Assessing investor response to information events using return and volume metrics.

I. INTRODUCTION

Empiricists interested in testing whether accounting disclosures convey incremental new information to the market typically use return-based event-study designs, which Beaver (1968) introduced into the accounting literature. Beaver also introduced volume-based event-study

designs, which are less common even though such tests yield insight into the extent to which announcements of accounting information prompt individual investors to act (i.e., to trade). Of course, researchers should use the metric that is conceptually most appropriate to the research question at hand. When the question of interest is whether information is associated with higher or lower prices (e.g., whether managers are inflating valuations by "hyping" stocks, as in Lang and Lundholm [2000]), then the associated design should employ directional price-based metrics (returns). If the question of interest is whether information directly affects trading decisions (e.g., Linsmeier et al.'s [2002] analysis of how market-risk disclosures dampen trading responses), then the associated design should use trading volume-metrics. However, in many cases, the focal question is simply whether investors respond to some disclosure. Large absolute returns or high volume of trading in a short window immediately surrounding the release of data suggest that the disclosed data affect, and are not simply associated with, investor decisions. Our study provides evidence on whether a return-based or trading-based metric leads to more powerful tests of investor response to information events.

Several recent studies focus on detecting the existence or nonexistence of short-window investor response to accounting-related information including: research and development spending disclosures (Boone and Raman 2002); 8-K filings (Carter and Soo 1999); earnings announcements and information released contemporaneously in the same disclosure (Francis et al. 2002); earnings announcements (Landsman and Maydew 2002); EDGAR 10-K filings (Qi et al. 2000); and real estate trust funds flow information (Vincent 1999). Of particular relevance to our analysis are studies that do not detect a statistically significant investor response to a hypothesized information event. Although the literature contains notable (widely cited) nonresponse findings such as Bernard and Stober's (1989) documentation of a nonresponse to cash flow information, and the absence of market response to various types of qualified audit opinions reported in Dodd et al. (1984), a more common outcome of a nonresponse "discovery" is an unpublishable or unfinished analysis. (1) Our analysis bears most directly on nonresponse settings because the reliability of such null "conclusions" is directly related to the underlying power of the research design.

Most short-window response studies use daily returns as their primary response metric. Variation in daily returns, however, is sizable, commonly as much as 4 percent of a firm's equity value. In the presence of such high variance, it is difficult to detect investor responses to information events that have modest valuation effects. Indeed, Cready and Mynatt (1991) document that even in sample sizes of near 600, the likelihood of detecting a half-percent shift in firm value occurring over a three-day event window is only about 60 percent, and they achieve this likelihood only by using an aggressive 20 percent Type I error rate (i.e., a 20 percent chance of rejecting the null hypothesis when it is true). The inherent noisiness of return metrics suggests investigating whether volume-based metrics increase the likelihood of correctly rejecting null hypotheses in low-power contexts. We examine this issue by evaluating return-based and volume-based metrics in earnings announcement periods. Assuming that the null hypothesis--that the mean investor response to earnings announcements is zero--is false for the population of earnings announcements, (2) this analysis yields insight into the relative power of tests based on returns vis-a-vis volume in detecting investor responses to earnings announcements.

Prior research on event-study methods identifies which of several alternative return response metrics (Brown and Warner 1980, 1985; Rohrbach and Chandra 1989), or which of several alternative volume response metrics (Cready and Ramanan 1991, 1995), is most likely to correctly reject the null hypothesis of no investor response to information events when the null hypothesis is false. In contrast to these prior studies that focused on either return-based or volume-based measures, we compare return- and volume-based metrics. That is, given a null hypothesis that information events lead to no investor response, do return-based or volume-based metrics provide the most powerful test of this null? We address this question using a framework similar to Brown and Warner (1980, 1985) and Dyckman et al. (1984). Specifically, from the grand sample of earnings announcements, we form many relatively small subsamples of earnings announcements, and then we empirically estimate each metric's sample-size-specific rejection rate. We find that in earnings announcement periods, where we assume that the null hypothesis of no reaction to earnings announcements is false, rejection rates are significantly higher for tests using volume-based metrics than for similar tests using return-based metrics.

Volume and return are complementary measures that capture different aspects of investor response to information events. Bamber and Cheon (1995) suggest that the relation between volume and return responses to earning announcements is closer to independence than it is to a "strong positive relation." Kandel and Pearson (1995) report that trading volume around earnings announcements is, on average, abnormally high even for earnings announcements that stimulate negligible return response. If volume and return measures are sufficiently complementary, then tests of market response based on both metrics should be more precise than tests based on either metric alone. We find that supplementing a return-based analysis with a volume-based analysis increases the likelihood of correctly rejecting the null hypothesis of investor nonresponse. The converse, however, is not true. Hence, our analysis suggests that when the focal research question is determining the presence or absence of an investor response to information events, and when power is critical (e.g., small sample or anticipated modest investor response), volume-based designs outperform return-based designs. Moreover, our results suggest that research concluding that there is no investor response to information events (i.e., that the information event does not convey incremental new information) would be more reliable if confirmed by evidence of insignificant abnormal trading volume, particularly if power is crucial.

In our evaluation of alternative price-response metrics, we also find that nondirectional return metrics outperform directional tests that condition the return sign on the earnings surprise. That is, when information on the likely direction of the return response to information events is noisy, simply ignoring this directional information leads to more powerful tests of investor response or nonresponse than do signed returned metrics. Because directional tests based on perfect knowledge of the direction of the return are more powerful than their nondirectional price-based competitors (Brown and Warner 1985), our results suggest that conventional earnings expectation measures suffer from substantial measurement error.

II. RESEARCH DESIGN

Most studies addressing the power of alternative market-response metrics seed existing values of the metric with specific permutations and then examine the ability of various statistical methods to detect the permutation correctly (by rejecting the null that no permutation is present). Unfortunately, such an approach is not feasible in cross-metric comparisons because a 1 percent return effect does not equal a 1 percent volume effect (or any other identifiable permutation of the volume effect). We address this problem by using earnings announcements, which we assume stimulate nonzero mean investor response (see footnote 2), as contemporaneous implicit return and volume seeds.

Limitations of Using Announcement-Defined Investor Responses

Using return and volume responses to actual earnings announcements leads to several limitations. First, we do not explicitly manipulate the magnitude of the return and volume inducements, so it is not possible to report rejection rates for 1 percent inducements, 2 percent inducements, and so forth, because earnings announcement returns and volumes implicitly define the inducements in a manner that the researcher cannot manipulate. However, we can probe rejection rate properties by varying sample size. The cross-sectional t-test statistic of the null hypothesis that the mean shift, U, equals zero is:

t = ([n.sup.1/2]) U/[sigma].

Since the numerator in this expression is the product of U, the mean shift, and n, the sample size, changing n holding U constant (assuming that U is nonzero) has the same directional effect on t as changing U holding n constant. Hence, in this context, observing how rejection rates behave when varying n is isomorphic to observing their behavior when varying U.

