Editor's note: Niels Schillewaert is managing partner and co-founder, and Tom De Ruyck is managing partner, of InSites Consulting, a Belgium-based research firm. Steven Debaere is a Ph.D. candidate at IESEG School of Management, Université Catholique de Lille (LEM, UMR CNRS 9221).
Despite the huge potential and popularity of online communities, many researchers still have doubts about how to successfully run them. While academic research has indicated that collaboration communities with customers accelerate innovation (Von Hippel and Katz 2002), more than half the research communities fail, as many suffer from a downturn in activity or churn (Gambetti and Graffigna 2014).
Concerns about quality are a threat to this novel form of collaboration and will only increase the uncertainty surrounding their effective use. While we know much about how to successfully run focus groups or in-depth interviews and ethnographies, we have much less knowledge about the drivers of quality and churn in online research communities. This article aims to contribute to that understanding by examining data from a study of 2,190 members in 10 communities across industries.
Benefits and experiences
Members of a community or a consumer consulting board interact with each other and a moderator to share their opinions and data via posts – text, visuals as well as videos. Actively participating in a community is demanding for consumers and it often requires consumers to spend considerable time and effort in completing tasks (e.g., taking pictures, completing shopping missions, doing interviews with peers, etc.) (Schillewaert and De Ruyck 2012). While there may be an extrinsic motivation for consumers to participate (e.g., cash incentives), they will only continue to do so if the community provides benefits and experiences that are valuable and interesting. Logically, community participants are the most important asset of research communities and determine the quality of a community and subsequently any business decisions based on them.
The quality of a community can be assessed by looking at its activation levels or churn, both in terms of quantity (participation) as well as in quality (contribution of content). Depending on the amount of effort community members put into participation and contribution, they can be classified as active or passive. Churn is collectively defined as passive participation and passive contribution. Churn is a threat to structural collaboration in communities for several reasons. Passive participation and contribution will result in shallow communities, superficial findings and inferior content. Churn is omnipresent and contagious as well. Research communities are dynamic social environments. Members not only mimic each other’s behavior but also use one another’s contribution as a proxy for the overall quality of the environment and adapt their own communication and contribution to it. In the end, non-active members are just a waste of resources and missed opportunities. Churn hampers co-creation and community managers must understand and minimize it.
To battle churn effectively, one first needs to understand its drivers. Only then can proper actions be taken. Extensive studies have explored what drives human behavior. Psychology scholars have identified past behavior (“what we did”) and environment (“where we find ourselves to be”) as effective predictors of future actions. In re-search communities, both types of drivers can be determined by analyzing participants’ posting behavior. Specifically, past behavior can be analyzed via the “recency, frequency, monetary” (RFM) framework and the environment is defined by the members who personify the community. The following sections explain why we believe these drivers are suitable potential churn predictors.
Size of the community. The word “community” refers to a group of people who share similar attitudes, in-terests or goals. For research purposes a community is mostly purposefully built and closed. A true consumer consulting board focusing on engagement and structural collaboration should therefore consist of an intimate sample of consumers interested in a specific topic. If communities are too big they become dysfunctional as wading through discussions for participants becomes cumbersome. Previous research on this topic has shown that having too large of a group of people and too many posts in a community can lead to lurking and lower activity (Schillewaert et al. 2011).
The recency, frequency, monetary framework. The RFM framework is a fundamental model originating from customer relationship management to identify churn behavior. The RFM model uses behavioral customer activity data over a specific time period to construct three key variables that characterize the quality of member activity.
The recency (how recently did a member post) dimension positively relates to churn. People who have recently contributed to a community feel they have done (part of) their duty and hence their predisposition for subsequent participation and the depth of an upcoming contribution may be lower. The frequency (number of posts over time) and monetary value (length of posts), on the other hand, have a negative relationship with near-future churn. This is because consumers with a high frequency of participation and contribution as well as those who made long posts in the past have made that behavior more habitual and developed a certain loyalty to the community. What also happens is that these consumers have observed similar behavior from peer participants, which they mimic. As RFM is valuable for churn identification, we believe it is very suited for community relationship management and to assess churn identification in communities.
Positive emotionality. Respondents are positively affected by the excitement levels of others. Enthusiasm and ambition generate positive vibes and make people work towards similar goals (Elliot and Thrash 2002). Emotions also work contagiously on the judgments and behaviors of others. In fact, Barsade (2002) found that a positive emotional state enhances cooperation between group members and improves perceptions of task performance, resulting in a constructive environment that creates the perfect atmosphere for community mem-bers to participate and contribute.
Negative emotions: swearing and anger. Negative feelings inhibit group performance and thus it is common practice to moderate and manage groups (e.g., focus groups, communities) in way that creates and promotes a positive atmosphere. Researchers always ask consumers to take a positive stance, to be critical but remain con-structive. When people are angry or stressed they can become less inclusive, may feel less motivated to com-plete cognitively demanding tasks and ultimately disengage from the community. On the other hand, as research-ers and marketers we want people to be critical and provide unfavorable feedback as well. So some form of negative expression should be allowed and may stir up the discussion. The question is, how far is too far? Can people use expressive language and swear or also rage and be angry and yet not inhibit the mood and larger goals of the community?
Study
InSites Consulting provided a dataset of 150,943 posts of 2,190 members in 10 branded communities, organized differently from 2011 until 2014. The communities were all managed by InSites Consulting for consumer insight purposes, using similar methodologies. The communities were spread across industries (e.g., CPG, media, tech-nology) as well as varying in terms of duration (from five to 32 months) and number of participants (from 71 to 436).
