This study aims to remind researchers that measurement errors and inappropriate inferences may result from improperly combining and adjusting certain Center for Research in Security Prices (CRSP) measures.
In addition to real-world working examples, the study uses earnings announcements data to examine the effects of improperly combining and adjusting CRSP measures.
This study assists researchers with the following two considerations when using CRSP data: stand-alone share prices adjusted with CRSP adjustment factors are inaccurate in the presence of property dividend, spin-off and rights offering events; and ignoring covertly missing stock returns may create misleading test results. The primary objectives of the study are to help researchers increase the integrity of their studies and the probability of publication.
Inadequate consideration for the two issues discussed in the paper may change the researcher’s statistical inferences.
Archival researchers who overtly address and discuss the existence of these issues achieve two important and related benefits. First, the researcher increases his or her credibility with editors and reviewers, which enhances the probability of a published study. Second, the researcher increases his or her perceived technical competency, which potentially affects promotion and tenure decisions, editorial membership decisions, co-authorship opportunities and other professional effects. Doctoral students will find this study to be particularly useful.
CitationDownload as .RIS
Emerald Publishing Limited
Copyright © 2018, Emerald Publishing Limited
Researchers frequently encounter tasks that require modifying variables from the Center for Research in Security Prices (CRSP) database. This study reminds researchers that measurement errors and inappropriate inferences may result from improperly adjusting the prices from CRSP and ignoring the implications of missing prices and stock returns. Moreover, the study aims to assist researchers in avoiding unnecessary reviewer and editorial skepticism, which reduces the probability of a publication.
The literature contains several studies which highlight biases in CRSP data (Guenther and Rosman, 1994; Shumway, 1997; Canina et al., 1998; Shumway and Warther, 1999; and Fisher et al., 2010). Other studies address errors in CRSP data (Rosenberg and Houglet, 1974; Bennin, 1980; Courtenay and Keller, 1994; and Elton et al., 2001). In contrast to past studies which highlight problems with CRSP data per se, the focal point of the current study is to address errors that can arise when archival researchers modify and adjust CRSP data to construct certain variables that are commonly used in the literature.
This study assists researchers with the following two considerations when using CRSP data:
Stand-alone share prices adjusted with CRSP adjustment factors are inaccurate in the presence of property dividend, spin-off and rights offering events.
Ignoring covertly missing stock returns may create misleading test results.
Archival researchers who explicitly address and discuss the existence of these issues increase their credibility with editors, reviewers and colleagues. Doctoral students will find this study to be particularly useful. A discussion of these two issues appears in the following sections, along with complementary statistics.
Center for Research in Security Prices adjustment factors
CRSP adjustment factors for share prices (CFACPR) enable the researchers to restate historical share prices for artificial reductions created by events such as stock splits and stock dividends. Although these events change share prices, they lack economic substance and create synthetic structural breaks, which undermine the comparability of share prices before and after the events. Unexpected cash dividends and other property dividends are examples of events which also change share prices. However, these events change the economic position of the firm; thus, the corresponding changes in share price are substantive and require no adjustment.
An important point is that using CRSP adjustment factors may create unintended consequences when the researcher’s sample contains events with and without economic substance. For example, assume that a researcher’s sample includes observations for property dividend, spin-off or rights offering events, along with observations for stock split or stock dividend events. Further assume that the researcher uses the CRSP cumulative adjustment factor for price (CFACPR) to accommodate the artificial decrease in share prices associated with stock split and stock dividend events. Herein lies the potential danger with the use of CRSP adjustment factors: the researcher adjusts share prices for all observations in the sample and unknowingly adjusts the share prices for property dividend, spin-off and rights offering events. These adjustments are problematic as CRSP adjustment factors for share price include the effects of property dividend, spin-off and rights offering events. Assuming that the researcher is using share price in a stand-alone manner (e.g. as a dependent variable for a regression model), then the adjustment of the historical share prices for these events creates erroneous sample observations. The key idea is that researchers must only use the CRSP adjustment factors to accommodate events which lack economic substance, such as stock splits and stock dividends.