The largest sample size we examine is 40. Although studies employing sample sizes of such small magnitude are uncommon, they are far from rare. In an analysis of event-study sample sizes employed in papers published in The Accounting Review during the years 1992-2001, we identified 35 studies employing short-window response metrics. Of these, 16 reported at least one substantive analysis of a subsample or data partition containing fewer than 50 observations. (3) Moreover, the fungibility of the sample size and mean shift effects in the above t-statistic calculation allows us to extrapolate our small sample size results to larger samples. For example, if we expect that the mean investor response to the announcement of firms' monthly sales data is a proportion p, say 20 percent, of the population earnings announcement response effect (U), then we can extrapolate the results reported here for a sample size of n earnings announcements to a sample size of n[(1/0.20).sup.2] for monthly sales announcements for which the mean investor response is 0.2U. Here, for instance, the results for a sample of 40 earnings announcements would correspond to the results of a sample of 1,000 (40[[1/0.20].sup.2]) monthly sales announcements. (4)

A second, related limitation is that because the earnings-announcement-period data identify the magnitudes of the underlying mean return and volume responses, our results directly pertain only to contexts in which the ratio of return to volume response is similar to that at earnings announcement dates. In other contexts where the ratio of return-to-volume response is greater than that for earnings announcements, the relative performance of volume-based tests should be lower than that reported here.

A third limitation is that the analysis assumes that earnings announcements cause population-wide price and trading responses. However, Kandel and Pearson (1995) find that approximately 10 percent of their sample of earnings announcements stimulate no significant price movement in the three days around the earnings announcement date. Bamber et al. (2000) suggest that prior research's conclusion that the return response is associated with earnings surprises is largely attributable to a relatively small subpopulation of earnings announcements. Such findings suggest that the earnings announcement population may consist of a mixture of unresponsive and responsive groups.

If the percentage of earnings announcements that stimulate price responses differs substantially from the percentage of earnings announcements that stimulate trading, then, ceteris paribus, the likelihood of rejecting the null hypothesis should be greater for the more broadly responsive metric than for the less broadly responsive metric.

At a purely unconditional level, the fact that differential responsiveness can explain any observed difference in rejection rates is irrelevant to our tests. Our analysis allows mean shifts to differ between the returns vis-a-vis volume metrics, so any responsiveness differences affect these mean shifts. That is, the overall population mean shift (U) is a function of both: (1) the percentage of the population that is responsive with respect to a given metric, q, and (2) the mean of the responsive subpopulation, [U.sup.r], because U = q[U.sup.r]. Thus, although an increase in response rate q necessarily increases U, an increase in U could be attributable to either: (1) a stronger reaction in the responsive subpopulation, or to (2) an increase in the response rate, q. Hence, our analysis does not provide direct evidence on relative cross-metric response rates (i.e., q's).

Differential responsiveness might matter, however, when researchers screen out likely zero response observations to increase the power of their tests, as Bamber et al. (2000) suggest occurs in Beaver (1968). In such cases the underlying mean for the post-screen subpopulation approaches [U.sup.r], which necessarily exceeds U. Effective screens should lead to higher rejection rates than reported here. (5) The generalizability of our cross-metric inferences depends on whether the research design differentially screens out zero response observations across return vs. volume metrics. If a screen eliminates similar numbers of zero (or near-zero) response observations for both metrics, then the power associated with each metric increases, but any differences in power across the two remain intact.

Response Metrics

We consider four distinct event-period response metrics: (1) percentage of outstanding shares traded, (2) number of transactions, (3) absolute return, and (4) return conditioned on the direction of the earnings news.

Beaver (1968) introduces the percentage of outstanding shares traded as a measure of market response. The specific measure we use is:

VSHR[S.sub.it] = ([V.sub.it] - E[V.sub.it])/[[sigma].sub.Vi]

where:

[V.sub.it] = the log of (100 times shares traded in firm i's stock on day t divided by the outstanding shares for firm i on day t, plus 0.000255); (6)

[[sigma].sub.Vi] = the standard deviation of the residuals from the market-volume regression used to determine EV; (7) and

E[V.sub.it] = the predicted level of [V.sub.it] from a first-order serial correlation regression of [V.sub.it] on [V.sub.MKT] where [V.sub.MKT] is the percentage of outstanding shares traded for all NYSE-listed firms in CRSP, estimated over days -55 to -6 and +6 to +55 relative to the announcement date. (8)

Cready and Ramanan (1991) indicate that market-adjusted volume outperforms simple mean-adjusted volume in detecting seeded trading increases. The serial correlation regression is similar to the methods Pincus (1983) and Cready and Mynatt (1991) used to correct for the strong positive serial correlation in daily volume. In addition, following Cready and Ramanan (1995), we use only the structural portion of the model to generate EV values in non-initial event test period days (e.g., days 0 and +1 when the test period is days -1 to +1).

Cready and Ramanan (1995) find that tests based on the number of trades, or transactions, outperform tests based on percentage of shares traded in detecting both seeded and announcement-period trading effects. Data on the number of transactions are now available in the ISSM and Trades and Quotes (TAQ) databases. Hence, we also consider a transactions-based metric:

VTRA[N.sub.it] = (T[R.sub.it] - PT[R.sub.it])/[[sigma].sub.TRi]

where:

T[R.sub.it] = the log of (number of transactions in firm i's stock on day t plus 1); (9)

PT[R.sub.it] = the predicted level of T[R.sub.it] from a first-order serial correlation regression of TR on T[R.sub.MKT] where T[R.sub.MKT] is based on the number of transactions for all NYSE firms in the ISSM database on that day; and

[[sigma].sub.TRi] = the standard deviation of the residuals from the market-transactions regression used to determine PTR.

May (1971) introduces absolute returns as a measure of market response to interim earnings announcements. Subsequent analysis by Rohrbach and Chandra (1989), Cready and Mynatt (1991), and Subramaniam (1995) indicate that absolute return is more powerful than Beaver's (1968) squared return metric in identifying unconditional price responses. In untabulated analyses, we confirm that absolute returns outperform squared returns in detecting announcement period price responses. The specific absolute-return response metric we employ is: (10)

ABRE[T.sub.it] = (|[R.sub.ikt] - [R.sub.kt]| - MEAN-ABRE[T.sub.is])/[[sigma].sub.ARiks]

where:

[R.sub.ikt] = the return on firm i, a member of CRSP size-decile k, on day t;

[R.sub.kt] = the CRSP size-decile k return on day t;

MEAN-ABRE[T.sub.is] = the mean value of |[R.sub.ikt] - [R.sub.kt]| over estimation periods (days -105 to -6 relative to the announcement date); and

[[sigma].sub.ARiks] = the standard deviation of |[R.sub.ikt] - [R.sub.kt]| estimated over estimation period s.

