The Effects of Firm Generated Content in Social Media on Customer Behavior: An Empirical Examination Ashish Kumar, Assistant Professor, School of Business, Aalto University Ram Bezawada, Associate Professor, School of Management, University at Buffalo Rishika Rishika, Clinical Assistant Professor, Darla Moore School of Business, University of South Carolina Ramkumar Janakiraman, Associate Professor, Darla Moore School of Business, University of South Carolina P.K. Kannan, Professor of Marketing, Robert H. Smith School of Business, University of Maryland

Social Media Spending and Dilemma

Social media spending is expected to grow faster than any other form of online marketing

• Forrester Research (2011)

Total social media advertising spending has increased worldwide ($17.74 billion in 2014 vs. $11.36 billion in 2013 - an increase of 56.2%)

• eMarketer (2015)

Marketers are concerned about measuring the returns on investment from social media

Research Questions

Systematic empirical examination of social media engagement on customer-firm relationship

• Leverage a unique dataset

• Use individual level data

• Connect customer social media engagement to their actual behavior

• Examine multiple aspects of customer-firm relationship for the same set

of customers

• Analyze firm generated content (FGC) of social media engagement

• Examine features/postings of the social media site

Research Issues

Develop a modeling framework for analyzing the effectiveness of FGC

• Causal effect of FGC • Operationalization of FGC

Relate the FGC to customer-firm relationship

• Customer spending (transaction side) • Cross-buying behavior (relationship side)

Consider the role of other marketing mix

• Television advertising • Email

Literature Review

Link between User Generated Contents (e.g., online reviews) and Outcomes (e.g., sales)

• Chevalier and Mayzlin (2006); Liu (2006); Godes and Silva (2012) • Blogs—Onishi and Manchanda (2012); Gopinath et al. (2013) • YouTube—Yoganarasimhan (2012) • eBay Germany—Algesheimer et al (2010) • Viral Marketing —Schulze et al. (2014) • Promotional Sales Campaign—Danaher and Dagger (2013)

Our Contribution • Our focus on FGC is more granular and we examine both promotional and non-

promotional marketing communications • Marketing mix variables; traditional advertising • Multiple dimensions of FGC— valence, receptivity, and susceptibility

Conceptual Framework

Firm Generated Content (FGC)

Vpost – Valence of posting (Naïve Bayes Algorithm/Classifier to classify the postings into positive, negative, and neutral sentiment) Rpost – Total number of likes/share/comment (Receptivity) FGCSuscept – FGC susceptibility of customer (construct from survey)

Model: DID (1)

Error Structure Heterogeneity

Handling Self-Selection Issue: Using Propensity Score Matching (PSM)

Model: Treatment Effect (TE)

Estimated on Random as well as Matched Sample

latent utility that customer derives by participating in the firm’s social media site and thereby choosing to receive FGC.

Identification

Unobserved factors can influence customers’ decision to engage and their subsequent behavior (i.e., endogeneity/self-selection issues)

• Propensity Score Matching • Two distinct customer groups

• Treatment (participate) and control (do not participate)

• Data from time periods pre and post launch of social media site

• Treatment effect model explicitly models participation

Model Estimation

Bayesian inference using MCMC methods • Hierarchical framework to capture heterogeneity

Conjugate priors for the parameters

50,000 draws for the MCMC with a burn-in of 40,000 • Monitor for convergence

PSM and TE for endogeneity issue

Research Context: Data

Large N.Y. state retailer • Operates multiple stores and specializes in selling wine and

similar products

August 2009: The retailer started having a presence in a major social networking site

• Marketing campaign that encouraged customers to join the social media site

• But, no incentives were given to join

Data: Sample Selection Multi-step process

• Survey to 5000 randomly selected customers from the firm’s database • Identify customers who participate in the firm’s social media page • Match the survey with transaction data of the retailer based on full

names/demographic/residence location information

Treatment group customers: 412 customers who are active on the firm’s social media site

Control group customers: do not participate in the firm’s social media page

• Use survey information and cross check with the social media site • Include customers who have similar demographics and loyalty as treatment group

customers • Matched selection based on Propensity Score

Data: Time Line

Firm Launches Social Media

Page

August 15, 2009

March 2011

Consumers’ Participate by Either Joining,

Contributing or Both

Sample Selection: Treatment and Control

Consumers

Firm Tracks the Activity on Social Media Site

February 2011

Survey Sent to Consumers

Merging Social Media with Transaction Data

Data: Variables Variable Description Variable Notation Variable Operationalization Mean SD

Social Media Participation Data and Traditional (TV Ad) & Digital (Email) Marketing Communications Data Firm Generated Content FGC Messages posted by the firm. (# of postings/week) 3.78 1.82 Customer Participation CustPar =1 if customer h participates in firm’s social media page at time t; 0 otherwise. Television Advertising TVAd Television ad based on gross rating point. (GRP/week) 40.79 26.42 Email Advertising Email Total number of emails sent by the firm that are opened by customer h during week t. (# of

emails opened/week/customer) 1.63 1.02

Customer Transaction Data Spending Behavior Spend Total dollar amount spent by customer h during week t on firm’s products.

