Download Metric Effectiveness and Use in Marketing

Metric Effectiveness and Use in Marketing-Mix Decisions: Correcting for Endogenous Selection Effects and Ex-Ante Expectations

Ofer Mintz Timothy J. Gilbride Imran S. Currim Peter Lenk *

April 2016

* Ofer Mintz ([email protected]) is Assistant Professor of Marketing, E. J. Ourso College of Business, Louisiana State University, Baton Rouge, LA 70803. Timothy J. Gilbride ([email protected]) is Associate Professor and Notre Dame Chair in Marketing, Mendoza College of Business, University of Notre Dame, Notre Dame, IN 46556. Imran S. Currim ([email protected]) is Chancellor’s Professor and Professor of Marketing, Paul Merage School of Business, University of California, Irvine, CA 92697. Peter Lenk ([email protected]) is Professor of Technology & Operations and Marketing, Ross School of Business, University of Michigan, Ann Arbor, MI 48109. The authors thank the UCI Paul Merage School of Business Dean’s Office for financial support and participants at the 2015 Marketing Strategy Meets Wall Street IV Conference at Singapore Management University at the 2015 Theory and Practice in Marketing Conference at Georgia State University for their useful feedback.

Metric Effectiveness and Use in Marketing-Mix Decisions: Correcting for Endogenous Selection Effects and Ex-Ante Expectations Abstract This study models the use and effectiveness of individual metrics for different marketingmix decisions in the context of managerial, firm, and industry characteristics. While there has been much progress on the development of metrics to improve marketing’s accountability, little or no research uses performance data on specific marketing mix decisions to assess the use and effectiveness of individual metrics. Our model adjusts for potential endogeneity bias in metric effectiveness due to correlated unobservables and selection effects. The selection effect model differs from past literature in that managers choose metrics based on their ex-ante expected effectiveness, as opposed to their ex-post effectiveness, which would be unknown to the manager when the decision has to be made. The main findings are (i) awareness and willingness to recommend are the metrics that are most often beneficial for managers to employ across different marketing-mix decisions, while target volume, net present value, and net profit are least effective, (ii) accounting for managerial, firm, and industry characteristics matters when assessing the effectiveness of metrics, and (iii) managers are more uncertain about the ex-ante effectiveness of marketing as compared to financial metrics, and this attenuates the use of marketing metrics. The implications generate methodological and managerial contributions. Keywords: Managerial Decision Making; Metrics; Endogenous Regression; Hierarchical Bayes; Rational Expectations

1

1. Introduction For more than a decade, the Marketing Science Institute (MSI) and the Institute for Study of Business Markets (ISBM) have continuously encouraged research on metrics to improve managerial marketing-mix decision making (e.g., see MSI Research Priorities 2002-2016; ISBM B-to-B Marketing Trends 2008-2014). In response, among other developments, a variety of metrics have been proposed for various marketing-mix decisions such as product development, pricing, promotion, and distribution (e.g., Farris et al. 2010). Researchers have also focused on managerial use of metrics, identifying conditions under which managers employ more versus less metrics and select certain types of metrics over others for different decisions (Mintz and Currim 2013). While these two developments are indeed useful, managerial use or selection of metrics is not synonymous with metric effectiveness, i.e., just because managers use more metrics or select certain metrics over others for a particular marketing-mix decision does not necessarily mean that the metrics selected are effective in improving the efficacy of that marketing-mix decision. Some managers could be selecting certain metrics which turn out to be less effective for their marketing-mix decision, or not selecting metrics which would be more effective. For instance, institutional inertia or firm culture may favor some metrics over others despite their relative benefits for certain marketing-mix decisions, or a lack of expertise may lead managers to select a less effective metric (Lehmann and Reibstein 2006). There remains an important gap in the metrics literature, between, on the one hand, proposing metrics and providing insights on the managerial use of metrics, and, on the other hand, providing insight on which metrics are effective in improving the outcomes of particular marketing-mix decisions. This gap has left managers and researchers unsure of which metrics perform best in different marketing-mix decisions (Stewart 2009). The gap is important to 2

address because firms rarely have a shortage of metrics but instead have difficulty understanding which metrics are best for a certain marketing-mix decision (Lehmann and Reibstein 2006). This paper investigates marketing-mix decision performance, metric use or selection, and effectiveness. We conceptualize the metric use and effectiveness problem as follows. A manager is faced with a marketing mix decision such as running a promotion, changing a price, starting a social media campaign, etc. Whether the ultimate outcome of that marketing mix decision exceeds, meets, or is less than the firm’s expectations is referred to as the performance of the marketing mix decision. While the marketing mix decision is being made, a manager uses or employs individual metrics as decision aids (e.g., for considering, benchmarking, or monitoring) to assist in their decisions. The effectiveness of the metric is the relationship between using the metric and the ultimate performance of the marketing mix decision. We propose a central research question: how do we determine the effectiveness of specific metrics employed for a particular marketing-mix decision? Any study investigating the impact of metrics on marketing-mix performance must deal with the fact that metrics are not randomly assigned to managers, but rather managers choose metrics based on how the managers think the metrics will perform in a particular setting. The empirical data required to answer our central research question must be sufficiently granular so that the measured performance relates to a particular marketing-mix decision and is not confounded by other marketing-mix decisions. It is also important to recognize that metric choice occurs before the result of the marketing-mix decision is known to the manager; i.e., metric choice is based on ex-ante beliefs about metric effectiveness before observing the metric’s effectiveness for the marketing-mix decision and not ex-post evaluations of the metric after observing the outcome of the marketing-mix decision.

3

Hence, addressing the key research question identified above involves three important modeling considerations: (i) modeling the performance of the marketing-mix decision while explicitly allowing for endogeneity of metric selection and effectiveness, (ii) correcting for the managers’ ex-ante expectations about metric performance before observing the outcome of the marketing-mix decision, and (iii) controlling for other characteristics of the setting in which the marketing-mix decision is made, i.e., characteristics of the manager, firm, and industry. Our proposed hierarchical Bayes (HB) model results in a system of equations which include (i) regression for the performance of a marketing-mix decision as a function of the metrics being used, (ii) multivariate correlated probit for the choice of metrics as a function of the managers ex-ante beliefs, and (iii) multivariate regression to describe the variation of ex-post metric performance across the population of managers. The model contributes to the current literature on endogeneity by incorporating ex-ante expectations into selection effect models. Endogeneity arises in the model because metric use is related to metric effectiveness (selection effects) and there are correlated unobservables. Luan and Sudhir (2010) and Park and Gupta (2012) and others propose statistical methods to deal with these types of endogeneity. In contrast, our approach is more similar to Manchanda, Rossi, and Chintagunta (2004) and Li and Tobias (2011) in that we explicitly model the relationship between metric use and effectiveness via additional equations and a correlated error structure. The modeling highlights the logical necessity of distinguishing between ex-ante expectations and ex-post realizations of endogenous parameters and provides a novel approach for dealing with them via a “weak-form” rational expectations framework (Pesaran and Weale 2006). In addition to addressing issues raised by MSI and ISBM, this research responds to calls advocating studies on metric use that model the complex relationships among metric use, effectiveness, and marketing-mix performance (e.g., Ambler 2003, Pauwels et al. 2009, Petersen 4

et al. 2009) in order to provide managerial guidelines to determine the right metrics for the right decision (Lehmann 2004, Stewart 2009, Wind 2008). Our empirical analysis reanalyzes the data collected by Mintz and Currim (2013). This data set includes information from 439 managers on 1,287 specific marketing-mix decisions, which metrics they employed for each decision, and the subsequent outcome of each decision based on several subjective performance measures. Hence, it differs from other work that only includes information on aggregate firm performance (e.g., return on assets or stock returns) and some aggregate measure of metric use. The level of detail available in Mintz and Currim’s data is necessary to relate individual metrics to specific decisions, as opposed to relating general metric use to overall firm performance, for the following reason. At any point in time, a firm typically makes several concurrent decisions about advertising, pricing, distribution, and new products for multiple product lines and may employ different metrics for each decision. Consequently, studies that rely on aggregate measures of firm-level financial performance are unable to link the individual metrics used for each decision to the outcome of that decision, or estimate the marginal contribution of employing a given metric for a particular decision. Although our study relies on managers’ subjective evaluations of marketing-mix performance outcomes, subjective evaluations are not arbitrary (e.g., see literature in the decision analysis field based on Keeney and Raiffa 1976). Multiple studies find method convergence, i.e., insignificant differences, between subjective and objective measures of performance outcomes (e.g., Dawes 1999, Dess and Robinson 1984, Germann, Lilien, and Rangaswamy 2013, O’Sullivan and Abela 2007, Venkatraman and Ramanujam 1987). The main empirical findings are the following. First, after controlling for covariates and correcting for endogeneity bias, we consistently find across different types of marketing-mix decisions that awareness and willingness to recommend act as “silver bullet” metrics that are associated with better performance; however, target volume, net present value (NPV), and net 5

profit act as “lead bullet” metrics which are associated with worse performance. Second, for each marketing-mix decision, the empirical findings allow us to infer which metrics are most effective and ineffective. For example, for price promotion decisions, we find that awareness, customer loyalty, and consideration set are significantly associated with improved performance, while NPV and share of voice are significantly associated with worse performance. Third, when assessing the effectiveness of individual metrics, we find that accounting for managerial, firm, and industry variables is important and not including such variables leads to unobserved heterogeneity biases. Fourth, when comparing marketing versus financial metrics1, we find that marketing metrics are on average more beneficial to marketing-mix decision outcomes, but managers appear more uncertain about the ex-ante effectiveness of marketing as compared to financial metrics, which attenuates the use of marketing metrics. Fifth, managers tend to use metrics that they view to be more effective. This result requires econometric methods to adjust for selection effect biases when evaluating metric effectiveness. Such managerial, methodological, and empirical contributions are important steps towards improving marketing’s accountability and managerial decision making.

2. Background and Related Literature When unobserved factors influence the use and effectiveness of a particular metric and the performance of a particular marketing-mix decision, this leads to the well-known endogeneity problem. This problem, often called “selection bias” (e.g., Heckman 1979), can lead to biased estimates of metric effectiveness if the endogenous metric selection process is ignored. There is a rich literature in labor econometrics in addressing such issues; Heckman and Vytlacil (1998) and

1

Marketing metrics are defined as metrics that are based on a customer mindset while financial metrics are defined as metrics that are either monetary or readily converted to monetary based outcomes (Mintz and Currim 2013).

6

Wooldridge (2003) offer instrumental variable techniques, Imbens and Wooldridge (2009) provide a recent survey. In an empirical context, models of metric effectiveness must confront two possible sources of endogeneity. The first is the observation that the use of a particular metric may be influenced by the expected effectiveness of the metric. For example, consider a simple linear model, yi = α + θimi + ei , where mi is an observed binary variable indicating whether or not the metric was used, yi is the measured performance (outcome) of the decision, and the regression parameter θi is the heterogeneous impact of using the metric on performance. Endogeneity occurs if the use of the metric mi depends on its effectiveness θi. A second source of endogeneity occurs when the explanatory variable, mi, is correlated with the error term, ei. This may occur when additional explanatory variables are correlated with mi but are omitted from the model. In a general context, Luan and Sudhir (2010) refer to the first source of endogeneity as “slope endogeneity” and the second as “intercept endogeneity.” There are two broad methods of dealing with slope and intercept endogeneity. In the first method, Manchanda et al. (2004) present a general framework for addressing slope endogeneity. Their method entails augmenting the outcome equation with another equation that explains the level of an explanatory variable as a function of a common parameter. Thus, we might represent the earlier linear model as yi=f1(α, θi, mi) and augment it with another equation mi = f2(θi, r); the presence of θi in multiple equations represents simultaneity in the data. This approach has been used by Van Diepen et al. (2009) in a direct mail application, Li et al. (2011) in a model of crossselling, as well as Manchanda et al. (2004) in their pharmaceutical industry application. However, this method requires some knowledge or a theory on the data generating mechanism, i.e., we must explicitly model the relationship between metric use and effectiveness. A second approach relies on statistical arguments to create instrumental variables and/or a control function 7

approach without making a substantive argument on the nature of the endogeneity. Luan and Sudhir (2010) detail a control function approach that addresses both slope and intercept endogeneity. Park and Gupta (2012) provide a useful overview of the issue and introduce a semiparametric method that can address slope or intercept endogeneity without identifying instrumental variables; their method relies on a semi-parametric representation of the endogenous regressor and then jointly modeling the regressor and the error terms. The model detailed in the next section uses several equations and explicitly considers the relationship between metric use and metric effectiveness. In this sense, it is similar to the approach suggested by Manchanda et al. (2004) to address slope endogeneity. However, the model also addresses intercept endogeneity via a correlated error structure between the different equations. The extended modeling framework distinguishes our approach from instrumental variable/control function adjustments to a single equation. We will show that the proposed model offers additional insights compared to the instrumental variable/control function methodology. Our approach is most similar to that of Li and Tobias (2011) who use a three equation model to determine earnings, the heterogeneous return to schooling, and the level of schooling; their model also includes a correlated error structure. Our model differs from Li and Tobias (2011) in three important ways. First, we consider multiple explanatory variables (which metrics were used) as opposed to a single explanatory variable (level of education). As documented by Mintz and Currim (2013), decision makers often employ more than one metric to make a decision. Second, our explanatory variables are binary as opposed to continuous. The control function approach of Luan and Sudhir (2010) does not consider dichotomous regressors and while the semi-parametric approach of Park and Gupta (2012) considers multinomial regressors, as they note, their model is not identified for binary variables. Li and Tobias (2011) also assume students know how much more money they would earn for each additional year of education 8

completed. Perfect foresight is an extreme assumption for students or managers. Rather, we assume the decision makers have rational expectations of the effectiveness of a particular metric, but that it may differ from the metric’s actual performance. Thus, our third departure is allowing for the ex-ante expectation of metric effectiveness to differ from the ex-post realization of metric effectiveness. This issue has not been addressed in the literature on slope endogeneity. The addition of ex-ante expectations, multiple binary endogenous regressors, and a complicated error covariance structure necessitates a new method for estimating model parameters.

