89% of convergence)) around the initial announcement date. BSR is the Boehmer et al. test (1991) and MSR is the Markov Switching Regression test (see Section 2.2).
BSR SAR SE t-stat SCAR (-3;+2) Var(SCAR) t-stat
-3 0.2017 0.0800 2.5226 3.9247 0.3668 10.7007
-2 0.3537 0.0776 4.5585
-1 1.8501 0.2443 7.5718
0 1.3269 0.2261 5.8685
1 0.1028 0.0794 1.2940
2 0.0894 0.0705 1.2690
MSR SAR SE t-stat SCAR (-3;+2) Var(SCAR) t-stat
-3 0.3803 0.1235 3.0788 5.8324 0.5361 10.8796
-2 0.4804 0.1191 4.0337
-1 2.6354 0.3750 7.0279
0 1.9153 0.2991 6.4046
1 0.2705 0.1234 2.1929
2 0.1504 0.1126 1.3361
44
Appendix 1 Standard deviation estimation and cross-sectional student test
The proposed MSR test relies on a two-state decomposition of the return generating process with state-dependent standard error estimation. As we use the estimated standard deviation of the low regime variance to standardize the abnormal return, we might think that we automatically increase the cross-sectional student t-stat. However, this is not true. We propose a proof of this below. More precisely, we study here, in the homoscedastic case, the consequences of using a systematically biased standard deviation on the behavior of the cross-sectional student test. We show that this is without any influence on the cross-sectional student t-stat, the increase in the cross-sectional standard deviation being strictly compensated by the increase in the cross-sectional average abnormal returns. Assume that ARi corresponds to the abnormal return of firm i on the event day and that the variance ( 2 ) of the AR is the same for each sample firm (homoscedasticity).
The cross-sectional test of Boehmer et al. (1991) is implemented in the following way : - first we standardize the abnormal return by dividing it by its standard deviation:
ti = ARi
- the cross-sectional average of the standardized abnormal return is given by the following expression:
ti = i =1 . n
n
- the variance of the average standardized abnormal return corresponds to:
(ti ti ) 2
i =1
t2 i
n 1 n
t2
i
45
- the t-stat following the Boehmer et al. (1991) approach is given by:
t stat = ti
t2
i
i =1
ARi
2
n
t stat =
i =1
ARi
i =1
ARi
2
n
)2
n 1 n
t stat =
1
1 n ( n 1)
AR
i =1
= ( ARi AR ) 2
AR n ( n 1)
( AR
i =1
AR ) 2
where AR= i =1
AR
n
, and therefore ti = AR .
Suppose now that the estimation of the abnormal return variance is biased downward with a constant c and corresponds to 2* = 2 c . The previous expressions can be written in the following way:
ARi ( 2 c)
ti =
46

i =1
ARi ( 2 c ) n
t stat =
i =1
ARi ( c )
2
i =1
ARi ( 2 c ) n )2
n 1 n
t stat =
1
1 AR ( c ) 1 n ( n 1) ( c )
i =1
= ( ARi AR ) 2
AR n ( n 1)
( AR
i =1
AR ) 2
where AR=
AR
i =1
, and therefore ti = AR c
The result is the same as the previous one. The estimation of the cross-sectional variance compensated the bias applied to the variance of the abnormal returns.
47
Appendix 2 Specification analysis

Appendix 2 presents average rejection rates for various test statistics (RANK (Corrado, 1989), BSR (Boehmer et al., 1991) and MSR (Markov Switching Regression test see Section 2.2)) for 250 portfolios of 50 securities at significance level of 5%. The test is realized under the null hypothesis of no abnormal returns. Securities are randomly chosen in the S&P 1500 Index universe during the period 1990/2000. Panel A presents the case of an undisturbed estimation window. Panel B introduces shocks during the estimation window. The symbol k represents the multiplicative factor applied to the variance during the estimation window and k, the one applied during the event window. The symbols k and k are used to generate abnormal returns respectively during the estimation period and the event window. AGAR is the level of average generated abnormal returns during the estimation period (0%, +/- 1%, +/- 2%, +/-4%). Specifics of the simulation procedure are described in Section 3.
Panel A Specification analysis without estimation window shocks

one tail test k'=0 k'=1 k'=2
left tail right tail left tail right tail left tail right tail
RANK BSR MSR
2.8% 4.0% 2.4%
3.4% 5.6% 7.6%
3.6% 4.8% 4.0%
12.00% 3.2% 5.2%
7.2% 6.0% 4.8%
10.0% 7.6% 6.0%
Panel B Specification analysis with estimation window shocks

