12 / Journal of Marketing, September 2014 12 / Journal of Marketing, September 2014

method. Only the intercepts were allowed to vary by latent approach employs a latent class specification with a finite

class; all the other parameter estimates were independent of number of support points for the intercept term (Heckman

latent class (i.e., did not vary by latent class). We maxi- and Singer 1984), so the estimates of the intercepts differ

mized the log-likelihood function of each equation using across unobserved latent classes (Chintagunta 2001; Wedel,

5,000 iterations of the expectation-maximization algorithm Kamakura, and Bockenholt 2000). To incorporate multiple

so that it converged to a global maximum independent of latent classes on the intercept term, we used the following

the starting values (Muthén and Muthén 2007). Our mea- generic form to replace Equations 1, 2, and 3:

sure of model fit, the modified Akaike information criterion (MAIC), suggested that the four-class solution for the inter-

cept was optimal because the MAIC reduced up to four (4)

∑ +ε k If { k } Xb f f ,

classes (regimes) and increased thereafter (Equation 1: k = 1 MAIC

1 regime = 501.82, MAIC 2 regime = 344.98, MAIC 3 where y f represents the generic form for dependent variable,

regime = 353.54, MAIC 4 regime = 153.54, MAIC 5 regime =

X f is the generic form for covariates, b represents the 164.09; Equation 2: MAIC 1 regime = 128.46, MAIC 2 regime = generic form for parameter estimates of covariates, and e f 139.21, MAIC 3 regime = 93.17, MAIC 4 regime = 85.73,

represents the residuals. The intercept, p, which is allowed MAIC 5 regime = 91.06; Equation 3: MAIC 1 regime = 1,386.37, to vary across latent classes, as noted previously, can take

MAIC 2 regime = 1,387.09, MAIC 3 regime = 1,324.22, MAIC 4 the value of a finite number of discrete (Bernoulli) support

regime = 1,302.54, MAIC 5 regime = 1,302.86). The entropy of points p 1 , ..., p K ; the indicator function I{} is 1 for all f Œ k,

separation of the four-class solution for each equation and 0 otherwise. In our estimation, we allow the number of

exceeded .85, indicating a good separation of the latent latent classes k to increase until the information criteria

classes (Celeux and Soromenho 1996). We provide the suggest otherwise (Wedel, Kamakura, and Bockenholt

results for the antecedent hypotheses (H 1a –H 2d ) in Table 3 2000).

and those for the consequences hypotheses (H 3 –H 5 ) in

Table 4.

Results

First-Stage Antecedent Model

We mean-centered all predictor variables except buyer–

H 1a suggests that buyer and seller concentration are key seller matching process and price discovery process, which

antecedents to a platform’s total customer orientation. As we operationalized as dummy variables. We estimated

shown in Table 3, this prediction is supported: the slopes of

TABLE 3 Latent Class Regression Results for Antecedent Model

Total Customer Customer Orientation Hypothesis Orientation Hypothesis Asymmetry a

Two- versus one-sidedness of matching process (MP) .18 (.31) .01 (.47) Dynamic versus static price discovery process (PD) .99** (.31) .21 (.48) Proportion of transaction-driven fee (TF) .01** (.002) .01* (.004)

Buyer concentration (BC) H 1a .15** (.06) H 2a .21** (.09) BC ¥ MP H 1b –.06** (.02) H 2b –.19** (.08) BC ¥ PD H 1c .24** (.06) H 2c .20** (.08) BC ¥ TF H 1d .001 (.005) H 2d –.004 (.008) Seller concentration (SC) H 1a .12** (.05) H 2a –.10 (.08) SC ¥ MP H 1b –.10** (.04) H 2b –.12** (.04)

SC ¥ PD H 1c .11** (.02) H 2c .07* (.04) SC ¥ TF H 1d –.005 (.04) H 2d –.001 (.001) Platform firm self-participation .22 (.45) .18 (.41) Platform firm’s IT capabilities .67** (.04) –.01 (.02) Industry complexity 1.39** (.21) .66** (.12) Firm size .29** (.03) –.02** (.004)

BC ¥ SC .05 (.08) .001 (.35) R-square .97 .89

*p < .05. **p < .01.

a Represents customer orientation asymmetry in favor of sellers relative to buyers. Notes: We used one-tailed tests for hypothesized effects and two-tailed tests for nonhypothesized effects. Only intercepts were allowed to vary

by latent classes; all other coefficients were independent of latent classes. Because the intercepts vary by classes and a four-class solution was optimal, we do not explicitly list the intercepts in this table, but we obtained the following intercepts for the four classes: total customer orientation (Class 1: –.92, p < .1; Class 2: 3.87, p < .01; Class 3: 5.56, p < .01; Class 4: 3.45, p < .01); customer orientation (Class 1: –.51, p > .1; Class 2: .23, p > .1; Class 3: 2.47, p < .01; Class 4: –3.04, p < .01). The R-square values reflect the fit of the overall latent class regression model.

