Abstract
The authors examine how strategy scholars have measured and tested industry effects. They report findings from three studies. First, they replicate the Dess, Ireland, and Hitt (1990) article on industry controls in strategic management research using a new sample of studies published during 2000 to 2009, finding that there has been a decrease in the proportion of articles that do not control for industry effects at all and at the same time noting a significant increase in the number of single-industry studies. Second, they employ a fine-grained content analysis of articles published in the Strategic Management Journal at three different points during the study period to identify the different ways that industry effects have been considered. Findings depict a myriad of highly diverse industry-level measures that researchers have applied. Third, they test the empirical implications of applying different measures of one particular industry characteristic, industry performance. They demonstrate that empirical findings and the interpretation of theoretical models can differ based on how industry effects are incorporated. Recommendations are offered for guiding future research about how to examine industry effects.
Introduction
Industry effects have long played an important role in strategic management research. For example, some researchers have shown that empirical results differ depending on industry membership (e.g., Christensen & Montgomery, 1981), and there has been a long-standing debate considering whether industry effects matter more or less than firm effects for explaining differences in firm performance (McGahan & Porter, 1997; Rumelt, 1991; Schmalensee, 1985). Given the early relevance of industry effects to strategic management theories and topics, such as the industrial economics framework pioneered by Porter (1980), Dess, Ireland, and Hitt (1990) reviewed the body of empirical work in strategic management to document how researchers were addressing the possibility that industry-level effects could influence results and findings. With the primary focus of strategy being the performance of individual firms relative to their peers along with the firm-level decisions and behaviors that lead to those performance outcomes, industry effects, when considered, are often included in empirical studies as statistical controls. In other studies, the characteristics of the industry play a primary theoretical role in predicting and explaining firm performance. In their broad review of the literature, which considered a range of impactful strategy article across numerous contexts, Dess et al. (from this point on referred to simply as DIH) concluded that from 1980 to 1988 there was “a lack of systematic consideration and treatment of potential industry effects, thereby leading to possible alternative explanations for reported results” (p. 7). Indeed, one fourth of the studies published between 1980 and 1988 did not consider industry effects at all. Others only used surrogate environmental measures or merely focused on one single industry, and very few examined multiple industry-level variables in their studies.
Now, more than 20 years have passed since the DIH review was reported. In that time the strategic management literature has exploded in terms of its size, diversity, and depth, while the methodological sophistication and rigor of research practices have simultaneously increased. With these developments, the methods by which strategy researchers measure and test for industry effects have likely changed relative to common practice in the 1980s, yet we know little about this possible evolution and what effects it might have on findings and theory development. If industry does indeed matter for explaining phenomena important to strategic management, then how do researchers capture it in their models? How has the practice of measuring and testing industry effects progressed in the field since the DIH review? Does the method of measuring industry effects matter—does the choice among available approaches have implications for empirical findings and theoretical development?
We consider the development and evolution in the use of industry measures in recent strategy research in three ways. First, we replicate the initial classification of industry proxies developed by DIH using studies from a more recent period, 2000 to 2009. We find a significant decrease in the proportion of articles that do not consider industry effects at all, but also a marked increase in the proportion of articles that rely on data from within a single industry. Second, we content analyze articles published in the Strategic Management Journal in 2000, 2005, and 2009 to more fully characterize the specific types of industry-level characteristics that researchers include in their empirical tests, finding substantial variety. Third, we conduct an empirical demonstration of how different approaches to measuring and testing industry characteristics can impact upon empirical results and their implications for theoretical understanding. Collectively, the findings reveal that approaches to measuring and testing industry effects have improved in some ways in the past 20 years but have also changed in ways that are potentially problematic. We also find that the choice of industry proxy has meaningful implications in that the various proxies available are not perfect substitutes for one another. The study highlights the need for researchers to understand how the methodological underpinnings of their industry measures could influence findings and the conclusions they draw.
Definition of Industry
Before beginning the study of industry effects in strategic management research, it is necessary to specify what we mean by the term industry. For a researcher, concern for the impact of industry on the performance and behavior of firms requires that they must first have an implicit understanding of the boundaries that distinguish entities that fall within the industry from those that exist in the broader environment. The most commonly accepted definition of an industry “is a group of firms producing products that are close substitutes for one another (Porter, 1980; Hitt[, Hoskisson, & Ireland, 2009])” (Forbes & Kirsch, 2011, p. 591). For example, the “beverage industry” would consist of all companies that manufacture, distribute, and sell drinks of all kinds, and the “automobile industry” could be defined as all companies that produce cars and trucks. In either case, the set of companies encompassed by the definition produces products that are reasonably close substitutes in terms of the characteristics of the products and services, the manner in which the products are produced and the services delivered, and the customer needs that are being met.
