Do your Findings Depend on your Data(base)? A Comparative Analysis and Replication Study Using the Three Most Widely Used Databases in International Business Research

Journal of International Management 22 (2016) 186–206

Contents lists available at ScienceDirect

Journal of International Management

Do your Findings Depend on your Data(base)? A Comparative Analysis and Replication Study Using the Three Most Widely Used Databases in International Business Research Jean B. McGuire a,⁎, Barclay E. James b, Andrew Papadopoulos c a b c

Rucks Department of Management, Louisiana State University, 2713 Business Education Complex, Baton Rouge, LA 70803, USA Universidad San Francisco de Quito USFQ, USFQ Business School, Campus Cumbayá, Quito, Ecuador Département de Stratégie, Responsabilité Sociale et Environnementale, École des Sciences de la Gestion (ESG UQAM), Université du Québec à Montréal, Canada

a r t i c l e

i n f o

Article history: Received 6 May 2015 Received in revised form 4 March 2016 Accepted 5 March 2016 Available online xxxx Keywords: Database International Replication Innovation Debt Governance

a b s t r a c t Theoretical and empirical advances in international business (IB) depend heavily on archival data obtained from databases. As there has been scant research on the implications of database choice in IB research, we analyzed data from the three most-widely used databases in the field: Compustat Global, Osiris and Worldscope. We examined data coverage across several geographic regions and countries, analyzed descriptive statistics and regression results, and replicated a study by O'Brien (2003), who theorized and found a negative firm performance effect of innovation strategy matched with high levels of debt governance. Based on our empirical results, we found the presence of what we call a “database effect” – researchers likely would come to a different conclusion based on the database used – particularly for developing country results. The database effect is confirmed using multiple estimation techniques and also in our replication study. Our replication study also found evidence that O'Brien's (2003) findings do indeed apply to countries outside of the United States. However, in China, we found statistically significant contrasting results in two databases, requiring a need for more theoretical reflection of firm governance of innovation in different institutional contexts. © 2016 Elsevier Inc. All rights reserved.

1. Introduction Over 50 years ago, Merton (1957) emphasized how empirical research informs and initiates theory. More recently, Ketchen et al. (2013) highlighted how differences in the way variables are defined and measured may have both theoretical and empirical implications. For management research, particularly in international business (IB), the use of different databases for firm-level data implies inherent differences in data definitions, data sources, firm and geographic coverage. Coupled with differences in national reporting requirements and accounting practices, it stands to reason that IB-related research heightens the potential adverse effects of choosing one database over another, even though the choice very often is one of database availability and cost.

⁎ Corresponding author. E-mail addresses: [email protected] (J.B. McGuire), [email protected] (B.E. James), [email protected] (A. Papadopoulos).

http://dx.doi.org/10.1016/j.intman.2016.03.001 1075-4253/© 2016 Elsevier Inc. All rights reserved.

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

187

The three most-widely used databases for IB research are Compustat Global (also called Global Vantage, distributed by Standard & Poor's), Osiris (distributed by Bureau van Dijk), and Worldscope (distributed by Thomson Reuters)1. To date, there has been little systematic attention given to the extent to which the data and results obtained using different databases are comparable generally and virtually none that we are aware of devoted to these three databases. In this study, we give substantial systematic attention to these three databases and describe the implications for IB research. There are a number of key reasons why the potential adverse effects of database choice are heightened for IB research. First, non-uniformity in reporting and inconsistency in data standardization procedures to ensure comparability across countries may create discrepancies when trying to replicate results using different databases. Even an assumption that data provided by different data sources are comparable within Western Europe or North America, with standardized reporting practices and more homogenous sample frames (e.g., FTSE- or NYSE-traded firms, S&P 500 firms), has been questioned. For example, in the United States (US), researchers have found differences in ownership data (Anderson and Lee, 1997), mutual fund data (Elton et al., 2001), and analysts' forecast data (Abarbanell and Lehavy, 2000) reported by different sources. Similarly, Yang et al. (2003) found differences in the data reported by Compustat North America and Valueline (which only covers US firms). Second, each of the three databases we examined uses different primary sources to compile its respective data (annual reports versus regulatory filing versus third-party suppliers). Compustat Global uses company-specific documentation and then standardizes the data to account for national differences in reporting practices (Standard and Poor's, 2011). Osiris makes use of a variety of data sources and providers, including other compilations of firm-level data that are unique to a particular country or industry (e.g., the Korea Information Service, Fitch Ratings, and Dun & Bradstreet TSR for Japan)—information is collected and processed by each provider. Osiris also makes use of World Vest Base (WVB) to obtain and analyze firm documents (Bureau van Dijk, 2011). Worldscope uses company-specific documentation and trained analysts to transfer the data into consistent formats for comparable cross-country analyses (Thomson Reuters, 2011). Third, firm and country samples available from Compustat Global, Osiris and Worldscope also differ despite each database claiming to be relatively comprehensive in terms of global firm “coverage”. Potential coverage is not identical and, as we discuss below, the actual coverage is far from identical. Country-specific differences in firm and country coverage may lead to biased conclusions, and inconsistent yearly samples may pose challenges for researchers using panel data. Alarmingly, one of the databases we examined entirely drops firms from inclusion in their database if they are subsequently delisted (from a stock exchange) or otherwise not subsequently covered. These differences in variable and firm data availability affect whether firms are included in the research sample, which introduces potential sample bias. These inconsistencies may be particularly consequential for the study of certain global geographic regions, which may receive less widespread and consistent coverage in secondary databases. To address these concerns, we analyzed the comparability of data obtained from Compustat Global, Osiris and Worldscope. We documented data coverage by specific global regions and countries, compared descriptive statistics for common variables, and estimated regressions using reasonable variations in the type of estimation techniques used and in the geographic regions chosen for these regressions. We undertook this analysis using samples of firms that were included in all three of the databases (what we call “matched samples”) and for samples of firms that would have been drawn if each of these databases were used on its own (what we call “unmatched samples”). This permitted us to compare the consistency of reporting across databases in the matched sample and to evaluate if database choice would yield different results through unmatched samples. Finally, we replicated a study by O'Brien (2003) on the firm performance implications of innovation strategy and debt governance. Our analyses allowed us to assess if and under which conditions a “database effect” occurred, which means that researchers likely would come to a different conclusion based on the database used. Highlights of our findings are the following. First, each database provides a substantially different quantity of firms across different geographic regions. Second, differences in methodology related to the type of estimation technique used for regression analyses can either heighten or temper the differences in how common firm-specific variables affect firm performance, measured by return on assets and by a firm's market-to-book ratio. Third, database choice may not only affect empirical results for samples of firms from particular geographic regions such as Latin America and East Asia, but also may determine whether a study is even feasible in these locations using a particular database because of the paucity of data points available. Finally, our replication analysis suggests that database choice can have significant implications for empirical results. Developed-country results showed consistency across databases while developing-country results did not show consistency, and thus we observed a database effect for developing countries. We also found in our replication study statistically significant results across two databases that directly contrast results hypothesized and found by O'Brien (2003) in the US and found by us (in this study) in other countries. Overall, our results suggest that (the growing) research on firms from developing countries may be particularly vulnerable and or prone to spurious empirical results due to database choice. Below, we discuss these inconsistencies in more detail along with their implications for theoretical and empirical IB research.

1 These databases are by far the dominant sources of secondary data. As we discuss below we identified 879 articles making use of these data sources in major academic business journals during the period 2005–2012. We do not consider more specialized databases such as the Securities Database Corporation (SDC) mergers and acquisitions database, nor the Thomson Reuters Asset 4 Social Responsibility database, nor country-specific databases such as the NEEDS database on Japanese firms. We did not include DataStream, another commonly used database as it is used almost exclusively in research in finance, accounting, and economics. Indeed, we found that only 1.5% of studies using DataStream were in management. Further, Datastream primarily was used for country-level data and access to Worldscope data (both are Thomson Reuters products).

188

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

2. Database use in previous IB research We found only two previous studies that have examined database choice for IB research; the first only compared two countries while the second only focused on firms from developed European countries. Ulbricht and Weiner (2005) compared data on US and Canadian firms using Compustat North America (not Compustat Global) and Worldscope. They found that a smaller sample size provided by Worldscope prior to 2007 implied a bias towards larger firms, though the larger coverage found in later years reduced the impact of database selection. In a second study, Garcia Lara et al. (2006) found systematic differences in country coverage using seven international databases. These sampling differences led to substantive differences in empirical results when a non-matched sample was used. Differences faded when a matched sample (firms included in all databases) was used. However, as Jones (2006) highlighted in a discussion of the Garcia Lara et al. (2006) study, data were only from 2004 and the study only included firms from developed European countries. Our study is much more comprehensive. In IB and related research, Compustat Global (also called Global Vantage), Osiris and Worldscope have been used extensively and have received wide acceptance. Table 1 illustrates the pervasive use of these three databases in articles published in major academic business journals that regularly publish international research in the fields of management, finance and accounting, economics and political science, and marketing over the 2005–2014 period. We identified 879 articles using one or more of these databases. Management is second only to finance and accounting in the use of these databases. In contrast to marketing and to finance and accounting where one database tends to dominate, use of all three databases is common in management: approximately 28% of management studies used Osiris, 24.5% used Compustat, and 19.5% used Worldscope. Studies using these three databases have investigated a variety of international management topics, including board of director diversity (Arnegger et al., 2014; Mulcahy and Linehan, 2014), corporate social responsibility (Jo et al., 2015; Lopatta and Kaspereit, 2014; Lourenço et al., 2012; Surroca et al., 2013), corporate governance (Busta et al., 2014; Desender et al., 2013; Hautz et al., 2013), research and development/innovation (Sambharya and Lee, 2014), expatriates (Cui et al., 2015), alliances (Kranenburg et al., 2014), and diversification (Hautz et al., 2013; Piaskowska and Trojanowski, 2014). 3. Methodology 3.1. Research questions and “advertised” database coverage To assess the implications of database choice, our analyses address the following research questions: 1. Are the data provided by each database comparable? To address these issues of consistency and reliability we compared data from a matched sample of firms from each database. 2. Are there systematic differences in each database's coverage of geographic regions, such as the size and characteristics of country-level samples? To assess these sampling issues we compared the samples obtained using each database (an “unmatched” sample). 3. How does the choice of database affect empirical results? We assessed the implications for empirical results in two ways. First, we used three multiple regression techniques on the full sample of firms and on regional samples. Second, we replicated a study by O'Brien (2003) on the firm performance implications of firm innovation strategy and debt governance using the three databases. We examined available documentation from each database provider that suggested that the firm and country samples available from each database likely differ. Each database describes its coverage differently. According to the Compustat Global data manual, the database includes over 90% of world market capitalization in over 80 countries (Standard and Poor's, 2011). Compustat Global documentation purports to provide data on over 28,500 companies in 80 countries. Osiris documentation suggests a broad target coverage, by leveraging both its data providers (e.g., Korea Information Service) and original firm documents, even going as far as to say that the firm “strives to cover all publicly listed companies worldwide” (Bureau van Dijk, 2011). Osiris asserts coverage on 55,000 listed companies (data on private companies is available from another Bureau van Dijk database). Worldscope claims coverage on 50,000 companies in 70 countries, describing its procedure as “targeted coverage”. The focus is Table 1 Sources of data used in international business research 2005–2014.

Management Marketing Finance/accounting Economics Political science Other Total Percent

Osiris

Compustat global

Worldscope

Total

Percent

64 20 66 31 29 20 230 26%

52 2 132 15 2 9 212 24%

85 5 296 26 10 15 437 50%

201 27 494 72 41 44 879

23% 3% 56% 8% 5% 5%

Data taken from a search of Business Source Complete. Classifications were based upon journal sponsor and or disciplinary focus (e.g. Academy of Management Journal; Journal of Finance; Journal of Accounting Research), except for multi-disciplinary journals such as Journal of International Business Studies or Journal of Business Research, where classification was made based upon the article title, abstract, and author affiliations.

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

189

on major developed and emerging markets, including countries represented in global indexes such as the FTSE, Dow Jones Global, and S&P Global. In selecting firms to be followed, Worldscope requires that firms must meet one or more criteria, including minimum numbers of broker estimates, market capitalization, and inclusion in global indexes. According to its documentation, Worldscope offers fundamental data on “the world's leading public and private companies” both on “as reported and standardized” bases (Thomson Reuters, 2011). However, as we show in this study, missing data significantly reduce the actual samples available, particularly for less consistently reported variables.