We calculate ABRET and then sum by day over the multiday response periods. For example, for a two-day period we determine ABRET values for each of the two days and then add these two values together to derive the two-day ABRET. (11)

Studies often incorporate the expected direction of the return effect into the research design. We incorporate expected direction into our analysis by conditioning returns on the sign of the difference between actual and expected earnings. Following Foster (1977), when actual earnings fall below expected earnings we multiply the return by negative one (i.e., a short sale) and perform statistical analysis using one-tailed tests on this signed return metric. If we have accurately measured the conditioning variable--in this case, unexpected earnings--then the observed return in direction-based designs incorporates additional information into the test procedure. In general, the more information upon which a test procedure is based, the greater is its power to reject false null hypotheses. However, if there is noise in the conditioning variable--that is, if "good news" announcements are misclassified as "bad news," and vice versa--then such conditioning can add noise to the design.

We examine the power of conditional return designs using the following metric:

URE[T.sub.it] = [I.sub.i] ([R.sub.ikt] - [R.sub.kt])/[[sigma].sub.URiks]

where:

[I.sub.i] = an indicator variable set to 1 if we predict the earnings announcement increases price, and is set to -1 if we predict the earnings announcement decreases price; and

[[sigma].sub.URikj] = the standard deviation of ([R.sub.ikt] - [R.sub.kt]) estimated over estimation period s.

In earnings announcement research, the earnings forecast error most commonly determines the indicator variable's sign. We use two alternative measures of forecast error: random-walk error based on year-ago quarterly earnings (RW-URET) and analyst forecast error based on the I/B/E/S mean forecast from the month preceding the earnings-announcement release month (AF-URET).

Grand Sample

The study employs a grand sample of 49,274 earnings announcements between January 1, 1983 and December 31, 1992. We include earnings announcements in this sample if (1) the firm is listed on the NYSE; (2) the earnings announcement date is available from Compustat; (3) return data are available for at least 80 of the 100 days in the day -105 to -6 period relative to the announcement date; (4) transactions and volume data are available for at least 80 of the 100 days from the time periods spanning day -55 to -6 and day +6 to +55 relative to the announcement date. We obtain return and volume data from CRSP and transactions data from the ISSM database. (12)

Table 1 presents descriptive information on the grand sample, partitioned by earnings announcement date into three time periods: 1983-1986, 1987-1989, and 1990-1992. The increase in sample size from 14,063 announcements in the initial 1983-1986 period to 19,311 in the 1990-1992 period is attributable to (1) a general increase in the number of NYSE-listed firms; (2) better trading-data coverage in the ISSM database (early years suffer from a large number of missing trading data); and (3) more accurate cross-referencing of firms by ISSM.

Table 1 indicates that daily non-earnings-announcement-period percentage volume decreases slightly from 0.23 percent to 0.22 percent of outstanding shares per day. Numbers of transactions and absolute return, however, both increase over the ten-year sample period, with the number of transactions increasing from 48.9 to 74.0 per day and daily absolute return increasing from 1.48 percent to 1.88 percent per day.

Earnings-announcement-period increases in percentage volume, number of transactions, and absolute returns are greatest in the 1990-1992 time period. Daily means for all three metrics decrease when we extend the event period from days -1 to 0, to days -1 to + 1. For the two trading measures this decrease is slight. However, the average daily absolute return effect decreases by 0.07 percentage points after including day +1.

Table 1 also reports descriptive statistics for random-walk and analyst-based earnings forecast errors. We calculate the random-walk error per share as the difference between the announced quarterly earnings per share (EPS) and year-ago EPS for that fiscal quarter. Requiring current- and prior-year EPS (from Compustat) reduced the number of firm announcements for which we have random-walk forecast error to 46,874. We used the I/B/E/S summary data set to compute analyst forecast error as the difference between I/B/E/S earnings and the mean I/B/E/S forecast from the month prior to the announcement month. (13) Because I/B/E/S covers fewer firms, we have analyst forecast errors for only 28,129 of the grand sample of earnings announcements.

Employing an investment strategy of buying stocks with positive forecast errors and selling short firms with negative forecast errors, based on foreknowledge of the sign of the random-walk forecast error on day -1, yields an average return of 0.45 percent per day over days -1 to 0 for the random-walk sample. Foreknowledge of the sign of the analyst forecast error on day -1 yields average daily returns of 0.39 percent over days -1 to 0. The lower return to foreknowledge of the analyst forecast error is attributable to selection bias in I/B/E/S coverage: the random-walk-based average return for only those firms for which I/B/E/S data are also available falls to 0.35 percent over days -1 to 0, which is less than the companion analyst forecast error return.

Statistical Test Procedures

To evaluate the statistical power of alternative metrics we use the grand sample to generate smaller test samples. We randomly generate (without replacement) 1,000 samples of 15 earnings announcements, 25 earnings announcements, and 40 earnings announcements from the grand sample. (Because of grand-sample-size restrictions, we generate only 500 samples of each of the three sample sizes when we compare the analyst forecast metrics.) We then test the null hypothesis that each metric has a zero mean response on each sample (j) using cross-sectional t-tests. (14)

Table 2 provides evidence that rejection rates for our test procedures and metrics are well-behaved. The table reports results of testing the hypotheses of a zero mean response using data from days -10 to -8 relative to the earnings announcement date, when the null hypothesis of no market response is likely true. (15) The rejection rates reported in Table 2 reveal no evidence of overrejection. Indeed, the tests appear somewhat conservative in that rejection rates tend to be too small rather than too large. In no instance does the rejection rate exceed its expected level under the null hypothesis (i.e., 5 percent in Panel A; 1 percent in Panel B).

III. RESULTS

Unconditional Response Metric Performance

Table 3 reports announcement period rejection rates for the VSHRS, VTRAN, and ABRET metrics by sample size using 5 percent (Panel A) and 1 percent (Panel B) significance levels. Rejection rates for the two trading metrics, VSHRS and VTRAN, significantly exceed their counterpart ABRET rates (p < 0.01). (16) The 5 percent significance level day -1 to 0 VSHRS metric rejection rates of 37.6 percent, 53.3 percent, and 72.0 percent for sample sizes of 15, 25, and 40, respectively, exceed their counterpart ABRET-based rejection rates by between 13.0 (sample size of 15) and 8.7 (sample size of 40) percentage points. The two-day 5 percent VTRAN rejection rates of 45.9 percent, 64.2 percent, and 84.8 percent exceed counterpart ABRET-based rejection rates by more than 20 percentage points. Table 3 shows similar rejection-rate patterns for tests based on 1 percent significance levels and three-day event periods. (17)

One primary means researchers can use to increase the power of their statistical tests is to increase the sample size. Therefore, one way of evaluating the implications of the differential cross-metric rejection percentages reported in Table 3 is to determine how much the sample size would have to increase for the ABRET-based rejection levels to equal their volume-based counterparts. The results reported in Table 3 suggest that researchers would need, for example, samples of 40 ABRETs around earnings announcements to attain rejection rates comparable to the rejection rates achieved by just 25 announcement-associated VTRANs. That is, replacing the ABRET metric with the VTRAN metric increases power to roughly the same extent as increasing sample size by 60 percent.