3. Model 3.1. Model Overview. Our model measures the effectiveness of various metrics controlling for the fact that managers are more likely to use metrics that they believe to be effective. The model specification has four equations. The first equation relates the performance of the marketing-mix decision to the metrics used for the decision and their individual effect on performance, which vary across the population. The second set of equations model the selection of the metrics and assumes that managers are more likely to use metrics with larger ex-ante effectiveness. Because managers select metrics before they evaluate the decisions, using the post-hoc evaluations of metric effectiveness would bias the selection sub-model. The error terms for the performance equation and the selection equations are correlated, thus allowing the model to adjust for endogenous selection effects and to group metrics based on the their similarity due to unobserved factors. The third equation describes the heterogeneity of ex-ante metric effectiveness across types of marketing decisions, managers, firms, and industries, and hence examines the settings in which metrics are more or less beneficial. The fourth equation imposes the weak form of rational expectations by relating ex-ante and ex-post expectations about metric effectiveness. The submodels use exclusion restrictions to identify the model. We estimate the four sub-models 9

simultaneously with Bayesian methods. We note the proposed modeling framework is novel because past theorizing on marketing metrics has not made a clear distinction between what makes a metric more effective, versus what makes a metric more likely to be used. 3.2. Model Specification. In our study, subjects are managers who evaluate the performance of multiple decisions where each decision concerns a different type of marketing-mix decision. Managers evaluated 1 to 10 decisions with an average of 2.9 decisions per manager. There are a total of D types of marketing-mix decisions. We assume that the decisions are exogenous to metric choice and effectiveness; i.e., managers do not select decisions based on their beliefs about metrics. Managers can use any combination of K possible metrics for each decision. Indexing the manager’s marketing-mix decisions and type of decision becomes slightly complex because managers evaluate different types of marketing decisions. We use i = 1, …, N to index managers in the study. Manager i evaluates ni decisions, which are indexed by j = 1, …, ni and index j only refers to the order of response and not the type of marketing decision. We introduce the notation d(i,j) = d to indicate that decision j for manager i is of decision type d. In the first equation, we measure the impact on marketing-mix performance of employing individual metrics. The dependent variable yij is an overall performance evaluation of decision j by manager i: 𝑲

𝑦𝑖𝑗 =

𝒙′𝑖𝑗 𝜷

1

+ ∑ 𝑚𝑖𝑑𝑘 𝜃𝑖𝑑𝑘 + 𝜀𝑖𝑗 for 𝑖 = 1, … , 𝑁; 𝑗 = 1, … , 𝑛𝑖 ; and 𝑑 = 𝑑(𝑖, 𝑗) 𝒌=𝟏

where xij are exogenous control variables that include characteristics of the manager and his or her firm and industry, and  are fixed effects. The observed indicator midk is one if manager i uses metric k for decision j of decision type d = d(i,j) and 0 otherwise. The random effect idk is the latent, ex-post measure of the effectiveness of metric k for manager i and marketing decision 10

d = d(i,j). The normally distributed random shocks ij have mean 0 and standard deviation Y. Li and Tobias (2011) assume a single, continuous midk, while in our context they are multiple, binary indicators. Next, we account for the drivers of metric-selection. Research to date on metric effectiveness treats metric use as exogenous, which implies that managers are assigned metrics or randomly select them. A more reasonable assumption is that manager i is more likely to use metric k if his or her ex-ante beliefs are that the metric will be effective. We use a multivariate probit model (Chib and Greenberg 1998) for the selection of metrics. Manager i uses metric k in decision j of decision type d = d(i,j) if the manager’s “utility” uidk for the metric is positive: 𝑚𝑖𝑑𝑘 = 1 if 𝑢𝑖𝑑𝑘 > 0, and 𝑚𝑖𝑑𝑘 = 0 if 𝑢𝑖𝑑𝑘 ≤ 0.

2

Managers select the metrics before observing the ex-post, latent metric effectiveness idk for the decision in Equation (1). We define the manager’s ex-ante belief about metric effectiveness as 𝜃̃𝑖𝑑𝑘 . The metric-selection equation is: 𝑢𝑖𝑑𝑘 = 𝒛′𝑖𝑗 𝜹𝑘 + 𝜌𝑘 𝜃̃𝑖𝑑𝑘 + 𝜈𝑖𝑑𝑘 for 𝑖 = 1, … , 𝑁; 𝑗 = 1, … , 𝑛𝑖 ; 𝑑 = 𝑑 (𝑖, 𝑗); and 𝑘 = 1, … , 𝐾

3

where zij are exogenous variables that are specific to the manager or firm; 𝜹𝑘 and 𝜌𝑘 are scaled by the ex-ante beliefs for metric k; and the normally distributed random shocks 𝜈𝑖𝑑𝑘 have mean 0 and variance one. The multivariate probit model constrains the variance to one to identify the model. The parameters ρk are restricted to be positive and relate ex-ante metric effectiveness to metric use: non-negative ρk imply that the manager is more likely to use metric k if he or she views it to be more effective. This parameter represents a selection propensity that is not captured elsewhere in the model. The presence of ρk in Equation (3) distinguishes this model from purely instrumental variable methods such as those discussed by Heckman and Vytlacil (1998) and Wooldridge (2003), also in the context of multiple treatment effects. Nonetheless, the 11

exogenous variables x in Equation (1) and z in Equation (3) have exclusion restrictions to identify the model by exogenous variation: z includes variables that are absent from x (discussed in more detail the next sub-section). The error terms in Equation (3) are correlated across metrics with covariance and correlation matrix U. The variances in the multivariate probit model are assumed to be one. Further, they are correlated with the random errors from Equation (1). The full correlation matrix

 and covariance matric  for Y and U are: 1 𝚺=[ 𝚺𝑈𝑌

𝚺𝑌𝑈 𝜎2 ] and 𝚵 = [ 𝑌 𝚺𝑈 𝜎𝑌 𝚺𝑈𝑌

𝜎𝑌 𝚺𝑌𝑈 ] 𝚺𝑈

4

where UY is a K vector of correlations between ij and idk where the correlations depend on the ′ metric and not on the type of decision, and 𝚺𝑌𝑈 = 𝚺𝑈𝑌 . If UY is non-zero, then metric selection

is endogenous. For example, if the cross equation correlations are positive, then managers whose utility for a metric are larger than its expectations (positive error terms) will tend to evaluate the decision higher than what one would expect based solely on Equation (1). The third sub-model describes the distribution of metrics’ realized effectiveness across managers and decisions. In other words, the third sub-model examines the settings in which individual metrics are more effective, i.e., associate with a beneficial or detrimental relationship with marketing performance: 𝜽𝑖𝑑 = 𝚽 ′𝒘𝑖𝑑 + 𝜼𝑖𝑑 for 𝑖 = 1, … , 𝑁; 𝑗 = 1, … , 𝑛𝑖 ; and 𝑑 = 𝑑(𝑖, 𝑗) where 𝜽𝑖𝑑 = (𝜃𝑖𝑑1 , … , 𝜃𝑖𝑑𝐾 )′ for d = d(i,j) is the latent, ex-post effectiveness for the K metrics from Equation (1). 𝒘𝑖𝑑 are exogenous variables that are specific to the type of marketing decision, the manager, firm, and industry.  is a matrix of regression coefficients. The multivariate normal random shocks 𝜼𝑖𝑗 have mean 0 and covariance matrix .

12

5

We use weak-form rational expectations theory to equate managers’ ex-ante expectations to the ex-post expectations across the population. Muth's (1961) strong form of rational expectations assumes that each agent’s distribution of ex-ante beliefs equals the true distribution of outcomes based only on public information. Private information does not play a role in this strong form of rational expectations. Pesaran and Weale (2006) propose a weak form of rational expectations where agents’ expected ex-ante and ex-post expectations are equal. The underlying assumption is that information about metric effectiveness is widely shared among managers with similar backgrounds and types of firms in similar industries, but managers can differ in their individual ex-ante beliefs of metric effectiveness. In other words, each manager has his or her beliefs about effectiveness; however, the information about metric effectiveness is well-diffused across managers, and these idiosyncratic beliefs average out across the population. In our model, the ex-ante metric effectiveness for manager i deviates from the population mean by a random shock 𝜁𝑖𝑘 .The ex-ante shocks would not be well separated from the errors terms in Equation (5) if there were unique random shocks for each manager, decision type, and metric. Therefore, we assume that the random shocks do not depend on decision type. However, weak-form rational expectations imply that 𝐸(𝜃̃𝑖𝑑𝑘 ) = 𝐸 (𝜃𝑖𝑑𝑘 ), or that on average ex-ante and ex-post metric effectiveness are equal across managers. Using the population level model for expost metric effectiveness in Equation (5) gives the model for ex-ante metric effectiveness: 2 ) 𝜃̃𝑖𝑑𝑘 = 𝒘′𝑖𝑑 𝝓𝑘 + 𝜁𝑖𝑘 where 𝜁𝑖𝑘 ~𝑁(0, 𝜎𝜁𝑘

6

for i = 1, …, N; j = 1, …, ni; d = d(i,j); k = 1, …, K, and 𝝓𝑘 is column k of  in Equation (5). Combining Equations (3) and (6) results in the estimation equation for metric use utilities: 𝑢𝑖𝑑𝑘 = 𝒛′𝑖𝑗 𝜹𝑘 + 𝜌𝑘 [𝒘′𝑖𝑑 𝝓𝑘 + 𝜁𝑖𝑘 ] + 𝜈𝑖𝑑𝑘

13

7

where the random shocks 𝜁𝑖𝑘 and 𝜈𝑖𝑑𝑘 are mutually independent. The rational expectations shock 𝜁𝑖𝑘 creates a correlation structure within subjects for each metric, while the unobserved utility 𝜈𝑖𝑑𝑘 imposes a correlation among metrics. The independent variables zij and wid have exclusion restrictions: most importantly zij does not include the type of marketing-mix decision while wid does. Specific exclusion restrictions will be discussed in the context of our empirical example in the data section. Equations (6) and (7) address a logical inconsistency in modeling endogenous relationships based on common parameters: when decisions and outcomes are sequenced (as here), decision makers must rely on ex-ante expectations as opposed to ex-post results. Bayesian inference requires prior distributions for the unknown parameters, and we use standard specifications: multivariate normal for regression coefficients, inverse Gamma for variances, and inverse Wishart for covariance matrices, except . We use the prior distribution of Barnard et al. (2000) for the correlation matrix  for the random shocks of Equations (1) and (3). Talhouk et al. (2012) provide an efficient algorithm for sampling from the posterior distribution of the correlation matrix. Also, we follow Lenk and Orme (2009) and impose the prior on the conditional error variance of Y given the metric utilities U instead of the error variance of Y. The Web Appendix presents the details of the prior distributions and details the full conditional distributions for the MCMC algorithm. 3.3. Identification. The reduced form model substitutes Equation (5) into Equation (1). Together with Equation (7), we obtain the 𝑦𝑖𝑗

′

= 𝒙′𝒊𝒋 𝜷 + 𝒎′𝒊𝒋 𝜱 𝒘𝒊𝒅 + 𝒎′𝒊𝒋 𝜼𝒊𝒅 + 𝜀𝑖𝑗

′ 𝑢𝑖𝑗𝑘 = 𝐳′𝐢𝐣 𝛅𝐤 + ρk 𝐰𝐢𝐝 𝛟𝐤 + 𝜌𝑘 𝜁𝑖𝑘 + 𝜈𝑖𝑑𝑘 𝑚𝑖𝑑𝑘 = 1 if 𝑢𝑖𝑑𝑘 > 0 and 𝑚𝑖𝑑𝑘 = 0 if 𝑢𝑖𝑑𝑘 ≤ 0.