The problem associated with CRSP adjustment factors potentially affects several studies, yet the narratives of most studies preclude the clear identification of the sources of the adjustment factors. For example, Kothari and Zimmerman (1995) indicate the use of earnings and returns data from Compustat and CRSP. The authors further state that “Earnings and price data are adjusted for stock splits, stock dividends and stock issues.” Presumably, the authors use the adjustment factor associated with the database from which they retrieve share price. However, the source of the share price, as well as the adjustment factor, is indeterminable. Similarly, Lipe (1986) uses data from Compustat and CRSP and indicates the use of “price at the end of March of year t, adjusted for stock splits and stock dividends.” Again, clearly identifying the source for the adjustment factor is not possible. One explanation for omitting the sources of the adjustment factors is that researchers assume that the same values appear in both databases (i.e. CRSP and Compustat). However, the clear identification of the source for the adjustment factors is crucial, as Compustat adjustment factors exclude the effects of property dividend, spin-off and rights offering events. It is worth noting that the primary reason that CRSP adjustment factors include the effects of property dividend, spin-off and rights offering events is to compute and report an accurate value for the stock return variable. The failure to adjust share prices for the effects of property dividend, spin-off and rights offering events will typically generate a large negative stock return, which is misleading.
Examples of studies clearly identifying CRSP as the source for the adjustment factors include (but are not limited to) Easton and Harris (1991), Easton et al. (1992), Brown and Pfeiffer (2007), Freeman, Koch and Li (2011) and Haggard et al. (2015). The use of CRSP adjustment factors should not automatically diminish the inferences from these studies, as the sample selection and exclusion processes are unique to each study and potentially exclude many of the problematic observations. More importantly, the overall objective of the current study is prospective in nature, where the aim is to inform and assist the research community in planning and avoiding any unintended consequences associated with the use of CRSP adjustment factors.
One obvious solution to the CRSP adjustment factor hazard is for the researcher to use share price and adjustment factor variables from Compustat when possible and practical. This limits any share price adjustments to stock split and stock dividend events. Alternatively, the researcher may use the CRSP events file to identify property dividend, spin-off and rights offering events in the sample and then exclude these observations from the sample or modify the adjustment factor for the impact of these events (i.e. modify using FACPR). A sensitivity test with and without these observations is also appropriate.
The relation between market value and book value is supported by theory. Thus, a regression of share price on book value per share provides a context to examine whether the cumulative adjustment factor includes or excludes substantive events will significantly affect the observed relation. Table I provides a firm size analysis of the proper and improper use of CRSP adjustment factors using a firm-specific (i.e. mnemonic identifier PERMCO) regression of share price on book value per share. Panels A and B provide mean and median regression coefficient analyses, respectively, while Panels C and D provide mean and median R2 analyses, respectively. The data for the regression analyses represent the intersection of the quarterly Compustat data set from 1976 to 2015 and the monthly CRSP data set from 1976 to 2015. Firm size quintiles use the CRSP capitalization files with an adjustment to collapse the decile rankings into quintile rankings. A total of 8,286 firms with at least 20 quarterly observations meet the following qualifications for sample membership: share price (PRC from CRSP) must be greater than or equal to $1; shares outstanding (SHROUT from CRSP) must be greater than 0; cumulative adjustment factors for price and shares (CFACPR and CFACSHR from CRSP) must be greater than 0; share code (SHRCD from CRSP) must be 10 or 11, representing ordinary common shares; foreign incorporation code (FIC from Compustat) must be “USA,” indicating US incorporation; and common equity (CEQQ from Compustat) must be greater than 0.
In general, Panels A and B from Table I indicate that the differences between the coefficients for improper or unconditional use of the factors and the coefficients for proper or conditional use are statistically significant. Paired-difference t-tests and signed-rank tests indicate that nine of the ten test statistics (i.e. the parametric t-statistic and the non-parametric Wilcoxon statistic) are significant at the 2 per cent level or less. The results for the R2 analyses appear in Panels C and D and are very similar to the regression coefficient analyses. Eight of the ten mean differences are significantly different from 0 at the 4 per cent level or less. All panels from Table I are consistent with the notion that book value per share is more relevant for the pricing of large firm shares than for small firm shares. For example, the mean proper use coefficient for large firms in Panel A is 2.077, which is noticeably larger than the mean proper use coefficients for small firms, which is 1.136. A similar pattern exists for the R2 values for small and large firms in Panel C. The generalization of the results to other samples is unknown. However, the reported differences are meaningful and call for researchers’ attention to a potential hazard when using CRSP adjustment factors.