3.2. Sample selection We found only twelve variables common to all three databases, and two of them (intangible assets and administrative expenses) were not consistently defined over the three databases and had substantial missing data for 2004. To illustrate, capital expenditures were available only from Compustat Global and Worldscope.2 Cost of goods sold was available only from Compustat Global and Osiris. We were surprised that few additional variables were reported in all three databases. For example, although we had thought that the number of employees would be a commonly reported variable, it was reported only by Osiris for the entire time period and by Worldscope for only the two later years that we sampled (i.e., 2008 and 2010), but only 17.4% of observations reported employee data in 2010. This provides clear support for our earlier assertion that the databases may have substantive implications for the questions researchers are able to address. It also illustrates the challenge of replicating existing studies making use of different data sources. An important part of our comparative analyses of databases was to develop two types of samples. The first, an unmatched sample, included all firms reporting data in each dataset. This analysis allowed us to assess whether the differing samples provided by the three databases (which might vary in terms of size, country composition, and characteristics of the country-level samples provided) influenced results. The second was a matched sample of firms appearing in all three databases. We used the most commonly populated firm-level identifier, Stock Exchange Daily Official List (SEDOL), to determine matched samples. This analysis allowed us to assess the reliability of the data over the three databases — for example, were descriptive statistics and results similar when identical samples were used? We collected data for the years 2004, 2006, 2008, and 2010. We accessed Compustat Global and Osiris through the Wharton Research Data Services or “WRDS” (provided by The University of Pennsylvania), and we accessed Worldscope through a desktopinstalled software online interface provided by Thomson Reuters. Database reporting procedures, namely variations in the definition of a year, calendar versus fiscal, suggested that biennial samples were more appropriate for the purpose of our sampling frame: to track the evolution of database coverage. Our starting point was Osiris, the database reporting the largest number of firms. Our criterion for country inclusion was a minimum of 30 firms in 2004, resulting in a sample of 66 countries3. We collected data on all firms4 from these 66 countries from the three databases for the selected years. We then constructed unmatched samples (all firms on which data was available in each database) and matched samples (firms appearing in all three databases in a given year). Our country sample was defined by 2004 coverage, possibly excluding countries on which sufficient data became available at a later date. However, a 66-country sample is large. Indeed, examining international research using secondary data published in Academy of Management Journal, Journal of International Business Studies, Global Strategy Journal, and Strategic Management Journal over the period 2010-2014 identified only two studies making use of a larger country sample. The practical implications of excluded countries are likely minimal. The 30-observation screen was based upon the number of firms listed as “covered”, in which a firm name or identification number was available. This potential sample differed significantly from the number of firms on which data was actually available. Firms with no reported data or significant missing data can account for as much as 50% or more of “reported” country coverage. Finally, given the likelihood of extremely small country-level samples for excluded countries it is questionable whether researchers would have been able to appropriately analyze data from a larger sample of countries, particularly in longitudinal research requiring multi-year samples or using nested models requiring a minimum country-level sample size. Our sampling procedure highlighted several important issues. First, the definition of what is a year is a critical consideration. For example, the WRDS interface allows a user to select either the calendar or fiscal year by which to report data. In other instances, the user must make a conscious effort to verify if the calendar or fiscal year convention is being used. For example, a firm reporting on March 31, 2006, may be included in a 2006-year sample because its fiscal year ends in 2006 or it may, in another instance, be included in a 2005 sample because nine of the twelve months occurred in 2005. In many instances, this may 2

Although listed as a data item in Osiris, all observations had missing data. For some analyses, we also separated these 66 countries into geographic regional groupings. The geographic regional groupings and 66 countries are the following: Africa and the Caribbean (Mauritius, Nigeria, South Africa, Bermuda, Cayman Islands, Jamaica, British Virgin Islands), East Asia and Oceania (Australia, Indonesia, Malaysia, New Zealand, Philippines, Singapore, South Korea, Taiwan, Thailand, Vietnam), Eastern Europe (Czech Republic, Latvia, Lithuania, Poland, Russia, Turkey), Latin America (Argentina, Brazil, Chile, Colombia, Ecuador, Mexico, Panama, Paraguay, Peru, Venezuela), Middle East (Egypt, Iraq, Israel, Jordan, Kuwait, Morocco, Oman, Saudi Arabia, United Arab Emirates), Western Europe (Austria, Belgium, Denmark, Finland, France, Germany, Greece, Ireland, Italy, Luxembourg, Netherlands, Norway, Portugal, Spain, Sweden, Switzerland, United Kingdom), We did not include China, Hong Kong, India, Japan in these regional analyses due to the unique characteristics of each market and the large size of the country samples compared to other countries in the regions. As will be discussed later, however, missing firm-level data reduced the number of countries included in our analyses. 4 We excluded firms in finance, insurance, and real estate, and public administration (i.e., Standard Industry Classification or “SIC” one-digit codes 6 and 9, respectively). These industries are commonly excluded from many IB and management studies, and Compustat Global does not include firm-level data for these industries. 3

190

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

not matter. However, in situations where there are major events (e.g., the “9/11” terrorist attacks or the 2008 stock market debacle) or changes to regulatory reporting methods (e.g., mark-to-market or fair value accounting), it is critically important to be cognizant of the year to which we are referring. Second, there is a need to carefully examine missing data. As we discuss below, usable data may not be available for a significant portion of firms included in the database. Researchers investigating a specific topic or region may find that one of these databases is a better source of a given data item. Third, the distinction between country of incorporation, country of listing, and country of headquarters determines in which country a database classifies a firm. Since firms may incorporate in a jurisdiction other than their country headquarters, where possible we used country of headquarters to assign firms to a country sample. This was not always possible for Osiris data (where search parameters were either country of incorporation or of trading) without verifying the headquarters address of every observation. The primary implication of this issue appeared to be inflation of the number of firms attributed to Caribbean locations, particularly the Cayman Islands. Fourth, and perhaps the most critical issue, concerns the treatment of delisted or unlisted firms. Worldscope's online access only database includes data on firms currently covered in the database. Data on firms not currently covered in Worldscope (due to delisting, bankruptcy, merger, or deletion from the database) are not available for prior years even though the firm might have been traded, and data reported, during an earlier period. Firms included in later years but not previously covered were listed as having missing data for earlier years. Therefore, the “raw” numbers of Worldscope firms (not considering missing data) were identical for each year. Similar issues did not appear in the other two databases. Due to increased coverage in later years, there were significant numbers of firms with no reported data for 2004 in Worldscope. Further, comparison with an earlier compact disc (CD) version of the 2004 Worldscope database showed that many firms followed in 2004 were no longer covered in later years, raising concerns regarding survivor bias and the reasons for dropped coverage. Specifically, we identified 5921 firms covered in the 2004 CD but not available through current online access. To assess the reasons for dropped coverage in Worldscope we accessed Osiris data (the database reporting the largest firm and country coverage) on these firms. Of these 5921 firms, 3210 (54%) were currently listed as covered by Osiris. Most of these firms were identified as “delisted”, suggesting that this is a major reason for dropped coverage in Worldscope. However, data on these subsequently delisted firms remain available in Osiris and could be included in analysis. Interestingly, 528 (nearly 10%) of the firms no longer covered by Worldscope remain actively traded and have data reported in Osiris. Given that listed and actual coverage may differ, we assessed the potential survivor bias (and reduction in sample size) by examining whether data on sales and assets for the delisted firms were actually available in Osiris. Data on assets was available for 2461 delisted firms in 2004, 1856 firms in 2006, and 1157 firms in 2008. Availability of data on sales was similar. Given that the current Worldscope sample was approximately 20,000 firms (see Table 2), this suggests a possible 10% reduction in bi-annual samples due to Worldscope's dropped coverage of currently listed and delisted firms. 4. Results 4.1. Variables, country coverage, and descriptive statistics In selecting variables for analyses we focused on data items available and identically defined in all three databases. We carefully examined the documentation provided by each provider (sometimes supplemented with personal communications) to identify variables consistently defined and measured in all three databases. Thus, our selection of variables was by necessity driven by pragmatic limitations, illustrating the limited scope of variables we could verify as consistently available and measured, in addition to theoretical concerns. After verifying the consistency of variable definitions, the following variables were selected: 1) 2) 3) 4) 5)

Size measures: total revenues/sales and total assets Available resources measures: cash to assets and earnings before interest and taxes (EBIT) Leverage measures: total liabilities to assets and long-term debt to total assets Strategic investment measures: research and development (R&D) expenses to total revenues Performance measures: return on assets (ROA) and market-to-book ratio

Table 2 illustrates country coverage. It shows sample characteristics for unmatched and matched samples for the four sample years using the numbers of firms on which data on assets or revenues/sales (two commonly reported and used variables) were available in each database. Table 2 therefore provides a more accurate illustration of the samples provided by each database, which can differ significantly from the number of firms advertised as followed or included in each database according to official database documentation. It is important to note that these samples still represent potential sizes because missing data for other variables diminish sample sizes when running regressions. Several trends emerge. For the total number of observations across all regions, in the unmatched sample Osiris and Worldscope generally provide higher numbers of observations. In general, each database showed a trend toward increasing numbers of observations, particularly between 2004 and 2006. This trend is strongest in Worldscope, where the number of observations increased up to 40% for some variables over the sample period. This led to significant numbers of firms with missing data in 2004 (over 30% in some regions). Table 2 also disaggregates the sample into commonly researched regions and countries. In general, there appears to be little difference in potential samples for Central Asia, Western Europe, and India. However, Osiris may offer researchers focusing on

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

191

Table 2 Sample characteristics for unmatched and matched. Compustat

Worldscope

2004

Africa/Caribbean Central Asia (ex, China, India) China India East Asia (ex, Japan) Japan Latin America Middle East Eastern Europe Western Europe Total

2006

2008

2010

Unmatched

Matched

U

M

U

M

U

312 949 1568 1323 3924 2869 604 299 437 4549 16,834

(201) (820) (1252) (689) (3254) (2330) (334) (234) (137) (2527) (11,778)

331 1053 1716 1443 4079 2969 599 377 502 4541 17,610

(232) (952) (1438) (1157) (3573) (2520) (352) (310) (179) (2889) (13,602)

322 (228) 292 1102 (987) 263 1925 (1490) 2755 1492 (1342) 1457 4245 (3858) 4810 2932 (2614) 2755 577 (355) 542 438 (376) 1065 502 (186) 393 4090 (2937) 3809 17,625 (14,373) 18,141 Total matched (3 years) Total unmatched (3 years)

2004

2006

M

U

M

U

M

(264) (262) (2705) (1455) (4684) (2755) (483) (986) (386) (3567) (17,547) 57300 70210

1059 235 1389 633 4052 2968 532 303 468 3222 14,861

(884) (148) (1137) (354) (3087) (2328) (301) (145) (121) (2510) (11,015)

1252 442 1431 1913 5109 3134 624 901 851 3583 19,240

(1008) (261) (1275) (1073) (3480) (2523) (341) (310) (180) (2889) (13,340)

Latin America, Japan, and the Middle East larger potential samples, but may provide smaller samples for African or Caribbean firms. Next, we turn to Tables 3a and 3b, which shows descriptive statistics for the unmatched (Table 3a) and matched (Table 3b) samples.5 Analysis of descriptive statistics for the unmatched sample suggests that Osiris tends to provide the largest number of firm observations. Compustat Global coverage remains relatively stable over the sample period, whereas Osiris and Worldscope show increasing sample size. This may significantly reduce the sample size in longitudinal research, and may have methodological implications for methodologies sensitive to the number of repeated firm observations in panel data.6 It is noteworthy that Compustat Global offers substantially smaller sample sizes for the market-to-book ratio. This may be an important consideration for researchers wishing to make use of this and perhaps other market-based performance measures. We supplemented observatory comparisons of these descriptive statistics with statistical tests of comparison of means for both the unmatched and matched samples. We performed a two-tailed “t-test” comparing the means for each variable with each database. We compared the values for 2010, which provided the largest sample sizes. For example, we first statistically compared the unmatched mean of revenues for Compustat with that of Osiris. Then, we statistically compared the unmatched mean of revenues for Compustat with that of Worldscope. Finally, we statistically compared the unmatched mean of revenues for Osiris with that of Worldscope. We went on to perform this three-step procedure for each variable for the unmatched and matched samples. So, we ended up with three t-tests for each variable for the unmatched samples and three t-tests for each variable for the matched samples. So, nine variables, three t-tests each for both unmatched and matched samples resulted in 54 total t-tests. To keep it simple, we note below here which databases had means that were statistically different from both other databases, at a significance level of at least 10%. We found the following:

Revenues Assets EBIT Cash Debt Liabilities R&D ROA Market-to-book

Unmatched

Matched

Compustat differed All differed Osiris differed All differed Osiris differed Osiris differed Compustat differed None differed Compustat differed

None differed None differed Worldscope differed All differed Osiris differed None differed None differed None differed Compustat differed

We note a few takeaways. First, we observe statistically significant mean differences for several variables across the databases in the unmatched samples. Given the greater consistency in the matched samples, differences in the unmatched samples likely reflect the differing samples provided by each database. The lack of significant differences between most of the matched samples 5

Examination of the data indicated the prevalence of extreme values. We therefore winsorized variables at three standard deviations above and below the mean. In a preliminary analysis, we found that the use of a “leading” ROA dependent variable, based on a regression similar to those in Table 4a, resulted in a reduction in sample size of about 5% for Compustat Global, 7% for Osiris, and 2% for Worldscope. Osiris and Worldscope show an increasing number of observations in later years (per Tables 3a and 3b), which suggests the likelihood of higher reductions in sample sizes for leading and lagging variables in Osiris and Worldscope compared to Compustat Global. We see this in the case of Osiris, yet for Worldscope we must also remember the deletion of firms subsequently dropped from coverage that we have discussed, which would likely decrease this reduction in sample size for Worldscope. Thus, what we find with a simple analysis based on a leading ROA dependent variable is not inconsistent with what we observe in Tables 3a and 3b coupled with Worldscope’s dropping of firms. 6

192

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

Table 2 Sample characteristics for unmatched and matched. Worldscope

Osiris

2008 U

2010 M

U

1376 (1013) 1664 460 (307) 1020 1814 (1387) 2542 2051 (1264) 3726 5559 (3794) 7943 3176 (2614) 3263 641 (345) 790 933 (376) 1147 919 (188) 1045 3659 (2900) 4621 20,588 (14,188) 27,761 Total Matched (3 years) Total Unmatched (3 years)

2004

2006

2008

2010

M

U

M

U

M

U

M

U

(906) (232) (1790) (1325) (4611) (2544) (349) (314) (168) (2865) (15,104) 53,647 82,450

1294 359 1470 1334 5321 3700 1495 1269 481 4748 21,471

(917) (195) (1214) (687) (3215) (2320) (333) (228) (137) (2532) (11,778)

1345 513 1670 2619 5843 3658 1486 1311 666 4939 24,050

(1004) (295) (1379) (1150) (3522) (2510) (346) (299) (176) (2826) (13,507)