14

8

where mij = (mij1, …, mijK)’ is the vector of metric indicators. The performance outcome equation for yij and the exclusion restrictions in the covariates identify the coefficients  and  and the error variances 2Y and . The metric use equations identify the coefficients  and the error correlation U. The identification of  from the performance outcome equation then identifies 2 its multiplier k in the metric utility equation. Since k is identified, then the variances 𝜎𝜁𝑘 of the ∗ ex-ante shocks 𝜁𝑖𝑘 are identified. In particular, if one defines k = ak;  = /a, and 𝜁𝑖𝑘 = ∗ 𝜁𝑖𝑘 /𝑎 where 𝑎 is a non-zero constant, then k k𝜁𝑖𝑘 = k k𝜁𝑖𝑘 , and the utility

equation is left unchanged. However, we cannot arbitrarily redefine  in the utility equation 2 without changing the density function in the rating equation. Therefore, k and 𝜎𝜁𝑘 are identified.

Simulation studies conducted by the authors, and available upon request, confirm the models ability to obtain identified parameter estimates.

4. Data 4.1. Data Collection and Variables. We test our model on the 1,287 marketing-mix decisions reported on by 439 managers from Mintz and Currim (2013). The managerial respondents were primarily obtained via LinkedIn professional organizations (81%); the remaining managers are MBA alumni of a U.S. west-coast university (19%). The questionnaire consisted of two sections. In the first section, managers indicated which of 10 types of marketing-mix decisions they had recently made.2 Then, for each type of decision made, managers indicated which of the 12 general marketing metrics and 12 general financial metrics listed in Figure 1.A were employed.3 Next, managers assessed the

2

Following Menon et al. (1999) the managers were told to select decisions “that were not so recent that performance evaluation was premature and not so long ago that memory about the decision and performance was fuzzy.” 3 Mintz and Currim (2013) also ask managers to indicate which of 3 specific to a marketing-mix decision marketing metrics and which of 3 specific to a marketing-mix decision financial metrics they employed for each decision. However, we focus solely on the 24 total general metrics because these metrics were suited across all the different

15

performance of each marketing-mix decision based on subjective measures taken from Jaworski and Kohli (1993), Moorman and Rust (1999), and Verhoef and Leeflang (2009). Following these works, we define performance of the marketing-mix activity in on our context based on a firm’s stated marketing, financial, and overall outcomes, relative to a firm’s stated objectives and to similar prior activities.4 In the second section of the questionnaire, managers answered questions on managerial, firm, and industry characteristics. Managerial characteristics include metric-based compensation, metric training level, functional area (marketing vs. not marketing), managerial level (top vs. mid-level management), managerial experience, and quantitative background. Firm characteristics include market orientation, strategic orientation (prospectors, analyzers, low-cost defenders, and differentiated defenders), organizational involvement in the decision, firm size, type of ownership (public vs. private), chief marketing officer (CMO) presence, recent business performance, and B2C vs. B2B and service vs. goods orientations. Industry characteristics include product life cycle stage (maturity/declining vs. introduction/growth), concentration (high vs. low), market growth, and market turbulence. Measures were taken from the extant literature; each variable is listed in Table 1 and further details on the definitions, operational measures, and published sources of these variables are provided in Web Appendix Table 1. Managerial, firm, and industry characteristics are likely to influence decision performance, which metrics are used, and metric effectiveness. For example, the use of a metric is expected to be a function of whether the manager is trained and compensated based on metric use (e.g., Mintz and Currim 2013). Likewise, metric effectiveness is expected to be a function of

types of marketing-mix decisions, while specific to a marketing-mix decision metrics were only suited to each type of marketing-mix decision, which limits their applicability to other types of decisions. 4 If the reader is interested in results of metric effectiveness for any of the performance measures individually (i.e., marketing, financial, overall, and relative to similar past activities), please feel free to contact the authors.

16

managerial characteristics since different metrics (e.g., marketing, financial) appeal to different managers (e.g., marketing and non-marketing managers, and more and less experienced managers). And performance of the marketing-mix decision (e.g., a new product development decision) is expected to be a function of characteristics of the industry (e.g., whether the industry is in a growth or decline stage) and the firm (e.g., the size of the firm competing in that industry). Hence, the inclusion of these controls in the performance, use, and effectiveness equations are useful to mitigate but not eliminate unobserved heterogeneity. The full set of managerial, firm, and industry controls, are presented in Table 1. Our survey data provides a rich source of information needed to identify parameters in the correlated random coefficient model proposed in the previous section. Table 1 also lists which variables were used in each equation in the model and therefore identifies the exclusion restrictions. So, for instance, the variables “metric-based compensation” and “metric-training” were included in the vector z in the metric use equation (3) and the vector w in the metric effectiveness equations (5) and (6); these two variables help to identify the use equation since they were excluded from the vector x in the decision outcome equation (1). The logic is that once the effect of “metric-based compensation” is controlled for in the use equation, whether a manager has metric-based compensation should not impact whether the ultimate marketing-mix decision performed well or not. As detailed in Li and Tobias (2011, p. 347), “what is most necessary for identification purposes is the existence of some variable affecting” the effectiveness of the metric 𝜃̃𝑖𝑑𝑘 , that is conditionally uncorrelated with metric use and the decision outcome. We have chosen the marketing-mix decision variables to serve this role. Once we include the metric’s effectiveness for making a particular type of decision, there should be no additional information in identifying the marketing-mix decision, wid. That is, if we expect “share of voice” to be effective for traditional advertising decisions, i.e., we know 𝜃̃𝑖𝑑𝑘 , then 17

knowing that we are making a traditional advertising decision is redundant for metric use and the decision outcome since the relevant information is already contained in 𝜃̃𝑖𝑑𝑘 . Our analysis of metric use and effectiveness is based on subjective measures of performance as opposed to objective measures such as stock market return, return on assets (ROA), etc. The primary substantive goal of this research is to understand if specific metrics are more effective for particular types of marketing decisions. Measures such as stock returns or Tobin’s Q aggregate performance over all decisions made by the firm in a particular time frame; statistically identifying the effect of one metric (“willingness to recommend”) on one particular type of decision (social media) seems problematic using aggregate data (e.g., see Katsikeas et al. 2016 review on marketing performance measures for similar discussion). Trying to control for all confounding factors (e.g., the mix of decisions that are made by a firm) in cross-sectional or even time series data to isolate the effect of one specific metric on one type of decision on aggregate firm performance would be a Herculean task. Ideally, there would be an objective measure of decision outcomes such as return on investment (ROI) or return on marketing investment (ROMI) that could be reported for all types of decisions. However, as Dess and Robinson (1984) argue, it would be very difficult for survey respondents to calculate the ROI or ROMI, for that calculation to be comparable between respondents, and for that methodology to be consistently applied to all types of decisions (e.g., pricing, distribution, sales force, etc.); any claim to objectivity would likely be illusory. One might be concerned that managers inflate the reported performance as either a demand effect or ego self-preservation; yet we find significant variation in the outcome measure both within managers and across decisions. While we recognize the subjectivity of our dependent measure, studies by Germann, Lilien, and Rangaswamy (2013), Germann et al. (2014), and O’Sullivan and Abela (2007) were able to test a subset their samples using both subjective and objective measures, and in each case obtained similar results. 18

4.2. Descriptive Statistics. Figure 1.A reports how often managers employed each metric overall and by type of marketingmix decision. One can see the significant variation of metric use across the different types of marketing-mix decisions. For example, the five metrics employed most often by managers in the dataset were target volume, awareness, total customers, ROI, and market share; but the five most employed metrics for traditional advertising decisions (awareness, marketing expenditures on branding, ROI, ROMI, and target volume) were considerably different than the five most employed metrics for pricing decisions (net profit, market share, target volume, perceived product quality, and total customers). Another example is the employment of net profit across decisions: managers employed the metric for 61% of pricing and 49% of new product development decisions, but only for 5% of social media and 3% of PR/sponsorship decisions. Two of our metrics, stock prices/returns and Tobin’s Q, were so rarely employed (less than 1% of the decisions) that we were forced to drop these metrics from our analysis. Figure 1.B provides averages of marketing-mix performance overall and by type of marketing-mix decision when a metric is employed in that decision. Similar to the aforementioned description of metric use, when examining this figure we can see significant variation of average marketing performance across marketing-mix decisions whenever a metric is employed compared to overall average. For example, the average marketing performance when economic value added (EVA) is employed is 5.32 (out of 7) but ranges from 3.59 for traditional advertising to 6.38 for PR/sponsorships, and the average marketing performance when share of voice is employed is 5.26 but ranges from 4.69 for sales force to 6.49 for new product developments decisions. This difference in decision outcome (performance) when the same metric is used in different decisions provides model-free evidence that metric effectiveness varies by which marketing-mix decision is being made. In addition, the performance measure has 19

considerable variation within and between subjects, which is inconsistent with demand effects where managers uniformly rate their projects highly. 5 When combining panels A and B of Figure 1, the descriptive statistics also show that although some metrics like share of voice, share of customer wallet, EVA, and customer lifetime value (CLV) are not employed that often, i.e., less than 15% of the decisions on average (Panel A), when they are employed average marketing-mix performance is greater than that for many of the other metrics (Panel B). In summary, metric use and effect on performance are expected to vary greatly depending on the type of marketing-mix decision. Analyses of the data shows no indication of multicollinearity based on variance inflation factor scores well below 6 (Hair et al. 1998) and over 99% of pairwise correlation coefficients being less than .40 (e.g., Leeflang et al. 2000). Loadings from exploratory factor analyses for all constructs6 were above 0.7, while coefficient alphas for all but three constructs were greater than 0.7 (market turbulence is .63, market growth is .66, managerial experience is .68). Common method bias is not detected based on the Lindell and Whitney (2001) test where we adjusted the correlation matrix by the lowest positive pairwise correlation value to create a partial-correlation adjusted matrix, and no resulting pairwise correlation lost significance. In addition, non-response bias is not found, based on the Armstrong and Overton (1977) test to compare early and late respondents scores on the included constructs.

5. Results The model detailed in Section 3 and the Web Appendix was estimated using MCMC methods. The algorithm was run for 100,000 iterations with the last 50,000 used to summarize the posterior moments of the parameters. Simulation studies were conducted to test the code, the 5

Also, our measure of metric effectiveness is the relative change in performance for using the metric and is not sensitive to biases in the absolute level of the performance ratings; i.e., adding a constant to the ratings will not change our metric effectiveness. 6 Full details on the questions used to create constructs are contained in Web Appendix Table 1.

20

recoverability of parameters, and the convergence properties of model parameters. Convergence of the actual data was assessed by examining the time series plots of selected parameters and reestimating the model with different random starting points. Selected parameter estimates are presented here and additional results are contained in the Web Appendix; full results are available from the authors. 5.1. Individual Metric Effectiveness. Figure 2.A displays boxplots of the posterior means of metric effectiveness for each subject and metric. The metrics are ordered by the average of the posterior means. Awareness, satisfaction and willingness to recommend have the largest average, with at least 75% of the subjects judging them to have positive effectiveness. In contrast, NPV, target volume, and ROS are judged to be least effective with negative scores. Figure 2.A shows that metric effectiveness varies considerably across managers. For instance, the mean effectiveness for CLV, EVA, and consideration set are almost zero, but range widely from -7 to 7. Managers are very diverse in their impression of metric effectiveness. Figure 2.B plots the average effectiveness by metric and type of decision, i.e., 𝜃̃𝑖𝑑𝑘 averaged over i. As in Figure 1.B, there is considerable variation of metric effectiveness by decision. Figure 2 masks the effectiveness of metrics for different contextual factors. In Table 2, we provide the parameter estimates for how the type of decision influences a metric’s effect on marketing-mix performance (a subset of the Φ matrix in equations (5), (6), and (7)). The coefficients in the table should be viewed as the impact of the type of decision on the effectiveness of an individual metric on marketing-mix performance for the average manager, firm, and industry, as we mean-center the continuous control variables and employ effects coding for the discrete control variables. In other words, the coefficients display whether a metric is associated with a significantly beneficial (bolded, positive coefficient), detrimental (bolded, 21

negative coefficient), or no effect (not bolded) on performance for that specific type of decision, for the average manager, firm, and industry. A coefficient was deemed to be significant if 97.5% of the MCMC draws from the posterior distribution were larger than 0 or 97.5% were less than zero. That is, the posterior distribution is shifted above or below 0. We summarize our empirical findings as follows. First, when looking broadly on the impact of financial and marketing metrics, we find that financial metrics are for a large part detrimental to performance for different types of marketing-mix decisions while marketing metrics have more of a mixed, beneficial and detrimental effect depending on the individual metric. Second, when examining individual metrics (i.e., by looking at each row in the table), two metrics, awareness and willingness to recommend, are found as “silver bullets” that are consistently associated with superior performance across different types of decisions. However, in contrast, we find three metrics, target volume, net present value (NPV), and net profit, are “lead bullets” associated with worse performance for most types of marketing-mix decisions. Third, we find certain metrics can have both a beneficial and detrimental association with performance, depending on the type of decision. For example, ROMI is associated with greater performance for internet advertising decisions, but worse performance for social media, pricing, and new product development decisions; while share of voice is associated with greater performance for social media and PR/sponsorship decisions, but worse performance for price promotions. Fourth, when examining the results for each type of decision (i.e., by looking at each individual column), we can suggest what are the “right” and “wrong” metrics for that decision. For instance, for pricing decisions, EVA, satisfaction, preference, and willingness to recommend are found as “right” metrics which are beneficial to performance, but NPV and target volume are found as “wrong” metrics that are detrimental. In addition, for internet advertising decisions, ROMI and awareness are “right” metrics but net profit, target volume, perceived quality, and 22