However, there are reasons to be cautious about this definition of industry. First, there are potential issues of internal consistency regarding other industry-level frameworks prevalent in the field of strategic management. For example, the Five Forces framework for analyzing the potential for abnormal returns in a particular industry specifically defines substitutes as being products from outside a focal industry that satisfy the same basic needs, reserving the term rivals to denote the set of companies that operates within an industry (Porter, 1979, 2008). This lack of consistency in terminology creates a risk of confusion that threatens our ability to build a solid body of theoretically sound knowledge.
Second, the Forbes and Kirsch (2011) definition requires the researcher to make a value judgment regarding what “close substitutes” are. Similarly, in the Porter (1979, 2008) framework, judgment must be used to delineate which firms are direct rivals inside the industry, distinct from substitutes, which are outside the industry. The result is that the population of firms that are determined to be in the industry can vary depending on how broadly a particular researcher interprets the term, leading to potential variance in the evaluation of industry conditions. Coca-Cola could be thought of as operating in the soft-drink industry or the beverage industry, depending on researchers’ judgment of what are close substitutes for Coca-Cola’s products or direct competitors to the Coca-Cola corporation. Both would be technically correct, but the decision between the two has serious implications for how a researcher would characterize the concentration, dynamism, resource intensity, and other industry-level characteristics. If authors are not explicit about their definition, then readers and future researchers who build on their work could encounter difficulties.
Third, problems with definitions can become magnified as dynamic environments and technological advances cause industry boundaries to blur (Bettis & Hitt, 1995; Hamel & Prahalad, 1994), such as when Voice Over IP (VOIP) technologies allow a company like Skype to precipitate an overlap between the previously distinct industries of Internet service providers and telephone companies. Where once a company like Apple may have been classified as operating in distinct industries of mobile communications and personal computers (among others), the march of technology has enabled the production of smartphones that now provide a reasonable substitute for a laptop or even a desktop computer. This eliminates the previously clear distinction between the industries. Here we focus on the implications of these definitional problems for researchers, but clearly they present problems as well for executives who need to identify who the competitors are and from which quarter the next threat to market share and profitability may arise.
Fourth, more complications arise with the emergence of companies producing new product classes for which it may be difficult to define the population of close substitutes (Forbes & Kirsch, 2011) or for diversified companies operating across multiple product classes (Rumelt, 1974; Villalonga, 2004). In the former case, it can be difficult to objectively determine when conditions have advanced sufficiently for a group of companies to have become a distinct industry (Aldrich & Ruef, 2006; Klepper & Graddy, 1990; Low & Abrahamson, 1997). To continue with the Apple example, at some point the technological convergence of mobile communications and personal computing may be said to have served as the genesis for a new “mobile computing” industry. However, there exist no clear guidelines to define at what point the new industry became a relevant construct. In the latter case, there can be issues associated with determining precisely the range of industries in which the diversified firm operates (Lang & Stulz, 1994; Villalonga, 2004) as well as with assessing how the differing conditions across the various industries combine and interact to affect the firm’s overall behavior and performance. Finally, defining diversification using resources rather than industry products provides an alternative view of how firms might create value and explain how firms appear to move laterally when they acquire or partner with firms having dissimilar industrial products but highly related knowledge resources (Chatterjee & Wernerfelt, 1991; Robins & Wiersema, 1995).
Last, some researchers suggest that industries are defined not by the products produced, but by the information and beliefs possessed by organizational leaders. For example, Sampler (1998) uses characteristics of information to offer a new industry definition, specifically that “firms possessing sufficient amounts of critical information for the same market (e.g., customers) define the industry boundary” (p. 349). Abrahamson and Fombrun (1994) build on the work of Porac and Thomas (1990) in suggesting a socially constructed view of industries wherein the boundaries are defined by managers’ perceptions of competitors and symbionts (other organizations that share some mutual benefits). This is consistent with the work of Hambrick and Fredrickson (2001), who argue that from the perspective of the practitioner, a thorough definition of the industry (what they call the “arena”) in which a company operates must go beyond simple product-based categorizations to also include clearly delineated market segments, geographical areas, core technologies, and value-creation strategies. These perceptions are largely unobservable for researchers who attempt to identify the population of organizations within a particular industry. Therefore, researchers tend to rely on more objective classification schemes such as Standard Industrial Classification (SIC) codes, which may or may not accurately reflect the way important decision makers within any given firm see their competition.
Overall, subjectivity exists in the concept of industry, with the definitions and boundaries dependent on the perspective of the particular observer. To examine industry effects in the conduct of strategic research, researchers are first tasked with carefully selecting the population of firms that fall within the boundaries of the industry or industries that are relevant to their studies and then with developing measures to operationalize the relevant characteristics of those industries. Given the diversity of perspective on the basic definition of the underlying construct, it comes as no surprise that there exists a wide variety of measures and empirical approaches used to capture industry effects. DIH characterized common practices for measuring and controlling industry effects in strategic management research 20 years ago, but we know little about how those practices have evolved since. To gain a better understanding of that evolution, in the following section we extend on their work by examining the current state of the art in industry measures and controls in detail. We also offer an empirical demonstration of the effect these choices can have on empirical results and thus on theory development.