We also explored the extent to which these sample size equivalencies hold in larger samples sizes by sampling with replacement. In these untabulated analyses, we reach a 95 percent rejection rate (5 percent probability of a Type II error) for sample sizes of 70, 55, and 85 for the VSHRS, VTRAN, and ABRET metrics, respectively. Hence, a 21 percent increase in sample size brings the ABRET-based test's power in line with the power of the VSHRS-based test, while a 51 percent increase in sample size brings it in line with the power of the VTRAN-based test. These percentage increases are somewhat smaller than those implied by the results reported in Table 3 for smaller sample sizes, but nonetheless reveal that the substantial power advantage of the VSHRS and VTRAN metrics extends beyond the range of sample sizes considered in Table 3.

Conditional (Signed) Return Analysis

Table 4 reports rejection rates for tests based on the RW-URET (Panel A) and AF-URET (Panel B) metrics for sample sizes of 15, 25, and 40 announcements. These conditional (signed) return metrics require earnings from the same quarter of the prior year (RW-URET) and analyst earnings forecasts (AF-URET). Data unavailability reduces the grand sample from which we form subsamples. For example, we can compute AF-URET measures for only 28,129 announcements. This makes it impossible to form 1,000 40-announcement subsamples without replacement, so we base AF-URET rejection rates on only 500 subsamples. Since data availability may be associated with systematic differences in announcement-period return and volume magnitudes, we also calculate rejection rates for VSHRS, VTRAN, and ABRET for the 500 RW-URET subsamples (Panel A) and for the 500 AF-URET subsamples (Panel B). Panels A and B of Table 4 show that rejection rates for the RW-URET and AF-URET conditional return metrics are considerably smaller than those for either of the counterpart trading-response metrics (p < 0.01). Indeed, rejection rates for the VTRAN metrics are at least twice as large as counterpart rejection rates for the RW-URET and AF-URET metrics.

Table 4 also reports that the rejection rates for the unconditional ABRET metric are significantly higher (p < 0.01) than the counterpart rates for the RW-URET and AF-URET metrics. Indeed, for the largest sample size of 40 announcements, the ABRET rejection rates exceed counterpart: (1) RW-URET rates by 27.2 (two-day period) and 31.7 (three-day period) percentage points, and (2) AF-URET rates by 31.6 (two-day period) and 32.4 (three-day period) percentage points. Thus, error in signing the earnings surprise outweighs the direction-based return metrics' potential to provide more powerful tests of market response to earnings announcements. (18)

Joint Price/Volume Tests

Bamber and Cheon's (1995) finding that volume and return are only modestly correlated at announcement dates suggests that a joint test of volume and return response may be more powerful than a simple examination of either volume or return exclusively. We investigate this possibility by conducting tests based on the joint outcome of volume and return response tests. In these joint volume and return tests we use a nominal 2.5 percent significance level rejection rule to independently evaluate whether each metric-specific test (i.e., volume or return) leads to rejection of the nonresponse null. We then reject the general hypothesis of no investor response to the information release if either the null of no volume response or the null of no return response is rejected. Based on Bonterroni inequalities, the bounded overall p-level for our two-test analysis of a single hypothesis is 5 percent. (19)

Panel A of Table 5 reports rejection rates when we employ ABRET in combination with VSHRS (ABRET&VSHRS). Panel B reports these rates for the ABRET and VTRAN (ABRET&VTRAN) combination. To facilitate comparisons, these panels also re-report the stand-alone 5 percent significance level rejection rates for VSHRS, VTRAN, and ABRET originally reported in Table 3. The ABRET&VSHRS rejection rates reported in Panel A exceed counterpart ABRET-only rejection rates for all three samples sizes, by between 7.3 and 9.8 percentage points. All differences are significant at the 1 percent level. The negligible differences between the ABRET&VSHRS rejection rates and the VSHRS-only rates suggest that their performances are comparable.

Panel B of Table 5 indicates that two-day accumulation period rejection rates for the ABRET&VTRAN test exceed counterpart ABRET-only rejection rates (based on a 5 percent rejection rule) by 12.0, 15.2, and 18.1 percentage points for the 15, 25, and 40 announcement samples, respectively. The three-day accumulation period rejection rates exhibit a similar pattern. All differences are significant at the 1 percent level, indicating that ABRET&VTRAN-based tests are more powerful than tests based on ABRET alone. However, Panel B also shows that VTRAN-only rejection rates are significantly higher (0.05 level or better) than ABRET&VTRAN-based tests. The difference in rejection rates narrows considerably as sample size increases, dropping to only 2.1 percentage points for the three-day-window, 40-announcement samples. In sum, our results suggest that although it is advantageous to supplement the absolute return response analyses with volume-based metrics, there is no obvious advantage to supplementing a volume-based test with a return-based test.

IV. CONCLUSION

Bamber and Cheon (1995, 440) conclude that because there is considerable independence between the market's return and trading responses at earnings announcement dates, "information content-oriented research should continue to examine both price and volume reactions." Our primary finding, that tests of investor response to earnings announcements are more likely to correctly reject the null hypothesis of no investor response based on measures of trading than are comparable tests based on returns, reinforces their conclusion. Our results further suggest that not only is volume a nonredundant measure of investor response to information events, but it is also a more powerful measure. Our analysis indicates that volume's power advantage can be sizable--in the settings we examine, the best-performing volume metric, the metric based on the number of transactions, increases power beyond that of the return metric equal to a 60-plus percent increase in sample size.

The importance of employing powerful test designs depends on the specific research context. When sample size is very large or the investor response is sizable, any reasonable design and metric is likely to reject a false null. However, when sample size is small or return responses are small, our analysis suggests that volume-based response metrics are more likely to detect the presence of an investor response than are return-based metrics. In addition, and as a general rule, in any setting where return-based analyses fail to reveal evidence of investor response, it is important to conduct powerful tests before concluding that no response is present. Hence, before concluding that investors do not respond to a public disclosure based on a returns analysis, we recommend that researchers confirm the nonresponse inference using a trading-based analysis.

Our findings, although consistent with the notion that trading response is a more widespread phenomenon than price response (Kandel and Pearson 1995), do not provide direct evidence on the relative incidence of price vis-a-vis trading response. Rejection-rate levels reflect (relative) mean shift magnitudes, and these magnitudes increase with both response frequency and response magnitude. Hence, our findings are attributable to either: (1) more earnings announcements stimulate trading responses than return responses, or (2) trading response magnitudes are simply larger (in a statistically detectable sense) than return response magnitudes.

