Environments, unobserved heterogeneity, and the effect of market orientation on outcomes for high-tech firms

As we elaborate subsequently, to address this need for theory refinement, we resort to accounting for unobserved heterogeneity. However, we do recognize that other approaches are equally feasible and mention three such approaches. First, scholars have sought to refine the market orientation construct (e.g., Jaworski et al. 2000; Narver et al. 2004). The most dominant approach here is to theorize market orientation as possessing a reactive and a proactive dimension (e.g., Atuahene-Gima et al. 2005; Narver et al. 2004). Second, scholars suggest that firm strategy type should moderate the relationship between market orientation and firm performance (e.g., Ketchen et al. 2007; Matsuno and Mentzer 2000). The reasoning here is that facets of the organizational environment determine the firm strategy type and in turn the efficacy of market orientation depends on strategy type (Matsuno and Mentzer 2000). Third, scholars have integrated the cultural (Narver and Slater 1990) and information (Kohli and Jaworski 1990) perspective on market orientation to refine theoretical insights from the market orientation literature (e.g., Hult et al. 2005; Ketchen et al. 2007).

Regimes arise due to differences in regression coefficients that relate explanatory variables to dependent variable(s). Clusters or strategic groups, in contrast, result from difference in the levels of variables. Thus, regimes cannot be obtained from clusters using a two-step procedure of first using median splits to define high/low values of a variable (e.g., firm size) and then running separate regression models in each cluster. Such an approach assumes that the regression coefficients are same for all large firms—which need not be the case (in fact it is not the case as in both studies we report large firms belong to two regimes and small firms belong to two regimes). Thus, from a conceptual standpoint how does one allow regression coefficients to vary across large (small) firms? And, from a methods standpoint, the two-step procedure is statistically inefficient.

This case, in which firm size benefits some and impairs other firms, offers fertile ground for latent-class analysis. Usually when faced with such a conundrum, researchers either reason why one effect is more prevalent in their sample or let the results determine which of the two effects dominates. With latent-class analysis, researchers may find regimes in which both effects are supported. This ability to capture richer patterns in the data enhances the relevance of latent-class analysis.

In some aspects identifying latent regimes for the relationships among market orientation, facets of the environment, and firm outcomes is similar to the notion of strategic groups, where firms within a strategic group in an industry pursue similar strategies and compete more intensely with each other than firms across industries (e.g., Dranove et al. 1998; McGee and Thomas 1986). However, there are two critical differences in our approach and the extant strategic groups research: (1) we focus on market orientation and related variables as opposed to all indicators of firm strategy as is prevalent in the strategic groups research and (2) we identify regimes (groups) by changes in the regression coefficients across groups as opposed to changes in means values across groups, as is prevalent in the strategic groups research (e.g., DeSarbo and Grewal 2008).

In their pioneering effort to examine these moderating effects, Jaworski and Kohli (1993) suggest that the value of market orientation increases as market dynamism increases, competitive intensity increases, and technological turbulence decreases. However, their results do not support for the three related hypotheses (Jaworski and Kohli 1993). Slater and Narver (1994) question the validity of Jaworski and Kohli’s (1993) performance measure and reinvestigate the moderating role of the environment. However, they reach the same conclusion and recommend that market-oriented firms “are best positioned for success under any environmental condition” (Slater and Narver 1994, p. 53). In contrast, in a recent study on economic crises, Grewal and Tansuhaj (2001) adopt the market orientation and environmental measures of Jaworski and Kohli (1993) and show that during economic crises, the three facets of the environment have the hypothesized interaction effects. Other research supports some of these hypotheses (e.g., Homburg and Pflesser 2000).

Aggregation bias in the moderating effect of market dynamism can arise due to the opposite direction of hypothesized effect in H2a and H2b.

We thank an anonymous reviewer for suggesting this two-stage conceptualization.