total customers are “wrong” metrics. Results in Figure 2.B and Table 2 are very similar; Figure 2.B differs only as a result of the actual mix of managers in our empirical sample. Finally, in Table 3 we provide the results of our control variables which account for the type of manager, firm, and industry (the remainder of the Φ matrix). Most importantly, the presence of significant coefficients demonstrates that accounting for these variables is important as they do in fact matter to whether metrics have a beneficial or detrimental impact on marketing-mix performance. For example, when examining an individual metric, i.e., CLV, we find the metric is significantly beneficial for marketing-mix performance for marketing (vs. non marketing) managers, who are working in private (vs. public), larger firms, with better recent performance, with a CMO (vs. without) in services (vs. goods), and more turbulent industries. Consequently, not accounting for the type of manager, firm, and industry may lead to omitted variable biases or measurement errors when assessing the effectiveness of an individual metric. 5.2. Combining Results on Metric Effectiveness and Use. An additional aspect of our analysis is that based on the metric’s perceived ex-ante effectiveness, we can analyze whether or not the metric will likely be used. In Table 4 we report the ρk multiplier scores from equation (7) for each metric. This parameter provides a model-based indicator of how a managers’ ex-ante beliefs about the impact of a metric on marketing performance determines the metric’s use in the decision, while accounting for managerial, firm, and industry characteristics. This type of analysis is not possible in instrumental variable/control value approaches to endogeneity. Larger values of ρk, for instance ρk > 1, mean that for a given value of 𝜃̃𝑖𝑑𝑘 , the metric is more likely to be used if 𝜃̃𝑖𝑑𝑘 > 0, and less likely to be used if 𝜃̃𝑖𝑑𝑘 < 0. In this sense, ρk magnifies the ex-ante effectiveness of the metric in the use equation. By contrast, when ρk < 1, it attenuates the role of ex-ante expectations. When 𝜃̃𝑖𝑑𝑘 > 0 and ρk < 1 the metric is less likely to be used compared to ρk > 1 23

Table 4 lists each of the 22 metrics, indicates whether it is a financial or marketing metric, and shows the posterior mean of ρk estimated in equation (7). We see that financial metrics are eight of the twelve metrics with ρk > 1 and that marketing metrics are eight of the ten metrics with ρk < 1. Thus, we find that the ex-ante effectiveness of financial metrics tends to be magnified in the use equation while it is attenuated for marketing metrics. Equation (6) specifies the model for the ex-ante effectiveness of a marketing metric, 𝜃̃𝑖𝑑𝑘 . The value of 𝜃̃𝑖𝑑𝑘 is multiplied times ρk in equation (7) in determining whether metric k is used in decision d by manager i. The last column in Table 4 contains the estimated standard deviation of the error shock from equation (6), i.e. 𝜎𝜁𝑘 . When 𝜎𝜁𝑘 is large, there is more uncertainty in 𝜃̃𝑖𝑑𝑘 . Table 4 indicates that there is a negative relationship between ρk and 𝜎𝜁𝑘 ; the correlation between these two parameters is -0.57. Consequently, we find that when there is more uncertainty in the ex-ante effectiveness of a metric, its effectiveness is attenuated in the decision of whether or not to use the metric. Overall, marketing metrics tend to have higher 𝜎𝜁𝑘 and lower ρk. Hence, managers appear more uncertain about the ex-ante effectiveness of marketing as compared to financial metrics, and this attenuates the use of marketing metrics. For marketing metrics, the average 𝜎𝜁𝑘 = 2.41 and average ρk = 0.70, compared to 𝜎𝜁𝑘 = 1.43 and ρk = 1.63 for the financial metrics. Therefore, managers appear to be more confident in assessing whether a financial metric will be effective in any given decision and rely on that assessment when deciding whether or not to use the metric. We explore possible reasons for these findings in the Conclusion section. 5.3. Analyzing the Error Covariance Matrix from the Use Equation. Our integrated model of metric effectiveness and use allows us to analyze the correlation structure between marketing metrics; this sort of analysis is not possible in instrumental variable 24

or control function approaches that focus only on statistically adjusting the performance equation. For this analysis we took the posterior mean for the correlation matrix U from equation (3) and performed several exploratory factor analyses with a Varimax rotation. Table 5 contains the results from a 5 factor solution that explains 75.3% of the variance in the original matrix; additional solutions were examined and yielded similar substantive results. Highlighted cells indicate scores that are relatively high, reading across each row of the table. Seven of the twelve marketing metrics as well as ROMI and marketing expenditures for branding load highly on factor one. This suggests that these marketing metrics share common, unobserved similarities in metric usage. The second factor seems to capture consideration set as well as “marketing specific” measures of profitability: segment profitability, CLV, and share of wallet. Traditional measures of profit, net profit, ROI, NPV and EVA load highly on the third factor. The fourth factor includes measures of “volume”: target volume, market share, and total customers. Return on sales (ROS) is the only metric that loads highly on the fifth factor. The results for factor one suggest that managers may have difficulty in selecting between a number of traditional marketing metrics, awareness, satisfaction, likeability, etc. when deciding on which metric to use. There may be an opportunity to either refine these metrics or to explain how these metrics capture different aspects of consumers’ experiences and which would be more appropriate for particular decisions. 5.4. Endogeneity and Heterogeneity. Additional analyses were conducted to illustrate the importance of endogeneity and heterogeneity. First, aggregate level models without the hierarchical Bayes components, equations (5) and (6), were estimated with and without corrections for endogeneity. The posterior means for the metric effectiveness parameters are presented in Table 6; note that unlike Table 2, the results are by metric, not metric-by-decision. The first column in Table 6 represents 25

results by just regressing decision outcome on the binary indicators for which metric was used. The second column contains the results when slope endogeneity is corrected with a “use” equation analogous to equation (3) and intercept endogeneity is modeled through a correlated error structure between the two equations. Comparing the two columns, we see that share of voice was significant in the OLS analysis but is no longer significant after correcting for endogeneity; target volume and total customers are significant in the endogeneity corrected model, and awareness switches from positive and significant to negative and significant in the corrected model. If endogeneity were not present, we would not see changes in the parameter estimates. The results in Table 6 suggest that few marketing metrics matter. This is in contrast to our results in Table 2 which control for manager, firm, industry, and decision characteristics. If we did not model observed and unobserved heterogeneity, we would draw very different conclusions about metric effectiveness. To investigate the presence of endogeneity in our full model, we first looked at the posterior value of syΣYU from equation (4). This vector captures the covariance between the performance equation (1) and the use equation (3); if there is no covariance between the outcome and use equation, then there is no intercept endogeneity. We find that nine of the 22 covariance terms have 95% of their posterior mass away from 0, or are significant. Interestingly, eight of the nine significant covariance terms are for marketing as opposed to financial metrics and of these eight, seven are negative. This result suggests that for seven of the marketing metrics there are factors that improve decision outcomes, but decrease the probability that the metric is used, and vice versa. Finally, we estimate the full model but with the restriction that k in equation (3) is equal to 0, this estimates the impact of slope endogeneity. With the restriction of k = 0, 48 of the 72 significant coefficients in Table 2 are no longer significant, while 23 additional significant

26

coefficients are estimated. Again, in the absence of slope endogeneity, we would anticipate no change in parameter estimates. Estimates for syΣYU and parameters when k = 0 are contained in the web appendix.

6. Conclusion The primary substantive goal of this paper is to assess whether particular metrics are more effective for specific types of marketing-mix decisions. It answers calls advocating research to help assess what are the right metrics at the right time in order to increase marketing accountability (e.g., MSI Research Priorities 2002-2016; ISBM B-to-B Marketing Trends 20082014; Lehmann 2004, Pauwels et al. 2009, Stewart 2009, Wind 2008). Metrics quantify trends or characteristics in order to explain phenomena, understand relationships, and decision results of future actions (Farris et al. 2010). Normatively, specific metrics are used because they are believed to enhance the efficacy of a particular decision, and ultimately firm performance. However, industry, firm, and/or managerial characteristics may affect that normative ideal. Because metrics are not randomly assigned, simple regression or analysis of variance techniques are not appropriate for relating effectiveness, or outcomes, to a particular metric. We must also consider factors leading to the selection of that metric. This paper proposes a modeling framework that integrates decision performance, metric use, and ex-ante expectations and ex-post realizations of metric effectiveness via multiple equations and a parametric error structure. A new Bayesian methodology for model estimation with multiple, binary endogenous regressors and weak form rational expectations is detailed. We believe this model structure and algorithmic development will be useful in applications beyond ours which concerns marketing metrics. This paper further develops the distinction between slope and intercept endogeneity in the marketing literature and adds to the methodologies that can be used to address them. As discussed by Park and Gupta (2012) and others, there is a 27

tradeoff between the “full-information approach” adopted here to model endogeneity compared to instrumental variable/control function methods. While the current method may offer insight into the underlying model structure, it is always possible that the model is mis-specified. In our case, we tested particular parameters (e.g., syΣYU) and compared results across specifications (rk = 0 versus rk > 0) supporting the veracity of our assumed data generating model. Managerially, our model-based results help shed light on the current state of metric use and effectiveness. The results show that awareness and willingness to recommend were beneficial across a number of marketing-mix decisions, while other marketing metrics were beneficial in some situations (i.e., share of voice for PR) and detrimental in others (i.e., share of voice in price promotions). It is interesting to note that awareness and willingness to recommend are “bookends” on the customer purchase journey signaling a customer’s initial learning of a product or brand and post-evaluation recommendation to others, demonstrating the importance of such customer mindset metrics to marketing-mix performance. In addition, we find that marketing metrics on average are more effective than financial metrics for making marketingmix decisions. It may be because financial outcomes were not as paramount to the respondents in our survey. It may be the case that calculating metrics such as NPV, EVA, ROI, etc. is neither straightforward nor easy in marketing-mix decisions such as advertising, promotion, distribution, etc. Or it may be that “raising awareness” was a primary goal for a marketing-mix decision but measuring the impact on profitability is still problematic. We find that managers were more uncertain in their assessments of the ex-ante effectiveness of marketing metrics (as measured by 𝜎𝜁𝑘 ), more hesitant to use them when they thought that they were effective (k < 1), and were less discerning in differentiating between specific metrics in their decisions of which one to use (factor analysis of U). So, while there are

28

differences in marketing metric effectiveness for different marketing-mix decisions, managers don’t seem to appreciate these differences in deciding on which metric to use, and/or are unable/unwilling to translate this knowledge into selecting the right metric for the right decision. Taken together, it appears there is a role for academics and consultants to enhance the knowledge and use of various metrics in marketing-mix decisions. For financial metrics, the challenge seems to be developing and/or applying metrics that link marketing activities to financial outcomes. For marketing metrics, understanding the nuances between differing metrics and gaining the individual and/or institutional confidence to implement and rely on these metrics seems necessary. We have identified two potential drawbacks to our current methodology; additional research on these topics would be useful. The first is our multiple equation/parametric approach to addressing endogeneity. Replicating the current study using instrumental variables/control functions would be beneficial from both a substantive and methodological standpoint. In particular, if the instrument-free method of Park and Gupta (2012) can be extended to binary, endogenous regressors this would be worthwhile, at least in verifying the performance equation. Second, while we have argued that it appears necessary to rely on subjective measures of performance to examine the relationship between individual metrics and individual marketingmix decisions, it may be possible to conduct field experiments in order to isolate the effect. Our model is flexible enough that a manager or researcher could just substitute their performance measures in place of ours to examine which metrics are more and less likely to be effective and used across the situations in which the decisions are made. Future research incorporating objective performance could substantiate recommendations for what are the “right metrics” for the “right decision.” We hope such future research will build on our efforts.

