روش های مطالعه رویدادی چند کشور

We provide the first simulation evidence of event-study test performance in multi-country non-US samples. The nonparametric rank and generalized sign tests are more powerful than two common parametric tests, especially in multi-day windows. The two nonparametric tests are mostly well specified, but neither is perfectly specified in all situations. The parametric standardized cross-sectional test can provide a useful robustness check but is less powerful than the nonparametric tests and rejects too often in single-market samples and when firm-specific events affect the market index. Local-currency market-model abnormal returns using national market indexes are sufficient.

Researchers use event-study methods to gauge the effects of information arrival on stock prices. The hypothesis tested is that information affects the value of stocks, on average, across firms with similar information arrival. Conclusions regarding the performance of event-study tests that appear in the methodological literature are based on simulations using data from single markets, especially the US, but the application of event-study methods to multi-country samples is growing rapidly. The suitability of specific event-study methods when applied to multi-country non-US samples has not been established in the methodological literature. This paper provides simulation evidence of the performance of several methods in such samples. Stock markets differ on many dimensions, e.g., size, liquidity, trading volume, market-making mechanisms, accounting standards, securities regulation, investor protection, ownership concentration, and corporate governance. Market characteristics can affect the statistical properties of stock returns (see Cole et al., 2008 and Hutson et al., 2008 as examples). We find that return distributions in non-US multi-country samples are non-normal, even at the portfolio level, to a greater degree than US-based studies report. In multi-country samples, where a mixture of distributions is present, the applicability of existing simulation evidence is an unexplored empirical question. Examining recent journal articles that report event studies on multi-country samples, we find that researchers tend to use simple methods for identifying a benchmark or “normal” return, primarily the single-index market model, with the market-adjusted return method also appearing repeatedly. For testing whether the average abnormal return differs from zero, the “crude dependence adjustment” (CDA) test by Brown and Warner, 1980 and Brown and Warner, 1985 is often used (see Bailey et al., 2006 and Aktas et al., 2007 as examples). A parametric test based on standardized abnormal returns, introduced by Patell, 1976 and Mikkelson and Partch, 1986 and modified by Boehmer et al. (1991) is also common. Several papers report nonparametric tests such as the rank test (Corrado, 1989) and the generalized sign test (Cowan, 1992), especially in conjunction with a parametric test (as in Harvey et al., 2004 and Behr and Güttler, 2008, among others). Nonparametric tests are naturally appealing for ill-behaved data, but in the absence of evidence cannot be assumed to be powerful and well specified. When a parametric and a nonparametric test are both reported in an article, they frequently lead to different inferences. Using the simulation approach pioneered by Brown and Warner, 1980 and Brown and Warner, 1985, we investigate the accuracy and power of statistical tests applied to market-model abnormal returns. Overall, we find that the generalized sign test (Cowan, 1992) and rank test (Corrado, 1989) are more powerful in simulation than the two commonly used parametric tests. The parametric tests also are well specified but less powerful than the nonparametric tests. In the presence of a large return variance increase on the event date, the nonparametric tests tend to reject too often, but their specification is better under a more moderate variance increase. The standardized cross-sectional test is well specified under a variance increase and is more powerful than the CDA test. We also examine test performance in samples that are potentially problematic for test specification or power. These include single-market samples, samples from the most concentrated national markets, and markets with the most non-normally distributed returns. The two nonparametric tests remain mostly well specified and powerful in these settings. The standardized cross-sectional test is less consistently well specified in single-market samples than in multi-country samples. We also examine the ability of tests to detect abnormal returns when the affected securities are potential “market movers.” This is when a stock can make up such a large fraction of its national market’s capitalization that the individual price effects of firm-specific information arrivals exert a significant influence on the market index. Thus, abnormal return calculations that use the national market index would deduct the part of the information effect included in the index return from the total information effect in the stock return, potentially reducing power. When we simulate such effects, we find that the rank and generalized sign tests continue to exhibit correct specification and good power. The standardized cross-sectional test, which uses the index return in estimating a security’s abnormal return variance, is not as reliably well specified in this situation. Aspects of multi-country event-study design, other than the selection of a test statistic, are also potentially important. First, many markets are characterized by high frequencies of missing returns due to non-trading. Our results show that a corrective procedure proposed in the literature, treating missing returns as zero returns, sometimes called the “lumped returns” procedure, produces somewhat worse event-study test performance compared to the more standard “trade to trade” method. The latter involves omitting missing-price days from calculations while accounting for the corresponding market-index returns when the stock eventually trades. Second, our results indicate that the use of a national market index, without incorporating an international or US index, is sufficient to produce well-specified and powerful tests of average stock-price effects. Third, the results suggest that for the types of stock-price reaction tests that we investigate, there is no need to convert returns from different markets into a common currency.

We examine the performance of event-study statistical tests applied to market-model abnormal trade-to-trade and lumped re- turns in simulations using actual return data on 48,258 ordinary share issues from 54 non-US markets over 1986–2006. In random samples, security abnormal returns, and even portfolio abnormal returns for 250-stock samples, depart substantially from a normal distribution. The simulation results show that four common tests tend to be well specified under most test conditions that we simulate. Two nonparametric tests, the generalized sign and rank tests, are the most powerful. The parametric standardized cross- sectional test is less powerful, and the other parametric test, based on Brown and Warner’s (1980, 1985) ‘‘crude dependence adjustment,” tends to be quite weak, especially in longer event windows, although its power increases somewhat in single- country samples. Although correctly specified in random samples, none of the three relatively powerful tests perfectly conforms to the nominal 5% significance level across various test conditions designed to check robustness. The standardized cross-sectional test tends to reject a true null hypothesis too often for longer windows in two situations: country-clustered samples and when stock-specific events move the national market index. The generalized sign test and the rank test (unless standardized) are sensitive to a large re- turn variance increase on the event date, but may perform well un- der more moderate variance increases. The rank test is sensitive to sample-wide extreme non-normality when testing a longer event window. However, most results under the null hypothesis show correct specification. Therefore, we recommend that at least two of the three tests be used and that any disagreement be interpreted with caution. The rank and generalized sign tests would be logical to use for balanced power and correct specification. If the research- er prefers to sacrifice some power for the sake of conservatism, the standardized cross-sectional test would provide a useful robust- ness check except where high cross-sectional correlation is likely (as in country clustering) and where firm-specific events are likely to move the market index. Apart from the selection of a test statistic, the results suggest that trade-to-trade returns and simple market-model methods of calculating abnormal returns with national market indexes, with- out converting to a common currency, work well. More elaborate methods do not improve test specification or power in the settings that we examine.