Finally, two significant caveats temper our conclusion that tests based on volume metrics outperform tests based on return metrics. First, our analysis applies to questions of the existence or nonexistence of some sort of investor response to information. If, however, a researcher is investigating a more subtle question, such as a theoretically expected valuation effect, or a specific form of investor behavior that is most directly revealed by their trading decisions, then the research should of course use the metric that is most appropriate. The second caveat is that the generalizability of our inferences to other information events depends on whether the relative magnitudes of volume and return response to earnings announcements typifies the response to other information events. While we have no ex ante reason to believe this is not the case, if other information events stimulate larger return response relative to trading response, then our conclusion that volume-based metrics provide more powerful tests may no longer apply.

TABLE 1
Descriptive Statistics for Grand Sample of Earnings Announcements
between 1/1/83 and 12/31/92 (a)

                              Announcement Date          Full

Descriptive Variable          1983-86  1987-89  1990-92  Sample

Number of earnings
 announcements                14,063   15,900   19,311   49,274

Mean market value of equity
 (in millions) (b)            $1,032   $1,358   $1,606   $1,359

Mean Estimation Period
 Statistics:
  Daily percentage
   volume (c)                   0.23     0.22     0.22     0.22
  Daily number of
   transactions (d)            48.9     55.8     74.0     61.0
  Daily absolute
   size-adjusted return (e)     1.48%    1.57%    1.88%    1.66%

Earnings Announcement Period
 Statistics:
Mean Increase in Percentage
 Volume:
  Day -1 to 0 average
   (in % points)                0.077    0.065    0.092    0.080
  Day -1 to -1 average
   (in % points)                0.065    0.071    0.092    0.078

Mean Increase in Number of
 Transactions
  Day -1 to 0 average            8.4     10.7     19.7     13.7
  Day -1 to +1 average           7.2      9.5     19.4     13.5

Mean Increase in Absolute
 Size-Adjusted Return
  Day -1 to 0 average
   (in % points)                 0.47     0.41     0.64     0.52
  Day -1 to +1 average
   (in % points)                 0.36     0.35     0.58     0.45

Mean Random-walk forecast
 error (per share) (f)          -$0.42    $0.10   -$0.06   -$0.11

Number of random-walk
 forecast errors                13,529   15,107   18,238   46,874

Mean Analyst forecast error
 (per share) (g)                -$0.16    $0.01   -$0.02   -$0.05

Number of analyst forecast
 errors                          7,395    9,699   11,035   28,129

Mean Random-Walk Signed
 Return (h)
  Day -1 to 0 average
   (in % points)                 0.47     0.37     0.49     0.45
  Day -1 to +1 average
   (in % points)                 0.34     0.30     0.42     0.36

Mean Analyst Signed
 Return (h)
  Day -1 to 0 average
   (in % points)                 0.36     0.29     0.51     0.39
  Day -1 to +1 average
   (in % points)                 0.27     0.24     0.41     0.32

(a) The grand sample includes 49,274 annual and quarterly earnings
announcements by NYSE-listed firms.

(b) Market value of equity equals share price multiplied by the
number of shares outstanding on trading day -106
relative to each announcement date.

(c) For each announcement, we calculate mean estimation period daily
percentage volume as the percentage of
outstanding shares traded each day per CRSP, averaged over trading days
-55 to -6 and +6 to +55 relative to
the announcement date.

(d) For each announcement, we calculate mean estimation period daily
number of transactions as the number of
transactions (i.e., trades) occurring in the firm's common stock,
calculated from the ISSM data base, averaged
over trading days -55 to -6 and +6 to +55 relative to the announcement
date.

(e) For each announcement, we calculate mean estimation period daily
absolute size-adjusted return as the absolute
value of the difference in the announcing firm's return and the
size-decile-matched CRSP portfolio return averaged
over trading days -105 to -6 relative to the announcement date.

(f) Random-walk forecast error is the difference between earnings
in quarter q and earnings for the same quarter of
the previous year.

(g) Analyst forecast error is the difference between actual quarterly
earnings (per I/B/E/S) and the mean I/B/E/S
earnings forecast for the month preceding the announcement.

(h) We multiply returns for negative earnings forecast error
announcements by -1.

TABLE 2
Rejection Rates for the Null Hypothesis of No Increase in the Mean
for Alternative Investor Response Metrics in Days -10 to -8 around
Earnings Announcement Dates (a)

Panel A: Rejection Rates across 1,000 Non-Overlapping
Samples Using 5 Percent, One-Tailed, Rejection Rule (b)

                      Two-Day                    Three-Day
               Accumulation Period (c)    Accumulation Period (d)

                  Sample Size                Sample Size
Response
Metric (a)     15       25       40       15       25       40

VSHRS         3.3%     2.8%     2.0%     2.6%     2.7%     2.0%
VTRAN         2.9      2.9      2.5      2.6      2.9      1.8
ABRET         2.7      3.3      2.8      2.2      2.9      3.0
URET          4.5      4.6      4.4      3.9      4.9      4.0

Panel B: Rejection Rates across 1,000 Non-Overlapping
Samples Using 1 Percent, One-Tailed, Rejection Rule (b)

                      Two-Day                    Three-Day
               Accumulation Period (c)    Accumulation Period (d)

                  Sample Size                Sample Size
Response
Metric (e)     15       25       40       15       25       40

VSHRS         0.5%     0.7%     0.3%     0.4%     0.6%     0.4%
VTRAN         0.6      0.7      0.6      0.5      0.6      0.6
ABRET         0.2      0.1      0.4      0.3      0.5      0.5
URET          0.6      0.5      0.8      0.8      0.6      0.5

(a) All samples are random without-replacement draws from the grand
sample of 49,871 annual and quarterly earnings announcements by
NYSE-listed firms between 1/1/83 and 12/31/92.

(b) We base rejection on a cross-sectional t-test of the null
hypothesis that the response metric mean is less than or
equal to 0.

(c) We accumulate each metric over days -10 and -9 relative
to the earnings announcement date.

(d) We accumulate each metric over days -10 through -8
relative to the earnings announcement date.

(e) Response metric definitions:

VSHRS = the actual percentage of outstanding shares traded,
        less the predicted percentage of outstanding shares
        traded (based on a time-series first-order serial
        correlation regression of daily percentage volume
        over days -60 to -11 and +11 to +60 relative to the
        announcement date volume, on the percentage of shares
        traded for all NYSE-listed firms), where this difference
        is divided by the standard deviation of the residuals
        from the regression (volumes are log-transformed);

VTRAN = the actual number of transactions, less the predicted
        number of transactions (based on a time-series
        first-order serial correlation regression of the
        daily number of transactions over days -60 to -11 and
        +11 to +60 relative to the announcement date, on the
        total number of all NYSE-listed firm transactions), where
        this difference is divided by the standard deviation of
        the residuals from the regression (the number of
        transactions is log-transformed);

ABRET = the absolute value of the difference between actual daily
        return and return on the CRSP size-matched decile
        portfolio for the announcing firm, less the mean of this
        same absolute difference over days -110 to -11 relative
        to the announcement date, where this difference is divided
        by the standard deviation of the absolute differences over
        days-110 to -11; and

URET = the difference between actual daily return and the return
       on the CRSP size-matched decile portfolio for the announcing
       firm, divided by the standard deviation of this difference
       over days -110 to -11 relative to the announcement date.