Assessment of Practices for Measuring and Testing Industry Effects
To best reflect a comprehensive understanding of the current state of the art in how industry effects are measured and tested, we adopt two different sampling approaches. In the first we replicate the process used by DIH in order to assess how the industry control approaches in the most impactful articles published between 2000 and 2009 compare to those used by the most impactful articles 20 years prior. DIH originally reported how industry was measured based on studies in the 1970s and 1980s, so their study serves as a natural baseline for determining how approaches may have evolved since that time. In the second sample we examine all empirical articles published in the Strategic Management Journal in 2000, 2005, and 2009. With this second sample, we catalog the kinds of industry characteristics that researchers have included in their studies. Using both approaches allows us to not only make an apples-to-apples comparison with the DIH results, but also to go beyond their conclusions and understand the industry measurement and testing practices in the body of strategic management literature as a whole.
Replication of Dess et al. (1990)
The DIH content analysis was motivated by the growing recognition that the best way to understand and predict the performance of firms was to combine the perspectives of strategic management and industrial organization (IO) economics into a more integrated theoretical framework (Bourgeois, 1984; Jemison, 1981; Porter, 1981, 1985). At the time, there was little sense of how widely that call for integration was being heeded, or in what manner. They argued that “potential effects of industry on performance should be measured and incorporated into strategic management research to avoid misleading interpretations” (Dess et al., 1990, p. 7). They intended their review of the literature to serve both as a normative evaluation of how often and how effectively that was being done, as well as a prescriptive guide to help future strategic management researchers improve their approach to industry controls.
Now, 20 years later, it is timely to identify how practices for measuring and testing industry effects have changed. How have researchers answered the call for more careful and explicit treatment of industry effects? We endeavor to answer that question by replicating the DIH study procedure using a review of the contemporary literature. For this purpose, we focused on three highly respected academic journals, the Strategic Management Journal (SMJ), Academy of Management Journal (AMJ), and the Journal of Management (JoM), that publish high-quality work in strategic management and have had a high impact in management and strategic management during the time periods considered by DIH up to the present (Podsakoff, Mackenzie, Bachrach, & Podsakoff, 2005). The timeframe for our study encompasses all articles published between 2000 and 2009, a time span that is similar in length to the 1980-1988 window studied by DIH. Like them, we focus only on the most impactful articles as determined by citation counts. DIH examine a sample of 40 articles published between 1980 and 1988. Clearly the volume of strategy research has increased dramatically between the late 1980s and today. In order for our sample to represent current scholarship at a similar scale we must consider a larger sample than 40 articles. We start by identifying the number of articles published in the SMJ as a rough proxy for research activity in the strategic management field. Between 1980 and 1988, the timeframe from which DIH drew their sample, the SMJ published a total of 348 articles for an average rate of 39 articles per year. Between 2000 and 2009, the timeframe for our sample, SMJ published a total of 708 articles for an average rate of 71 articles per year, representing an increase of 82%. For our sample to represent a similar proportion of the total body of scholarship, a minimum of 73 articles (1.82 × 40) are needed in our sample.
The sample was determined by ranking all articles published in SMJ, AMJ, and JoM between 2000 and 2009 based on citation counts reports in the ISI Web of Science, a popular source for determining impact and citation counts in strategic management research (Bergh, Perry, & Hanke, 2006). We then reduced the sampling frame to just the articles that are strategy related (e.g., address firm performance, firm strategy, or any common strategic topic such as diversification, resources and capabilities, innovation, etc., determined by examining abstracts). Those that were not empirical in nature were discarded. The final sample included the 75 most impactful empirical strategy articles. Of the 75, 45 articles came from the Strategic Management Journal, 29 from the Academy of Management Journal, and 1 from the Journal of Management.
For the sake of comparison, we use the same typology of industry controls as DIH: single-industry studies, where all sample companies operate in the same industry; multiple industry controls, where the authors include more than one variable meant to capture some dimension of industry conditions; quasi-industry controls where the authors used a single variable to capture industry conditions, or a perceptual measure of industry conditions, or had a market-based measure of performance; and articles where there were no industry controls. Table 1 reports representative samples of articles falling in each category, along with a summary of each and details on how industry effects were controlled. 1
Representative Examples of Articles Using Various Types of Industry Controls in the Most Frequently Cited Strategy Research (2000-2009).
Note: Complete table available from the authors on request. SMJ = Strategic Management Journal; AMJ = Academy of Management Journal; JoM = Journal of Management; SIC = Standard Industrial Classification.