29

References Ambler, T. 2003. Marketing and the Bottom Line: The Marketing Metrics to Pump Up Cash Flow 2nd ed. London, FT Prentice Hall. Armstrong, J. S., T. S. Overton. 1977. Estimating Nonresponse Bias in Mail Surveys. J. Mark. Res. 14(3) 396–402. Barnard, J., R. McCulloch, X.-L. Meng. 2000. Modeling Covariance Matrices in terms of Standard Deviations and Correlations, with Application to Shrinkage. Stat. Sin. 10(4) 1281–1312. Chib, S., E. Greenberg. 1998. Analysis of Multivariate Probit Models. Biometrika 85(2) 347– 361. Dawes, J. 1999. The Relationship between Subjective and Objective Company Performance Measures in Market Oriented Research: Further Empirical Evidence. Mark. Bull. (10) 65–75, Research Note 3. Dess, G. G., R. B. Robinson. 1984. Measuring Organizational Performance in the Absence of Objective Measures: The Case of the Privately-held Firm and Conglomerate Business Unit. Strateg. Manag. J. 5(3) 265–273. Farris, P. W., N. T. Bendle, P. E. Pfeifer, D. J. Reibstein. 2010. Marketing Metrics: The Definitive Guide to Measuring Marketing Performance 2nd ed. Upper Saddle River, New Jersey, Wharton School Publishing. Germann, F., G. L. Lilien, L. Fiedler, M. Kraus. 2014. Do Retailers Benefit from Deploying Customer Analytics? J. Retail. 90(4) 587–593. Germann, F., G. L. Lilien, A. Rangaswamy. 2013. Performance Implications of Deploying Marketing Analytics. Int. J. Res. Mark. 30(2) 114–128. Hair, J. F., R. E. Anderson, R. L. Tatham, W. C. Black. 1998. Multivariate Data Analysis. Upper Saddle River, New Jersey, Prentice Hall. Heckman, J. J. 1979. Sample Selection Bias as a Specification Error. Econometrica 47(1) 153– 161. Heckman, J., E. Vytlacil. 1998. Instrumental Variables Methods for the Correlated Random Coefficient Model: Estimating the Average Rate of Return to Schooling When the Return is Correlated with Schooling. J. Hum. Resour. 33(4) 974–987. Imbens, G. W., J. M. Wooldridge. 2009. Recent Developments in the Econometrics of Program Evaluation. J. Econ. Lit. 47(1) 5–86. Jaworski, B. J., A. K. Kohli. 1993. Market Orientation: Antecedents and Consequences. J. Mark. 57(3) 53–70. 30

Katsikeas, C. S., N. A. Morgan, L. C. Leonidou, G. T. M. Hult. 2016. Assessing Performance Outcomes in Marketing. J. Mark. 80(2) 1–20. Keeney, R. L., H. Raiffa. 1976. Decision Analysis with Multiple Conflicting Objectives. New York, Wiley & Sons. Leeflang, P. S. H., D. R. Wittink, M. Wedel, P. A. Naert. 2000. Building Models for Marketing Decisions. Dordrecht, Kluwer Academic Publishers. Lehmann, D. R. 2004. Metrics for Making Marketing Matter. J. Mark. 68(4) 73–75. Lehmann, D. R., D. J. Reibstein. 2006. Marketing Metrics and Financial Performance. Cambridge, Massachusetts, Marketing Science Institute. Lenk, P., B. Orme. 2009. The Value of Informative Priors in Bayesian Inference with Sparse Data. J. Mark. Res. 46(6) 832–845. Li, M., J. L. Tobias. 2011. Bayesian Inference in a Correlated Random Coefficients Model: Modeling Causal Effect Heterogeneity with an Application to Heterogeneous Returns to Schooling. J. Econom. 162(2) 345–361. Li, S., B. Sun, A. L. Montgomery. 2011. Cross-Selling the Right Product to the Right Customer at the Right Time. J. Mark. Res. 48(4) 683–700. Lindell, M. K., D. J. Whitney. 2001. Accounting for Common Method Variance in CrossSelectional Research Designs. J. Appl. Psychol. 86(1) 114–121. Luan, Y. J., K. Sudhir. 2010. Forecasting Marketing-Mix Responsiveness for New Products. J. Mark. Res. 47(3) 444–457. Manchanda, P., P. E. Rossi, P. K. Chintagunta. 2004. Response Modeling with Nonrandom Marketing-Mix Variables. J. Mark. Res. 41(4) 467–478. Marketing Science Institute (2002, 2004, 2006, 2008, 2010, 2012), Research Priorities: A Guide to MSI Research Programs and Procedures, Cambridge, Massachusetts: Marketing Science Institute. Menon, A., S. G. Bharadwaj, P. T. Adidam, S. W. Edison. 1999. Antecedents and Consequences of Marketing Strategy Making: A Model and a Test. J. Mark. 63(2) 18–40. Mintz, O., I. S. Currim. 2013. What Drives Managerial Use of Marketing and Financial Metrics and Does Metric Use Affect Performance of Marketing-Mix Activities? J. Mark. 77(2) 17–40. Moorman, C., R. T. Rust. 1999. The Role of Marketing. J. Mark. 63(4) 180–197. Muth, J. F. 1961. Rational Expectations and the Theory of Price Movements. Econometrica 29(3) 315–335. 31

O’Sullivan, D., A. V. Abela. 2007. Marketing Performance Measurement Ability and Firm Performance. J. Mark. 71(2) 79–93. Park, S., S. Gupta. 2012. Handling Endogenous Regressors by Joint Estimation Using Copulas. Mark. Sci. 31(4) 567–586. Pauwels, K., T. Ambler, B. H. Clark, et al. 2009. Dashboards as a Service: Why, What, How, and What Research Is Needed? J. Serv. Res. 12(2) 175–189. Pesaran, M. H., M. Weale. 2006. Chapter 14 Survey Expectations. C. W. J. G. and A. T. G. Elliott, ed. Handb. Econ. Forecast. Elsevier, 715–776. Petersen, J. A., L. McAlister, D. J. Reibstein, R. S. Winer, V. Kumar, G. Atkinson. 2009. Choosing the Right Metrics to Maximize Profitability and Shareholder Value. J. Retail. 85(1) 95–111. Stewart, D. W. 2009. Marketing Accountability: Linking Marketing Actions to Financial Results. J. Bus. Res. 62(6) 636–643. Talhouk, A., A. Doucet, K. Murphy. 2012. Efficient Bayesian Inference for Multivariate Probit Models With Sparse Inverse Correlation Matrices. J. Comput. Graph. Stat. 21(3) 739– 757. The Institute for the Study of Business Markets (2008, 2010, 2012), “B-To-B Marketing Trends,” University Park, Pennsylvania: The Institute for the Study of Business Markets. van Diepen, M., B. Donkers, P. H. Franses. 2009. Dynamic and Competitive Effects of Direct Mailings: A Charitable Giving Application. J. Mark. Res. 46(1) 120–133. Venkatraman, N., V. Ramanujam. 1987. Measurement of Business Economic Performance: An Examination of Method Convergence. J. Manag. 13(1) 109–122. Verhoef, P. C., P. S. H. Leeflang. 2009. Understanding the Marketing Department’s Influence Within the Firm. J. Mark. 73(2) 14–37. Wind, Y. 2008. A Plan to Invent the Marketing We Need Today. MIT Sloan Manag. Rev. 49(4) 21–28. Wooldridge, J. M. 2003. Further Results on Instrumental Variables Estimation of Average Treatment Effects in the Correlated Random Coefficient Model. Econ. Lett. 79(2) 185– 191.

32

Figure 1. Descriptive Statistics of Metric Use and Effectiveness by Marketing-Mix Decision. A. Proportion of times the metric was used for each decision. B. Average rating of decision if metric was used.

Use Proportion

A

0.8

Sales Force Pricing NPD PR Price Promotion Average Traditional Ad D2C Internet Ad Social Media Distribution

0.7 0.6 0.5

0.4 0.3 0.2 0.1

Tobin's Q Stock Price or Return Consideration EVA Share of Voice NPV CLV Share of Wallet Likeability Preference ROS Segment Profitability Loyalty Satisfaction ROMI Expenditures Quality Recommend Market Share Net Profit ROI Total Customers Awareness Target Volume

0

Performance

B

6.5

PR

6.0

Price Promotion Social Media

5.5

NPD Sales Force

5.0

Average

4.5

D2C

EVA

Share of Voice

Segment Profitability

Share of Wallet

Consideration Set

ROS

Satisfaction

Market Share

CLV

Loyalty

Quality

Net Profit

Recommend

Total Customers

Likeability

Awareness

NPV

Target Volume

Preference

Pricing

ROI

3.5

ROMI

Distribution

Expenditures

4.0

Legend: PR=PR/sponsorships, NPD = New Product Development, D2C=direct to consumer

33

Internet Ad Traditional Ad

Awareness Satisfaction Recommend Likeability Loyalty CLV EVA Consideration Preference Share of Wallet Share of Voice Market Share Quality ROMI Segment Profit Expend ROI Total Customers Net Profit ROS Target Volume NPV

Average Effectiveness

Figure 2. Posterior Distribution of Metric Effectiveness. A. Boxplots of posterior means, B. Averages of posterior means by decision. A

B

1.5

1.0

0.5

0.0

-0.5

-1.0

-1.5

-2.0

-2.5

-3.0

Metric

34

Table 1. Manager, Firm, Industry, and Decision Characteristics Used in Models Equation 1: Decision Performance

Equation 3: Metric Use

Equations 5 & 6: Metric Effectiveness

z z

w w

Managerial Metric Orientation Metric-based Compensation Metric Training Managerial Characteristics Marketer (vs. Non-Marketer)

x

z

w

TMT (vs. mid-level manager)

x

z

w

Work Experience Quantitative Background Firm Strategy Market Orientation Analyzer (vs. Prospectors) Low-Cost Defender (vs. Prospectors) Differentiated Defender (vs. Prospectors) Organization Involvement Firm Characteristics Firm Size Public Ownership (vs. Private) CMO Presence Recent Business Performance B2C (vs. B2B) Services (vs. Goods) Industry Characteristics Mature/Declining Product Life Cycle (vs. Intro/Growth) Industry Concentration Industry Growth Industry Turbulence Marketing-mix Decision Traditional Advertising Internet Advertising Direct to Consumer Social Media Price Promotions Pricing New Product Development Sales Force Distribution PR/Sponsorships

x

z z

w w

x

z z z

w w w

z

w

z

w

x x

z z z z z z

w w w w w w

x

z

w

x x

z z z

w w w

x

w w w w w w w w w w

35

Int. Adv.

Direct to Cons.

Social Media

Price Promo.

Pricing

New Prod Dev.

Sales Force

Distribution

PR / Sponsor.

Type of Metric

Trad. Adv.

Table 2. Type of Decision’s Impact on Metric Effectiveness

-0.74

-0.63

-0.55

-1.03

-0.18

0.03

-0.25

-0.38

-0.27

-1.07

-0.16

-0.01

-0.08

-0.50

-0.29

-0.33

0.05

-0.20

-0.54

-0.38

-0.33

-0.14

-0.09

-0.67

0.30

0.11

-0.49

0.59

0.46

-1.26

0.29

0.36

0.10

-0.47

-0.19

-0.47

-0.79

-0.33

-0.17

-0.26

-1.57

-0.69

-0.91

-1.80

-0.67

-0.27

0.23

-1.09

-1.03

-1.38

0.06

1.08

0.42

0.19

-0.03

1.41

2.51

1.24

0.53

0.08

0.44

0.16

0.01

0.09

-0.25

-0.30

-0.49

-0.30

-0.17

0.15

-0.68

-0.76

-0.72

-0.94

-0.11

-0.39

-0.26

-0.27

-0.38

-1.11

-0.25

0.33

-0.01

-0.02

-0.53

-0.05

-0.02

0.28

0.28

-0.69

-0.60

-0.73

-0.13

-0.71

-0.20

-0.09

-0.23

-0.58

-0.23

-1.19

-0.09

-0.44

-0.55

-0.52

-0.09

0.12

0.14

-0.01

-0.24

-0.60

Awareness

1.41

1.16

1.16

1.26

0.79

0.47

0.71

0.73

0.47

1.44

Satisfaction

0.17

0.48

0.43

0.24

0.24

1.14

0.88

0.62

1.41

-0.28

Likeability

0.03

0.24

-0.03

0.87

-0.01

-0.64

0.15

-0.22

0.34

0.25

Preference

-0.08

-0.39

-0.20

-0.14

-0.34

0.79

0.12

-0.06

-1.62

0.07

Willingness to Recommend

0.55

0.73

1.00

0.84

0.88

0.79

0.78

0.82

1.16

0.65

Loyalty

0.53

0.42

0.90

0.19

1.50

0.29

0.43

0.37

-0.36

0.59

Quality

-0.71

-0.80

-0.38

-0.18

-0.21

0.53

0.06

0.27

0.33

-0.09

Consideration Set

-1.06

0.84

0.28

1.14

1.35

0.59

1.28

0.56

-0.31

0.10

Total Customers

-0.67

-0.73

-0.70

-0.57

-0.02

-0.43

-0.34

-0.05

0.12

-0.59

Share of Wallet

-0.47

-0.17

-0.29

-1.31

0.08

0.40

0.24

0.06

-0.30

-0.65

Share of Voice

0.50

0.17

-0.30

0.84

-2.85

-0.01

-0.18

-0.14

0.25

1.14

Metric

Financial Metrics

Net Profit Return on Investment (ROI) Return on Sales (ROS) Return on Mktg. Investment (ROMI) Net Present Value (NPV) Economic Value Added (EVA) Mktg. Expend. on Branding Target Volume Segment Profitability Customer Lifetime Value (CLV)

Marketing Metrics

Market Share

Notes: Bolded numbers indicated significant coefficient; Marketing metrics are defined as metrics that are based on a customer mindset while financial metrics are defined as metrics that are either monetary or readily converted to monetary based outcomes (Mintz and Currim 2013).