1281 (983) 1518 577 (350) 463 1667 (1432) 2323 2751 (1332) 2962 5783 (3789) 7096 3590 (2590) 3298 1333 (350) 1023 1118 (363) 947 601 (185) 594 4426 (2918) 3888 23,127 (14,292) 24,112 Total Matched (3 years) Total Unmatched (3 years)

M (1498) (427) (2307) (2725) (6762) (3283) (992) (919) (592) (3772) (23,277) 62,854 92,760

suggests that researchers may be able to fill in observations with missing data from one database with data from another. Second, we observe that the performance measures (i.e., ROA and market-to-book) appear to be similar (with the exception of Compustat Global's market-to-book value). This suggests that average performance values remain relatively consistent even with firm-level differences. Third, our finding of a lower mean (and median) revenue and asset values in Osiris relative to the other two databases indicates that Osiris provides more data on smaller firms. Finally, the reduction in significant differences in the matched samples suggests that the firm-level data reported by the three databases is relatively consistent. Overall, this “univariate” analysis comparing the three databases shows some differences across databases, but also some similarities. We next discuss how Compustat Global, Osiris and Worldscope compare based on “multivariate” (i.e., regression) analyses. First, we do a general assessment of database effects using variations of unmatched and matched samples, variations in estimation techniques, and variations of geographical scope. Second, we assess database effects by replicating an existing study. 4.2. Regression analysis for general assessment of database effects 4.2.1. Regression methodology for general assessment of database effects Our third research question was to assess how the choice of database affects empirical results. Our aim is to understand the conditions under which we observe a database effect, which would occur if researchers likely would come to a different conclusion based on the database used. For our general assessment of database effects, we regressed two common firm performance measures (ROA and market-to-book ratio) on the set of variables shown in Tables 3a and 3b, with the exclusion of liabilities to assets in favor of only one debt measure, debt to assets. ROA is a commonly used accounting-based performance measure while market-to-book is a commonly used market-based performance measure (Gentry and Shen, 2010; Hult et al., 2008; Richard et al., 2009). This database effect approach is a straightforward assessment using multiple estimation techniques across several global geographic regions to compare coefficient estimates and statistical significance of all firm-level variables that may affect firm performance. We used several different samples: 1) unmatched samples of all countries, 2) matched samples of all countries, and 3) unmatched and matched samples of regional groupings of countries (i.e., Western Europe, Latin America, Eastern Europe, East Asia). In addition to the influence of sampling, the choice of methodology may affect the extent to which database choice influences empirical results. Therefore, we used three estimation techniques. First, we used ordinary least squares (OLS) regression with country, industry, and year fixed effects. Given the variety of country-level variables used in IB research (e.g. culture, economic development, legal and political variables), we included country fixed effects to account for these country-level characteristics. Second, to more adeptly address outlier concerns, we used robust regression7 with country, industry, and year fixed effects. Finally, we used a multi-level model with industry and year fixed effects and random country intercepts, addressing the nested nature of the data across countries. This methodology is becoming particularly visible in management and other IB-related research (Peterson et al., 2012). 4.2.2. Regression results for general assessment of database effects Tables 4a and 4b show results for the unmatched and matched samples for all countries, using three different estimation techniques, for ROA (Table 4a) and market-to-book (Table 4b) models, respectively. Instances in which the results were parallel for all three databases (statistically significant at least at the 10% level in the same direction or non-significant) were classified as “no discrepancy” (coefficients are underlined), as the conclusions drawn would have been the same for all databases. Instances where one or more of the databases generated statistically significant results but the other database(s) generated non-significant results, but in the same direction, 7 Robust regression weights observations differently by estimating the Cook's D values and eliminating those observations with Cook’s D values greater than one; this iterative process continues using Huber weighting and bi-weighting, resulting in the most influential observations eliminated while more extreme observations are down-weighted.

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

193

Table 3a Descriptive statistics for unmatched samples. Compustat

Revenue Count Mean Median Std dev Assets Count Mean Median Std dev EBIT Count Mean Median Std dev Cash Count Mean Median Std dev Debt/assets Count Mean Median Std dev Liabilities/ assets Count Mean Median Std dev R&D/revenue Count Mean Median Std dev ROA Count Mean Median Std dev Market/book value Count Mean Median Std dev

Osiris

Worldscope

2004

2006

2008

2010

2004

2006

2008

2010

2004

2006

2008

2010

18,498 587 84 1786

19,525 624 94 1874

19,511 723 118 2045

17,352 768 138 2087

22,867 438 61 1404

25,516 499 68 1576

24,646 530 78 1607

23,282 618 96 1752

15,866 584 90 1750

20,511 543 76 1728

22,214 616 90 1862

23,219 642 98 1882

18,491 669 100 2164

19,539 719 113 2238

19,509 825 142 2348

17,312 930 176 2518

23,108 493 71.2 1619

26,032 578 79.1 1900

24,884 619 92.2 1922

23,578 756 122.1 2131

15,884 679 107 2175

20,522 630 90 2083

22,193 716 108 2190

23,167 809 127 2383

18,484 19,475 19,506 17,328 23,093 26,043 25,155 23,754 15,539 19,951 21,728 22,810 48.33 52.84 59.79 62.55 37.08 45.01 42.46 54.83 52.92 54.05 49.80 60.00 4.87 5.63 6.38 7.54 3.67 4.05 3.63 5.28 6.25 5.71 5.32 6.83 182 192 217 208 155 183 185 204 202 204 213 216 17,992 19,047 19,165 17,587 22,558 25,374 24,253 23,083 15,978 20,632 22,365 23,380 0.15 0.15 0.15 0.17 0.14 0.14 0.14 0.15 0.17 0.17 0.16 0.18 0.10 0.11 0.11 0.12 0.09 0.09 0.09 0.10 0.11 0.11 0.10 0.12 0.14 0.14 0.14 0.15 0.14 0.14 0.14 0.15 0.17 0.18 0.18 0.18 18,315 19,436 19,491 17,957 18,595 20,959 20,136 19,055 15,953 20,644 22,382 23,347 0.12 0.11 0.11 0.12 0.18 0.20 0.21 0.21 0.12 0.14 0.13 0.12 0.06 0.05 0.05 0.05 0.10 0.10 0.09 0.09 0.05 0.05 0.04 0.04 0.50 0.43 0.25 0.44 1.19 1.47 2.11 2.22 0.54 2.86 1.93 1.10

18,414 19,611 19,642 18,065 23,123 26,096 24,870 23,821 15,995 20,667 22,397 23,392 0.60 0.56 0.53 0.55 0.63 0.63 0.62 0.63 0.60 0.57 0.61 0.58 0.52 0.50 0.50 0.48 0.52 0.51 0.51 0.49 0.50 0.49 0.49 0.46 2.02 1.71 1.25 1.83 2.33 2.71 2.66 3.34 4.10 3.37 5.00 3.14 5121 0.28 0.02 3.12

5721 0.45 0.02 4.87

6461 0.30 0.01 2.84

6272 0.28 0.02 3.05

7608 0.20 0.01 2.14

9154 0.22 0.01 2.13

9145 0.17 0.01 1.82

9183 0.15 0.01 1.56

5971 0.24 0.01 3.60

7725 0.20 0.01 2.54

8206 0.19 0.01 2.59

8927 0.19 0.01 2.70

16,597 18,512 19,210 17,799 23,217 26,203 25,054 23,805 15,928 20,607 22,380 23,367 −0.04 −0.02 −0.05 −0.05 −0.04 −0.04 −0.09 −0.02 0.00 −0.02 −0.05 −0.03 0.03 0.04 0.03 0.03 0.03 0.03 0.03 0.03 0.03 0.04 0.03 0.03 3.60 1.85 1.83 2.98 2.72 1.77 3.06 3.07 2.78 1.58 2.25 3.11

13,115 3 1 38

15,423 3 1 64

16,702 1 0 27

16,629 2 1 43

15,495 17,585 18,611 21,477 13,023 16,783 19,146 20,390 1.84 1.98 1.00 1.49 1.55 1.59 1.01 1.44 0.61 0.78 0.46 0.60 0.63 0.78 0.47 0.64 42 46 5 21 23.96 10.70 8.18 11

Revenues, assets, and EBIT in millions of US$. Cash, debt, liabilities, R&D, ROA, and market-to-book are ratios.

were classified as “low discrepancy” (coefficients are bolded and in italics). In this case, findings might have been reported as significant using one database, but insignificant using others. We classified as a “moderate discrepancy” (coefficients are in italics) instances where one or more databases generated statistically significant coefficients, while others reported non-significant coefficients in the opposite direction. Here again, researchers using different databases might have drawn different conclusions. Finally, a “severe discrepancy” (coefficients are bolded and underlined) occurred when the databases generated statistically significant results but in different directions. Thus, we did not simply assess the statistical significance of differences in reported results but sought to verify the consistency and repeatability of results across databases and how the conclusions drawn from such empirical analyses may differ. These results suggest that methodology does matter and, as might be expected, greater differences exist in unmatched samples, although, in a few instances, differences were found only in matched samples. Greater differences were found for regressions using OLS and multi-level estimations. Even after the data was winsorized to remove extreme observations, use of robust regression to control for the effects of extreme observations significantly reduces differences in results compared to OLS. Interestingly, we also see differences when multi-level models are used, suggesting that methods accounting for the nested nature of the data may make results more sensitive to differences in the country-level samples provided by each database. In general, we tend to see greater differences in Table 4b, with market performance (i.e., market-to-book) as the dependent variable. There is also greater similarity in the matched sample where robust regression leads to greater consistency in research findings.

194

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

Table 3b Descriptive statistics for matched samples. Compustat

Revenue Count Mean Median Std dev Assets Count Mean Median Std dev EBIT Count Mean Median Std dev Cash Count Mean Median Std dev Debt/assets Count Mean Median Std dev Liabilities/ assets Count Mean Median Std dev R&D/revenue Count Mean Median Std dev ROA Count Mean Median Std dev Market/book value Count Mean Median Std dev

Osiris

Worldscope

2004

2006

2008

2010

2004

2006

2008

2010

2004

2006

2008

2010

11,681 683 112 1953

14,210 674 110 1960

15,241 765 135 2107

14,560 805 151 2127

11,559 584 100 1639

13,900 656 114 1824

14,906 673 127 1790

14,638 788 158 1968

11,653 682 113 1927

14,184 670 110 1938

15,236 755 135 2065

14,770 811 154 2102

11,681 771 130 2332

14,215 763 128 2313

15,236 852 158 2370

14,531 959 190 2533

11,684 637 116 1848

14,178 728 126 2123

15,176 752 141 2087

14,834 923 191 2318

11,675 770 129 2336

14,209 752 126 2276

15,244 845 157 2353

14,760 978 193 2569

11,690 14,197 15,264 14,560 11,698 14,200 15,264 14,926 11,489 13,979 15,038 14,611 56 57 63 64 48.84 58.50 54.62 66.0 59 63 60 71 6.76 6.87 7.23 8.38 6.21 7.20 6.62 9.06 7.27 7.80 7.80 10.11 195 195 222 208 175 204 207 221 214 221 229 234 11,516 13,952 15,010 14,681 11,512 13,943 14,915 14,639 11,772 14,325 15,410 14,943 0.15 0.15 0.15 0.17 0.15 0.15 0.15 0.16 0.17 0.17 0.17 0.18 0.11 0.11 0.11 0.13 0.11 0.10 0.10 0.12 0.12 0.11 0.11 0.13 0.14 0.14 0.14 0.15 0.14 0.14 0.14 0.15 0.17 0.18 0.17 0.18 11,731 14,231 15,268 14,991 0.11 0.11 0.11 0.11 0.05 0.05 0.05 0.05 0.56 0.21 0.23 0.39

9,831 11,910 12,622 12,287 11,765 14,326 15,406 14,927 0.15 0.15 0.14 0.15 0.12 0.12 0.11 0.11 0.09 0.09 0.09 0.09 0.05 0.05 0.05 0.05 0.69 0.62 0.26 0.34 0.62 0.77 0.25 0.31

11,747 14,311 15,369 15,057 11,771 14,319 15,308 15,043 11,774 14,329 15,414 14,947 0.54 0.53 0.51 0.51 0.56 0.55 0.54 0.54 0.57 0.53 0.53 0.54 0.50 0.50 0.49 0.47 0.52 0.51 0.51 0.49 0.49 0.49 0.49 0.46 1.47 1.41 0.9 1.16 1.55 1.42 1.49 1.95 3.87 1.45 2.78 2.74 3,769 0.26 0.02 3.27

4,584 0.35 0.02 4.04

5,468 0.25 0.01 2.77

5,401 0.27 0.01 2.83

4,375 0.21 0.01 2.42

5,859 0.20 0.01 2.03

6,526 0.15 0.01 1.68

6,608 0.15 0.01 1.58

4,629 0.22 0.01 3.20

5,676 0.20 0.01 2.48

6,175 0.21 0.01 2.87

6,143 0.20 0.01 2.88

10,826 13,723 15,200 14,892 11,772 14,322 15,305 15,040 11,749 14,303 15,411 14,940 −0.03 −0.01 −0.03 −0.02 0.00 −0.03 −0.04 −0.02 0.01 −0.03 −0.04 −0.02 0.03 0.04 0.03 0.03 0.03 0.04 0.03 0.03 0.03 0.04 0.03 0.03 4.33 1.96 1 1.38 2.8 1.55 1 1.4 2.77 1.66 1.42 1.38

9318 2 1 22

12,016 3 1 63

13,863 1 0 7

14,208 2 1 44

9480 11,515 13,176 14,450 10,373 12,753 14,335 14,065 1.42 2.02 0.99 1.34 1.29 1.53 0.97 1.34 0.63 0.79 0.48 0.64 0.66 0.81 0.49 0.67 16 53 5 15 12.0 10.0 5.9 4.9

Revenues, assets, and EBIT in millions of US$. Cash, debt, liabilities, R&D, ROA, and market-to-book are ratios.