($/week/customer) 14.34 7.26

Cross-buying Behavior CrossBuy Total number of distinct categories customer h purchased during week t. (# of distinct categories bought/week/customer)

3.23 1.24

Customer Experience CExp Duration of the relationship of customer h with the firm until week t. (# of weeks of relationship with focal firm/customer)

94.26 17.89

Promotion Depth Index PromD Weighted average of price cuts availed by customer h for week t across all product items purchased. (cents/ml/week)

.38 .65

Cross Category Promotion CrossP Proportion of categories bought on promotion by customer h for week t. (proportion of categories bought on promotion/week/customer)

.46 .21

Survey Data Time Spent on Social Network TimeOnSocialNet Time spent on social networks (from survey). (minutes/customer) 41.54 8.50 Online Social Profile OProf Number of online social profiles (# online social profile /customer) 2.47 1.15 Age Age Age of customer h. (years/customer) 45.46 18.52 Distance Dist Distance between the customer h and the focal store. (miles/customer) 5.17 8.06 Attitudinal Variables TechS

These constructs are measured via survey.

3.69 .80 SocialNet 2.77 1.22 PrivCon 1.55 .51 IMov 4.08 .65 SMov 2.88 .76 EscMov 2.12 .80

OEnt 2.77 .78 Gender Gender =1 if customer h is male; 0 otherwise. 47% Race Race =1 if customer h is white; 0 otherwise. 86% Note: Summary statistics for customer transaction data are unconditional on purchase occasions. The summary statistics are based on a matched pair of 412 customers from each of the treatment and the control groups.

Model Free Evidence Average Difference in Differences: Purchase Behavior of Treatment vs.

Control Groups

The difference in the spending/cross-buying/profitability between treatment and control group customers is not significant in the pre period but is significant in the post participation period.

Outcome Variable Groups Post-FGC Pre-FGC Difference Difference-in-Differences Spending ($) Treatment Group 15.96 13.52 2.44**

(.92) 1.26** Control Group 14.53 13.35 1.18 (1.01)

Cross-buying Treatment Group 3.74 3.08 .66** (.25) .46** Control Group 3.15 2.95 .20 (.19)

Profit ($) Treatment Group 5.91 4.79 1.12** (.43) 1.02** Control Group 4.78 4.68 .10 (.09)

Results: DID Variables Spending Cross-buying

TCust .0803 .0941 FGC .9851 .1291 TCust×FGC 1.3233** .4987** TCust×FGC×TVAd .0057** .0001** TCust×FGC×Email .0991** .0047** TCust×FGC×CExp .1686** .0797** TCust×FGC×TechS .2753** .0303** TCust×FGC×SocialNet .1906** .1523** TVAd .1373** .1733** Email .9268** .1242** CExp .7579** .336** TechS -.5285** -.2062** SocialNet -.5906** -.3596** PromoD 3.9708** --- CrossP --- .6949** Dist -.3019* -.0316 Gender .9660 .4133 Race .7214 .2093 Age .2778 .3005 Intercept 6.2314* 2.6182** Serial-correlation .0187* .0482 Cross-correlation .4765** Heterogeneity # of Observations 140080 Sample Size 824 LMD -56313.75 **p≤.01 (Parameter is significant at 99% level i.e., the 99% confidence interval does not contain 0) *p≤.05 (Parameter is significant at 95% level i.e., the 95% confidence interval does not contain 0) The sample size of 824 customers is based on a matched pair of 412 customers from each of the treatment and the control groups.

Results: TE Model Variables Spending Cross-buying

FGC .1835** .1103** FGC×TVAd .0621** .0019** FGC×Email .1755** .1137** FGC×CExp .0450** .1228** FGC×TechS .0918** .0281** FGC×SocialNet .1091** .0956** TVAd .1082** .1218** Email 1.0258** .8954** CExp .5853** .2055** TechS -.1537** -.0975** SocialNet -.1976** -.4158** PromoD 2.2350** --- CrossP --- .5707** Dist -.0394* -.0385 Gender .9690 .1114 Race 1.0673 .1102 Age .6973 .1844 Intercept 6.6342* 1.2867* Serial-Correlation .0076* .0128 Cross-Correlation .4124* Heterogeneity # of Observations 66980 Sample Size 788 LMD -75752.36 **p≤.01 (Parameter is significant at 99% level i.e., the 99% confidence interval does not contain 0) *p≤.05 (Parameter is significant at 95% level i.e., the 95% confidence interval does not contain 0) The sample size of 824 customers is based on a matched pair of 412 customers from each of the treatment and the control groups.