Customer Orientation Structure / 13

TABLE 4

process) are operationalized as dummy variables, which,

Antecedent Model: Effects of Buyer and Seller

unlike other predictor variables, are not mean-centered. As

Concentration at Multiple Combinations of Buyer–

such, instead of an overall main effect, we separately exam-

Seller Matching Process and Price Discovery

ine the effects of buyer (seller) concentration at each of the

Process

two levels (0, 1) of buyer–seller matching process and price discovery process. Applying Wald tests (Wooldridge 2010),

A: Effect of Buyer Concentration on Total Customer Orientation

in Table 4, Panels A and B, we provide separate effects of buyer and seller concentration at each level of the two pre-

One-Sided Two-Sided

dictor variables.

Buyer–Seller Buyer–Seller

We find (see Table 4, Panel A) that the effect of buyer

Matching Matching

Process (0) Process (1)

concentration is positive and statistically significant at each of the 0 and 1 combinations of the two predictor variables

Static price .15* (.08) .09** (.04) (one-sided matching process [0] and static price discovery discovery process (0)

Dynamic price .39*** (.14) .33* (.17) process [0]: b = .15, p < .10; two-sided matching process discovery process (1)

[1] and static price discovery process [0]: b = .09, p < .05; one-sided matching process [0] and dynamic price discov-

B: Effect of Seller Orientation on Total Customer

ery process [1]: b = .39, p < .01; two-sided matching

Orientation

process [1] and dynamic price discovery process [1]: b =

One-Sided Two-Sided

.33, p < .10). In Table 4, Panel B, the effect of seller con-

Buyer–Seller Buyer–Seller

centration is positive and statistically significant at three of

Matching Matching

the four combinations of the two predictor variables (one-

sided matching process [0] and static price discovery Static price .12** (.05) .02 (.12)

Process (0) Process (1)

process [0]: b = .12, p < .05; two-sided matching process discovery process (0)

[1] and static price discovery process [0]: b = .02, p > .10; Dynamic price .23* (.12) .13* (.07)

one-sided matching process [0] and dynamic price discov- discovery process (1)

ery process [1]: b = .23, p < .10; two-sided matching

C: Effect of Buyer Concentration on Customer Orienta-

process [1] and dynamic price discovery process [1]: b =

tion Asymmetry Toward Sellers Relative to Buyers

.13, p < .10). Overall, H 1a is supported in seven of eight cases. Substantively, these results imply that a platform

One-Sided Two-Sided

firm’s total orientation is crafted partly in response to cus-

Buyer–Seller Buyer–Seller

Matching Matching

tomer concentration, as predicted by H 1a .

Process (0) Process (1)

In H 1b we predicted that the positive influence of concentration on total customer orientation would be weaker

Static price .12*** (.04) –.02 (.16) discovery process (0)

for two- than for one-sided matching processes. As we Dynamic price .41** (.19) .22** (.11)

depict in Table 3, the interaction of the matching process discovery process (1)

with buyer concentration (b = –.06, p < .01) and with seller concentration (b = –.10, p < .01) were both negative, in sup-

D: Effect of Seller Concentration on Customer Orienta- tion Asymmetry Toward Sellers Relative to Buyers

port of H 1b. We graphically depict these interaction effects in Figure 3, Panels A and B, by plotting the effects (slopes)

One-Sided Two-Sided

of buyer concentration on total customer orientation for dif-

Buyer–Seller Buyer–Seller

ferent levels of the moderator—namely, for two-sided and

Matching Matching

one-sided matching processes. We also find support for H 1c ,

which proposes that the interactions of dynamic price dis- Static price –.10** (.04) –.03* (.016)

Process (0) Process (1)

covery process with buyer concentration (b = .24, p < .01) discovery process (0)

and seller concentration (b = .11, p < .01) both enhance Dynamic price –.22* (.12) –.15** (.07)

total customer orientation (Table 3; for a graphical depic- discovery process (1)

tion of these interactions, see Figure 3, Panels C and D).

*p < .10.