Although our descriptive statistics in Tables 2, 3a and 3b suggested that Osiris provided the largest sample size, the number of firms included in each regression depends on the extent to which complete data is available for a given observation. The number of observations used in the regressions in Tables 4a, 4b, 4c, and 4d suggests that Worldscope generally provides the largest samples in multivariate analysis. This suggests that the larger scope of coverage provided by Osiris is associated with less complete firm-level data. Thus, the smaller potential sample size found in Worldscope and Compustat Global may be offset by more complete firm-level data, and thus fewer deleted observations due to missing data. In addition to overall sample size, it is important to consider the size of country-level samples, particularly when using multilevel or nested estimations. Table 5 illustrates the number of countries which would have been included in the ROA analysis (Table 4a, unmatched sample) at a given minimum country-level sample threshold using each database. Results from Table 5 indicate that Worldscope generally provides larger samples (i.e., more complete firm-level data) in a larger number of countries. Particularly when researchers want large country-level samples (e.g., 100 firms or more), use of Worldscope may offer advantages. The differences in regional coverage noted earlier may also influence the size of country-level samples. Research often focuses on a specific geographic region. To assess whether database effects vary by region, we conducted separate analysis for Western Europe, Latin America, Eastern Europe, and East Asia. Tables 4c (ROA) and 4d (market-to-book) show the results. In Table 4d, for example, Compustat Global and Worldscope provided extremely small sample sizes for Latin

Table 4a Return on asset models. OLS: extreme values excluded

Independent variables

Compustat

Osiris

Worldscope

Compustat

Osiris

Worldscope

Compustat

Osiris

Worldscope

Revenues Total assets EBIT Cash/total assets Debt/total assets R&D/revenues Constant Year fixed effects Industry fixed effects Country fixed effects Observations R-squared

9.23E−07 −4.91E−06 0.000197⁎⁎⁎ −0.124⁎⁎⁎ −0.374⁎⁎⁎ −0.0180⁎⁎⁎

−1.87E−06 −0.00000448 0.000247⁎⁎⁎

4.08E−07 −6.20E−06 0.000215⁎⁎⁎ −0.148⁎⁎⁎ −0.388⁎⁎⁎ −0.0187⁎⁎⁎

7.70e−07⁎⁎ −6.82e−06⁎⁎⁎ 0.000129⁎⁎⁎ 0.0757⁎⁎⁎ −0.0605⁎⁎⁎ −0.0547⁎⁎⁎ 0.0835⁎⁎

7.11e−07⁎⁎ −7.92e−06⁎⁎⁎ 0.000131⁎⁎⁎ 0.0804⁎⁎⁎ −0.0838⁎⁎⁎ −0.0318⁎⁎⁎ 0.144⁎⁎⁎

3.56E−06 −0.00000642 0.000192⁎⁎⁎

−2.21E−06 −4.20E−06 0.000225⁎⁎⁎ −0.0373 −0.352⁎⁎⁎ −0.136⁎⁎⁎ 0.0935⁎

Yes Yes Yes 21734 0.96

Yes Yes Yes 28003 0.659

Yes Yes Yes 29418 0.762

−0.0699 −0.302⁎⁎⁎ −0.0246⁎⁎⁎ 0.127 Yes Yes No 21734 59 groups

1.31E−06 −9.00E−06 0.000228⁎⁎⁎ −0.117⁎ −0.298⁎⁎⁎ −0.0236⁎⁎⁎

0.189 Yes Yes Yes 29418 0.012

3.00E−07 −7.02e−06⁎⁎⁎ 0.000127⁎⁎⁎ 0.0973⁎⁎⁎ −0.0810⁎⁎⁎ −0.0896⁎⁎⁎ 0.129⁎⁎

Yes Yes No 28005 56 groups

−0.0134 Yes Yes No 29418 59 groups

Matched Revenues Total assets EBIT Cash/total assets Debt/total assets R&D/revenues Constant Year fixed effects Industry fixed effects Country fixed effects Observations R-squared No discrepancy Low discrepancy Moderate discrepancy Severe discrepancy

OLS: robust

Multilevel

0.248 Yes Yes Yes 21734 0.022

−0.0166 −0.230⁎⁎⁎ −0.0366⁎⁎⁎ 0.164 Yes Yes Yes 28005 0.034

6.55E−07 −4.49E−06 0.000170⁎⁎⁎ −0.110⁎ −0.361⁎⁎⁎ −0.0173⁎⁎⁎

1.87E−06 −9.05e−06⁎⁎⁎ 0.000226⁎⁎⁎ −0.0472⁎⁎ −0.146⁎⁎⁎ −0.0316⁎⁎⁎

9.45E−07 −6.64E −06 0.000205⁎⁎⁎ −0.147⁎⁎⁎ −0.358⁎⁎⁎ −0.0156⁎⁎⁎

3.88E−07 −6.60e−06⁎⁎⁎ 0.000119⁎⁎⁎ 0.0955⁎⁎⁎ −0.0801⁎⁎⁎ −0.0281⁎⁎⁎

9.68e−07⁎⁎⁎ −6.33e−06⁎⁎⁎ 0.000118⁎⁎⁎ 0.0733⁎⁎⁎ −0.0600⁎⁎⁎ −0.0548⁎⁎⁎

8.63e−07⁎⁎⁎ −7.22e−06⁎⁎⁎ 0.000120⁎⁎⁎ 0.0755⁎⁎⁎ −0.0801⁎⁎⁎ −0.0326⁎⁎⁎

2.54E−06 −5.92E−06 0.000176⁎⁎ −0.0474 −0.257⁎⁎⁎ −0.0229⁎⁎⁎

2.96E−07 −7.73e−06⁎⁎ 0.000192⁎⁎⁎ 0.0153 −0.239⁎⁎⁎ −0.139⁎⁎⁎

3.53E−06 −1.29E−05 0.000255⁎⁎⁎ −0.097 −0.206⁎⁎ −0.0251⁎⁎

0.141 Yes Yes Yes 17926 0.018

−1.14E−01 Yes Yes Yes 18996 0.054

0.0197 Yes Yes Yes 21612 0.022

0.0173 Yes Yes Yes 17923 0.716

0.0304 Yes Yes Yes 18995 0.642

0.0677 Yes Yes Yes 21612 0.809

0.122 Yes Yes No 17926 54 groups

0.0119 Yes Yes No 18996 50 groups

−0.0204 Yes Yes No 21612 59 groups

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

Unmatched

statistically signficant, same direction, for all three databases statistically significant in one direction and non-statistically significant in same direction statistically significant in one direction and non-statistically significant in opposite direction statistically signficant in opposite directions

⁎⁎⁎ p b 0.01. ⁎⁎ p b 0.05. ⁎ p b 0.1.

195

196

Table 4b Market-to-book models. OLS: extreme values excluded

Independent variables

Compustat

Revenues Total assets EBIT Cash/total assets Debt/total assets R&D/revenues Constant Year fixed effects Industry fixed effects Country fixed effects Observations R-squared

3.92E−04⁎⁎ −6.1E−04⁎⁎⁎

Matched Revenues Total assets EBIT Cash/total assets Debt/total assets R&D/revenues Constant Year fixed effects Industry fixed effects Country fixed effects Observations R-squared No discrepancy Low discrepancy Moderate discrepancy Severe discrepancy ⁎⁎⁎ p b 0.01. ⁎⁎ p b 0.05. ⁎ p b 0.1.

−3.10E−04 3.58⁎⁎⁎ 11.6⁎⁎⁎ −0.01441 1.134 Yes Yes Yes 19229 0.106 5.09E−04⁎⁎⁎ −6.35E−04⁎⁎⁎ −0.00199 4.275⁎⁎⁎ 13.45⁎⁎⁎ −0.08578 1.788 Yes Yes Yes 16334 0.13

Osiris 1.54E−05 −8.27e−05⁎⁎⁎ 0.000961⁎⁎⁎ 3.058⁎⁎⁎ 1.177⁎⁎⁎ 0.0707⁎⁎⁎ 0.58 Yes Yes Yes 23766 0.079

1.10E−05 −7.20e−05⁎⁎⁎ 0.000933⁎⁎⁎ 2.871⁎⁎⁎ 1.216⁎⁎⁎ 0.0880⁎⁎⁎ −1.057 Yes Yes Yes 16744 0.088

OLS: robust

Multilevel

Worldscope

Compustat

Osiris

Worldscope

Compustat

Osiris

Worldscope

−1.71E−05 −5.42E−05 0.000749⁎⁎⁎ 3.430⁎⁎⁎

−7.88E−06⁎⁎⁎ −2.58e−05⁎⁎⁎ 0.000779⁎⁎⁎ 1.134⁎⁎⁎ −0.573⁎⁎⁎ 0.08583⁎⁎⁎

−0.0000032 −2.24e−05⁎⁎⁎ 0.000605⁎⁎⁎ 1.013⁎⁎⁎ −0.486⁎⁎⁎ 0.00606⁎⁎⁎ 0.844⁎⁎⁎

Yes Yes Yes 23765 0.401

Yes Yes Yes 25475 0.338

−1.40E−04 −3.55E−04 0.00159 −1.370 12.09 0.1085 2.042 Yes Yes No 19229 55 groups

−2.22E−05 −9.38e−05⁎⁎⁎ 0.00124⁎⁎⁎ 2.666⁎⁎⁎

0.254 Yes Yes Yes 19228 0.475

−2.40E−06 −2.10e−05⁎⁎⁎ 0.000684⁎⁎⁎ 0.949⁎⁎⁎ −0.379⁎⁎⁎ 0.132⁎⁎⁎ 0.979⁎⁎⁎

0.322 0.271⁎⁎⁎ 0.497⁎⁎⁎ Yes Yes No 23766 54 groups

−2.39E−05 −1.87E−05 −0.000101 9.395 −2.74 0.0445 1.199⁎⁎ Yes Yes No 25476 59 groups

−1.01E−05⁎⁎⁎ −2.16E−05⁎⁎⁎ 0.000766⁎⁎⁎ 1.125⁎⁎⁎ −0.603⁎⁎⁎ 0.0821⁎⁎⁎ 2.841⁎⁎⁎

−5.17e−06⁎⁎ −1.55e−05⁎⁎⁎ 0.000676⁎⁎⁎ 0.929⁎⁎⁎ −0.441⁎⁎⁎ 0.0667⁎⁎⁎ 2.605⁎⁎⁎

−2.85E−06 −2.08e−05⁎⁎⁎ 0.000600⁎⁎⁎ 1.005⁎⁎⁎ −0.559⁎⁎⁎ 0.00485⁎⁎⁎ 1.278⁎⁎⁎

−2.85E−05 −9.14e−05⁎⁎⁎ 0.00127⁎⁎⁎ 2.596⁎⁎⁎

−3.29E−05 −8.00E−05 0.000326 9.091⁎

Yes Yes Yes 16333 0.457

Yes Yes Yes 16743 0.383

Yes Yes Yes 19884 0.364

0.221 0.485⁎⁎⁎ 0.772⁎⁎⁎ Yes Yes No 16744 50 groups

−2.471 0.00837 0.959 Yes Yes No 19887 58 groups

−0.0578 0.0176 −0.0447 Yes Yes Yes 25476 0.058 −1.58E−05 −5.40E−05 0.000780⁎⁎⁎ 3.881⁎⁎⁎ 0.343 0.00334 −1.076 Yes Yes Yes 19887 0.123

−8.05E−06 −4.967E−04⁎ 0.00088 0.768 12.96 0.143 1.702 Yes Yes No 16334 50 groups

statistically signficant, same direction, for all three databases statistically significant in one direction and non-statistically significant in same direction statistically significant in one direction and non-statistically significant in opposite direction statistically signficant in opposite directions

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

Unmatched

Table 4c Return on asset models by region. Western Europe

Independent variables

Compustat

Osiris

Worldscope

Compustat

Osiris

Worldscope

Compustat

Osiris

Worldscope

Compustat

Osiris

Worldscope

Revenues Total assets EBIT Cash/total assets Debt/total assets R&D/revenues Constant Year fixed effects Industry fixed Effects Country fixed effects Observations R-squared