Missing event-period returns
A subtle, yet important, issue for researchers is the management of missing event-period returns, which appear overtly and covertly in the CRSP database. Overtly missing returns appear with traditional missing value indicators such as the period (.), as well as the period followed by the letters B or E (i.e. .B and .E). Covertly missing returns appear in the CRSP database as seemingly valid returns (i.e. with actual numbers instead of missing value indicators). However, these returns are artificial, as CRSP reports a stock return based on the change in the bid-ask price midpoint when the beginning or ending (or both) prices are missing. Missing prices typically indicate an absence of trading, which is the hallmark for thinly traded shares. CRSP uses zero or negative price values (i.e. PRC = 0 or PRC < 0) to identify the covertly missing returns. If one or both of the closing prices for a return period is/are missing and the closing values for bid-ask prices are missing, then CRSP overtly reports the missing return (i.e. the period). The key idea is that covertly, as well as overtly, missing returns are likely to be present in the researchers’ sample.
Researchers generally choose one of three strategies for managing missing returns:
Use the returns as reported by CRSP, that is, omit the overtly missing returns and include the covertly missing returns.
Use trade-to-trade returns, where the change in non-missing share prices from actual trading determines the return and excludes both the overt and covert missing returns.
Use lumped returns, where zeros replace any missing return (i.e. both overt and covert).
The least preferable strategy is the use of CRSP reported returns due to the absence of actual trading. Although changes in the bid-ask midpoint reflect new information, the absence of trading suggests that the costs of trading on the new information are greater than or equal to the benefits of trading on the new information. A common pitfall for trade-to-trade and lumped returns is that they both generate multi-period returns in the presence of missing prices. Moreover, trade-to-trade and lumped returns may generate the same return for certain return periods.
For example, consider a price series with four daily observations which begins with $10, followed by two missing prices, and terminates with $11. Both trade-to-trade and lumped returns will generate a three-day return of 10 per cent, although each method uses a slightly different approach. Trade-to-trade returns simply discard or ignore missing returns, while lumped returns substitute zeros for missing returns. Thus, both methods report a multi-period return for the third day. While the two methods generate the same three-day returns, both possible two-day returns (i.e. ending price for Days 1-3 and Days 2-4) will differ between the two methods. The first two-day return period contains a beginning price of $10, and the ending price is missing, while the beginning price is missing for the second two-day return period and the ending price is $11. Trade-to-trade returns require non-missing prices at the end of the event period, and otherwise generate missing event-period returns. In contrast, lumped returns substitute zeros for the missing returns for the first two-day return period. Thus, the trade-to-trade return is only available for the second two-day return period, while the lumped return is available for both two-day return periods. This example highlights an important advantage of the lumped return method, which is the preservation of sample observations by replacing missing returns with zeros.
Most, if not all studies, in the accounting literature ignore the potential hazards associated with missing event-period returns. The authors reviewed 46 narrow window event studies (i.e. 10 trading days or less) published in five nationally recognized academic accounting journals from 2005 through 2014. With the exception of the method of accumulation (i.e. cumulative abnormal returns or buy and hold abnormal returns), none of the 46 studies explicitly discuss the construction of the event-period returns. In contrast, a crude search of the Business Source Complete database finds that event studies in the finance literature commonly discuss the construction of event-period returns (Maynes and Rumsey, 1993; Lesmond et al., 1999; Bartholdy et al., 2007; and Campbell et al., 2010). The absence of any discussion concerning the construction of event-period returns leaves editors, reviewers and readers to make implicit assumptions about the competency of the researcher and the integrity of his or her data.