TABLE 3
Rejection Rates for the Null Hypothesis of No Increase in the Mean
for Alternative Investor Response Metrics in Days -1 to +1 Relative
to Earnings Announcement Dates (a)

Panel A: Rejection Rates across 1,000 Non-Overlapping
Samples Using 5 Percent, One-Tailed, Rejection Rule (b)

                    Two-Day                Three-Day
             Accumulation Period (c)  Accumulation Period (d)

                  Sample Size            Sample Size
Response
Metric (e)    15      25      40      15      25      40

VSHRS        37.6%   53.3%   72.0%   41.2%   57.1%   77.7%
VTRAN        45.9    64.2    84.8    48.7    68.6    87.1
ABRET        24.6    43.5    63.3    27.8    47.5    68.1

All VSHRS and VTRAN rejection rates significantly exceed
their counterpart ABRET rejection percentages at the 1
percent significance level using McNemar's [chi square] test.

Panel B: Rejection Rates across 1,000 Non-Overlapping
Samples Using 1 Percent, One-Tailed, Rejection Rule (b)

                    Two-Day                Three-Day
             Accumulation Period (c)  Accumulation Period (d)

                   Sample Size            Sample Size
Response
Metric (E)    15      25      40      15      25      40

VSHRS        14.0%   25.9%   43.2%   17.3%   30.0%   51.2%
VTRAN        17.3    33.5    59.5    20.3    37.6    66.1
ABRET         3.9    12.7    30.1     5.9    14.5    34.2

All VSHRS and VTRAN rejection rates significantly exceed
their counterpart ABRET rejection percentages at the 1
percent significance level using McNemar's [chi square] test.

(a) All samples are random without-replacement draws from
the grand sample of 49,274 annual and quarterly earnings
announcements by NYSE-listed firms occurring between 1/1/83
and 12/31/92.

(b) We base rejection on a cross-sectional t-test of the
null hypothesis that the response metric mean is less than
or equal to 0.

(c) We accumulate each metric over days -1 and 0 relative
to the earnings announcement date.

(d) We accumulate each metric over days -1 through +1
relative to the earnings announcement date.

(e) Response metric definitions:

VSHRS = the actual percentage of outstanding shares traded,
        less the predicted percentage of outstanding shares
        traded (based on a time-series first-order serial
        correlation regression of daily percentage volume
        over days -55 to -6 and +6 to +55 relative to the
        announcement date volume, on the percentage of
        shares traded for all NYSE-listed firms), where this
        difference is divided by the standard deviation of
        the residuals from the regression (volumes are
        log-transformed);

VTRAN = the actual number of transactions, less the predicted
        number of transactions (based on a time-series
        first-order serial correlation regression of the daily
        number of transactions over days -55 to -6 and +6 to +55
        relative to the announcement date, on the total number of
        all NYSE-listed firm transactions), where this difference
        is divided by the standard deviation of the residuals from
        the regression (the number of transactions is
        log-transformed); and

ABRET = the absolute value of the difference between actual daily
        return and return on the CRSP size-matched decile portfolio
        for the announcing firm, less the mean of this same absolute
        difference over days -105 to -6 relative to the announcement
        date, where this difference is divided by the standard
        deviation of the absolute differences over days -105 to -6.

TABLE 4
Rejection Rates (5 Percent p-level) for the Null Hypothesis
of No Increase in the Mean for Conditional (Signed) Unexpected
Return vs. Unconditional Trading and Absolute Return Response
Metrics in Days -1 to +1 Relative to Earnings Announcement Dates (a)

Panel A: Rejection Rates by Sample Size for Unexpected Return
Conditioned on Sign of Change in Earnings Per Share (RW-URET),
Transactions-Based (VTRAN), Percentage Volume-Based (VSHRS),
and Absolute Return-Based (ABRET) Metrics for 500 Non-Overlapping
Samples (b)

                    Two-Day                  Three-Day
             Accumulation Period (c)   Accumulation Period (d)

                  Sample Size               Sample Size
Response
Metric (e)    15      25      40        15      25      40

VSHRS        37.9%   53.8%   72.6%     40.2%   57.7%   78.0%
VTRAN        44.9    65.2    85.2      47.7    70.1    87.7
ABRET        23.7    43.4    63.1      26.9    47.7    67.7
RW-URET      18.8    25.4    35.9      18.5    25.3    36.0

All VTRAN, VSHRS, and ABRET rejection rates significantly
exceed their counterpart RW-URET rejection percentages at
the 1 percent significance level using McNemar's [chi
square] test.

Panel B: Rejection Rates by Sample Size for Unexpected Return
Conditioned on Analyst Earnings Forecast Error (AF-URET), VTRAN,
VSHRS, and ABRET Metrics for 500 Non-Overlapping Samples (b)

                    Two-Day                  Three-Day
             Accumulation Period (c)   Accumulation Period (d)

                  Sample Size               Sample Size
Response
Metric (e)    15      25      40        15      25      40

VSHRS        38.9%   54.6%   76.8%     41.3%   61.0%   80.8%
VTRAN        45.2    67.0    86.6      48.2    71.4    89.0
ABRET        26.9    46.5    71.0      32.8    51.9    72.4
AF-URET      21.8    283.0   39.4      20.7    26.9    38.0

All VTRAN, VSHRS, and ABRET rejection rates significantly exceed
their counterpart UR-W rejection percentages at the 1 percent
significance level using McNemar's [chi square] test.

(a) All samples are random draws from the grand sample of 28,129
annual and quarterly earnings announcements for which analysts
forecasts are available, by NYSE-listed firms occurring between
1/1/83 and 12/31/92.

(b) We base rejection on a cross-sectional t-test of the null
hypothesis that the response metric mean is less than or
equal to 0.

(c) We accumulate each metric over days -1 and 0 relative to
the earnings announcement date.

(d) We accumulate each metric over days -1 through 1 relative
to the earnings announcement date.

(e) The VSHRS, VTRAN, and ABRET response metrics are
described in the notes to Table 3. RW-URET is URET,
as described in Table 3, multiplied by -1 if actual
quarterly earnings is less than earnings for the same
quarter of the previous year, and multiplied by 1
otherwise. AF-URET is URET multiplied by -1 if actual
quarterly earnings (per I/B/E/S) is less than the mean
I/B/E/S earnings forecast for the month preceding the
announcement, and multiplied by 1 otherwise.