Table 2 compares our findings for articles published between 2000 and 2009 to those of DIH for articles published from 1980 and 1988. The contrast is striking. Firmly in the category of improved performance, the number of influential articles that had no industry controls has dropped precipitously from 25% to just 8%. That is a dramatic improvement and evidence that strategic management researchers have more consistently adopted industry factors when testing their theories. Even more remarkable is the increase in the number of single-industry studies, from less than 13% to over 45%. Nearly half of the most impactful empirical articles published in strategy over the previous decade were conducted in a single-industry setting.
Comparison of Industry Control Practices in the Most Impactful Articles, 1980-1988 Versus 2000-2009.
There is one other common method of industry controls in the 2000-2009 sample of articles that is worth noting and is absent from DIH discussion: the use of industry dummy variables. In this practice, a series of dummy variables is created for each industry in the sample, with a particular observation taking the value of 1 for the variable representing that company’s industry and 0 for the others. The dummies are then included in the regressions, with the exception of one that must be omitted to keep the regression equation from being misspecified. The coefficients on each dummy variable then represent the marginal effect of being in that industry on the dependent variable of interest compared to being in the omitted industry. Of the 75 articles in our sample, 21 included some variation of industry dummies.
Review of all Strategic Management Journal Articles in 2000, 2005, and 2009
In the second study, we consider articles appearing in the Strategic Management Journal during 3 years spread across our sampling frame by using an unrestricted content analysis to document all possible approaches to defining, measuring, and testing industry effects. This comparison provides a more fine-grained understanding of recent developments, reduces our reliance on a sample derived from impact, and allows for comparisons of how industry measures and testing practices in the most impactful papers compare to the broader population. This second sample consisted of all empirical SMJ articles in 2000, 2005, and 2009 for a total of 164 articles.
Two authors split the task of coding each article according to the DIH typology and recording the specific industry measures and testing procedures that were used. A subsample of articles was coded by both with 100% agreement. Table 3 shows the year-by-year breakdown of articles according to the DIH categories. Two observations are noteworthy. First, the frequency with which articles fell into each of the categories was relatively similar across all 3 years, which suggests that the use of particular industry effects measures tend to be stable over time. Second, the frequency with which all SMJ articles fell into each category is consistent with what was found in the replication study for the 75 most impactful articles between 2000 and 2009 (29 of which were published in SMJ). We would expect the work published in such a highly regarded outlet to exhibit a standard of quality consistent with the state of the art, and we find evidence of that here.
Comparison of Industry Control Practices in All Strategic Management Journal Articles, 2000, 2005, and 2009.
For those articles that included industry measures in the empirical models (those classified as either multiple control variables or quasi-industry controls in the DIH schema), we then counted the frequency with which each particular measure was implemented. After reviewing all the measures, we categorized them into two basic types: those that measure general industry characteristics such as munificence, dynamism, and rivalry (reported in Table 4) and those that deal directly with the impact of industry-level performance (reported in Table 5). As shown in Table 4, the most frequently used industry characteristic measurements are industry dummies, competition, dynamism, munificence, and resource intensity. It is likely that the specific industry characteristics that can and should be included in models will depend in large part on the specifics of the particular article. For example, the strength of rivalry in an industry might be critically important to a study on the formation and evolution of horizontal alliances but would have less salience in an article on internal organizational communication processes. Because context is such an important element behind the selection of those industry measures, we offer Table 4 as an illustration of the kinds of characteristics being included in various studies and as a reference for researchers to determine what if any might be appropriate for their particular research. Clearly, multiple and diverse approaches have been developed to represent the industry construct, and theoretical frameworks and logic will be necessary to guide the selection of particular variables.
Industry Characteristic Controls in All Strategic Management Journal Articles, 2000, 2005, and 2009.
Note: The sum of frequencies of different industry controls in the articles published each year may exceed 100% because some studies used multiple measures. SIC = Standard Industrial Classification.
Industry Performance Controls in All Strategic Management Journal Articles, 2000, 2005, and 2009.
Note: This table represents the frequency with which the different industry-level performance measures were incorporated relative to the entire population of empirical studies. ROA = return on assets; ROE = return on equity; EPS = earnings per share.
The issue of measuring and testing industry-level performance effects has more general implications. Any multi-industry study that uses firm-level performance as either a dependent or independent variable must contend with the fact that a net profit or return on assets that would be considered exceptionally high in some industries could be considered subpar in others (e.g., average profitability differs across industries). Therefore, in those studies that use firm performance as either a dependent or independent variable and that include observations across multiple industries, it becomes essential to consider industry average level performance. Our content analysis indicates that the prevailing approaches to measuring industry-level performance are either to include the industry average level of the particular performance measure (whether it be return on assets [ROA], return on equity [ROE], profitability, or something else) as an independent variable in the regression equations (Iyengar & Zampelli, 2009; Tanriverdi & Venkatraman, 2005) or to adjust the firm-level performance measure by subtracting from it the industry average value, thereby creating a difference score (Love & Nitin, 2005; Roberts & Dowling, 2002). The choice between these approaches has potential implications, which we explore next.