An alternate model specification is to use a three-way interaction model where the firm size and organizational innovativeness qualify the two-way interaction of market orientation with market dynamism and technological turbulence. The concomitant variable latent class analysis we use offers at least four advantages over three-way interaction models: (1) Foremost, latent class analysis provides greater flexibility than a three-way interaction model. Specifically, a three way interaction model qualifies the two-way interaction in only one way and what it precludes is the possibility of finding (a) the opposite effect, (b) null effect, and (c) change in magnitude of the two-way interaction being qualified by the three-way interaction. (2) When there are more than one three-way interactions, the latent class models allows these to be introduced in a linear manner on the concomitant profiling term (i.e., the logit part of the model). (3) The latent class model allows us to model unobserved heterogeneity—which can be seen as being represented by the intercept term in the logit part of the model. More importantly, one can run the model with any third-way interaction variable by only having an intercept term in the logit part of the model to see how unobserved heterogeneity might impact the results (e.g., Hutchinson et al. 2000). (4) Three-way interaction, especially when there are multiple three-way interactions (we would have one for firm size and another for organizational innovativeness) induce the possibility of multicollinearity concerns that result in statistically inefficient parameter estimated (e.g., Mason and Perreault 1991). These multicollinearity concerns arise as to qualify one two-way interaction with a three-way interaction one has to introduce additional two-way interaction to have a complete model specification. With experimental data and orthogonal designs the multicollinearity issue can be managed—but as illustrated by Hutchinson et al. (2000), latent class analysis can be used to model unobserved heterogeneity.

To assess whether sample selection is an issue, we compare respondents with nonrespondents in terms of sales. We collect data from the Compustat and Dun and Bradstreet databases. Independent sample t-tests for firms for which we could gather data indicate no differences among (1) respondents and nonrespondents in Study 1 (t = 1.121, p > .24), (2) respondents and non-respondents in Study 2 (t = 1.145, p > .20), and (3) respondents in Studies 1 and 2 (t = 1.127, p > .26). Thus, sample selection does not seem to be an issue.

Consistent with domain based sampling (e.g., Nunnally and Bernstein 1994), we treat the original scales as the population of items and the items retained in the final construct as a sample from the population (i.e., domain of the construct). The final scales consist of six, three, four, and four items for market orientation, technological turbulence, market dynamism, and firm performance, respectively, in Study 1, and six, four, four, and four items for the same constructs, respectively, in Study 2.

The BIC values for the single-regime model (CFA₁) and two-regime models in which factor loadings (CFA_2λ), error variances (CFA_2θ), and both factor loadings and error variances (CFA_2λθ) vary across the latent classes are 9683, 9785, 9765, and 9788, in Study 1 and 11012, 11138, 11101, and 11211, respectively, in Study 2.

In Study 1, the ICOMP values for the five regimes are 592, 566, 557, 550, and 558. In Study 2, they are 616, 590, 567, 542, and 581, respectively.

We seek to acquire a sense of the degree of unobserved heterogeneity and the amount of improvement in fit in going from an aggregate to a disaggregate model and therefore examine an R ² type measure of the fit of the latent-class model with the aggregate model. We compute overall and regime-specific R ² values for the latent-class models, such that the regime-specific R ² values are based on the mean square error (MSE). Regime-specific MSE equals the difference between observed Y and predicted Y, that is, E(Y|X). The posterior regime membership probabilities quantify the contribution of a specific case to the error in that regime. Similarly, we calculate the overall R ² for the latent-class model based on MSE (regime memberships do not enter the equation in assessing the overall R ²). In Study 1, the aggregate model R ² is .10, and the R ² values for the latent-class model are .28, .53, .91, and .95 for the four regimes, with an overall model R ² of .79. For Study 2, the R ² values are .20 for the aggregate model and .63, .18, .78, and .19 for the four regimes, with an overall R ² value of .93. These results show that substantial unobserved heterogeneity exist in our data.

We have SIC information for 160 firms in Study 1 and 185 firms in Study 2. The relevant SIC codes are 28 (chemicals and allied products), 33 (primary metal industries), 34 (fabricated metal and machinery), 35 (industrial machinery and computer equipments), 36 (electronics), 37 (transportation equipment), 38 (measurement instruments), 39 (miscellaneous manufacturing industries), 42 (motor freight transportation and warehousing), 48 (communications), 49 (electric, gas, and sanitary services), 50 (wholesale of durable goods), 67 (holding and other investment offices), 73 (business services and engineering), 79 (amusement and recreation), and 87 (management services). Four SIC codes include more than six firms in both studies, namely, 31, 51, 30, and 34, which contain 35, 36, 38, and 73 firms, respectively, in Study 1 and 36, 59, 43, and 34 firms, respectively, in Study 2. For these SIC codes, we conduct χ² tests to determine if the firms are proportional in regime sizes and in the aggregate population. In Study 1, SIC code 35 firms are more likely to belong to regime 3 (p < .07) and SIC code 38 firms are less likely to belong to regime 1 and more likely to belong to regime 2 (p < .12). In Study 2, we find that SIC code 36 firms are more likely to belong to regime 2 and less likely to belong to regime 1 (p < .11).