36

Table 3. Managerial, Firm and Industry Characteristics’ Impact on Metric Effectiveness

Marketing Metrics

0.53 -0.42 0.10 0.13 -0.41 0.01 0.39 0.10 0.20 -0.21 -0.08 0.10 0.10 -0.03 -0.10 -0.20 -0.01 -0.36 0.08 0.00 0.11 -1.07 0.25 0.20 0.13 0.07 -0.20 0.10 -1.70 0.68 0.28 -0.65 0.87 0.48 -0.26 2.29 -1.40 -0.57 0.72 0.31 0.48 -0.64

0.01 0.38 -0.42 0.05 0.21 -0.05 0.14 -0.05 0.18 0.15 -0.06 -0.18 -0.01 0.15 0.25 -0.01 -0.12 -0.74

-0.24 0.16 0.23 -0.07 0.06 0.03 0.18 -0.37 -0.01 0.08 0.13 1.06 -0.04 0.43 -0.05

0.41 -0.01 -0.22 -0.45 0.23 -0.44 -0.06 0.56 0.18 0.24 0.19 0.74 0.27 -0.97 -0.39

0.27 -0.22 -0.15 0.30 0.02 -0.29 -0.05 0.07 -0.08 0.20 -0.21 0.55 -0.22 -0.54 0.07 0.29 -0.22 -1.60 0.12 0.13 0.33

37

0.18 -0.17 -0.06 -0.02 0.29 0.08 -0.05 -0.10 -0.20 0.34 -0.19 -0.14 0.10 0.19 0.34

-0.21 0.35 0.43 0.11 -0.19 0.01 0.25 0.10 -0.32 -0.31 0.20 1.82 0.12 -0.11 0.03

0.29 0.14 -1.50 -0.24 0.06 -0.13 -0.42 -0.51 0.14 0.50 -0.25 -0.42 -0.20 0.01 0.19

0.04 0.11 0.49 -0.20 -0.18 -0.07 -0.08 0.12 0.08 -0.05 -0.03 0.87 -0.20 -0.35 0.06

0.02 -0.17 0.48 -0.01 -0.05 -0.28 -0.36 -0.12 0.18 0.06 0.12 -0.81 -0.11 0.06 0.72

PLC (Mat. /Decl.) Ind. Conc. Ind. Growth Ind. Turb.

Services

B2C

Rec. Bus. Perf.

0.27 -0.42 0.48 0.29 0.19 -1.00

CMO

-0.48 0.09 -0.20 -0.33 -0.23 1.42

Industry

Low Cost Defend1 Diff. Defend1 Org. Invol. Firm Size (ln) Public Owner

Analyzer1

Financial Metrics

0.10 -0.40 0.22 0.12 -0.18 0.35 Net Profit 0.06 -0.01 0.10 -0.03 0.08 0.12 ROI 0.06 0.29 0.13 -0.41 -0.14 -0.23 ROS 0.00 0.09 -0.63 -0.33 -0.01 -0.02 ROMI 0.04 0.52 -0.90 -1.04 -1.10 -1.02 NPV 0.52 -0.21 0.89 -0.07 0.54 -0.18 EVA Mktg. Brand -0.09 0.05 -0.10 -0.13 0.01 -0.18 Expend. 0.00 0.10 0.09 0.27 -0.08 0.18 Target Vol. Cust. Seg. Profit 0.19 -0.07 0.23 0.34 0.02 0.26 -0.03 -0.12 0.53 0.04 -0.06 -0.13 CLV -0.23 0.10 -0.27 -0.24 0.07 0.00 Market Share 0.09 0.03 0.06 -0.18 0.17 -0.03 Awareness -0.44 -0.15 -0.22 0.37 -0.20 -0.32 Satisfaction -0.24 0.40 0.07 0.07 -0.20 -0.34 Likeability 0.49 -0.54 0.57 -0.14 0.34 0.41 Preference -0.45 0.37 -0.40 -0.26 0.06 -0.02 Will. to Rec. 0.08 -0.04 0.05 -0.14 -0.12 0.49 Loyalty 0.26 -0.29 0.17 0.29 -0.08 0.33 Quality Consideration Set 0.34 -0.33 0.07 0.69 -0.34 0.22 Total Customers 0.17 0.17 0.01 -0.09 0.10 0.04 -0.11 -0.50 0.04 -0.04 -0.16 -0.31 Shr. of Wallet -0.09 0.17 0.08 0.31 -0.10 0.12 Shr. of Voice Note: Bolded numbers indicated significant coefficient.

Market Orient.

Quant.

Firm Work Exp.

TMT

Marketer

Metric

Manager Metric Comp. Metric Train.

Characteristic

-0.17 0.31 0.11 -0.08 0.18 0.14 -0.02 0.08 -0.30 0.24 0.51 0.20 0.76 -0.42 0.48 -0.31 0.35 -0.31

0.15 -0.17 0.37 0.16 0.25 -0.93

-0.06 0.08 -0.04 0.17 -0.01 -0.79

-0.05 0.11 -0.03 -0.09 -0.38 0.18

0.13 -0.16 -0.03 0.07 0.00 0.06 0.17 0.17 0.08 0.06 0.08 0.24 0.36 -0.31 -0.11 0.06 -0.09 -0.54 -0.11 -0.10 -0.13

-0.38 0.28 0.47 -0.24 -0.13 -0.04 -0.01 -0.06 -0.05 0.16 -0.05 0.32 0.09 -0.05 0.07

-0.06 -0.22 -0.05 -0.06 -0.14 -0.05 0.26 -0.01 -0.10 -0.03 0.09 -0.84 -0.01 0.40 -0.05

-0.02 0.21 0.24 -0.22 -0.22 0.10 0.57 -0.21 -0.25 0.03 0.11 0.03 -0.14 -0.02 0.32

-0.05 0.11 0.02 0.14 -0.03 0.29 0.15 -0.01 0.04 0.37 -0.45 -0.27 0.03 0.18 -0.22

0.07 -0.25 0.59 -0.05 0.06 -0.12 0.29 -0.15 0.07 0.02 -0.08 -0.14 -0.06 0.09 0.05

Table 4. ρ Multiplier for Metric Use and Uncertainty in Ex-Ante Effectiveness Posterior Mean Type of 𝝈𝜻𝒌 Metric ρk Metric Preference Marketing 0.20 2.98 Willingness to Recommend Marketing 0.29 2.73 Loyalty Marketing 0.31 3.03 Consideration Set Marketing 0.36 9.47 Customer Segment Profitability Financial 0.40 3.04 Satisfaction Marketing 0.41 1.81 Quality Marketing 0.49 1.75 Likeability Marketing 0.56 1.36 EVA Financial 0.79 3.74 Share of Voice Marketing 1.00 1.53 Total Customers Marketing 1.03 1.12 NPV Financial 1.03 1.69 ROS Financial 1.04 1.48 Awareness Marketing 1.06 0.66 Share of Customer Wallet Marketing 1.14 1.79 CLV Financial 1.23 1.38 Market Share Marketing 1.53 0.66 Marketing Expenditures on Branding Financial 1.91 0.60 ROMI Financial 2.06 0.94 Target Volume Financial 2.26 0.46 ROI Financial 2.27 0.52 Net Profit Financial 2.69 0.46 Note: Marketing metrics are defined as metrics that are based on a customer mindset while financial metrics are defined as metrics that are either monetary or readily converted to monetary based outcomes (Mintz and Currim 2013).

38

Table 5. Factor Analysis of ΣU Metric Satisfaction Likeability Awareness Loyalty Recommend Preference Marketing Expenditures on Branding Quality ROMI Consideration Set Share of Wallet CLV Segment Profitability NPV ROI EVA Net Profit Target Volume Total Customers Market Share ROS Share of Voice Eigenvalue Percent

Factor 1 .835 .811 .789 .766 .758 .749

Factor 2 .023 .099 .138 .213 .334 .432

Factor 3 .267 .363 .162 .121 .082 .257

Factor 4 .278 .027 -.028 .290 .147 -.013

Factor 5 -.065 .050 -.066 .012 .177 .199

Communality 0.850 0.803 0.672 0.731 0.746 0.853

.693

.439

-.071

.097

.157

0.712

.668 .663

.322 -.039

.105 .462

.235 .165

.327 .052

0.723 0.684

.416

.746

-.036

-.059

.012

0.734

.223 .109 .113

.849 .794 .616

.186 .191 .189

.103 .231 .503

-.126 -.206 .298

0.831 0.774 0.769

.175 .217 .389 .273 .080 .178

.610 .055 .275 .213 .055 .074

.710 .832 .659 .535 .418 .010

-.016 .124 .303 .413 .711 .808

.009 .151 -.077 .221 .174 -.251

0.907 0.781 0.759 0.626 0.720 0.753

.229 .350 .288 5.911 26.870

.418 -.162 .457 3.871 17.594

.121 .313 .400 3.019 13.721

.505 .050 .173 2.320 10.546

.317 .753 -.492 1.444 6.566

0.597 0.816 0.724

Factors were rotated using a Varimax procedure to improve interpretability. Eigenvalues and communalities calculated using the rotated factor solution.

39

Table 6. Posterior Means of Metric Effectiveness in Models without Heterogeneity Metric No Endogeneity Endogeneity Net Profit 0.00 0.07 Return on Investment (ROI) 0.04 0.71 Return on Sales (ROS) 0.09 0.35 Return on Mktg. Investment (ROMI) 0.02 -0.09 Net Present Value (NPV) -0.21 -0.31 Economic Value Added (EVA) 0.05 -0.24 Mktg. Expend. on Branding -0.17 -0.69 Target Volume 0.09 0.81 Segment Profitability 0.28 0.51 Customer Lifetime Value (CLV) 0.10 0.01 Market Share 0.00 -0.31 Awareness 0.13 -0.71 Satisfaction 0.10 -0.31 Likeability -0.03 -0.28 Preference -0.08 -0.37 Willingness to Recommend 0.07 -0.41 Loyalty 0.01 -0.13 Quality 0.06 0.14 Consideration Set 0.08 0.09 Total Customers 0.05 0.76 Share of Wallet 0.18 -0.09 Share of Voice -0.41 0.27 Note: Bolded numbers indicate significant coefficient

40

Web Appendix The appendix details the full conditional distribution for the MCMC algorithm. We use the notation “|Rest” for the distribution of parameter  given all observables and all other parameters except . It is convenient to rewrite manager i’s utility in Equation (3) as a vector: 𝒖𝑖 = 𝚫′𝒛𝑖 + 𝑹′𝜽𝑖 + 𝝂𝑗 for 𝑖 = 1, … , 𝑛

9

where 𝚫 = [𝜹1 , … , 𝜹𝐽 ], and 𝑹 = 𝑑𝑖𝑎𝑔[𝜌1 , … , 𝜌𝐽 ], a matrix with 0 on the off-diagonals and 𝜌𝑗 for the (j,j) element. The entire data for the latent utilities can be written by stacking the transpose of Equation (9): 𝑼 = 𝒁𝚫 + 𝚯𝑹 + 𝚬𝑼

10

where row i is ui’ for U, zi’ for Z, i’ for , and i’ for U. Similarly, Equation (5) can be compactly written for all observations: 𝚯 = 𝑾𝚽 + 𝚬𝚯

11

where row i is wi’ for W, and i’ for 𝚬𝚯 . Equation (1): Full Conditional of 𝜷 𝜷|Rest

𝑁(𝝁𝛽𝑛 , 𝑽𝛽𝑛 )

~

𝑛

𝑽𝛽𝑛

=

−1 (𝑽𝛽0

+ 𝜎

−2 ∑

12 −1

𝒙𝑖 𝒙′𝑖 )

𝑖=1 𝑛

𝝁𝛽𝑛

=

−1 𝑽𝛽𝑛 (𝑽𝛽0 𝝁𝛽0 + 𝜎 −2 ∑ 𝒙𝑖 [𝑦𝑖 − 𝒎′𝑖 𝜽𝑖 ]) 𝑖=1

Equations (1), (3) and (5): Full Conditional of 𝜽𝒊 𝜽𝑖 |𝑅𝑒𝑠𝑡

~

𝑽𝜃𝑖

=

𝝁𝜃𝑖

=

𝑁 (𝝁𝜃𝑖 , 𝑽𝜃𝑖 ) (𝚲−1 + 𝜎 −2 𝒎𝑖 𝒎′𝑖 + 𝑹′𝚺 −1 𝑹)

13 −1

𝑽𝜃𝑖 (𝚲−1 𝚽 ′ 𝒘𝑖 + 𝜎 −2 𝒎𝑖 [𝑦𝑖 − 𝒙′𝑖 𝜷] + 𝑹′𝚺 −1 [𝒖𝑖 − 𝚫′𝒛𝑖 ])

1

Equation (1): Full conditional of 𝝈𝟐 𝜎2 𝑟𝜎𝑛

~ =

𝑠𝜎𝑛

= 𝑠𝜎0 + ∑(𝑦𝑖 − 𝒙′𝑖 𝜷 + 𝒎′𝑖 𝜽𝑖 )