AGAR=0% one tail test k'=0 k=1 RANK BSR MSR k=2 RANK BSR MSR 2.8% 3.6% 2.4% 5.6% 6.4% 9.6% 11.2% 6.0% 4.0% 6.4% 4.8% 6.4% 11.6% 6.8% 5.6% 8.4% 6.0% 6.4% 2.8% 5.2% 2.8% 4.4% 3.2% 5.6% 10.8% 8.0% 6.8% 4.0% 2.8% 4.8% 10.4% 5.6% 5.6% 8.4% 5.2% 6.0% k'=1 k'=2
AGAR = +1% one tail test k'=0 k=1 RANK BSR MSR k=2 RANK BSR MSR 8.0% 9.2% 5.6% 5.2% 4.4% 5.2% 4.4% 8.0% 4.0% 2.8% 2.8% 4.4%
k'=1
k'=2
AGAR = -1% one tail test k'=0 k=1
k'=1
k'=2
12.8% 8.4% 3.6%
6.0% 4.4% 6.8%
12.0% 6.4% 4.8%
6.4% 2.8% 3.6%
RANK BSR MSR k=2
3.6% 3.6% 2.8%
4.0% 6.8% 6.8%
6.8% 3.6% 2.0%
7.2% 5.6% 6.8%
9.2% 4.8% 4.4%
8.8% 4.4% 5.2%
14.4% 8.8% 7.6%
4.8% 3.2% 5.2%
12.4% 4.0% 3.2%
12.8% 6.0% 8.0%
RANK BSR MSR
4.4% 4.0% 4.0%
3.6% 3.2% 4.4%
9.2% 6.8% 4.8%
9.2% 6.0% 6.4%
10.0% 4.4% 3.6%
7.6% 5.2% 7.2%
48
k'=1
k'=2
AGAR = - 2% one tail test k'=0 k=1
k'=1
k'=2
11.2% 7.2% 4.8%
6.0% 5.6% 8.0%
18.0% 7.2% 5.6%
4.8% 1.6% 4.0%
RANK BSR MSR k=2
4.0% 5.2% 3.2%
3.2% 6.0% 4.8%
6.8% 4.0% 3.2%
10.4% 8.0% 8.4%
8.4% 4.0% 3.6%
12.0% 8.0% 7.2%
12.8% 6.8% 3.6%
4.8% 2.8% 4.4%
13.2% 8.4% 5.6%
7.2% 4.8% 6.8%
RANK BSR MSR
3.6% 4.0% 3.6%
5.6% 7.6% 8.0%
7.6% 4.0% 4.0%
7.2% 5.6% 3.6%
8.4% 2.4% 2.0%
10.0% 6.0% 5.6%
k'=1
k'=2
AGAR = - 4% one tail test k'=0 k=1
k'=1
k'=2
15.6% 11.2% 5.6%
3.6% 2.0% 6.0%
18.0% 6.8% 3.6%
5.6% 4.4% 7.2%
RANK BSR MSR k=2
2.0% 2.4% 3.2%
9.2% 15.2% 8.4%
5.6% 3.6% 4.0%
13.6% 12.8% 8.4%
5.6% 3.2% 3.6%
15.6% 8.0% 6.4%
17.6% 13.6% 6.8%
4.4% 2.4% 6.4%
18.0% 7.6% 5.6%
4.0% 1.6% 5.2%
RANK BSR MSR
1.6% 3.6% 4.0%
4.8% 13.2% 5.6%
5.2% 3.6% 2.4%
9.6% 8.4% 6.4%
8.0% 5.2% 6.0%
17.2% 11.5% 8.0%
49
Appendix 3 Power analysis

Appendix 3 presents average rejection rates for various test statistics (RANK (Corrado, 1989), BSR (Boehmer et al., 1991) and MSR (Markov Switching Regression test see Section 2.2)) for 250 portfolios of 50 securities at significance level of 5%. The test is realized under the null hypothesis of no abnormal returns (two tails test). Securities are randomly chosen in the S&P 1500 Index universe during the period 1990/2000. The abnormal performances injected into the event day have an average of 0.5% or 1%. Panel A presents the case of an undisturbed estimation window. Panel B introduces perturbations during the estimation window. The symbol k represents the multiplicative factor applied to the variance during the estimation window and k, the one applied during the event window. The symbols k and k are used to generate abnormal returns respectively during the estimation period and the event window. AGAR is the level of average generated abnormal returns during the estimation period (0%, +/- 1%, +/- 2%, +/-4%). Specifics of the simulation procedure are described in Section 3.
Panel A Power analysis without estimation window shocks

Abnormal performance and induced variance
k'=0 k'=1 k'=2 0.50% 1.00% 0.50% 1.00% 0.50% 1.00%
RANK BSR MSR
47.6% 98.0% 36.0% 68.8% 27.2% 61.2% 42.8% 91.6% 27.6% 59.6% 18.4% 48.0% 48.8% 93.6% 30.0% 61.2% 19.2% 50.4%
Panel B Power analysis with estimation window shocks