TABLE 5
Power of Joint Return/Volume Tests to Reject the Null Hypothesis
of No Investor Response in Days -1 to +1 around Earnings Announcement
Dates (a)

Panel A: Rejection Rates for Joint Return and Volume Test of the
Null Hypotheses of No Investor Response (ABRET&VSHRS) across 1,000
Non-Overlapping Samples Using 1 Percent, One-Tailed, Rejection Rule (b)

                        Two-Day                   Three-Day
                 Accumulation Period (c)  Accumulation Period (d)

                       Sample Size               Sample Size
Test Metric/
Procedure (e)   15      25       40       15       25       40

ABRET&VSHRS   32.1%   50.8%    73.2%    36.1%    57.3%    77.9%
VSHRS only    37.6 ** 53.3     72.0     41.2 **  57.1     77.7
ABRET only    24.6 ** 43.5 **  63.3 **  27.8 **  47.5 **  68.1 **

Panel B: Rejection Rates for Joint Return and Volume Test of the
Null Hypotheses of No Investor Response (ABRET&VTRAN) across 1,000
Non-Overlapping Samples Using 5 Percent, One-Tailed, Rejection Rule (b)

                        Two-Day                   Three-Day
                 Accumulation Period (c)  Accumulation Period (d)

                       Sample Size               Sample Size
Test Metric/
Procedure (e)   15      25       40       15       25       40

ABRET&VTRAN   36.6%   58.7%    81.4%    41.4%    64.2%    85.0%
VTRAN only    45.9 ** 64.2 **  84.8 **  48.7 **  68.6 **  87.1 *
ABRET only    24.6 ** 43.5 **  63.3 **  27.8 **  47.5 **  68.1 **

*, ** Rejection rate differs from ABRET&VSHRS (Panel A) or ABRET&VTRAN
(Panel B) rejection rate at the 0.05 and 0.01 levels, respectively,
using McNemar's [chi square] test.

(a) All samples are random without-replacement draws from the grand
sample of 49,274 annual and quarterly earnings announcements by
NYSE-listed firms occurring between 1/1/83 and 12/31/92.

(b) We base rejection on a cross-sectional t-test of the null
hypothesis that the response metric mean is less than or
equal to 0.

(c) We accumulate each metric over days -1 and 0 relative to
the earnings announcement date.

(d) We accumulate each metric over days -1 through +1 relative
to the earnings announcement date.

(e) Table 3 defines the response metrics.

We thank Linda Bamber, Ken Gaver, Ray Pfeiffer, and workshop participants at University of Georgia, University of Kansas, and Tulane University for their valuable comments on this paper. Expected earnings measures are computed from data from The Institutional Brokers Estimate System (I/B/E/S), which is a service of I/B/E/S International Inc. The I/B/E/S data has been provided as part of a broad academic program to encourage earnings expectations research.

Submitted April 2001

Accepted May 2002

(1) Greenwald (1976), Lindsay (1994), and Bamber et al. (2000) discuss the peer review process's inherent bias against studies that fail to reject the null hypothesis.

(2) Despite the evidence in Bamber et al. (2000) that most earnings announcements do not appear to stimulate investor responses, our assumption that mean investor response is not zero is reasonable--as long as the responses are not zero for at least one earnings announcement, the mean response is nonzero.

(3) Details of this analysis are available from the authors.

(4) This correspondence holds exactly when the event does not affect the estimated standard deviation ([sigma]), such as when tests estimate the standard deviation over a nonannouncement estimation period. In contrast, in tests that use the event period to estimate the standard deviation, the sample size necessary to yield an equivalent rejection rate level changes directly with the degree to which the event affects the estimated standard deviation. If the percentage impact on the estimated standard deviation is similar for return and volume metrics, then this error will not materially affect our inferences about differential performance across metrics.

(5) Since screens reduce power by reducing the sample size, a strict improvement in power is guaranteed only if the screen never erroneously eliminates a responsive observation.

(6) The log-transformation mitigates the skewness in the volume distribution (Ajinkya and Jain 1989). To avoid taking the log of zero we add the constant (0.000255) following Cready and Mynatt (1991) and Richardson et al. (1986).

(7) We do not use the standard deviation of the predicted values here or in the subsequent transactions-based metric because the calculation of such deviations adds a great deal of computational complexity. Moreover, because the number of estimation-period observations is large and mean market-level trading is unlikely to differ much from the mean level of trading in the estimation period, the standard deviation of the residuals is unlikely to differ much from the standard deviation of the prediction errors.

(8) We use before- and after-event estimation periods because trading increases with time (Cready 1988). Results, not reported here, are very similar when we use a pre-announcement estimation period in conjunction with a trend variable.

(9) The log-transformation mitigates the skewness in the transaction distribution (Cready and Ramanan 1995) and we add 1 to avoid taking the log of zero.

(10) Market-adjusted returns and abnormal returns relative to beta-matched portfolios led to similar inferences.

(11) Alternatively, we could recalculate ABRET using two-day returns, expectations, and standard errors. Preliminary analysis, however, indicated that such multiday accumulations did not attain as high rejection rates as the day-by-day accumulation. This may reflect that any price reversal effects over the two days are captured as separate positive influences in the daily absolute returns, but are netted in multiday accumulations.

(12) The Institute for the Study of Security Markets based at the University of Memphis produced the ISSM data. The data set provides transaction and quotation records from January 1, 1983, through December 31, 1992, for all NYSE-listed, AMEX-listed, and (beginning in 1990) OTC-listed securities, for trades occurring on the NYSE, AMEX, and regional exchanges as well as the NASDAQ OTC market.

(13) We also examine the most recent analyst forecasts and the mean of those forecasts updated within 45 days before the announcement. Untabulated analyses revealed that these alternative expectation measures lead to slightly lower rejection rates, on average.

(14) A more powerful alternative is to use normal-distribution-based tests similar to those described in Patell (1976) and Marais (1984, 1985). Such tests, however, overreject the null hypothesis when the underlying distributions are kurtotic, which is the case for absolute returns. Hence, we follow Cready and Mynatt (1991) and use cross-sectional rather than normal-distribution-based tests. Moreover, in analyses not reported here, application of normal-distribution tests across all metrics yielded similar relative rejection rate differences, except that the differences between the VSHRS rejection rates and the ABRET rejection rates narrowed, while the differences between the VSHRS rejection rates and the VTRAN rejection rates widened, relative to those reported here.

(15) We adjust the estimation periods for these tests accordingly. Specifically, we use data from days -110 to -11 to estimate expected absolute return and standard deviation of both absolute and unexpected returns. We use data from days -60 to -11 and 11 to 60 to estimate expected volume, expected number of transactions, and standard deviations of unexpected volume and unexpected transactions.

(16) Consistent with Cready and Ramanan (1995), VTRAN rejection rates also exceed counterpart VSHRS rejection rates at the 1 percent level in all instances except for the three-day I percent level analysis, where the difference is significant at only the 5 percent level.