However, we did not find support for moderating effects of

**p < .05. ***p < .01.

the proportion of transaction-driven fees (H 1d ) on the

Notes: We conducted Wald tests to assess the statistical significance

effects of buyer (b = .001, p > .10) or seller (b = –.005, p >

of each effect (each effect is a linear combination of esti-

.10) concentration on total customer orientation. These

mated coefficients of either Equation 1 or Equation 2). Num- bers in parentheses are standard errors of Wald coefficients.

results support our broad contention that platforms rely on total orientation to manage dependence on powerful market participants, in light of their particular transaction attributes

both buyer (b = .15, p < .01) and seller (b = .12, p < .01)

(e.g., matching process).

concentration on total customer orientation are positive. We In H 2a , we propose that buyer (seller) concentration would interpret these main effects of buyer (seller) concentration in

enhance (reduce) orientation asymmetry toward sellers rela- light of the fact that two of our theoretically relevant predic-

tive to buyers; we found support for buyer concentration (b = tors (i.e., buyer–seller matching process and price discovery

.21, p < .01), but although the effect of seller concentration

14 / Journal of Marketing, September 2014

FIGURE 3 Statistically Significant Two-Way Interactions in Antecedent Model: Total Customer Orientation as Dependent Variable

A: Joint Effect of Buyer Concentration and Matching B: Joint Effect of Seller Concentration and Matching Process on Total Customer Orientation

Process on Total Customer Orientation

ss = .12** O .15

o n –.05

p tr

Sl n –.10

Sl c n –.10

Buyer Concentration (± 1 SD) Seller Concentration (± 1 SD)

One-sided matching process

Two-sided matching process

C: Joint Effect of Buyer Concentration and Price D: Joint Effect of Seller Concentration and Price Discovery Process on Total Customer Orientation

Discovery Process on Total Customer Orientation

C ss = .23*

Buyer Concentration (± 1 SD) Seller Concentration (± 1 SD)

Dynamic price discovery

Static price discovery

*p < .10. **p < .05. ***p < .01. Notes: SS = slope size.

was in the expected direction, it was not significant (b = –.10, [1]: b = .41, p < .05; two-sided matching process [1] and p > .10). In the absence of mean-centering the two predictor

dynamic price discovery process [1]: b = .22, p < .05). In variables (buyer–seller matching process and price discov-

Table 4, Panel D, as we hypothesized, we find that the ery process operationalized as dummy variables), we pro-

effect of seller concentration is negative and statistically vide further details of the effects of buyer (seller) concen-

significant at all four combinations of the two predictor tration at each of the two levels (0, 1) of buyer–seller

variables (one-sided matching process [0] and static price matching process and price discovery process. As we discovery process [0]: b = –.10, p < .05; two-sided matching hypothesized, we find that the effect of buyer concentration

is positive and statistically significant at three of the four 0 process [1] and static price discovery process [0]: b = –.03, and 1 combinations of the predictor variables (one-sided

p < .10; one-sided matching process [0] and dynamic price matching process [0] and static price discovery process [0]:

discovery process [1]: b = –.22, p < .10; two-sided matching

b = .12, p < .01; two-sided matching process [1] and static process [1] and dynamic price discovery process [1]: b = –.15, price discovery process [0]: b = –.02, p > .10; one-sided

p < .05). Thus, H 2a is supported in seven of eight cases. matching process [0] and dynamic price discovery process

These effects lend credence to our assertion that platform

Customer Orientation Structure / 15

firms deliberately cultivate orientation asymmetry to bal- ance their dependence on powerful market participants.

In support of H 2b , two- versus one-sidedness in the buyer–seller matching process negatively moderated the effects of buyer concentration on customer orientation asymmetry in favor of sellers over buyers (b = –.19, p < .01). The interaction involving seller concentration and the two-sided matching process (b = –.12, p < .01) was also negative (for a graphical depiction, see Figure 4, Panels A

and B). Also in support of H 2c , both buyer (b = .20, p < .01)

and seller (b = .07, p < .05) concentration interacted posi- tively with dynamic price discovery (Table 3 and Figure 4,

Panels C and D). However, we did not find support for H 2d ,

which proposes a moderating effect of transaction-driven

16 / Journal of Marketing, September 2014

fees on the relationships of buyer (b = –.004, p > .1) and seller (b = –.001, p > .1) concentration with orientation asymmetry. Collectively, these results suggest that orientation asymmetry enables dependence balancing when culti- vated selectively in light of a focal platform’s transaction attributes (e.g., matching process).

12 / Journal of Marketing, September 2014 12 / Journal of Marketing, September 2014