Empirical Demonstration
Given the reported diversity in approaches for measuring and testing industry effects, we now test if how one chooses to measure and test industry constructs have any meaningful impact on empirical results. We focus on the choice between two very specific approaches: including industry performance as a separate variable in the regression model versus using industry performance to normalize firm performance by calculating a difference score. These two approaches may not be substitutes for one another, as there are potentially important methodological differences between them. In particular, the use of difference scores for measuring and controlling for average industry profitability could be problematic. Studies of difference scores have found them to be less reliable and meaningful than their components (Bergh & Fairbank, 2002; Edwards, 1994a, 1994b). We therefore tested whether the choice between these two alternative approaches to controlling for industry performance have implications for findings and theory. In particular, we compare whether representing average industry performance as a variable all of its own or as a difference score (the performance of a firm in a sample minus average industry performance) yields different empirical results.
We apply these different approaches to a sample of divestitures. There is a developing literature that suggests that while divested units begin their life independent of the parent, they nonetheless are endowed with the parent’s resources, routines, cultures, and mental models (Moschieri, 2011; Semadeni & Cannella, 2011). As a result, divested units could exhibit levels of performance similar to those of their parents, an inheritance effect (some studies demonstrate this relationship for spinout firms; see Agarwal, Echambadi, Franco, & Sarkar, 2004; Franco & Filson, 2006; Phillips, 2002). Divestitures provide an especially appropriate laboratory to consider the effect of different industry performance measures because there are potentially two different industries to include (when the divested unit operates in a different industry than the parent). We drew a sample of divested businesses announced between 1990 and 2003. 2
Table 6 reports the findings from the two approaches for measuring and testing industry average performance using the sample of divested businesses. The dependent variable is divested business return on assets. In the first model, parent and divested unit industry average performance (in this case ROA) are included as variables in the model. In the second, industry performance is included by subtracting the industry average ROA from the relevant firm ROA. Other firm-specific factors that might have an effect on spinoff ROA are also considered in the model, including spinoff firm size, price-to-earnings ratio, debt-to-equity ratio, and the year in which the divestiture took place.
Results of Ordinary Least Squares (OLS) Moderated Regression Analysis for Divested Firm Performance: The Influence of Inheritance and Leadership Origin.
Note: ROA = return on assets. *p < .05. **p < .01.
As shown in Table 6, tests of the two different approaches for including industry-level ROA produced different empirical findings. When using the difference between firm and industry average ROA, the well-documented and theoretically supported relationship between parent and child performance appears to be absent. When industry average ROA is included as a variable on its own instead, that relationship changes and becomes statistically significant. Further, we developed the model through introducing a possibly critical mechanism that strengthens the performance inheritance effect: whether the divested business is led by a former manager of the parent firm. Prior research reports that inheritance is transmitted through employees that leave the parent to lead the offspring entity. How we account for industry performance has a significant impact on those results as well. When we account for industry average performance by including ROA variables in the regression, having a legacy president from the parent has a negative direct effect on divested firm performance but does not modify the parent-spinoff performance relationship. When accounting for industry ROA by normalizing the firm-level performance variables, it appears that having a spinoff president from the parent does in fact make the parent-spinoff performance relationship more positive.
This example demonstrates that the way we measure an industry-level factor can have a significant impact on empirical results and could lead to different conclusions and inferences for theoretical and knowledge development.
Discussion
The field of strategic management is a dynamic pursuit, in terms of how theoretical models change and grow over time and in terms of the empirical techniques that are used to tease out the effects of those models. This article reports the evolution of practice of industry measures, with an emphasis on increasing understanding of how measures and methods have changed over time and what the implications of those changing practices might be.
We find that researchers’ approach to measuring industry effects has changed over the past 20 years. Compared to the findings of DIH, we find that more researchers are incorporating industry measures into their empirics, but simultaneously more are also focusing on single-industry studies. It is complicated to evaluate the legacy of DIH based on this analysis. On one hand, DIH were writing at a time when other authors were already beginning to see the importance of controlling for industry effects. It is not surprising, then, that their article has had a significant impact. It received 174 citations according to the Web of Science database since its publication in 1990 (through 2011), compared to an average of only 52 citations per article for the total population of work published in the Journal of Management in 1990. Clearly the article reached and influenced a large audience compared to its peers, and thus it is reasonable to conclude that at least some of the shift in research protocol has come about because of their work. On the other hand, even though the proportion of articles with no industry controls has been reduced by two thirds, the growing reliance on single-industry studies is potentially problematic. Overall, however, the DIH work likely led to significant improvements in practice by raising awareness of the importance of controlling for industry-level factors in strategic management research and by offering a guide as to how to do so.