𝐼𝐺 (𝑟𝜎𝑛 ⁄2 , 𝑠𝜎𝑛 ⁄2) 𝑟𝜎0 + 𝑛

14

𝑛

2

𝑖=1

Equations (2) and (3): Impute the Latent Utility ui We sequentially generate uij from truncated, conditional normal distributions. Define ui(j) to be the vector of latent utilities ui with row j deleted; 𝝁𝑖 = 𝚫′𝒛𝑖 + 𝑅′𝜃𝑖 ; ij to be row j of 𝝁𝑖 ; 𝝁𝑖(𝑗) to be 𝝁𝑖 with row j deleted; jj to be the (j,j) element of ; (jj) to be  with row j and column j deleted; and j(j) to be row j of  with column j deleted. For j = 1, …, J sequentially generate 𝑢𝑖𝑗 |𝑅𝑒𝑠𝑡

~

𝑁(𝜇𝑖𝑗|(𝑗) , Σ𝑗|(𝑗) )𝜒[(2𝑚𝑖𝑗 − 1)𝑢𝑖𝑗 > 0]

𝜇𝑖𝑗|(𝑗)

=

−1 [𝒖𝑖(𝑗) − 𝝁𝑖(𝑗) ] 𝜇𝑖𝑗 + 𝚺𝑗(𝑗) 𝚺(𝑗𝑗)

Σ𝑗|(𝑗)

=

−1 ′ Σ𝑗𝑗 − 𝚺𝑗(𝑗) 𝚺(𝑗𝑗) 𝚺𝑗(𝑗)

15

We use the normal, inverse cdf transform to generate from the truncated normal distribution. In particular, let F be the univariate, normal cumulative distribution function (cdf) with means and variances as in Equation (15). If mij = 1, then uij is positive. The cumulative, truncated normal cdf for uij is: G(u) = [F(u) – F(0)] / [1 – F(0)]. If mij = 0, then uij is negative, and its cumulative distribution function if G(u) = F(u)/F(0). Then G(uij) has a uniform distribution. Therefore, to generate uij first generate a uniform random variable  and invert G(uij) = : 𝑢𝑖𝑗 = {

𝐹 −1 [𝐹 (0) + 𝜉 {1 − 𝐹(0)}] 𝐹 −1 [𝜉𝐹(0)]

if 𝑚𝑖𝑗 > 0 if 𝑚𝑖𝑗 ≤ 0.

Equations (3) and (10): Full conditional of  vec(𝚫′)|Rest ~ 𝑽Δ𝑛 = 𝝁Δ𝑛 =

𝑁(𝝁Δ𝑛 , 𝑽Δ𝑛 ) + 𝒁′𝒁⨂𝚺 −1 ]−1 −1 𝑽Δ𝑛 {𝑽−1 Δ0 𝝁Δ0 + (𝒁′⨂𝚺 )vec[(𝑼 − 𝚯𝑹)′]} [𝑽−1 Δ0

2

16

where the Kronecker product is ⨂, and “vec” stacks the columns of the matrix into a vector. Equations (3) and (10): Full Conditional of j Define  = (1,…, J)’ and i = diag(i), the matrix with zero off-diagonal elements and i on the diagonal. The full conditional is: 𝐽

𝝆|Rest

17

𝑁(𝝁𝜌𝑛 , 𝑽𝜌𝑛 ) ∏ 𝜒[𝜌𝑗 > 0]

~

𝑗=1 𝑛

𝑽𝜌𝑛

−1

−1 [𝑽𝜌0 + ∑ 𝚨′𝑖 𝚺 −1 𝚨𝑖 ]

=

𝑖=1 𝑛

𝝁𝜌𝑛

−1 𝑽𝜌𝑛 [𝑽𝜌0 𝝁𝜌0 + ∑ 𝚨′𝑖 𝚺 −1 (𝐮𝑖 − 𝚫′𝒛𝑖 )] .

=

𝑖=1

These coefficients are constrained to be positive. Define  = (1,…, J)’ and (j) to be  without the row j. Next, we sequentially generate each j from their conditional normal distribution in order to include the truncation. Define (j) and (j) to be  and 𝝁𝜌𝑛 without the row j. Define Vjj to be the (j,j) element of 𝑽𝜌𝑛 ; μj as the jth element of μρn;V(jj) to be 𝑽𝜌𝑛 after removing row j and column j; and Vj(j) to be row j of 𝑽𝜌𝑛 after deleting column j. For j = 1, …, J sequentially generate j from conditional normal distributions: 𝜌𝑗 |𝑅𝑒𝑠𝑡

~

𝑁(𝜇𝑗|(𝑗) , 𝑉𝑗|(𝑗) )𝜒(𝜌𝑗 > 0)

𝜇𝑗|(𝑗)

=

𝜇𝑗 + 𝑽𝑗(𝑗) 𝑽−1 (𝑗𝑗) [𝝆(𝑗) − 𝝁(𝑗) ]

𝑉𝑗|(𝑗)

=

′ 𝑉𝑗𝑗 − 𝑽𝑗(𝑗) 𝑽−1 (𝑗𝑗) 𝑽𝑗(𝑗) .

We use the inverse cdf transform to generate the truncated normal random variables: 𝜌𝑗 = 𝐹 −1 [𝐹 (0) + 𝜉 {1 − 𝐹(0)}] where  is a uniform random variables and F is the normal cdf for Equations (17) and (18).

3

18

Equations (3) and (10): Full Conditional of  The algorithm for generating  adapts the parameter expansion missing data procedure from Talhouk, Doucet, and Murphy (2012). The first step is to generate the “missing” variance parameters: 𝐽 + 1 Σ𝑗𝑗 𝑑𝑗2 ~𝐼𝐺 ( , ) for 𝑗 = 1, … , 𝐽 2 2

19

where Σ𝑗𝑗 is the (j,j) element of 𝚺 −1 . Define the matrix D = diag(d1, …, dJ) with zeros on the offdiagonals and dj in element (j,j). Set bi = Dui and B as the stacked vector of bi. The covariance matrix of bi is  = DD, and the variances are no longer equal to one. The key point of the method is that given the prior for 𝚺 in (Barnard et al. 2000) and the definition of D in Equation (19), the full conditional distribution of  given D is inverted Wishart with known parameters: 𝛀|Rest ~ 𝑓Ω𝑛 = −1 𝑮Ω𝑛 =

𝐼𝑊 (𝑓Ω𝑛 , 𝑮−1 Ω𝑛 ) 𝐽−1+𝑛 𝑰𝐽 + (𝑩 − 𝒁𝚫𝑫 − 𝚯𝑹𝑫)′(𝑩 − 𝒁𝚫𝑫 − 𝚯𝑹𝑫)

20

where IJ is a J by J identity matrix. Finally, we obtain the correlation matrix 𝚺 from the “covariance” matrix  by dividing by the “standard deviations” j = jj1/2 where jj is the (j,j) element of  :  = -1-1 where  = diag(1 , …,J ). Equation (5) and (11): Full Conditional of  vec(𝚽′)|Rest 𝑽Φ𝑛 𝝁Φ𝑛

~ 𝑁(𝝁Φ𝑛 , 𝑽Φ𝑛 ) −1 [𝑽Φ0 + 𝑾′𝑾⨂𝚲−1 ]−1 = −1 = 𝑽Φ𝑛 [𝑽−1 Φ0 𝝁Φ0 + (𝑾′⨂𝚲 )vec(𝚯′)]

21

Equation (5) and (11): Full conditional of  𝚲|Rest ~ 𝑓Λ𝑛 = −1 𝑮Λ𝑛 =

𝐼𝑊 (𝑓Λ𝑛 , 𝑮−1 Λ𝑛 ) 𝑓Λ0 + 𝑛 −1 𝑮Λ0 + (𝚯 − 𝑾𝚽)′(𝚯 − 𝑾𝚽)

4

22

Web Appendix Table 1. Definition of Constructs and Operational Measures Construct Basis

Definition and Operational Measures

Market Orientation (Deshpande & Farley 1998; Kohli & Jaworski 1990; Verhoef & Leeflang 2009)

Definition: The extent to which a firm measures, monitors, and communicates customer needs and experiences throughout the firm and whether the firm’s strategy is based on this information. Measures: How strongly do you agree or disagree with each of the following statements: (1 = strongly disagree, 7 = strongly agree)  Our business objectives are driven primarily by customer satisfaction  We constantly monitor our level of commitment and orientation to serving customer needs  We freely communicate information about our successful and unsuccessful customer experiences throughout all business functions  Our strategy for competitive advantage is based on our understanding of customer needs  We measure customer satisfaction systematically and frequently  We have routine or regular measures for customer service  We are more customer focused than our competitors  I believe this business exists primarily to serve customers Definition: The strategy which a firm employs to compete in an industry or market, categorized based on two dominant frameworks of strategic orientation, the Miles and Snow (1978) typology which focuses on the firm’s intended rate of productmarket change, and the Porter (1980) typology, which focuses on the firm’s differentiation or cost advantage. Measures: Please select one of the following descriptions that best characterizes your organization:  Prospectors: These firms are frequently the first-to-market with new product or service concepts. They do not hesitate to enter new market segments in which there appears to be an opportunity. These firms concentrate on offering products that push performance boundaries. Their proposition is an offer of the most innovative product, whether it is based on substantial performance improvement or cost reduction.  Analyzers: These firms are seldom first-in with new products or services or first to enter emerging market segments. However, by monitoring market activity, they can be early followers with a better targeting strategy, increased customer benefits, or lower costs.  Low-Cost Defenders: These firms attempt to maintain a relatively stable domain by aggressively protecting their product market position. They rarely are at the forefront of product of service development; instead, they focus on producing goods or services as efficiently as possible. In general, these firms focus on increasing share in existing markets by providing products at the best prices.  Differentiated Defenders: These firms attempt to maintain a relatively stable domain by aggressively protecting their product market position. They rarely are at the forefront of product or service development; instead, they focus on providing superior service and/or product quality. Their prices are typically higher than the industry average. Definition: The extent to which a firm’s marketing-mix decision or action is based on involvement of a wide range of managers across functions. Measures: How strongly do you agree or disagree with each of the following statements: (1 = strongly disagree, 7 = strongly agree)  This marketing action was a real company-wide effort  People from all over the organization were involved in this marketing action  A wide range of departments or functions in the company got involved in this marketing action Definition: The importance of metrics in a manager’s compensation package. Measures: Please indicate how important each metric type is related to your compensation package: (1= not at all important, 7 = extremely important)  Overall Metrics  Marketing Metrics  Financial Metrics Definition: A manager’s level of training on the use of metrics. Measures: Please indicate your level of training with metrics (can be through work or educational experiences): (1= much less than average amount of training, 7 = much more than average amount of training)  Overall Metrics

Strategic Orientation (Olson, Slater, & Hult 2005; Slater & Olson 2000)

Organizational Involvement (Noble & Mokwa 1999)

Metric-based Compensation

Metric-based Training

5

α

Mean

St. Dev.

0.86

5.03

1.14

0.29

---

0.24

---

0.12

---

0.35

---

0.94

3.80

1.70

0.82

4.90

1.50

0.94

4.45

1.68

N/A

Functional Area (Finkelstein, Hambrick, & Cannella 2009)

Managerial Level (Finkelstein, Hambrick, & Cannella 2009)

Managerial Experience

Quantitative Background

Firm Size Type of Ownership (Verhoef & Leeflang 2009) CMO Presence Recent Business Performance (Jaworski & Kohli 1993)

B2B vs. B2C (Verhoef & Leeflang 2009) Goods vs. Services (Verhoef & Leeflang 2009) Product Life Cycle (Deshpande & Zaltman 1982) Industry Concentration (Kuester, Homburg, & Robertson 1999) Market Growth (Homburg, Workman, & Krohmer 1999)

Market Turbulence (Miller, Burke, & Glick 1998)

 Marketing Metrics  Financial Metrics Definition: Whether a manager works in the marketing department. Measures: Please indicate your job title: CEO/Owner, CMO, C-Level (Other than Marketing), SVP/VP of Marketing, SVP/VP Sales, SVP/VP (Other than Marketing and Sales), Director of Marketing, Director of Sales, Brand Manager, Marketing Manager, Product Manager, Sales Manager, Other (Please list) Definition: Whether a manager is (a) VP-level or higher (e.g., SVP, C-level or Owner) or (b) lower than VP-level (e.g., Director, Manager). Measures: Please indicate your job title: CEO/Owner, CMO, C-Level (Other than Marketing), SVP/VP of Marketing, SVP/VP Sales, SVP/VP (Other than Marketing and Sales), Director of Marketing, Director of Sales, Brand Manager, Marketing Manager, Product Manager, Sales Manager, Other (Please list) Definition: A manager’s experience in number of years as a manager, at the firm, and in the current position. Measures: How many years of managerial experience do you have? How many years have you been working for this company? How many years have you been working at your current position? Definition: A manager’s qualitative/quantitative orientation based on education and work experience. Measures: Please rate your qualitative/quantitative background: (1 = entirely qualitative, 7 = entirely quantitative)  Overall orientation  Educational Background  Work Experience Background Definition: The number of full-time employees in a firm. Measure: Approximately how many full-time employees does your firm have? Definition: Whether a firm is publicly traded or privately held. Measure: Is your firm publicly traded? Definition: Whether a firm employs a Chief Marketing Officer (CMO). Measure: Does your firm employ a Chief Marketing Officer (CMO)? Definition: A business unit’s overall performance last year, relative to its own expectations and its competitors’ performance. Measures: To what extent did the overall performance of the business unit meet expectations last year: (1= poor, 7=excellent) To what extent did the overall performance of your business unit relative to your major competitors meet expectations last year: (1= poor, 7=excellent) Definition: The extent to which a manager’s sales come from B2B or B2C markets. Measure: Please indicate the extent to which your sales come from B2B or B2C markets: (1 = mostly B2B, 7 = mostly B2C) Definition: The extent to which a manager’s sales come from goods or services markets. Measure: Please indicate the extent to which your sales come from goods or services markets: (1 = mostly goods, 7 = mostly services) Definition: The stage of the product life cycle. Measure: At which one of the following stages would you place your product? (shown in a product life cycle diagram, introductory, growth, maturity, decline) Definition: The percentage of sales the four largest businesses competing in a market control. Measure: Approximately what percentage of sales does the largest 4 competing businesses in your market control?  0-50%, 51-100% Definition: The average annual growth or decline of the company and the industry over the last three years. Measure: Over the last three years, what was the average annual market growth or decline for your company? Over the last three years, what was the average annual market growth or decline for your industry? Definition: The rate at which products or services become obsolete, the ease of forecasting consumer preferences, and how often a firm needs to change its marketing and production/service technology to keep up with competitors and/or consumer preferences.