A natural question arises concerning how the signs (±), frequency (N), magnitude (mean) and volatility (standard deviation) vary between the three missing return methods. Table II examines this variation using ten portfolios partitioned on trading volume to analyze the three-day raw returns surrounding earnings announcements from the years 1976 through 2015. Missing prices and returns are more likely for firms with low trading volume than firms with high trading volume. Thus, the mean monthly trading volume for the 12 months prior to each earnings announcement forms the basis for the ten trading volume portfolios. Panel A from Table II reports the descriptive statistics for all observations (i.e. positive, negative and zero signs) associated with each missing returns method. As expected, the frequency of trade-to-trade returns is noticeably lower than the frequencies for lumped and CRSP returns for all volume portfolios. The lower frequency of trade-to-trade returns is most notable for the low-volume portfolios, which reflects thinly traded shares with numerous missing returns. Combining positively and negatively signed returns naturally leads to an expected neutral mean result. The results in Panel A are consistent with this idea, as the mean returns for each of the three methods is near zero. Finally, the volatility of the three methods, measured using the standard deviation, is the lowest for the CRSP method and the highest for the trade-to-trade method. This is consistent with CRSP substituting the bid-ask average for missing prices, which dampens the volatility of the returns. The high volatility associated with the trade-to-trade and lumped returns is consistent with the multi-period returns created using both methods. Finally, substituting zeros for missing returns reduces the volatility of the lumped returns relative to the trade-to-trade returns. An unreported analysis comparing the variances of the three missing return methods (i.e. an F-test) indicates that the variances for all ten trade-to-trade portfolios are significantly larger than the corresponding variances for the lumped and CRSP return portfolios. The smaller variances for the lumped and CRSP returns are important as they inflate the statistical significance of the return.
Panel B from Table II reports the positively signed returns, where the results are very similar to those previously reported for all observations, with two exceptions. First, the magnitudes of the returns are noticeably larger than those reported for all observations, which is consistent with the positive signs of the returns. Second, the mean trade-to-trade returns range from 5.7 to 6.9 per cent, while the means for the lumped returns are slightly lower and range from 5.2 to 6.3 per cent. The means of the CRSP returns are noticeably lower, as the means range from 3.6 to 4.5 per cent. It is worth noting that the frequency of lumped returns is notably smaller, instead of larger, than the frequency of trade-to-trade returns. This result reflects the presence of numerous zero-value three-day lumped returns in the previous panel, which contains all observations (i.e. positive, negative and zeros). As in the previous panel, the standard deviations for the trade-to-trade and lumped return portfolios are significantly larger than the standard deviations for the CRSP return portfolios.
The results reported for the negatively signed returns in Panel C are slightly smaller in magnitude, but otherwise very consistent with the results reported for the positively-signed returns. The smaller magnitudes for the negatively signed returns reflect the truncated range of realizations, where the (mathematically) largest possibility is slightly less than zero, while the (mathematically) smallest realization is negative (−1 or −100 per cent). As in the previous panels, the standard deviations for the trade-to-trade and lumped return portfolios are significantly larger than those for the standard deviations for the CRSP return portfolios.
Table III examines the mean paired differences from the previous table between the three possible pairs of missing returns using ten trading volume portfolios. Panel A reports the frequency and mean paired differences for all observations regardless of the sign for the return and for the difference. The number of observations is similar across the three methods due to the paired nature of the analysis. As expected, pooling the positive and negative differences results in near zero mean differences for most of the ten portfolios.
The results in Panel B from Table III reflect all returns, both positive and negative:
where trade-to-trade returns are greater than lumped returns;
where trade-to-trade returns are greater than CRSP returns; and
where lumped returns are greater than CRSP returns.
In general, the results are very similar for all three paired-comparisons. All ten volume portfolios contain statistically significant mean paired differences at less than 1 per cent level, and the magnitude of the mean paired-differences is fairly consistent across the three paired comparisons. The number of observations for both comparisons involving trade-to-trade returns is noticeably smaller, which reflects the requirement of non-missing ending prices for a return observation.
The results in Panel C from Table III reflect all instances:
where trade-to-trade returns are less than lumped returns;
where trade-to-trade returns are less than CRSP returns; and
where lumped returns are less than CRSP returns.
The frequencies and mean paired differences for all volume portfolios are very similar to those reported in Panel B. Again, all ten volume portfolios contain statistically significant paired differences at less than 1 per cent level. Also, the number of observations for the two comparisons which include trade-to-trade returns is lower due to the requirement of non-missing ending prices for the return calculation. Finally, the magnitudes of the mean paired differences are similar across the three methods.