AGAR=0% k=1 RANK BSR MSR k=2 RANK BSR MSR
k'=0 k'=1 k'=2 0.50% 1.00% 0.50% 1.00% 0.50% 1.00%
42.0% 94.4% 29.2% 73.2% 27.2% 62.0% 37.6% 89.2% 19.6% 66.0% 18.8% 50.8% 46.8% 90.8% 23.2% 66.8% 17.6% 54.0%
44.8% 95.6% 28.4% 77.2% 22.0% 57.2% 36.0% 88.4% 20.0% 65.6% 14.8% 40.0% 43.2% 89.2% 24.0% 67.2% 14.0% 43.6%
50
AGAR= +1% k'=0 k'=1 k'=2 0.50% 1.00% 0.50% 1.00% 0.50% 1.00% k=1 RANK BSR MSR k=2 RANK BSR MSR 46.0% 94.0% 31.2% 72.8% 20.8% 57.2% 38.4% 87.6% 20.8% 60.4% 16.4% 43.2% 46.8% 91.2% 26.8% 66.4% 17.6% 48.0% 45.2% 95.2% 27.2% 71.2% 23.2% 59.2% 41.6% 85.2% 16.8% 62.4% 14.0% 50.4% 48.0% 92.4% 23.2% 67.2% 16.4% 52.8%
AGAR= -1% k'=0 k'=1 k'=2 0.50% 1.00% 0.50% 1.00% 0.50% 1.00% k=1 RANK BSR MSR k=2 RANK BSR MSR 53.6% 96.0% 33.2% 76.4% 27.2% 66.4% 51.2% 92.4% 27.6% 69.2% 17.2% 55.2% 54.8% 92.8% 26.8% 67.2% 18.8% 54.8% 51.2% 97.2% 36.8% 83.2% 28.8% 66.0% 46.8% 92.8% 28.0% 75.2% 16.8% 50.4% 50.4% 94.8% 28.0% 76.4% 18.0% 51.2%
51
Appendix 4 Proportion-data regression
The dependent variable of Section 3 regressions is the proportion (Pi) of the 250 simulations which rejects the null hypothesis. The regression analysis of proportion data, as shown in Greene (2002, p. 886), raises a concern of heteroscedasticity. The observed
Pi is an estimate of the population quantity,
i = F(' X i) . If we treat this problem as
sampling from Bernoulli population, then we have:
Pi = F(' X i)+ i = i + i
where:
(A4.1)
E[ i]=0 , Var[ i]=
i(1 i)
ni
(A4.2)
This heteroscedastic regression format suggests that the parameters could be estimated by a nonlinear weighted least squares regression. But the author proposes a simpler way to proceed. Since the function F(' X i) is strictly monotonic, it has an inverse.
F 1(Pi)= F 1( i + i) ' X i + i
This equation produces an heteroscedastic linear regression:
(A4.3)
F 1(Pi)= Zi = ' X i +ui

where:
(A2.4)
E[ui]=0 , Var[ui]= Fi(1 Fi)

ni i2
The inverse function for the logistic model is easy to obtain. If
(A4.5)
i=
then:
exp(' X i) 1+exp(' X i)
(A4.6)
Ln( i )= ' X i 1 i
(A4.7)
Weighted least squares regression produced the minimum 2 estimator of . Since the weights are function of the unknown parameters, a two-step procedure is called for.
Event Study under Noisy Estimation Period, April 10 52
Simple least squares at the first step produces a consistent but inefficient estimate. The weights for the logit model based on the first step estimates are then:
exp(' X i) Wi =ni i(1i) with i =1+exp(' Xi)

and can be used for weighted least squares in the second step procedure.
(A4.8)
53

SSRN Id559007 PDF

Transféré par

Informations du document

Description originale:

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

SSRN Id559007 PDF

Transféré par

Droits d'auteur :

Formats disponibles

Event Study under Noisy Estimation Period

by Nihat Aktas, Eric de Bodt, Jean-Gabriel Cousin April 10, 2003

Address Voice Fax e-mail

See Chapter 4 of Campbell et al. (1997) for a review.

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

2. EVENT STUDY METHODOLOGY

2.1. Return generating model

Event Study under Noisy Estimation Period, April 10

N: number of firms in the sample;

Event Study under Noisy Estimation Period, April 10

S j : standard deviation of firm js abnormal returns during the estimation period;

SR jE : standardized abnormal returns of firm j at the event date, corrected in order

The Boehmer et al. (1991) standardized residuals test (BOEHMER)

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

The Corrado (1989) rank test (RANK)

where the standard error, S(K), is :

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

The Markov Switching Regression test (MSR)

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

P( S t = 1 S t 1 = 1) = p11 P( S t = 1 S t 1 = 2) = p21 P( S t = 2 S t 1 = 1) = p12 P( S t = 2 S t 1 = 2) = p22

Hamilton (1994) (p. 688-689) presents an Expectationminimization algorithm that proves to be

numerically very efficient.

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

mixed distribution, which is not taken into account here.

Event Study under Noisy Estimation Period, April 10

3. SPECIFICATION AND POWER TESTS

Center for Research in Security Prices (Chicago University).

Event Study under Noisy Estimation Period, April 10

3.2. Sample generation

Event Study under Noisy Estimation Period, April 10

3.3. Event generation

Event Study under Noisy Estimation Period, April 10

R'i,0 = Ri,0 +(Ri, X Ri)

methods (BOEHMER, RANK and MSR) are highly powerful.

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

4. THE CASE OF MERGERS AND ACQUISITIONS

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

0.050 0.025 0.000 -0.025 0 1.0

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Event Study under Noisy Estimation Period, April 10

Mean -0.0006 -0.0007 -0.0008 0.0004 0.0002 0.0001