1.16E−06 −4.39E−06 0.000199⁎⁎⁎ −0.264⁎⁎⁎ −0.536⁎⁎⁎ −0.0178⁎⁎⁎

−4.29E−06 3.46E−06 0.000159⁎⁎⁎ −0.316⁎⁎⁎ −0.742⁎⁎⁎ −0.0203⁎⁎⁎

−1.56E−06 −1.32E−06 0.000170⁎⁎⁎ −0.237⁎⁎⁎ −0.552⁎⁎⁎ −0.0158⁎⁎⁎

4.43E−06 −1.50e−05⁎⁎⁎ 0.000128⁎⁎⁎

−5.29E−06 1.29E−06 5.53E−05 −0.0902 −0.251⁎

−2.62E−07 −9.58e−06⁎⁎⁎ 0.000144⁎⁎⁎ 0.116⁎⁎ −0.137⁎⁎⁎

1.87E−06 −1.85e−05⁎⁎⁎ 0.000162⁎⁎⁎ 0.340⁎⁎⁎ −0.239⁎⁎⁎

0.31 Yes Yes

0.0823 Yes Yes

−0.126 0.0359 Yes Yes

−0.154 Yes Yes

0.187 Yes Yes

1.30E−06 −6.64e−06⁎⁎⁎ 0.000225⁎⁎⁎ 0.124⁎⁎⁎ −0.101⁎⁎⁎ −0.381⁎⁎⁎ 0.0713⁎

Yes Yes

−0.0112 0.0622⁎ Yes Yes

1.34E−06 −5.42e−06⁎⁎⁎ 0.000158⁎⁎⁎ 0.117⁎⁎⁎ −0.111⁎⁎⁎ −0.00650⁎⁎⁎

−8.69E−06 −3.07e−05⁎⁎ 0.000913⁎⁎⁎ 2.318⁎⁎⁎ −0.799⁎⁎⁎

−0.0404 0.132 Yes Yes

−1.50E−06 −8.19e−06⁎⁎ 0.000136⁎⁎⁎ 0.213⁎⁎⁎ −0.129⁎⁎⁎ −1.101⁎⁎⁎

0.185 Yes Yes

0.0000462 −5.18E−05 0.000432 −0.113 −0.0635 2.064 0.235 Yes Yes

Yes Yes

0.0501 0.919⁎⁎ Yes Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

4792 0.077

3524 0.329

4781 0.28

120 0.543

39 0.683

323 0.11

269 0.28

270 0.592

698 0.273

12347 0.066

19154 0.067

13879 0.098

−5.85E-07 −0.00000228 0.000152⁎⁎ −0.225⁎⁎⁎ −0.595⁎⁎⁎ −0.0175⁎⁎⁎ 0.22 Yes Yes

−6.13E-06 4.15E-06 0.000146⁎⁎⁎ −0.301⁎⁎⁎ −0.827⁎⁎⁎ −0.0172⁎⁎⁎ 0.338 Yes Yes

−2.99E-06 −8.13E-07 0.000167⁎⁎⁎ −0.241⁎⁎⁎ −0.612⁎⁎⁎ −0.0115⁎⁎⁎ 0.226⁎⁎ Yes Yes

8.25E-06 −1.29e-05⁎ 7.99e-05⁎⁎ 0.273⁎⁎ −0.120⁎⁎⁎ −1.781⁎⁎ 0.216⁎⁎ Yes Yes

0.0000959 −0.000111 0.00134 0.608 0.368 −1.247 0.0575 Yes Yes

0.000125 −0.0000862 0.000295 0.128 0.548 1.551 0.629 Yes Yes

3.87E-06 −1.53e-05⁎⁎⁎ 0.000155⁎⁎⁎ 0.0693 −0.181⁎⁎⁎ −0.228 0.0493 Yes Yes

3.30E-06 −1.56e-05⁎⁎ 0.000164⁎⁎⁎ 0.0988 −0.150⁎⁎⁎ 0.46 0.0534 Yes Yes

3.66E-06 −1.57e-05⁎⁎⁎ 0.000161⁎⁎⁎ 0.0137 −0.160⁎⁎⁎ −0.994⁎⁎⁎ 0.00015 Yes Yes

1.48E-06 −5.74e-06⁎⁎⁎ 0.000147⁎⁎⁎ 0.113⁎⁎⁎ −0.0962⁎⁎⁎ −0.00512⁎⁎⁎ 0.0513⁎ Yes Yes

1.87E-06 −7.00e-06⁎⁎⁎ 0.000188⁎⁎⁎ 0.0820⁎⁎⁎ −0.0795⁎⁎⁎ −0.106⁎⁎⁎ 0.0887⁎⁎ Yes Yes

1.11E-06 −3.62e-05⁎⁎⁎ 0.000943⁎⁎⁎ 2.156⁎⁎⁎ −0.618⁎⁎⁎ 0.0326 1.331 Yes Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

3477 0.102

2442 0.458

4089 0.267

68 0.708

20 0.975

191 0.163

189 0.378

128 0.348

241 0.413

10800 0.08

13214 0.072

11102 0.18

Matched Revenues Total assets EBIT Cash/total assets Debt/total assets R&D/revenues Constant Year fixed effects Industry fixed effects Country fixed effects Observations R-squared

Latin America

0.0988 −0.127⁎⁎⁎ −1.619⁎⁎⁎ 0.315⁎⁎

Eastern Europe

East Asia

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

Unmatched

Industry and country dummies included in all UNMATCHED models. No discrepancy Low discrepancy Moderate discrepancy Severe discrepancy

statistically signficant, same direction, for all three databases statistically significant in one direction and non-statistically significant in same direction statistically significant in one direction and non-statistically significant in opposite direction statistically signficant in opposite directions

⁎⁎⁎ p b 0.01. ⁎⁎ p b 0.05. ⁎ p b 0.1.

197

198

Table 4d Market-to-book models by region. Western Europe

Independent variables

Compustat

Osiris

Worldscope

Latin America Compustat

Osiris

Worldscope

Eastern Europe Compustat

Osiris

Worldscope

Compustat

East Asia Osiris

Worldscope

Revenues Total assets EBIT Cash/total assets Debt/total assets R&D/revenues Constant Year fixed effects Industry fixed effects Country fixed effects Observations R-squared

−2.83E−05 −7.8E−05 9.84E−04⁎⁎⁎ 3.178⁎⁎⁎ 0.502⁎⁎⁎ 0.0645⁎⁎⁎

−1.93E−05 −9.67e−05⁎⁎⁎ 0.00112⁎⁎⁎ 3.752⁎⁎⁎ 0.909⁎⁎⁎ 0.0684⁎⁎⁎

−1.72E−03 −7.30E−03 −0.0109 −178.1 4.654e+02⁎⁎⁎

−0.00169 0.00108 0.0024 14.83⁎

−0.0029 0.00176 −0.0129 65.78⁎⁎⁎

−2.26E−05 −6.36E−05 .0006665⁎

−2.11E−05 −0.000115 0.000855⁎

0.1468 .−931⁎

Yes Yes Yes 4497 0.217

0.0378 1.004⁎⁎ Yes Yes Yes 230 0.168

−0.425 −0.362 2.518 −0.106 Yes Yes Yes 246 0.111

−0.576 1.792⁎⁎⁎ Yes Yes Yes 595 0.135

−5.86E−06 −3.16e−05⁎ 0.000903⁎⁎⁎ 1.976⁎⁎⁎ −0.857⁎⁎⁎ 0.579⁎⁎⁎

−8.69E−06 −3.07e−05⁎⁎ 0.000913⁎⁎⁎ 2.318⁎⁎⁎ −0.799⁎⁎⁎

−2243 134.8 Yes Yes Yes 110 0.345

−9.688 −123.4 −7.573 Yes Yes Yes 293 0.283

−1.33E−05 −3,643E−05⁎⁎⁎ .001258⁎⁎⁎ 2.242⁎⁎⁎ −0.6023⁎⁎⁎ 0.480⁎⁎⁎

−0.648 Yes Yes Yes 3323 0.206

5.086 −41.61⁎ 4.243 Yes Yes Yes 25 0.855

−5.18E−05 −2.49E−05 0.000415 1.185⁎⁎ −0.464⁎⁎

0.388 Yes Yes Yes 4429 0.223

−2.09E−05 −6.82e−05⁎⁎⁎ 0.000717⁎⁎⁎ 3.560⁎⁎⁎ 0.559⁎⁎⁎ 0.0251⁎⁎⁎ 0.727⁎

0.9535 Yes Yes Yes 11546 0.166

0.738 Yes Yes Yes 16103 0.039

Yes Yes Yes 13879 0.098

Matched Revenues Total assets EBIT Cash/total assets Debt/total assets R&D/revenues Constant Year fixed effects Industry fixed effects Country fixed effects Observations R-squared

−2.39E−05 −7.845E−05⁎⁎⁎ 9.34E−04⁎⁎⁎ 3.309⁎⁎⁎ 0.677⁎⁎⁎ 0.0561⁎⁎⁎ 0.4471 Yes Yes Yes 3329 0.249

−1.54E−05 −9.77e−05⁎⁎⁎ 0.00111⁎⁎⁎ 3.675⁎⁎⁎ 1.083⁎⁎⁎ 0.0733⁎⁎⁎ 1.168⁎⁎ Yes Yes Yes 2355 0.248

−2.02E−05 −7.26e−05⁎⁎⁎ 0.000757⁎⁎⁎ 3.518⁎⁎⁎ 0.815⁎⁎⁎ 0.00444 −0.0505 Yes Yes Yes 3980 0.211

0.011 −0.0126 −0.0767 42.32 5.420e+02⁎⁎⁎ −3.02E+03 37.78 Yes Yes Yes 65 0.435

−0.0147 0.0115 −0.0640 35.33 −27.21 0 −23.22 Yes Yes Yes 18 1

−0.00683 0.0035 −0.0070 69.37⁎⁎ −45.83 −118.1 14.11 Yes Yes Yes 184 0.372

1.20E−05 −1.01E−04 .0007517⁎ −0.614 −0.539 12.23⁎⁎⁎ 0.773 Yes Yes Yes 174 0.212

2.04E−05 −0.000158⁎ 0.000815 −0.658 −0.893 −15.3 1.534 Yes Yes Yes 122 0.174

1.05E−05 −7.92E−05 0.000561⁎ −0.521 −0.866 −2.294 0.533 Yes Yes Yes 231 0.157

−1.18E−05 −3.6.0E−05⁎⁎⁎ 1.239E−03⁎⁎⁎ 2.308⁎⁎⁎ −0.626⁎⁎⁎ 0.498⁎⁎⁎ 0.957 Yes Yes Yes 10215 0.165

−7.80E−06 −2.47E−05 0.000872⁎⁎⁎ 1.685⁎⁎⁎ −0.830⁎⁎⁎ 0.481⁎⁎ 0.641 Yes Yes Yes 11474 0.033

1.11E−06 −3.62e−05⁎⁎⁎ 0.000943⁎⁎⁎ 2.156⁎⁎⁎ −0.618⁎⁎⁎ 0.0326 1.331 Yes Yes Yes 11102 0.18

0.0501 0.919⁎⁎

Industry and country dummies included in all UNMATCHED models. No discrepancy Low discrepancy Moderate discrepancy Severe discrepancy

⁎⁎⁎ p b 0.01. ⁎⁎ p b 0.05. ⁎ p b 0.1.

statistically signficant, same direction, for all three databases statistically significant in one direction and non-statistically significant in same direction statistically significant in one direction and non-statistically significant in opposite direction statistically signficant in opposite directions

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

Unmatched

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

199

America. Further, results depend upon the specific variables selected, as illustrated by the differences in number of observations in Tables 4c (ROA) and 4d (market-to-book). These results suggest that in addition to the effect on empirical results, choice of database may influence the number and type of countries included in analyses. 4.3. Regression analysis for replication study 4.3.1. Regression methodology for replication study Our second approach to assess database effects is through replication of an existing study. As noted in our discussion of variable selection, the lack of overlap among the variables provided in each database and the use of more specialized databases (e.g. SDC mergers and acquisitions database, patent data databases) severely constrained the choice of potential studies to replicate. Given these limitations, we identified O'Brien (2003) as a feasible study to replicate because it showed substantial overlap with the data available from our three databases. It also addresses a topic of importance to IB research: governance of R&D investment (Bah and Dumontier, 2001; Bougheas, 2004; Criscuolo et al., 2005; David et al., 2008; Hillier et al., 2010; Hundley et al., 1996). We replicated the third hypothesis of O'Brien, which states the following: “There will be a negative interaction between leverage and the importance of innovation to the firm's strategy with respect to their impact on firm performance” (2003: 421). This hypothesis is based on a relatively long history in management research that has been devoted to understanding how to best govern R&D investment — O'Brien (2003) innovation measure is based on R&D investment. Investment in R&D is important to a firm's innovation strategy and drives competitive advantage, though R&D investment creates firm-specific assets that are difficult to govern. Two governance modes, equity and debt, have distinct features that make equity generally more appropriate to govern R&D investment than debt. Equity compared to debt is more flexible to earnings disturbances and provides better monitoring of managerial investments, both of which are helpful to govern R&D. Several variants of this general relationship have been tested empirically with results published in several management and IB journals. Balakrishnan and Fox (1993), for example, found high R&D investment to be negatively related to firm leverage. David et al. (2008) is a more nuanced examination using a Japanese sample of the difference between bank loan and bond debt for governing R&D investment and implications for firm performance. Hundley et al. (1996) compared Japanese and US companies with respect to how profitability and liquidity affect R&D investment. Hillier et al. (2011) examined various factors that affect the sensitivity of R&D investment to cash flow, such as the level of development of a country's financial system. O'Brien (2003) tests leverage as the dependent variable in his first two hypotheses, while his third hypothesis – the one we test – uses performance as the dependent variable. It tests the performance implications of innovation matched with high leverage, which we think may be more interesting to the broad IB readership. In our replication, we aimed to match O'Brien (2003) methodology by extending his sampling of US firms to sampling across a wide variety of countries and regions using the Compustat Global, Osiris and Worldscope databases. Therefore, we are able to examine if and where the “innovation-leverage-performance” relationship is most and least consistent across databases. We also are able to examine if and where this relationship is similar to and different from what O'Brien (2003) found in the US, which was a negative performance effect of high innovation matched with high leverage. This is important because cross-national differences in institutional contexts have significant implications for the role of creditors. For example, in “bank-based” systems such as Japan and Germany, close relations between creditors and firms have important implications for the use of debt to fund risky investments such as R&D (Demirguc-Kunt and Maksimovic, 2002; Lee, 2012; Levine, 2002; Shiow-Ying and Yu, 2012). James and McGuire (in press) found, for example, that high levels of bank debt combined with high levels of R&D investment was associated with higher performance in bank-based systems but lower performance in market-based systems. Therefore, although our primary purpose is to examine database effects, our cross-national sample also allows us to examine cross-country differences, an essential and fundamental pillar of IB research. Our data sources and the multi-country nature of our sample required us to deviate slightly from O'Brien (2003) analysis; we did not have data on two control variables, advertising expenditures and tangible assets, and thus these two variables were not included in our analysis. O'Brien (2003), whose data source was Compustat (not Compustat Global), created a variable called Innovation that was based on the weighted average of percentile scores of R&D investment at the business segment level. Because R&D data were not available at the business segment level in the databases we compared, our innovation variable was calculated at the firm level, with percentile scores calculated by industry within country. Unavailability of some variables and lack of business segment level data highlight limitations on the availability of data from Compustat Global, Osiris and Worldscope. Our variables mimic those in O'Brien (2003). Our dependent variable of firm performance is Market:Book, measured as the market value of the firm divided by the book value of total assets, in which market value of the firm is the book value of debt plus the market value of equity. Then, this quantity was natural log-transformed. Innovation is measured as the percentile score of R&D investment for each firm by industry (two-digit SIC) within country. This differs slightly from O'Brien (2003) in that we measure at the firm rather than at the business segment level and we use two-digit SIC codes, which we had available for all three databases. Leverage is measured as long-term debt divided by the market value of the firm. Then, 0.1 was added to this quantity for zero Leverage values per O'Brien (2003) and was natural log-transformed. Consistent with O'Brien (2003), our control variables are the following. Size is measured as the natural logarithm of total assets. ROA is measured as return on assets— O'Brien (2003) calls this variable “profitability”. Capital Intensity is measured as total assets divided by total revenues. R&D Intensity is measured as R&D expenses divided by revenues. At the industry level, again consistent with O'Brien (2003), Industry M:B is