(17) Rejection rates are smaller, but follow a similar cross-metric pattern when we use one-day (i.e., day -1, day 0) event periods.

(18) Brown and Warner (1985) find that tests based on time-series standard errors are more powerful than the cross-sectional tests we employ. In further analyses, we determined conditional return metric rejection rates using a time-series standard error test like the method Dyckman et al. (1984) used. As expected, rejection rates increase using this test. Moreover, for 15 announcement samples the conditional returns' rejection rates exceed counterpart ABRET-based rejection rates (significant at the 0.01 level). However, as sample size increases, the advantage of time-series standard errors dissipates rapidly. For samples of 25 announcements, rejection rates for the conditional returns time-series based tests are statistically indistinguishable from the ABRET-based rates reported in Table 4. For 40 announcement samples the conditional returns' time-series rejection rates are substantially smaller than the reported ABRET-based rates.

(19) The parametric test of choice here is a Hotelling [T.sup.2] test of the hypothesis that the vector of means is zero. In untabulated analyses we found that rejection rates based on the Hotelling test were consistently lower than those reported in Table 5. The Hotelling tests are less powerful than the Bonferroni approach because in the Hotelling test the alternative hypothesis is inherently two-tailed. Here, however, the alternative hypothesis is one-tailed (that the means exceed zero). The experiment-wide Bonferroni approach exploits these directional alternative hypotheses.

REFERENCES

Ajinkya, B., and P. Jain. 1989. The behavior of daily stock market trading volume. Journal of Accounting and Economics 11 (November): 331-359.

Bamber, L. S., and Y. S. Cheon. 1995. Differential price and volume reactions to accounting earnings announcements. The Accounting Review 70 (July): 417-441.

--, T. E. Christensen, and K. M. Gaver. 2000. Do we really know what we think we know? A case study of seminal research and its subsequent overgeneralization. Accounting, Organizations and Society 25 (February): 103-129.

Beaver, W. H. 1968. The information content of annual earnings announcements. Journal of Accounting Research (Supplement): 67-92.

Bernard, V. L., and T. L. Stober. 1989. The nature and amount of information in cash flows and accruals. The Accounting Review 64 (October): 624-652.

Boone, J. P., and K. K. Raman. 2002. Does the market fixate on reported earnings for R&D firms? Working paper, Mississippi State University, Mississippi State, MS.

Brown, S. J., and J. B. Warner. 1980. Measuring security price performance. Journal of Financial Economics 8 (September): 205-258.

--, and --. 1985. Using daily stock returns: The case of event studies. Journal of Financial Economics 14 (March): 3-31.

Carter, M., and B. Soo. 1999. The relevance of form 8-K reports. Journal of Accounting Research 37 (Spring): 119-132.

Cready, W. M. 1988. Information value and investor wealth: The case of earnings announcements. Journal of Accounting Research 26 (Spring): 1-27.

--, and P. Mynatt. 1991. The information content of annual reports: A price and trading response analysis. The Accounting Review 66 (April): 291-312.

--, and R. Ramanan. 1991. The power of tests employing log-transformed volume in detecting abnormal trading. Journal of Accounting and Economics 14 (June): 203-214.

--, and --. 1995. Detecting trading response using transactions-based research designs. Review of Quantitative Finance and Accounting 5 (June): 203-221.

Dodd, P., N. Dopuch, R. Holthausen, and R. Leftwich. 1984. Qualified audit opinions and stock prices: Information content, announcement dates, and concurrent disclosures. Journal of Accounting and Economics 6 (April): 3-38.

Dyckman, T., D. Philbrick, and J. Stephan. 1984. A comparison of event study methodologies using daily stock returns: A simulation approach. Journal of Accounting Research 22 (Supplement): 1-30.

Foster, G. 1977. Quarterly accounting data: Time-series properties and predictive-ability results. The Accounting Review 52 (January): 1-21.

Francis, J., K. Schipper, and L. Vincent. 2002. Expanded disclosures and the increased usefulness of earnings announcements. The Accounting Review 77 (July): 515-546.

Greenwald, A. 1976. Consequences of prejudice against the null hypothesis. Psychological Bulletin 82 (January): 1-20.

Kandel, E., and N. Pearson. 1995. Differential interpretation of public signals and trade in speculative markets. Journal of Political Economy 103 (August): 831-872.

Landsman, W., and E. Maydew. 2002. Beaver (1968) revisited: Has the information content of quarterly earnings announcements declined in the past three decades? Journal of Accounting Research 40 (June): 797-808.

Lang, M., and R. Lundholm. 2000. Voluntary disclosure and equity offerings: Reducing information asymmetry or hyping the stock? Contemporary Accounting Research 17 (Winter): 623-662.

Lindsay, R. 1994. Publication system biases associated with the statistical testing paradigm. Contemporary Accounting Research 11 (Summer): 33-57.

Linsmeier, T., D. Thornton, M. Venkatachalam, and M. Welker. 2002. The effect of mandated market risk disclosures on trading volume sensitivity to interest rate, exchange rate, and commodity price movements. The Accounting Review 77 (April): 343-377.

Marais, M. L. 1984. An application of the bootstrap method to the analysis of squared, standardized market model prediction errors. Journal of Accounting Research 22 (Supplement): 34-54.

--. 1985. Some computer applications of computer intensive statistical methods to empirical research in accounting. Ph.D. dissertation, Stanford University.

May, R. G. 1971. The influence of quarterly earnings announcements on investor decisions as reflected in common stock price changes. Journal of Accounting Research 11 (Supplement): 119-163.

Patell, J. 1976. Corporate forecasts of earnings per share and stock price behavior: Empirical tests. Journal of Accounting Research 14 (Autumn): 246-276.

Pincus, M. 1983. Information characteristics of earnings announcements and stock market behavior. Journal of Accounting Research 23 (Spring): 166-183.

Qi, D., W. Wu, and I. Haw. 2000. The incremental information content of SEC 10-K reports filed under the EDGAR system. Journal of Accounting, Auditing, and Finance 15 (Winter): 25-46.

Richardson, G., S. Sefcik, and R. Thompson. 1986. A test of dividend irrelevance using volume reactions to a change in dividend policy. Journal of Financial Economics 15 (December): 313-333.

Rohrbach, K., and R. Chandra. 1989. The power of Beaver's U against a variance increase in market model residuals. Journal of Accounting Research 27 (Spring): 145-155.

Subramaniam, C. 1995. Detecting information content of corporate announcements using variance increases: A methodological study. Journal of Accounting, Auditing, and Finance (Fall): 415-430.

Vincent, L. 1999. The information content of funds from operations (FFO) for real estate investment trusts (REITs). Journal of Accounting and Economics 26 (January): 69-104.

William M. Cready
Louisiana State University

David N. Hurtt
Western Michigan University