Variables Participation PrivCon -.1766** IMov .2865** TechS .0521* SMov .0771** OEnt .0509** TimeOnSocialNet 1.1770** Gender .0095 Race -.0297 Age -.0043 Intercept -.4516* Cross-Correlation (Spending, Participation)

.1426**

Cross-Correlation (Cross-buying, Participation)

.0414

**p≤.01 (Parameter is significant at 99% level i.e., the 99% confidence interval does not contain 0) *p≤.05 (Parameter is significant at 95% level i.e., the 95% confidence interval does not contain 0)

Discussion of Results

FGC has positive and significant effect on customer spending and cross buying

The interaction effects between FGC and television advertising and the interaction effects between FGC and email marketing are positive for customer spending and cross-buying

FGC has a greater effect on more experienced customers for both customer spending and cross-buying

FGC has a higher effect on customers who are more tech-savvy and those who are more prone to using social media.

Implication: Sensitivity Analysis Type of Marketing Communication

Overall First six months

Last six months

Spending Cross- buying

Spending Cross- buying

Spending Cross- buying

Social Media (FGC) .0140 (.0027)

.0587 (.0035)

.0060 (.0007)

.0441 (.0061)

.0108 (.0023)

.0493 (.0049)

Digital Media (Email) .0913 (.0215)

.0433 (.0165)

.0799 (.0137)

.0409 (.0043)

.0888 (.0358)

.0408 (.0202)

Traditional Media (TV Ad) .1010 (.0745)

.0507 (.0038)

.0949 (.0023)

.0457 (.0202)

.0968 (.0097)

.0532 (.0085)

Notes: Calculation of elasticity is based on the results of the TE model. First 6 months refers to the period from August 2009 to January 2010 (first six months since inception of the firm’s social media page). Last six months refers to the period from October 2010 to March 2011 (last 6 months of our data’s post-FGC period). Standard errors are shown in parentheses.

The elasticities of FGC in the last six months with respect to spending and cross-buying are comparable to the corresponding elasticities from the first six months. This suggests that the effect of FGC is not purely due to novelty factors and implies that FGC can remain effective after the inception of a firm’s social media page

Implication: Simulation Study

Synergistic Effects of FGC with TV Advertising and Email Marketing

Simulation Variables in the Model % Change in Spending

% Change in Cross-buying

Traditional Ad (TVAd) and FGC Main effects of FGC and TV 2.08 (.23)

1.06 (.14)

Main effects of FGC and TV and interaction effect between FGC and TV

3.11 (1.33)

1.90 (.85)

Incremental change 1.03 .84 Digital Ad (Email) and FGC Main effect of FGC and email 3.25

(.54) 1.54 (.35)

Main effects of FGC and email and interaction effect between FGC and email

5.27 (.68)

2.76 (1.18)

Incremental change 2.02 1.22 Notes: The base case is the TE model with the main effect of FGC but without TV and Email. All the models include the control variables that we presented in Equations (5)-(7). Standard errors are shown in parentheses.

Traditional and New Communication Media have Synergistic Effects with FGC

Implication: Dimensions of FGC Elasticity Analysis: Effect of the Three Dimensions of FGC on Customer Behavior

Focal Dimension of FGC Spending Cross-buying

Valence .0128 (.0068)

.0285 (.0061)

Receptivity .0188 (.0020)

.0864 (.0162)

Susceptibility .0092 (.0032)

.0076 (.0028)

Notes: Calculation of elasticity is based on the results of the TE model. Standard errors are shown in parentheses.

FGC receptivity has the greatest impact followed by FGC valence and susceptibility

Managerial Implication

Embrace Social Media • To connect, communicate, and nurture relationships with customers

Utilize Synergies across Media

• Exploit synergistic relationship between social media and other media used for marketing communication—(TV and emails)

Encourage Customer Participation • FGC has a greater impact on customers who have a greater tenure with the firm

Monitor FGC popularity

• FGC receptivity has the greatest impact.

Conclusion

Our study offers key insights into the impact of FGC and contributes to both theory and practice

FGC helps in nurturing customer-firm relationship • Transaction side (Spending) • Relationship side (Cross buying)

Limitation and Future Research

Examining the effectiveness of firm generated content across multiple social media(e.g., Facebook and Twitter)

Effect of FGC across multiple channels

Extension to other contexts (such as search goods, consumer packaged goods)

Detailed message level analysis (e.g., informative vs. persuasive)

Thank You.