200

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206 Table 5 Sample size in multi-variate analysis. Country sample cut-off +1 =+20 =+30 =+50 =+100 =+200

Osiris

Compustat

Worldscope

50 34 30 27 17 9

54 32 31 26 15 12

59 41 35 30 22 13

measured as average Market:Book by industry within country. Industry-ROA is measured as average ROA by industry within country. We also include year and industry fixed effects and country fixed effects when the sample is a multi-country sample. Consistent with O'Brien (2003), we estimate using OLS with a lagged dependent variable included as an independent variable. 4.3.2. Regression results for replication study Tables 6a, 6b and 6c show the results of our replication study. Table 6a shows results based on a sample of all countries, which is based on the 66 countries included for representation that we discussed above. Results are also shown for individual samples of the three developed countries with the largest economies as measured by nominal growth domestic product: Germany, Japan and the United Kingdom (UK). Table 6a (and Tables 6b and 6c) shows the results from each database side-by-side within each geographic group. To keep it simple, we focus the discussion of our results exclusively on the interaction term Innovation*Leverage, while excluding comparative discussion of control and other main effect variables. In Table 6a, for the all-country sample, the interaction term Innovation*Leverage is negative and statistically significant at commonly accepted levels in each database. Therefore, without any country nuance, which means lumping all countries together, we observe results consistent with O'Brien (2003) across databases. Next, we observe results for individual developed countries. For Germany, Japan and the UK, the interaction term Innovation*Leverage is also negative and statistically significant except for the Worldscope UK sample, in which the interaction term is negative but not statistically significant. For these three developed countries, there appears to be database consistency across databases. However, we lose this consistency across databases when we observe results for samples of developing countries. Table 6b shows results for “BRIC” countries—Brazil, Russia, India and China. For Brazil, we observe only one statistically significant result for the key interaction term Innovation*Leverage, which occurs using Compustat Global. For Russia, no database shows a statistically significant result. For India, Compustat Global and Osiris show a negative but not statistically significant result while Worldscope shows a negative and statistically significant result. For both Brazil and India, researchers would likely come to a different conclusion about support for the hypothesis depending on database used, and thus we observe a database effect for Brazil and India. For China, the key interaction term Innovation*Leverage is negative and not statistically significant for Compustat Global but positive and statistically significant for Osiris and Worldscope, which suggest a database effect for China, too. And quite interestingly, two of the three databases show a statistically significant result that is the direct opposite of O'Brien (2003) third hypothesis applied to China. This contrasting result challenges the ubiquity of the innovation-leverage-performance relationship, suggests boundary conditions on its application, and prompts additional theory to explain geographic-based differences. Table 6c shows the replication results for the four regions we used in our general assessment of database effects—Western Europe, Latin America, Eastern Europe and East Asia. For Western Europe, results are consistent across databases, with negative and statistically significant coefficients for the interaction term Innovation*Leverage using Compustat Global, Osiris and Worldscope. For Latin America, we observe a coefficient that is negative and statistically significant using Compustat, negative and not statistically significant using Osiris, and positive and not statistically significant using Worldscope. Therefore, for Latin America, a researcher very likely would come to a different conclusion about support for the hypothesis depending on database used. For Eastern Europe, all interaction term coefficients are positive, and only using Compustat Global, the coefficient is statistically significant. Therefore, for Eastern Europe, we observe a database effect. We also observe statistically significant support using Compustat Global for the opposite relationship of the tested hypothesis. Just as we found in China, O'Brien (2003) hypothesized result does not play out in Eastern Europe. For East Asia, all interaction term coefficients are negative and statistically significant across the databases. Taken together, our replication results show less consistency across databases for developing than for developed countries. We observed this contrast in consistency when we sampled individual countries and when we sampled geographic regions. In other words, we found more of a database effect in Table 6b results (Brazil, Russia, India, China) than in Table 6a results (Germany, Japan, UK). Further, we found more of a database effect for Latin America and Eastern Europe than for Western Europe and East Asia. Finally, for China and Eastern Europe we found results that were opposite to what was hypothesized. Instead of the combination of high innovation and high leverage having a negative performance effect (as was hypothesized), we found results showing a statistically significant positive performance effect of high innovation combined with high leverage in Eastern Europe using one databases (Compustat Global) and in China using two databases (Osiris and Worldscope). 5. Discussion and conclusion Our objective in this study was to assess if and how the choice of a particular database would lead to inconsistent empirical results. We engaged in a relatively comprehensive assessment of the three most-widely used databases in IB research: Compustat

Table 6a Replication regression results for an all-country and developed-country samples. All countries

Germany

Japan

United Kingdom

(O1)

(W1)

(C2)

(O2)

(W2)

(C3)

(O3)

(W3)

(C4)

(O4)

(W4)

Compu

Osiris

World

Compu

Osiris

World

Compu

Osiris

World

Compu

Osiris

World

Capital intensity

0.33⁎⁎⁎ (0.00) −0.00 (0.00) −0.00⁎⁎⁎ (0.00) −0.00⁎⁎⁎

0.55⁎⁎⁎ (0.00) 0.01⁎⁎⁎ (0.00) 0.00⁎⁎⁎ (0.00) −0.00⁎⁎⁎

0.58⁎⁎⁎ (0.00) −0.00⁎⁎ (0.00) 0.00⁎⁎⁎ (0.00) −0.00⁎⁎

0.55⁎⁎⁎ (0.02) −0.01 (0.01) −0.12⁎⁎⁎ (0.02) −0.00⁎⁎

0.58⁎⁎⁎ (0.02) −0.03⁎⁎⁎ (0.01) −0.05 (0.05) −0.00⁎

0.65⁎⁎⁎ (0.01) 0.02⁎⁎⁎ (0.00) 0.08⁎⁎⁎ (0.03) −0.00⁎⁎

(0.00) 0.00⁎⁎⁎

(0.00) 0.00 (0.00) 0.36⁎⁎⁎ (0.01) 0.00⁎ (0.00) −0.14⁎⁎⁎

(0.00) 0.03⁎⁎ (0.01) 0.02 (0.44) −1.70 (1.60) −0.11⁎⁎⁎

(0.00) 0.20⁎⁎⁎

(0.00) 0.15⁎⁎⁎

(0.05) 0.08 (0.43) 5.35 (3.80) 0.10⁎⁎⁎

(0.04) −2.15 (2.23) −15.50 (14.54) −0.07⁎⁎⁎

(0.04) 2.72⁎⁎ (1.09) −2.22⁎⁎ (1.05) −0.05⁎⁎⁎

0.55⁎⁎⁎ (0.02) 0.02⁎⁎ (0.01) −0.15⁎⁎⁎ (0.02) −0.00 (0.00) 0.00 (0.00) −2.09 (1.53) −4.92 (3.72) −0.16⁎⁎⁎

Constant

(0.01) 0.00 (0.02) −0.03⁎⁎ (0.01) −0.47⁎⁎⁎

(0.01) −0.00 (0.02) −0.05⁎⁎⁎ (0.01) −0.68⁎⁎⁎ (0.16) Yes Yes Yes 45,827 0.57

(0.30) Yes Yes Yes 1,402 0.53

(0.08) Yes Yes Yes 8,741 0.70

(0.73) Yes Yes Yes 2,689 0.46

(0.02) −0.24⁎⁎ (0.10) −0.27⁎⁎⁎ (0.06) −0.55 (0.41) Yes Yes Yes 2,106 0.47

(0.03) 0.04 (0.11) −0.09 (0.06) −0.81⁎⁎

(0.11) Yes Yes Yes 35,580 0.56

(0.01) −0.02 (0.03) −0.05⁎⁎⁎ (0.02) −1.66 (1.07) Yes Yes Yes 9,322 0.68

(0.01) −0.02 (0.03) −0.06⁎⁎⁎ (0.02) −0.32⁎⁎⁎

(0.14 Yes Yes Yes 39,571 0.49

(0.03) −0.04 (0.09) −0.09⁎ (0.05) −0.37 (0.30) Yes Yes Yes 1,404 0.59

(0.01) −0.14⁎⁎⁎ (0.03) −0.11⁎⁎⁎ (0.02) −7.99⁎⁎⁎

Industry dummies Country dummies Year dummies Observations R-squared

(0.03) −0.32⁎⁎⁎ (0.09) −0.30⁎⁎⁎ (0.05) 0.36 (0.43) Yes Yes Yes 1,191 0.59

(0.00) 1.37 (2.44) 11.44 (19.51) −0.00 (0.02) −0.06 (0.09) −0.17⁎⁎⁎ (0.05) −6.72⁎⁎⁎

0.54⁎⁎⁎ (0.02) 0.04⁎⁎⁎ (0.01) −0.09⁎⁎⁎ (0.03) 0.00 (0.00) 0.00 (0.00) 0.18 (0.96) −0.03 (0.24) −0.13⁎⁎⁎

(0.01) −0.02 (0.02) −0.09⁎⁎⁎ (0.01) −4.66⁎⁎⁎

(0.01) 0.37 (0.62) 0.71 (1.10) −0.05⁎ (0.03) −0.07 (0.10) −0.10⁎ (0.05) −5.59⁎⁎⁎

(0.00) −0.01 (0.01) −0.79 (1.62) 1.29 (1.39) −0.10⁎⁎⁎

0.51⁎⁎⁎ (0.02) 0.02⁎⁎ (0.01) −0.08⁎⁎⁎ (0.02) −0.00 (0.00) −0.00⁎⁎

(0.00) 0.55⁎⁎⁎ (0.01) 0.00⁎⁎ (0.00) −0.12⁎⁎⁎

(0.00) −0.00 (0.00) 0.37⁎⁎⁎ (0.01) −0.00 (0.00) −0.18⁎⁎⁎

0.56⁎⁎⁎ (0.01) 0.03⁎⁎⁎ (0.00) −0.07⁎⁎⁎ (0.03) −0.00 (0.00) 0.16⁎⁎⁎

0.64⁎⁎⁎ (0.01) 0.02⁎⁎⁎ (0.00) 0.09⁎⁎⁎ (0.03) −0.01⁎⁎⁎

R&D intensity

0.41⁎⁎⁎ (0.02) −0.02⁎⁎ (0.01) −0.05 (0.06) −0.00 (0.00) 0.03⁎⁎⁎

Lagged dep. var. Size ROA

Industry M:B Industry ROA Leverage Innovation Innovation*Leverage

(0.19) Yes Yes Yes 7,940 0.64

(0.39) Yes Yes Yes 2,152 0.50

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

(C1)

Notes: “Compu”, “Osiris” and “World” refer to the databases of Compustat Global, Osiris and Worldscope, respectively. Likewise, Columns C1, O1, and W1 refer to a column “1” of Compustat Global, Osiris and Worldscope, respectively. C1, O1 and W1 sample firms from all countries (i.e., 66 countries). C2, O2 and W2 sample firms from Japan. C3, O3 and W3 sample firms from Germany. C4, O4 and W4 sample firms from the United Kingdom. Standard errors in parentheses. ⁎⁎⁎ p b 0.01. ⁎⁎ p b 0.05. ⁎ p b 0.1.