Overall, the results in Table III indicate that the magnitudes and variances of event-period returns are significantly different for the three common methods for managing missing returns. Moreover, these results have implications for many streams of accounting research event studies including earnings response coefficients, post-earnings announcement drift, responses to regulatory filings, changes in analysts’ earnings estimates, etc.
The CRSP database enjoys widespread use in the academic community. Despite the precision of CRSP data, researchers may inadvertently generate imprecise measurements when modifying and adjusting CRSP variables. This study calls the researchers’ attention to two potential hazards concerning the use of CRSP data. First, across-the-board adjustments to share prices using CRSP adjustment factors create inaccurate values for any price series containing property dividend, spin-off or rights offering events. The key is that CRSP adjustment factors include the effects of property dividend, spin-off and rights offering events, as well as stock split and stock dividend events. Second, CRSP uses bid-ask midpoints to replace missing prices, and including the corresponding returns in the researcher’s sample understates the true magnitude and variance of the returns.
Addressing the issues in this study increases trust between editorial staff and researchers, and a candid discussion of the issues advances the researcher’s credibility. A good practice for researchers is to disclose the sensitivity of their results to the use of any alternative computational methods.
Improper versus proper use of CRSP adjustment factors
|Comparative analyses||Firm size (MVE) quintile|
|(Smallest) 1||2||3||4||(Largest) 5|
|Panel A – mean coefficient analysis (α1)|
|Number of firm-specific regressions||810||1,354||1,617||1,908||2,597|
|Mean improper use coefficient||1.196||1.430||1.595||1.666||2.029|
|Mean proper use coefficient||1.136||1.411||1.579||1.675||2.077|
|Paired difference t-statistic||4.411||2.211||1.345||−2.371||−8.216|
|Panel B – median coefficient analysis (α1)|
|Number of firm-specific regressions||810||1,354||1,617||1,908||2,597|
|Median improper use coefficient||1.002||1.169||1.307||1.368||1.660|
|Median proper use coefficient||0.939||1.140||1.304||1.389||1.705|
|Paired difference W-statistic||7,489||13,838||10,923||5,535||−23,067|
|Panel C – mean R2 analysis|
|Number of firm-specific regressions||810||1,354||1,617||1,908||2,597|
|Mean improper use coefficient||31.4%||34.9%||37.1%||39.5%||49.3%|
|Mean proper use coefficient||30.6%||34.3%||36.8%||39.4%||49.8%|
|Paired difference t-statistic||2.979||4.080||2.099||0.196||−4.380|
|Panel D – median R2 analysis|
|Number of firm-specific regressions||810||1,354||1,617||1,908||2,597|
|Mean improper use coefficient||26.4%||32.1%||34.7%||37.4%||53.0%|
|Mean proper use coefficient||25.3%||31.2%||34.5%||37.1%||53.9%|
|Paired difference W-statistic||5,810||12,895||11,424||5,577||−3,443|
Merged CRSP and Compustat data (1994-2015); firm-specific OLS regression model: ; The table reports the firm-specific coefficient and R2 value for each quintile/portfolio of firm size, where the CRSP capitalization deciles (combined to create quintiles) define memberships for the portfolios; using the CRSP cumulative adjustment factor for price (CFACPR) in an unconditional and universal manner (i.e. PRC/CFACPR for all observations) represents the improper use of the factor. In contrast, the proper use of the factor requires adjusting price for observations which occur prior to the date of stock split and stock dividend events (i.e. PRC/CFACPR only for stock splits and stock dividends); share price is the absolute value of the closing price from CRSP (mnemonic PRC); book value is common/ordinary equity from quarterly Compustat (mnemonic CEQQ) with positive values from US firms (i.e. mnemonic FIC = USA); other sample restrictions: share price (PRC) must be greater than or equal to $1; shares outstanding (SHROUT) must be greater than zero; CRSP adjustment factors for price (CFACPR) and shares outstanding (CFACSHR) must be greater than zero; and share code from CRSP must be 10 or 11, representing ordinary common shares; the shares outstanding in the denominator of the independent variable (SHROUT) require division by 1,000 to create a measurement basis consistent with the numerator (i.e. millions). An additional adjustment to the SHROUT variable occurs with multiplication by the cumulative adjustment factor for shares (CFACSHR); the W-statistic is the Wilcoxon test statistic from the non-parametric signed-rank test; the sample excludes extreme observations which generate adjusted values for the dependent and independent variables, which are less than the 0.5 percentile and greater than the 99.5 percentile; each firm-specific regression uses a minimum of 20 observations (i.e. 20 quarters), and the total number of firms is 8,286