6

N/A

0.54

---

N/A

0.58

---

0.68

9.54

5.68

0.85

4.31

1.11

N/A

5.35

---

N/A

0.22

---

N/A

0.29

---

0.84

5.34

1.30

N/A

2.91

---

N/A

4.68

---

N/A

0.55

---

N/A

0.43

---

0.66

5.23

1.87

0.63

4.29

1.07

Marketing-mix Decision (Menon et al. 1999)

Metrics Use (Ambler 2003; Ambler et al. 2004; Barwise & Farley 2003; Du et al. 2007; Farris et al. 2010; Hoffman & Fodor 2010; Lehmann & Reibstein 2006; Pauwels et al. 2009; Srinivasan et al. 2010) Marketing-mix Activity Performance (Jaworski & Kohli 1993; Moorman & Rust 1999; Verhoef & Leeflang 2009)

Measures: How strongly do you agree or disagree with each of the following statements (1 = strongly disagree, 7 = strongly agree): ® = reverse scored  Products/services become obsolete very slowly in your firm’s principal industry ®  Your firm seldom needs to change its marketing practices to keep up with competitors ®  Consumer demand and preferences are very easy to forecast in your firm’s principal industry ®  Your firm must frequently change its production/service technology to keep up with competitors and/or consumer preferences Definition: A major marketing-mix decision undertaken not so recently that performance evaluation is premature and not so long ago that memory of the decision and its performance is fuzzy. Measures: Please indicate which types of major marketing decisions you have undertaken (or implemented) that (1) were not so recent that performance evaluation is premature and (2) not so long ago that memory about the decision and performance is fuzzy:  Traditional Advertising (i.e., TV, Magazine, Radio, etc.)  Internet Advertising (i.e., Banner Ads, Display Ads, SEO, etc.)  Direct to Consumer (i.e., Emails, CRM, Direct mail, etc.)  Social Media (i.e., Twitter, Facebook, MySpace, etc.)  Price Promotions  Pricing  New Product Development  Sales Force  Distribution  PR/Sponsorships Metric Use Definition: A metric is defined to be used in a marketing-mix decision if a manager employed the metric as a decision aid when making the marketing-mix decision. Measure: Please indicate if you used any of the following MARKETING or FINANCIAL metrics when making your marketing-mix decision: See Table 2 for the list of metrics.

Definition: The performance of a marketing-mix activity is defined based on a firm’s stated marketing, financial, and overall outcomes, relative to a firm’s stated objectives and to similar prior decisions. Measures: Relative to your firm’s stated objectives, how is the last major marketing activity undertaken performing overall? (1=much worse, 7=much better) Relative to similar prior marketing activities you've undertaken, how is the last major marketing activity undertaken performing? (1=much worse, 7=much better; N/A if unsure or never undertook activity) Relative to your firm’s stated objectives, how is the last major marketing activity undertaken performing on: (1=much worse, and, 7=much better; N/A if unsure)  Customer satisfaction  Profitability  Customer loyalty  Sales  Market share  ROI Note: The first 3 columns in the table are taken from (Mintz and Currim 2013)

7

0.11 0.12 0.17 0.11 0.10 0.05 0.08 0.11 0.04 0.12

---------------------

N/A

---

---

0.94

4.90

1.06

N/A

Web Web Appendix Table 2. Covariance between Performance and Use Equation, syΣYU Net Profit 0.09 ROI 0.16 ROS 0.05 ROMI -0.15 NPV 0.02 EVA -0.11 Mktg. Brand Expend. -0.29 Target Vol. 0.31 Cust. Seg. Profit 0.15 CLV 0.02 Market Share -0.03 Awareness -0.80 Satisfaction -0.61 Likeability -0.67 Preference -0.53 Will. to Rec. -0.65 Loyalty -0.53 Quality -0.46 Consideration Set -0.29 Total Customers 0.26 Shr. of Wallet -0.11 Shr. of Voice -0.22 Note: Bolded numbers indicated significant covariance.

8

Target Volume Segment Profitability Customer Lifetime Value (CLV) Market Share

Pricing

New Prod Dev.

Sales Force

Distribution

PR / Sponsor.

0.11

0.97

0.58

1.48

0.34

0.19

0.50

0.46

-0.05

2.30

0.62

1.03

0.45

0.77

0.82

0.90

0.55

0.34

2.16

1.35

0.44

1.20

1.08

0.28

1.44

0.51 -0.19

0.83

0.34 -1.06

-0.01

0.22

0.63

0.37

0.38

0.94

0.71

1.58

1.22 -0.09

0.07 -0.33 -0.85

-0.29

1.52 -1.57

-0.87 -0.57 -0.93 -0.26 0.78

2.18

3.45

4.50

0.58

0.61

3.19

2.09

0.33 -1.14

0.59

0.63

0.56

0.51

2.13

0.49

0.78

0.64

3.28

1.31

0.12

0.35

0.11

0.44

0.29

0.03

0.43

0.23 -0.34

0.33 -0.25

1.14

-0.77

0.31 -0.04

-0.42

-1.28

-0.42 -0.39 -0.14

-0.66

-0.32 -1.12

-0.33

1.00 -0.34 -0.89 -0.64 -0.47 -0.51 -0.24 -0.08

0.40

-0.93

0.25

0.90 -0.02

0.88

0.81

0.80

1.18

0.72

0.23

-0.07

0.39

0.35 -0.06

0.47

0.14

-1.62

0.59

1.19

Satisfaction

0.56

0.28 -0.09 -0.33

-0.06 -0.03 -0.19

0.75

0.14

1.36

1.07

0.81

-0.38 -0.38

Awareness

Likeability Marketing Metrics

Price Promo.

Return on Investment (ROI) Return on Sales (ROS) Return on Mktg. Investment (ROMI) Net Present Value (NPV) Economic Value Added (EVA) Mktg. Expend. on Branding

Social Media

Financial Metrics

Net Profit

Direct to Cons.

Metric

Trad. Adv.

Type of Metric

Int. Adv.

Web Appendix Table 3. Type of Decision’s Impact on Metric Effectiveness when r = 0.

1.23

-0.12

1.50 -0.50

Preference

0.96

0.25

1.04

1.05

1.77

1.54

1.03

0.66

-1.65

0.25

Willingness to Recommend

0.85

1.32

1.27

0.82

0.67

1.06

1.26

0.79

1.53

1.02

Loyalty

0.01

0.73

0.94

0.26

0.91 -0.64

0.53

0.33

0.21

0.27

Quality

0.43 -0.68 -0.24

0.42

0.20

0.66

0.20

0.36

0.58

0.83

Consideration Set

1.23

2.17

2.51

0.90

5.09

1.40

1.61

3.34

-0.20 -0.07 -0.63

-0.55

-0.67 -0.17

1.23

3.58

Total Customers

-0.80 -0.88 -0.37 -0.34

Share of Wallet

-0.75

0.66

0.06 -0.04

Share of Voice

0.33

0.70

0.56

0.30

Note: Bolded numbers indicated significant coefficient.

9

0.02

1.30

0.96

0.67

1.92 -0.14

-1.52

0.18

1.89

1.60

-0.36 -0.34

Web Appendix References Ambler, T. 2003. Marketing and the Bottom Line: The Marketing Metrics to Pump Up Cash Flow 2nd ed. London, FT Prentice Hall. Ambler, T., F. Kokkinaki, S. Puntoni. 2004. Assessing Marketing Performance: Reasons for Metrics Selection. J. Mark. Manag. 20(3) 475–498. Barwise, P., J. U. Farley. 2004. Marketing Metrics: Status of Six Metrics in Five Countries. Eur. Manag. J. 22(3) 257–262. Deshpande, R., J. U. Farley. 1998. Measuring Market Orientation: Generalization and Synthesis. J. Mark.-Focus. Manag. 2(3) 213–232. Deshpande, R., G. Zaltman. 1982. Factors Affecting the Use of Market Research Information: A Path Analysis. J. Mark. Res. 19(1) 14–31. Du, R. Y., W. A. Kamakura, C. F. Mela. 2007. Size and Share of Customer Wallet. J. Mark. 71(2) 94–113. Farris, P. W., N. T. Bendle, P. E. Pfeifer, D. J. Reibstein. 2010. Marketing Metrics: The Definitive Guide to Measuring Marketing Performance 2nd ed. Upper Saddle River, New Jersey, Wharton School Publishing. Finkelstein, S., D. C. Hambrick, A. A. Cannella. 2009. Strategic Leadership: Theory and Research on Executives, Top Management Teams, and Boards. Oxford, UK, Oxford University Press. Hoffman, D. L., M. Fodor. 2010. Can You Measure the ROI of Your Social Media Marketing? MIT Sloan Manag. Rev. 52(1) 41–49. Homburg, C., J. P. Workman Jr., H. Krohmer. 1999. Marketing’s Influence Within the Firm. J. Mark. 63(2) 1–17. Jaworski, B. J., A. K. Kohli. 1993. Market Orientation: Antecedents and Consequences. J. Mark. 57(3) 53–70. Kohli, A. K., B. J. Jaworski. 1990. Market Orientation: The Construct, Research Propositions, and Managerial Implications. J. Mark. 54(2) 1–18. Kuester, S., C. Homburg, T. S. Robertson. 1999. Retaliatory Behavior to New Product Entry. J. Mark. 63(4) 90–106. Lehmann, D. R., D. J. Reibstein. 2006. Marketing Metrics and Financial Performance. Cambridge, Massachusetts, Marketing Science Institute. Menon, A., S. G. Bharadwaj, P. T. Adidam, S. W. Edison. 1999. Antecedents and Consequences of Marketing Strategy Making: A Model and a Test. J. Mark. 63(2) 18–40. 10

Miller, C. C., L. M. Burke, W. H. Glick. 1998. Cognitive Diversity among Upper-Echelon Executives: Implications for Strategic Decision Processes. Strateg. Manag. J. 19(1) 39– 58. Mintz, O., I. S. Currim. 2013. What Drives Managerial Use of Marketing and Financial Metrics and Does Metric Use Affect Performance of Marketing-Mix Activities? J. Mark. 77(2) 17–40. Moorman, C., R. T. Rust. 1999. The Role of Marketing. J. Mark. 63(4) 180–197. Noble, C. H., M. P. Mokwa. 1999. Implementing Marketing Strategies: Developing and Testing a Managerial Theory. J. Mark. 63(4) 57–73. Olson, E. M., S. F. Slater, G. T. M. Hult. 2005. The Performance Implications of Fit Among Business Strategy, Marketing Organization Structure, and Strategic Behavior. J. Mark. 69(3) 49–65. Pauwels, K., T. Ambler, B. H. Clark, et al. 2009. Dashboards as a Service: Why, What, How, and What Research Is Needed? J. Serv. Res. 12(2) 175–189. Slater, S. F., E. M. Olson. 2000. Strategy Type and Performance: The Influence of Sales Force Management. Strateg. Manag. J. 21(8) 813–829. Srinivasan, S., M. Vanhuele, K. Pauwels. 2010. Mind-Set Metrics in Market Response Models: An Integrative Approach. J. Mark. Res. 47(4) 672–684. Verhoef, P. C., P. S. H. Leeflang. 2009. Understanding the Marketing Department’s Influence Within the Firm. J. Mark. 73(2) 14–37.

11