We also find that researchers have a broad range of industry-level measures that they might choose to include in their empirical models depending on the specifics of the theory and setting being examined. We find a surprising diversity in approaches to accounting for something as straightforward as industry-level performance, including the choice of performance metric as well as the way in which the industry-level value is incorporated into the models, and we demonstrate how such a seemingly minor choice can have a significant impact on our empirical results and thus our understanding of theoretical phenomenon. Each of these findings carries potential implications for current research.
First, the shift toward single-industry studies has potential implications for the ability to generalize findings to other settings, given that empirical results either supporting or refuting a particular theoretical position may be due at least in part to the unique combination of conditions that exist in that particular industry. For example, several of the single-industry studies in our sample were conducted in the chemicals or banking industries (Ahuja, 2000; Ahuja & Katila, 2001; Ahuja & Lampert, 2001; Christmann, 2000; Chung, Singh, & Lee, 2000; Deephouse, 2000; Richard, 2000; Zajac, Kraatz, & Bresser, 2000), industries that exhibit high economies of scale, high entry barriers, and relatively high levels of concentration. It is possible that those characteristics that affect the dynamics of alliance formation are different for banks than they would be in other industries (Chung et al., 2000) or that the performance effects of environmental management techniques are unique to the chemical industry because of the special impact that the industry can have on the environment and the consequent public and governmental scrutiny (Christmann, 2000).
This pattern is potentially a double-edged sword in that single-industry studies not only allow a tighter focus on firm- or corporate-level effects without concern for industry-level confounds, but also do not provide a direct inference to a larger sample of industries. This development requires that we recognize the potential trade-offs by acknowledging that the boundaries of our understanding may be more restrictive than would otherwise be the case. As Bansal (2005) points out in her study encompassing the Canadian forestry, mining, and oil industries, the meaning and importance of central variables in some models may vary widely across industries, leading her to conclude that “the dependent variable, model, and findings should not be generalized without due consideration of these limitations” (p. 204). Across all single-industry studies published in SMJ in 2000, 2005, and 2009, 42% of the authors explicitly recognized the potential limitations to the generalizability of their findings. In the other 58%, the authors were mute on the issue. Only one study tested generalizability by means of a second study that includes multiple industries (Shaw, Gupta, & Delery, 2000). Warnings about generalizability are only of value if the next article in the stream acknowledges and accounts for the fact that theory that seemed to be supported in a single-industry study may be idiosyncratic to that industry setting. Therefore, researchers need to be more mindful about the approach they take to build on the theory they proposed and tested in these single-industry studies.
Second, in comparison with single-industry studies, studies encompassing multiple industries gain in generalizability but suffer from less precise control of industry effects. Examining our sample of articles that included multiple industry variables reveals a wide variance in the specific variables chosen to proxy for relevant industry conditions. In their study of entry order and entry mode in foreign direct investment into the United States, Chang and Rosenzweig (2001) use a fairly robust set of industry variables, including shipment growth, change of globalization, home market concentration, U.S. market concentration, and the existence of trade barriers. Following Dess and Beard (1984), Subramaniam and Youndt (2005) include industry munificence, dynamism, and complexity in their study of the interrelationship among human capital, organizational capital, social capital, and innovation. Table 4 demonstrates a catalog of the myriad industry-level characteristics and variables that previous researchers have chosen to study and future research can choose to include. The primary issue that arises is the question of selection: Of the nearly infinite dimensions of industry conditions to choose from, which are the most relevant to the theory under consideration? What are the implications of choosing some industry variables for inclusion in preference to others? Would results be different if we used a different set of industry measures? If so, what does that mean for the strength and durability of our theories?
Third, industry dummy variables, as an alternative for representing industry effects in a multi-industry study, seem like an elegant solution. They provide a direct method for accounting for potential differences among industries without the trouble of parsing out which particular characteristics are most likely to be relevant. However, industry dummies also have some limitations. First, there is a concern about their impact on statistical power. When dealing with a sample of companies across many different industries, including a dummy variable for each can quickly lead to problems with the ratio of observations to model variables, where the ratio falls as the number of model variables increases. A second concern pertains to the levels of analysis. Our sample includes examples of articles that include industry dummies at the four-digit SIC level (McWilliams & Siegel, 2000), at the two-digit SIC level (Krishnan, Martin, & Noorderhaven, 2006; Lu & Beamish, 2001), and at the sector level of manufacturing versus service (Lane, Salk, & Lyles, 2001; McGrath, 2001; Park & Luo, 2001; Peng & Luo, 2000). The coarser the dummy variable categories, the more likely there are to be intra-industry or intra-sector differences that are theoretically and empirically relevant but that are not accounted for. Also there is no reason to believe that SIC codes are always the optimal scheme for classifying industries. 3 Despite the fact that the Office of Management and Budget (OMB) discontinued use of SIC codes in favor of the North American Industry Classification System (NAICS) in 1997, our review indicates that researchers still overwhelmingly rely on the more outdated system. This is likely due to the conservative nature of academics. Once a particular way of doing things is established, either theoretically or empirically, it often propagates without significant question or criticism (Lane, Koka, & Pathak, 2006; Vandenberg, 2006). Indeed, researchers can even adopt incomplete views of theories, leading to frameworks and views that may be quite different than the initial formulation (Mizruchi & Fein, 1999). Many researchers continue to use SIC codes because that is what the articles they cite use, but they have neglected other potentially more promising taxonomies such as the Global Industry Classification Standard (GICS), the Industry Classification Benchmark (ICB), and the International Standard Industrial Classification (ISIC). Each of the alternative systems differs in its specifics, and these differences could be relevant to researchers.