Descriptive statistics for three-day raw returns around earnings announcement dates (1976-2015) (719,347 events, 20,619 firms)
|Volume portfolio||Trade-to-trade returns||Lumped returns||CRSP returns|
|Panel A – all observations (positive, zero and negative)|
|Panel B – positive event-period return observations|
|Panel C – negative event-period return observations|
The events are earnings announcement dates from Compustat (mnemonic RDQ) during the period from January 1976 through November 2015, and the event window begins with the day prior to the earnings announcement date (−1) and ends with the day following the earnings announcement date (+1); the event-period returns are three-day raw (i.e., unadjusted) returns from the CRSP database; trade-to-trade returns represent the changes in share prices from actual trading and exclude overtly missing returns (i.e. missing values), as well as covertly missing returns (i.e. returns using changes in prices from the bid-ask average prices); lumped returns are essentially trade-to-trade returns with zeros substituted for any missing return (overt or covert); and CRSP returns are the returns as reported by CRSP (i.e. mnemonic RET)
Paired difference t-tests for three-day event-period returns (1976-2015)
|Trade-to-trade returns vs lumped returns||Trade-to-trade returns vs CRSP returns||Lumped returns vs CRSP returns|
|Volume portfolio||N||Mean difference||N||Mean difference||N||Mean difference|
|Panel A – all observations|
|Panel B – positive difference observations|
|Panel C – negative difference observations|
indicate statistical significance at the 1%, 5% and 10% levels, respectively
The events are earnings announcement dates from Compustat (mnemonic RDQ) during the January 1976 through November 2015, and the event window begins with the day prior to the earnings announcement date and ends with the day following the earnings announcement date (i.e. −1 and +1); The event-period returns are three-day raw (i.e. unadjusted) returns from the CRSP database; the t-tests assess the mean differences in paired observations from two return measurements, where a firm identifier (PERMCO) and an event date define a pair, and the return is one of three types: trade-to-trade, lumped or CRSP. Thus, the three possible pairings are: trade-to-trade returns vs lumped returns; trade-to-trade returns vs CRSP returns; and lumped returns vs CRSP returns; the standard errors include adjustments for clusters of observations in time and across firms (i.e. repeated measures of time and firms) using the method proposed by Gow et al. (2010); positive differences refer to the following three outcomes: trade-to-trade returns > lumped returns; trade-to-trade returns > CRSP returns; and lumped returns > CRSP returns. In contrast, negative returns refer to the following three outcomes: trade-to-trade returns < lumped returns; trade-to-trade returns < CRSP returns; and lumped returns < CRSP returns. Note that the sum of the number of observations for positive and negative differences reported in Panels B and C is less than the total number of observations reported in Panel A due to the equivalence of the three measures for several observations (i.e. some observations generate zero differences and are therefore neither positive nor negative)
References to stock splits are for transactions that reduce share prices. Although potentially relevant, the discussion excludes events which increase share prices, such as reverse stock splits. References to adjustment factors are for share prices instead of shares outstanding.
This list is not exhaustive and only serves as an example.
Although shares which do not trade for a given period generate no true return for the period, the bid and ask prices may change despite the absence of trading.
Hereafter, all references to missing returns include both overt and covert missing returns.
The focal point of the current study is upon raw return rather than abnormal returns. It is worth noting that event studies frequently use abnormal returns, and abnormal trade-to-trade returns require the use of market index returns (see Campbell et al., 2010).
The five journals and their respective number of event studies are The Accounting Review (16); Contemporary Accounting Research (11); Journal of Accounting and Economics (9); Journal of Accounting Research (7); and Review of Accounting Studies (3).
Bartholdy, J., Olson, D. and Peare, P. (2007), “Conducting event studies on a small stock exchange”, European Journal of Finance, Vol. 13 No. 3, pp. 227-252.