201

202

Table 6b Replication regression results for developing-country samples—“BRIC” countries. Brazil

Size ROA Capital intensity R&D intensity Industry M:B Industry ROA Leverage

(O1)

(W1)

(C2)

(O2)

(W2)

(C3)

(O3)

(W3)

(C4)

(O4)

(W4)

Compu

Osiris

World

Compu

Osiris

World

Compu

Osiris

World

Compu

Osiris

World

0.25⁎⁎⁎ (0.03) 0.01 (0.05) −0.22⁎⁎⁎ (0.04) 0.00⁎⁎⁎

0.36⁎⁎⁎ (0.05) 0.15⁎⁎⁎

0.73⁎⁎⁎ (0.03) −0.02 (0.03) −0.10⁎⁎⁎ (0.02) 0.00 (0.000) −0.36 (6.505) 0.44⁎⁎⁎ (0.14) −2.63 (2.32) −0.04 (0.06) −0.09 (0.36) −0.14 (0.33) 0.41 (0.31) Yes Yes Yes 512 0.70

0.11 (0.07) 0.08 (0.10) −0.33 (1.08) −0.24⁎ (0.137) 434.97 (264.31) 1.39 (1.14) −20.19 (22.50) −0.00 (0.20) −0.37 (0.70) −0.28 (0.47) 0.08 (1.39) Yes Yes Yes 45 0.85

0.21⁎⁎⁎ (0.06) 0.08⁎⁎

0.41⁎⁎⁎ (0.05) 0.04 (0.03) 0.96⁎⁎ (0.42) −0.00 (0.000) 21.44 (36.37) 0.33 (0.74) −8.30 (22.23) −0.09 (0.07) −0.52 (0.35) −0.28 (0.23) 0.80 (0.52) Yes Yes Yes 331 0.51

0.33⁎⁎⁎ (0.01) 0.06⁎⁎⁎

0.50⁎⁎⁎ (0.02) −0.02 (0.01) −0.04⁎⁎⁎ (0.01) 0.00 (0.000) 0.59 (0.72) 2.06 (2.03) 12.96 (16.39) −0.35⁎⁎⁎

0.52⁎⁎⁎ (0.01) 0.02⁎⁎⁎

0.52⁎⁎⁎ (0.01) −0.18⁎⁎⁎

0.62⁎⁎⁎ (0.02) −0.13⁎⁎⁎

(0.01) −0.13⁎⁎ (0.06) 0.00 (0.000) −0.12 (0.34) −0.13 (0.412) −0.13 (0.18) −0.25⁎⁎⁎

(0.01) −0.00⁎⁎⁎ (0.000 0.00⁎⁎⁎

(0.01) 0.00⁎⁎⁎ (0.00) −0.00⁎

(0.000) −4.62⁎

(0.000) 0.13 (0.74) 0.52 (0.35) −1.15 (0.72) −0.26⁎⁎⁎

0.72⁎⁎⁎ (0.01) −0.12⁎⁎⁎ (0.01) 0.00⁎⁎⁎ (0.00) −0.00 (0.000) −0.54 (0.48) 0.38 (24.55) 0.00 (2679.14) −0.16⁎⁎⁎

(0.000) −2.99 (9.38) 0.51 (0.63) 0.59 (1.22) −0.87⁎⁎⁎

Constant Industry dummies Country dummies Year dummies Observations R-squared

(1.08) Yes Yes Yes 481 0.57

Innovation*Leverage

China

(C1)

(0.12) −0.85⁎ (0.45) −1.04⁎⁎⁎ (0.32) −4.66⁎⁎⁎

Innovation

India

(0.05) −0.19 (0.28) −0.00 (0.007) 9.12 (9.60) 2.85 (3.20) −0.14 (0.21) −0.56⁎⁎⁎ (0.10) −0.87 (0.63) −0.46 (0.83) −3.12 (1.91) Yes Yes Yes 241 0.55

(0.04) 1.56⁎⁎⁎ (0.58) −0.00⁎⁎ (0.002) 28.24 (25.22) 2.08⁎⁎ (0.85) −9.00 (20.98) −0.21⁎⁎⁎ (0.08) −0.47 (0.34) −0.35 (0.21) −0.29 (1.00) Yes Yes Yes 193 0.65

(0.01) 1.88⁎⁎⁎ (0.16) −0.00 (0.000) 0.02 (0.03) 0.58 (0.42) −11.60 (15.004) −0.12⁎⁎⁎ (0.02) −0.02 (0.06) −0.03 (0.04) −4.62⁎⁎⁎ (0.67) Yes Yes Yes 2,152 0.56

(0.03) −0.06 (0.09) −0.01 (0.06) −0.89 (0.74) Yes Yes Yes 1,278 0.58

(0.02) −0.10⁎ (0.06) −0.12⁎⁎⁎ (0.04) −0.73⁎⁎⁎ (0.27) Yes Yes Yes 3,588 0.52

Notes: “BRIC” refers to Brazil, Russia, India and China. “Compu”, “Osiris” and “World” refer to the databases of Compustat Global, Osiris and Worldscope, respectively. Likewise, Columns C1, O1, and W1 refer to a column “1” of Compustat Global, Osiris and Worldscope, respectively. C1, O1 and W1 sample firms from Brazil. C2, O2 and W2 sample firms from Russia. C3, O3 and W3 sample firms from India. C4, O4 and W4 sample firms from China. Standard errors in parentheses. ⁎⁎⁎ p b 0.01. ⁎⁎ p b 0.05. ⁎ p b 0.1.

(2.762) −2.07 (471.82) −0.05 −0.09⁎⁎⁎ (0.02) −0.19 (0.24) −0.04 (0.12) −5.31 (223.00) Yes Yes Yes 3,226 0.69

(0.02) 0.12 (0.09) 0.31⁎⁎⁎ (0.06) 0.59⁎⁎⁎ (0.17) Yes Yes Yes 2,797 0.68

(0.02) 0.06 (0.09) 0.13⁎⁎⁎ (0.05) 0.90 (169.23) Yes Yes Yes 3,938 0.73

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

Lagged dep. var.

Russia

Table 6c Replication regression results for geographic region samples. Western Europe

Latin America

Eastern Europe

East Asia

(O1)

(W1)

(C2)

(O2)

(W2)

(C3)

(O3)

(W3)

(C4)

(O4)

(W4)

Compu

Osiris

World

Compu

Osiris

World

Compu

Osiris

World

Compu

Osiris

World

0.53⁎⁎⁎ (0.01) 0.01⁎⁎⁎ (0.00) −0.09⁎⁎⁎ (0.01) −0.00 (0.00) −0.00 (0.00) 0.40⁎⁎⁎ (0.03) 0.00 (0.00) −0.14⁎⁎⁎

0.57⁎⁎⁎ (0.01) 0.00 (0.00) −0.12⁎⁎⁎ (0.02) −0.00⁎

0.30⁎⁎⁎ (0.02) 0.01 (0.03) −0.19⁎⁎⁎ (0.023 0.00⁎⁎⁎

−0.00 (0.01) −0.05⁎⁎⁎ (0.02) 1.50⁎⁎⁎ (0.19) 0.00 (0.01) 0.16 (0.10) 0.96⁎⁎⁎ (0.12) −2.14⁎⁎ (0.92) −0.20⁎⁎⁎

0.34⁎⁎⁎ (0.03) −0.01 (0.02) 0.18 (0.26) −0.00 (0.00) −8.49 (8.16) 0.53⁎⁎⁎ (0.09) −0.67 (0.76) −0.23⁎⁎⁎

0.47⁎⁎⁎ (0.02) −0.01 (0.01) 0.65⁎⁎⁎ (0.13) −0.00 (0.00) 1.27 (0.96) 0.53⁎⁎⁎ (0.08) −0.68 (0.44) −0.09⁎⁎⁎

0.59⁎⁎⁎ (0.01) 0.02⁎⁎⁎ (0.00) −0.02⁎⁎⁎ (0.00) 0.00⁎⁎

0.59⁎⁎⁎ (0.01) 0.02⁎⁎⁎ (0.00) 0.05⁎⁎⁎ (0.01) 0.00⁎⁎⁎

(0.00) −2.57 (6.44) 0.36⁎⁎⁎ (0.09) −0.19 (0.29) −0.46⁎⁎⁎

0.72⁎⁎⁎ (0.02) 0.01 (0.01) −0.08⁎⁎⁎ (0.02) −0.00 (0.00) 1.47 (3.96) 0.30⁎⁎⁎ (0.05) 0.00 (0.00) −0.07⁎⁎

0.48⁎⁎⁎ (0.01) 0.02⁎⁎⁎ (0.00) 0.01⁎⁎⁎ (0.00) −0.00⁎⁎⁎

(0.00) 0.00 (0.00) 0.36⁎⁎⁎ (0.024) 0.09⁎⁎⁎ (0.02) −0.12⁎⁎⁎

0.45⁎⁎⁎ (0.03) 0.09⁎⁎⁎ (0.02) 0.02 (0.03) 0.00 (0.00) 6.32 (5.11) 0.29⁎⁎⁎ (0.08) 0.00 (0.00) −0.36⁎⁎⁎

(0.00) 0.01 (0.01) 0.39⁎⁎⁎ (0.02) 0.02 (0.02) −0.11⁎⁎⁎

(0.01) −0.08⁎⁎ (0.04) −0.12⁎⁎⁎ (0.02) −0.22⁎

(0.06) −1.08⁎⁎⁎ (0.28) −1.17⁎⁎⁎ (0.21) −4.88⁎⁎⁎

(0.04) −0.61 (0.39) −0.30 (0.38) −1.35⁎⁎

(0.13) Yes Yes Yes 8,904 0.59

(0.77) Yes Yes Yes 1,094 0.53

(0.68) Yes Yes Yes 818 0.59

(0.043-) 0.24 (0.18) 0.21⁎⁎ (0.11) −0.20 (0.29) Yes Yes Yes 770 0.49

(0.04) 0.13 (0.15) 0.10 (0.10) 0.25 (0.58) Yes Yes Yes 689 0.57

(0.03) 0.04 (0.1) 0.04 (0.07) 0.06 (0.20) Yes Yes Yes 1,442 0.55

(0.01) −0.00 (0.03) −0.06⁎⁎⁎ (0.02) −0.35⁎⁎⁎

(0.17) Yes Yes Yes 9,358 0.52

(0.03) −0.06 (0.254) 0.01 (0.19) −0.56 (0.35) Yes Yes Yes 1,474 0.70

(0.01) −0.03 (0.03) −0.05⁎⁎⁎ (0.02) −0.37⁎⁎

Industry dummies Country dummies Year dummies Observations R-squared

(0.01) −0.15⁎⁎⁎ (0.04) −0.18⁎⁎⁎ (0.02) −0.27 (0.26) Yes Yes Yes 8,601 0.55

(0.00) 0.30⁎⁎⁎ (0.05) 0.50⁎⁎⁎ (0.03) −0.02 (0.02) −0.00 (0.01) −0.08⁎⁎ (0.03) −0.09⁎⁎⁎ (0.02) −6.75⁎⁎⁎

(0.00) 0.16⁎⁎⁎ (0.05) 0.40⁎⁎⁎ (0.02) 0.01 (0.02) −0.12⁎⁎⁎

Constant

0.45⁎⁎⁎ (0.01) 0.01⁎⁎⁎ (0.00) −0.07⁎⁎⁎ (0.02) −0.00 (0.00) −0.00 (0.00) 0.41⁎⁎⁎ (0.03) 0.04 (0.04) −0.02 (0.01) −0.14⁎⁎⁎ (0.04) −0.17⁎⁎⁎ (0.02) −6.30⁎⁎⁎

(0.22) Yes Yes Yes 17,036 0.52

(0.18) Yes Yes Yes 17,065 0.57

(0.07) Yes Yes Yes 19,609 0.58

Lagged dep. var. Size ROA Capital intensity R&D intensity Industry M:B Industry ROA Leverage Innovation Innovation*Leverage

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

(C1)

Notes: “Compu”, “Osiris” and “World” refer to the databases of Compustat Global, Osiris and Worldscope, respectively. Likewise, Columns C1, O1, and W1 refer to a column “1” of Compustat Global, Osiris and Worldscope, respectively. C1, O1 and W1 sample firms from Western Europe. C2, O2 and W2 sample firms from Latin America. C3, O3 and W3 sample firms from Eastern Europe. C4, O4 and W4 sample firms from East Asia. Standard errors in parentheses. ⁎⁎⁎ p b 0.01. ⁎⁎ p b 0.05. ⁎ p b 0.1.

203

204

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

Global, Osiris and Worldscope. Our comparative analysis was driven by three research questions. First, are the data provided by each database comparable? Second, are there systematic differences in each database's coverage of geographic regions, such as the size and characteristics of country-level samples? Third, how does the choice of database affect empirical results? We found answers to these research questions, and we also learned other things important to understanding how database choice may lead to spurious or even contradictory conclusions. Broadly, we found that the choice of database influences empirical results much more for developing than for developed countries, with some minor caveats. Yet, overall, the results obtained from both our replication study and from our general assessment of database effects suggest that more danger lurks in drawing conclusions from results obtained by researchers using one of these databases for regional and or individual country samples in which the countries are relatively less developed. Examples include the region of Latin America or the country of India. The danger is being too confident in results appearing to support a particular hypothesis or too dismissive of results appearing to show no support or even contrasting support for said hypothesis. Our study also has the normative implication that researchers should place more confidence in results sampling developed countries but less when investigating firms in developing countries. It would seem that depending on the particular country or region being sampled, in addition to standard confidence intervals for coefficients based on properties of the sample itself, an additional confidence interval based on the particular country sample should also be considered. This thought, of course, is an abstraction from the statistical principals that result in confidence intervals. But, given the increasing amount of research focused on developing country firms, it is critically important to be aware of database effects that are more likely to be driving results for developing countries relative to developed countries. Being too confident or too dismissive of results obtained from a particular database also applies to the methodology of the study. We found that methodology matters: database differences arise in empirical results across estimation techniques, especially when using multi-level models. Also, it is important to adequately screen for outliers where, even after winsorizing data, OLS and multi-level or nested models reflected the impact of extreme values. Moreover, considering the size of the country-level samples provided as differences in the country coverage of a database, notwithstanding the prevalence of missing data, the sheer number of countries on which a sufficient sample is available varies across database. Given the growing prominence of these methodologies, database choice may become an increasingly critical issue. In our replication study, not only did we find differences across databases for developing countries, we also found statistically significant results that directly contrasted with O'Brien (2003) hypothesis. For China, in particular, two out of the three databases showed a statistically significant result with a sign opposite to that hypothesized by O'Brien (2003). This finding is remarkable but understandable. The governance role of debt, particularly for investments such as R&D, is subject to a country's institutional context. The typical managerial constraints of debt which limit flexibility to effectively govern R&D investment often are lacking in less developed countries due to a less developed legal and regulatory framework which can limit creditors' access to information and hamper monitoring of the large shareholdings typical of developing markets (Dharwadkar et al., 2000). In state-owned firms in less developed countries, the constraints and associated discipline that debt puts on managers may be lacking even further, since the state often acts as a guarantor of additional financing if the firm becomes financially distressed (Lu et al., 2005). In China, banks themselves are often state-owned (Firth et al., 2008), thus exacerbating any lack of managerial discipline from debt in a state-owned firm. Le and O'Brien (2010), who examined the combination of state ownership and debt levels in China, note that weak bankruptcy laws diminish the influence that banks have over managers of borrowing firms. These institutional characteristics of developing countries and of state-owned firms help explain why we found geographic locations where the “innovation-leverage-performance” relationship is entirely consistent with and directly contrasting with what O'Brien (2003) hypothesized and found using a US-only sample. We also found that database choice appears to matter more for unmatched rather than matched samples, which reflects research practice; researchers typically do not limit firm-level observations to those included in multiple databases, so researchers effectively use unmatched samples. This finding alone highlights the biases inherent in selecting a particular database because it is readily available or not prohibitively costly to obtain access to its data. Other findings are also noteworthy. We found greater differences in both sample size and empirical results using our market-based measure of firm performance. Data smoothing and statistical techniques that address the impact of outliers reduce database effects. Although we found no instances of “severe discrepancies”, researchers would likely have come to different conclusions regarding empirical results depending upon the database used. In addition, not only did we find geographic coverage differences among databases, we also observed differences in how some variables affected firm performance across different geographic regions. Though our study is largely descriptive, a normative implication is that researchers must consider data availability for specific variables and regions. These differences in geographic coverage have implications for research studies that sample firms from multiple countries as well as studies that focus on specific regions or specific countries; underrepresentation in particular countries and or regions biases results in a multi-country or an individual-country study. The extent to which that bias occurs is dependent on numerous factors such as researcher choices of country, industry and methodology, as we have shown in this study. Indeed, in a study comparing alliance databases, Schilling emphasized that because alliance databases only report a sample of actual alliances, “it would be risky to rely on any single database” (2009: 257). Several other implications arise from our study. First, for many commonly reported variables, researchers may be able to make use of multiple databases to replace observations missing from one data source. Before doing so, however, researchers should compare reported values and variable definitions for consistency. Second, results from the unmatched samples suggest that sampling issues should be considered. Database choice is most critical when researchers wish to focus on specific countries, regions or variables. If researchers have access to multiple databases, they should examine the data truly available in each data source, not