Finally, we find from our empirical demonstration that it is not just the choice of industry measures that matters, but also the method used to incorporate the measure. In our example, simply changing the way industry-level ROA is operationalized in the model determines whether or not the theoretical expectations were confirmed. A similar example comes from Villalonga (2004), where changing the way in which industry boundaries are defined causes the diversification discount to turn into a diversification premium. Given the sensitivity of results to these refined details of industry measures, we must be exceedingly cautious in designing our studies and communicating the implications of our choices. Prior research on difference scores provides guidance with respect to choosing among some of the alternative approaches to measuring and testing industry effects. Typically, as noted previously, single component measures have more conceptual meaning, validity, and reliability than does the computed difference between them (Bergh & Fairbank, 2002; Edwards, 1994a, 1994b). Thus, when considered relative to industry effects, the difference score literature would support the use of a variable like average industry-level return on assets over a variable that is the difference between average industry-level returns on assets and a firm’s own return on assets. Empirical results can be more easily interpreted and do not require assumptions regarding measurement reliability and possible correlations between the variables if one remains in the analysis.
Recommendations
The findings from our study have strong implications for future strategic research. Based on what we have reviewed, assessed, and demonstrated in this study, we can make specific recommendations not only on the issues surrounding the measurements and tests of industry effects, but also on the issues relating to defining industry before deciding how to approach it. In the initial phases of an empirical study there is a significant amount of researcher judgment involved in determining what close substitutes for a particular company’s products are (or who the rivals are), dealing with the dynamic element of changing technologies that might shift industry boundaries over time, and best characterizing industry effects when firms are diversified. As such, subjectivity exists in how a particular researcher defines industry in the context of their work. To reduce the potential impact of this subjectivity on the development of knowledge in our field, we offer two recommendations. The first is that authors explicitly describe the decision processes that led them to define an industry in a particular way. Transparency about this process allows future researchers to understand the industrial context of different studies, understand how the choice of industry definition could have influenced the study results, and evaluate whether they should make use of the same industry definition.
The second is that where appropriate researchers rely on the precedent of previously published work when defining their industries. Failing to do so risks the accumulation of disconnected and incompatible studies from which it is difficult to build an integrated understanding of phenomena. Of course “where appropriate” is a significant caveat. Shifting boundaries may render previous definitions inapplicable, as might industry definitions that were not carefully constructed and well considered in the first place.
Once researchers choose a particular definition of industry, they then face the question of how best to measure the relevant measures and techniques in order to control for industry effects. Our findings show that strategic management researchers continue to utilize a wide range of techniques to that end, each with their own strengths and weaknesses, with no sign of convergence on a single approach. Table 7 lists the different categories of control methodologies identified by DIH and briefly lists some of their strengths and weaknesses. In addition, we propose some additional recommendations relating to the choice of control methodologies.
Pros and Cons of Various Industry Control Methods.
In terms of parsimoniously capturing all potentially relevant industry-level effects, industry dummy variables offer critically important advantages. They allow the researcher to account for industry characteristics without the risk of including irrelevant variables or excluding relevant ones, which is a danger when choosing among the myriad industry characteristic variables that are extant in the literature. There are three caveats, however, that apply to this recommendation. One, researchers must be careful to make sure that adding the number of dummy variables necessary to account for all of the industries in a particular sample does not lead to an undesirable ratio of variables to observations. Two, careful consideration must be given to how the industries are classified. We strongly encourage researchers to familiarize themselves with the different coding systems discussed earlier, evaluate their relative strengths, choose the one that makes the most sense in the theoretical context of their particular work, and then be explicit about how and why that particular system is selected. Third, thought must be given to levels of analysis. Using SIC codes as an example, creating dummy variables at the two-digit level leads to more aggregation compared to a four-digit dummy, potentially washing out important industry-level effects. Overall, we highly recommend that the choice of measurement approach aligns closely with how theoretical effects are conceptualized and tested.