Bennin, R. (1980), “Error rates in CRSP and compustat: a second look”, Journal of Finance, Vol. 35 No. 5, pp. 1267-1271.
Brown, R. and Pfeiffer, R. (2007), “Causes and consequences of the relation between split-adjusted share prices and subsequent stock returns”, Journal of Business Finance & Accounting, Vol. 34 Nos 1/2, pp. 292-312.
Campbell, C., Cowan, A. and Salotti, V. (2010), “Multi-country event-study methods”, Journal of Banking and Finance, Vol. 34 No. 12, pp. 3078-3090.
Canina, L., Michaely, R., Thaler, R. and Womack, K. (1998), “Caveat compounder: a warning about using the daily CRSP equal-weighted index to compute long-run excess returns”, Journal of Finance, Vol. 53 No. 1, pp. 403-416.
Courtenay, S. and Keller, S. (1994), “Errors in databases revisited: an examination of the CRSP shares outstanding data”, The Accounting Review, Vol. 69 No. 1, pp. 285-291.
Easton, P. and Harris, T. (1991), “Earnings as an explanatory variable for returns”, Journal of Accounting Research, Vol. 29 No. 1, pp. 19-36.
Easton, P., Harris, T. and Ohlson, J. (1992), “Aggregate accounting earnings can explain most of security returns: the case of long run return intervals”, Journal of Accounting and Economics, Vol. 15 Nos 2/3, pp. 119-142.
Elton, E., Gruber, M. and Blake, C. (2001), “A first look at the accuracy of the CRSP mutual fund database and a comparison of the CRSP and morningstar mutual fund databases”, Journal of Finance, Vol. 56 No. 6, pp. 2415-2430.
Fisher, L., Weaver, D. and Webb, G. (2010), “Removing biases in computed returns”, Review of Quantitative Finance and Accounting, Vol. 35 No. 2, pp. 137-161.
Freeman, R., Koch, A. and Li, H. (2011), “Can historical returns-earnings relations predict price responses to earnings news?”, Review of Quantitative Finance and Accounting, Vol. 37 No. 1, pp. 35-62.
Gow, I., Ormazabal, G. and Taylor, D. (2010), “Correcting for cross-sectional and time-series dependence in accounting research”, The Accounting Review, Vol. 85 No. 2, pp. 483-512.
Guenther, D. and Rosman, A. (1994), “Differences between compustat and CRSP SIC codes and related effects on research”, Journal of Accounting and Economics, Vol. 18 No. 1, pp. 115-128.
Haggard, K., Walkup, B. and Xi, Y. (2015), “Short-term performance of US-bound Chinese IPOs”, Financial Review, Vol. 50 No. 1, pp. 121-141.
Kothari, S. and Zimmerman, J. (1995), “Price and return models”, Journal of Accounting and Economics, Vol. 20 No. 2, pp. 155-192.
Lesmond, D., Ogden, J. and Trzcinka, C. (1999), “A new estimate of transaction costs”, Review of Financial Studies, Vol. 12 No. 5, pp. 1113-1141.
Lipe, R. (1986), “The information contained in the components of earnings”, Journal of Accounting Research, Vol. 24, pp. 37-64.
Maynes, E. and Rumsey, J. (1993), “Conducting event studies with thinly traded stocks”, Journal of Banking and Finance, Vol. 17 No. 1, pp. 145-157.
Rosenberg, B. and Houglet, M. (1974), “Error rates in CRSP and Compustat databases and their implications”, Journal of Finance, Vol. 29 No. 4, pp. 1303-1310.
Shumway, T. (1997), “The delisting bias in CRSP data”, Journal of Finance, Vol. 52 No. 1, pp. 327-340.
Shumway, T. and Warther, V. (1999), “The delisting bias in CRSP’s NASDAQ data its implication for the size effect”, Journal of Finance, Vol. 54 No. 6, pp. 2361-2379.
The authors are grateful for the helpful comments of Jim Angel, Tom Dyckman, Oscar Varela and Faith Xie, the participants at the 2015 Midwest American Accounting Association Conference, and two anonymous reviewers. All errors and omissions are solely the responsibility of the authors.