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

205

just “advertised” to be available, to select the most appropriate resource(s) for their research. Although more subject to missing values, Osiris coverage of smaller firms may be an advantage for some researchers. Third, variable choice matters. Several variables, specifically debt and R&D, are more sensitive to database choice than other variables. Further, results for regressions predicting market-to-book are more sensitive to database choice (and have greater instances of missing data) than those predicting ROA. Although extrapolating from these differences must necessarily be speculative, choice of database may have more significant implications for research that incorporates less commonly reported variables. Fourth, database choice may significantly influence the data items available to researchers and therefore the range of research topics they are able to address. For example, Osiris does not provide usable data on capital expenditures, but it is the only source of data on employees. Finally, differences in database construction must also be considered. As noted earlier, Worldscope generally provided more complete firm-level data. Deletion of firms no longer included in the database may have advantages for panel data and use of lagged or leading variables. However, researchers should note possible survivor bias in Worldscope. The increase in the number of firms covered by both Worldscope and Osiris may pose a challenge to researchers conducting longitudinal analyses or wishing to use lagged or lead variables as firms are added or dropped from the sample. From a more general perspective, our study raises several issues regarding the development of IB research as the questions a researcher is able to address depend upon the data available. Researchers without access to specialized databases such as SDC Spectrum (mergers and acquisitions data), the Center for Research in Security Prices or “CRSP” (equity market data), audit analytics (audit data), the Institutional Brokers' Estimate System, the Bureau van Dijk Zephyr database on venture capital, or “IBES” (analyst forecasts) would be constrained in their choice of research topics. For example, international research on mergers and acquisitions, such as Lai et al. (2012), overwhelmingly rely upon the SDC Spectrum database. Similarly, research on venture capital, a topic of growing international interest, is highly dependent upon either hand collection of data of secondary data such as the Zephyr database (Tykvová and Schertler, 2014). Other secondary databases focus on specific industries such as the Thomson Reuters Recap database on alliance activity in biotech (Rogbeer et al., 2014). Even without access to such specialized databases, methodological issues may pose serious obstacles. However, that only twelve variables were available on the three databases providing comprehensive data was surprising and suggests that database availability may have significant implications for the questions a researcher can address. It also suggests that replication of prior research using different data sources may be virtually impossible. Although similar issues arise in North American research, there is significantly greater overlap in firm coverage and variable availability. For example, most commonly used financial data are available through Compustat North America, Osiris and other commonly used data sources. Ownership, executive compensation, and board of director data are available from several sources (e.g. Standard and Poor's Execucomp, Edgar (SEC), Osiris, Board Analyst, and Risk Metrics/Institutional Shareholder Services (ISS)). We did not examine other country- or region- specific databases such as the China Stock Market and Accounting Research (CSMAR) Database or the Nikkei Economic Electronic Database System (NEEDS), which provides financial data for Japanese firms. Country-specific databases may make use of unique data sources making use of differing standardization procedures which may lead to additional empirical differences. Indeed, database discrepancy (i.e., inconsistent results across databases) may be more acute when multiple country-specific data sources are used. We also did not assess the consistency between primary data (e.g., firm regulatory filings) and the data reported in these secondary sources. Differences in the results obtained may imply differences in the guidance offered to practitioners. In interpreting the practical implications of research findings, managers may need to consider differences in data coverage. For example, research using Osiris may provide better guidance to managers in Latin America, Japan, and the Middle East due to its greater sample size. Our replication analysis suggests that managers may receive differing advice regarding the use of debt and equity to fund research and development depending on the country. References Abarbanell, J.S., Lehavy, R., 2000. Differences in commercial database reported earnings: Implications for empirical research. Available at Social Science Research Network: http://ssrn.com/abstract=228918. Anderson, R., Lee, S., 1997. Ownership studies: the data source does matter. J. Financ. Quant. Anal. 32 (2), 311–329. Arnegger, M., Hofmann, C., Pull, K., Vetter, K., 2014. Firm size and board diversity. J. Manag. Gov. 18 (4), 1109–1135. Bah, R., Dumontier, P., 2001. R&D intensity and corporate financial policy: some international evidence. J. Bus. Financ. Account. 28 (5/6), 671–692. Balakrishnan, S., Fox, I., 1993. Asset specificity, firm heterogeneity and captial structure. Strateg. Manag. J. 14 (1), 3–16. Bougheas, S., 2004. Internal vs external financing of R&D. Small Bus. Econ. 22 (1), 11–17. Bureau van Dijk, 2011. Osiris Product User Guide. https://osiris.bvdep.com/version-2011104/cgi/template.dll?tpl=help/helpframe.tpl&head=userguide&page= help&context=6749I. Busta, I., Sinani, E., Thomsen, S., 2014. Ownership concentration and market value of European banks. J. Manag. Gov. 18 (1), 159–183. Criscuolo, P., Narula, R., Verspagen, B., 2005. The role of home and host country innovation systems in R&D internalization: a patent citation analysis. Econ. Innov. New Technol. 14, 417–433. Cui, L., Li, Y., Meyer, K., Li, Z., 2015. Leadership experience meets ownership structure: returnee managers and internationalization of emerging economy firms. Manag. Int. Rev. 55 (3), 355–387. David, P., O'Brien, J.P., Yoshikawa, T., 2008. The implications of debt heterogeneity for R&D investment and firm performance. Acad. Manag. J. 51 (1), 165–181. Demirguc-Kunt, A., Maksimovic, V., 2002. Funding growth in bank-based and market-based financial systems: evidence from firm-level data. J. Financ. Econ. 65 (3), 337–363. Desender, K.A., Aguilera, R.V., Crespi, R., Garcia-cestona, M., 2013. When does ownership matter? Board characteristics and behavior. Strateg. Manag. J. 34 (7), 823–842. Dharwadkar, R., George, G., Brandes, R., 2000. Privatization in emerging economies: an agency theory perspective. Acad. Manag. Rev. 25 (3), 650–669. Elton, E., Gruber, M., 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. J. Financ. 55 (6), 2415–2430. Firth, M., Lin, C., Wong, S.M.L., 2008. Leverage and investment under a state-owned bank lending environment: evidence from China. J. Corp. Financ. 14 (5), 642–653.

206

J.B. McGuire et al. / Journal of International Management 22 (2016) 186–206

Garcia Lara, J., Garcia Osma, B., Albornoz Nogeur, B., 2006. Effects of database choice on international accounting research. Abacus 42 (3-4), 426–454. Gentry, R.J.P., Shen, W.P., 2010. The relationship between accounting and market measures of firm financial performance: how strong is it? J. Manag. Issues 22 (4), 514–530. Hautz, J., Mayer, M.C.J., Stadler, C., 2013. Ownership identity and concentration: a study of their joint impact on corporate diversification. Br. J. Manag. 24 (1), 102–126. Hillier, D., Pindado, J., Queiroz, V.d., Torre, C.d.l., 2010. The impact of country-level corporate governance on research and development. J. Int. Bus. Stud. 42 (1), 76–98. Hillier, D., Pindado, J., Queiroz, V.D., Torre, C.D.L., 2011. The impact of country-level corporate governance on research and development. J. Int. Bus. Stud. 42 (1), 76–98. Hult, G.T.M., Ketchen, D.J., Griffith, D.A., Finnegan, C.A., Gonzalez-padron, T., Harmancioglu, N., Huang, Y., Talay, M.B., Cavusgil, S.T., 2008. Data equivalence in crosscultural international business research: assessment and guidelines. J. Int. Bus. Stud. 39 (6), 1027–1044. Hundley, G., Jacobson, C., Park, S.H., 1996. Effects of profitability and liquidity on R&D intensity: Japanese and U.S. companies compared. Acad. Manag. J. 39 (6), 1659–1674. James, B.E., McGuire, J.B., 2016. Transactional-institutional fit: corporate governance of R&D investment in different institutional contexts. J. Bus. Res. (in press). Jo, H., Kim, H., Park, K., 2015. Corporate environmental responsibility and firm performance in the financial services sector. J. Bus. Ethics 131 (2), 257–284. Jones, S., 2006. Discussion of Garcia Lara, Garcia Osma, and Gill de Albornoz Noguer. Abacus 42 (3-4), 455–560. Ketchen, D.J., Ireland, R.D., Baker, L.T., 2013. The use of archival proxies in strategic management studies: castles made of sand? Organ. Res. Methods 16 (1), 32–42. Kranenburg, H., Hagedoorn, J., Lorenz-Orlean, S., 2014. Distance costs and the degree of inter-partner involvement in international relational-based technology alliances. Glob. Strateg. J. 4 (4), 280–291. Lai, J., Chen, L., Chang, S., 2012. The board mechanism and entry mode choice. J. Int. Manag. 18, 379–392. Le, T.V., O'Brien, J.P., 2010. Can two wrongs make a right? state ownership and debt in a transition economy. J. Manag. Stud. 47 (7), 1297–1316. Lee, B.-S., 2012. Bank-based and market-based financial systems: time-series evidence. Pac. Basin Financ. J. 20 (2), 173–197. Levine, R., 2002. Bank-based or market-based financial systems: which is better? J. Financ. Intermed. 11 (4), 398–428. Lopatta, K., Kaspereit, T., 2014. The world capital markets' perception of sustainability and the impact of the financial crisis. J. Bus. Ethics 122, 475–500. Lourenço, I., Branco, M., Curto, J., Eugénio, T., 2012. How does the market value corporate sustainability performance? J. Bus. Ethics 108, 417–428. Lu, D., Thangavelu, S.M., Hu, Q., 2005. Biased lending and non-performing loans in China's banking sector. J. Dev. Stud. 41 (6), 1071–1091. Merton, R., 1957. Social Theory and Social Structure. The Free Press, New York. Mulcahy, M., Linehan, C., 2014. Females and precarious board positions: further evidence of the glass cliff. Br. J. Manag. 25 (3), 425–438. O'Brien, J.P., 2003. The capital structure implications of pursuing a strategy of innovation. Strateg. Manag. J. 24 (5), 415–431. Peterson, M., Arregle, J.-L., Martin, X., 2012. Multilevel models in international business research. J. Int. Bus. Stud. 43 (5), 451–457. Piaskowska, D., Trojanowski, G., 2014. Twice as smart? the importance of managers' rormative-years' international experience for their international orientation and foreign acquisition decisions. Br. J. Manag. 25 (1), 40–57. Richard, P.J., Devinney, T.M., Yip, G.S., Johnson, G., 2009. Measuring organizational performance: towards methodological best practice. J. Manag. 35 (3), 718–804. Rogbeer, S., Almahendra, R., Ambos, B., 2014. Open-innovation effectiveness: when does the macro design of alliance portfolios matter? J. Int. Manag. 20, 464–477. Sambharya, R., Lee, J., 2014. Renewing dynamic capabilities globally: an empirical study of the world's largest MNCs. Manag. Int. Rev. 54 (2), 137–169. Shiow-Ying, W., Yu, J., 2012. International evidence on relationship banking in mergers and acquisitions. Int. Res. J. Financ. Econ. 89, 32–45. Standard and Poor's, 2011. Standard & Poor’s Compustat Global database. Understanding the Data. Surroca, J., Tribo, J.A., Zahra, S.A., 2013. Stakeholder pressure on MNEs and the transfer of socially irresponsible practices to subsidiaries. Acad. Manag. J. 56 (2), 549–572. Thomson Reuters, 2011. Worldscope Database Datatype Definitions Guide. Tykvová, T., Schertler, A., 2014. Does syndication with local venture capitalists moderate the effects of geographical and institutional distance? J. Int. Manag. 20, 406–420. Ulbricht, N., Weiner, C., 2005. Worldscope meets Compustat: A Comparison of Financial Databases (December 14, 2005). Available at SSRN: http://ssrn.com/abstract= 871169. Yang, D.C., Vasarhelyi, M.A., Liu, C., 2003. A note on the using of accounting databases. Ind. Manag. Data Syst. 103 (3), 204–210.