If sample size or other concerns make industry dummies problematic, the researcher’s next choice would likely be to include multiple industry-level variables to control for the most theoretically relevant aspects of the environment. However the choice of which variables to include must be carefully considered and “guided by prior theory and empirical research in the substantive area” (Aiken & West, 1991). There must also be a high level of transparency on the part of researchers to spell out their logic behind the inclusion of some industry measure and the exclusion of others when dealing with multi-industry studies (Atinc, Simmering, & Kroll, 2012; Becker, 2005), and we must always be wary of adding more and more controls simply for the sake of appearance. As Carlson and Wu (2012) point out, “adding control variables does not make a study more rigorous” (p. 431) and can actually have the reverse effect of harming the meaning of the findings by diluting statistical power and reducing the remaining variance to be explained by predictor variables. We further recommend that if multiple industry-level variables are to be included in empirical models, that they enter as standalone variables rather than being used to “normalize” firm-level values. The potential problems of validity, reliability, and interpretation associated with difference scores make them a risky choice, and one that we have demonstrated can have a significant impact on empirical results.
It is also appropriate to highlight the risks associated with the proliferation of single-industry studies in strategic management research. While they offer numerous strengths and impressive advantages, the more prevalent they become, the greater the risk that future work will build on findings that could very well be industry specific. In response, we recommend that researchers undertake a wider debate on the role and value of replication studies in management research (Singh, Ang, & Leong, 2003). Current editorial policy at the top journals combined with a concern about article impact among colleagues and management departments strongly discourage researchers from conducting replication studies (Geuens, 2011). Hubbard, Vetter, and Little (1998) found that in a sample of 701 empirical strategic management journal articles there were zero direct replications and only 37 (5.3%) studies that they characterized as “replications with extensions.” Mezias and Regnier (2007) attribute the lack of published replication studies to a perceived lack of prestige associated with replications as well as “the emphasis placed on originality of contribution by the publication, promotion and tenure processes” (p. 286). Under the best of circumstances, that tendency can lead to limitations in knowledge development as researchers build future theory on the foundation of previous findings that may have been simply an artifact of a one-shot study (Evanschitzky, Baugarth, Hubbard, & Armstong, 2007). The potential for problems becomes even greater when so much of our work is being conducted within single industries where prevailing conditions might lead to a theory being either confirmed or undermined when the conclusion could be different in a different industry. In order to build a more externally valid body of scientific understanding we must be willing to reward and publish work that takes theories that have been established in one industry and applies them across others. Finding evidence that supports industry-specific findings lends confidence that the theory can be generalized and built upon, and finding no support gives us valuable information about the boundaries of the theory. Our continued insistence on breaking new theoretical ground as the gateway criteria for being published in top journals risks developing a proliferation of models that, truth be told, are largely industry specific.
However, taking a devil’s advocate position, industry effects simply may not matter in the context of a given theoretical model or research design. If they do not correlate with the outcome variables or the predictors, then they may simply be adding noise, be redundant, and introduce possible error to the analytical equations. Although research models need to be guided by theory, recent research on control variables suggests a more critical view to including variables and to carefully assessing each variable before blindly accepting it in the model, leading some to recommend, “when in doubt, leave it out” (Atinc et al., 2012; Carlson & Wu, 2012). Clearly, reasoned argument and empirical evidence should be used to guide decision making regarding whether to include industry effects in an empirical model and if so, which particular approach is most appropriate given the question, data characteristics, and purpose of the tests.
As with any study, ours carries limitations to the findings. One, by focusing only on the most impactful articles and those published in premier outlets for strategy research, we are unable to draw specific conclusions regarding the broader body of lower tier research. Although research in premier outlets appears to influence research in other outlets, we nonetheless have a limited sample of journals in our study and the findings apply directly to them. Two, we recognize that any empirical test involves compromise. There is no such thing as the perfectly designed study, and we do not mean for our suggestions here to be taken as a license for reviewers to demand ever more onerous sets of industry controls “just in case.” Three, while the results of our empirical example are suggestive, they are based on a single sample. It is possible, although we believe highly unlikely, that the sensitivity to how industry ROA was incorporated is idiosyncratic to this particular data set.
In conclusion, the results of our studies indicate that common practice for measuring and testing industry has changed over the past 20 years and that those changes bring with them serious implications for the empirical and theoretical underpinnings of our field. We also find that the choice of a particular approach to measuring industry profitability can lead to different empirical findings and influence inferences and interpretations. Researchers therefore must not only give great care and consideration to their choice of industry measures, but they must explicitly acknowledge the ways in which those choices may impact their work. We call on our research community to reconsider the value of replication studies designed to explore the limitations of our work with respect to industry effects and measures. In terms of replication, is the sensitivity to how industry ROA is implemented, which we demonstrate here, merely the idiosyncratic result of a one-shot study? Or are the results of other empirical models similarly affected by such minor changes in industry measures? What happens when the theoretical models that are supported in single-industry studies are applied more widely? Are the results in fact generalizable despite the narrow context of the original studies? For that matter, how do the results from multi-industry studies, regardless of industry measures included, hold up when applied to single-industry settings? These questions offer us a vast opportunity to explore more fully the effect of industry in strategic management research. We believe it is our obligation to define the boundaries of our understanding carefully in order for future research to be able to build on the solid foundation we create.
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.