To assess churn prediction, a predictive modeling methodology was used that applies logistic regression on past data to predict a community member’s churn probability.
Figure 1 depicts the conceptual model tested. The variables used in our model were operationalized as follows:
- The churn variables were determined in different ways. The participation dimension is calculated as the mem-ber’s participation rate in total active community topics. The contribution dimension relies on the text-mining software Linguistic Inquiry and Word Count (LIWC) (Pennebaker 2007). Text analytics represents a powerful tool to easily operationalize churn drivers and churn variables via analyzing posting behavior. Contribution is operationalized by means of the average amount of cognitive words a participant used per post. This variable is indicative of quality and makes a reliable measure for member contribution (Ludwig, Ruyter and Mahr 2014).
- Size of the community is, of course, the number of members active in the community.
- The RFM framework is operationalized via “the number of days since the last post of a member” for the recency variable (R), “the number of topics a member participated in” relates to frequency (F) and “the average word count members use per post” is a proxy for the monetary variable (M).
- LIWC also contains word categories (positive, negative, angry and swear words) for all the corresponding emotional dimensions of the churn drivers.
- Because the communities varied in sector as well as length we controlled for these variables in identifying churn. In fact, it can be expected that, for example, studies in media or technology are intrinsically more or less attractive than CPG/FMCG studies for consumers to participate in. Similarly, shorter or longer communities may also have a given impact on churn.
Results and implications
Table 1 provides an overview of the importance and significance of the drivers for churn.
Big size breeds big problems. Smaller communities with an intimate sample of participants lead to higher levels of activation and higher-quality contributions or, collectively, less churn. When communities have more active participants, threads become crowded and consumers believe everything has been said, hence their participation and number of posts drop. Also, threads may become very text-heavy and participants do not want to read all the comments of their peers and thus may start lurking (merely browsing) or contributing with very short replies, often repeating or confirming what has already been said.
Less recency, less frequency, shorter posts = lower-level members. Recency and frequency of participa-tion had the expected effect on churn as well. The less recently a member of the community has posted, the higher the chance of passive participation and contribution in the near future. When people have not been active for a long time or they feel their duty is done, they feel less urgency to actively interact with the community. The frequency and regularity with which a consumer participates in the community is also important for activity levels. Frequent participants will continue to keep their quantity of participation up but also keep contributing in a mean-ingful way. Monetary value or the sheer amount of words members contribute in one period does not significantly predict future participation. Quantity of words, however, does significantly lower passive contribution – people’s lengthy contributions are a form of leading indicator for meaningful contributions in the subsequent phases of a community.
Pursue positivity. The better the tone of voice in emotional writing in the community, the lower the probability for churn in terms of both passive participation and passive contribution. This confirms that a supportive community environment encourages participation and collaboration.
Allow cursing and avoid raging. The impact of using negative wording seems a little more complicated. Our results show that the use of swear words is not detrimental to the quantity of participation, well on the contrary. When the tone of voice is dominated by anger, however, we found the chance of lower-quality contributions to increase. So it seems that cursing and the use of bad language is tolerable but a delicate balance is needed to ensure communities do not become poisoned by the raging anger of some members.
Conclusions
The fact that we can adequately identify community churn based on recency, frequency and tone of voice allows for effective and proactive community management and moderation. Our results show that an important role for the research designers and moderators of a consumer consulting board is to make sure members stay engaged. In case of churn, they can be individually targeted based on their behavior (RFM) and/or they can be subtly man-aged in positive ways to avoid damaging their content contribution.
Past frequency of participation is by far the best predictor for future participation. This implies that getting off to a good start is important when launching a community. Also, frequency can be used to estimate participation for upcoming tasks and used for active response management as well as additional recruitment efforts. For good contributions in the form of high-quality content, the length of a member’s posts is important. Community managers should stimulate people to participate in great detail as they will build the habit of doing so as well as set examples for peers. These findings are important and confirm the need for the development of high-quality engagement techniques and moderation. Communities are not just about technology but rather about understanding humans and their social relations. The experience of participating in a consumer consulting board should be a positive (brand) touchpoint experience for consumers and not one that causes them to disengage.
Another important aspect for practitioners running communities is to delicately determine the size of the com-munity. Overcrowded and big communities require more attention and it is worthwhile to consider having sub-communities or rooms within a bigger whole rather than one big open platform (which is more panel-like).
Finally, our findings indicate that moderators must closely monitor the community’s emotional level, which can be easily done using the automatic sentiment analysis capabilities of existing text-mining software. Additionally, moderators must not act in a polarized way when it comes to emotions being expressed. Not everything with a negative connotation is necessarily bad for the community. Moderators must strive for a positive environment within the community and avoid allowing too much anger to overshadow the proceedings but can allow some cursing as long as it doesn’t get out of hand.
REFERENCES
Barsade, S. (2002), “The ripple effect: emotional contagion and its influence on group behavior,” Administrative Science Quarterly, 47, 644-675.
Elliot, Andrew J. and Todd M. Thrash (2002), “Approach-avoidance motivation in personality: approach and avoidance temperaments and goals.,” Journal of Personality and Social Psychology, 82 (5), 804–18.
Gambetti, R.C. and G. Graffigna (2014), “Value co-creation between the ‘inside’and the ‘outside’ of a company – insights from a brand community failure,” Marketing Theory.
Ludwig, S., K. de Ruyter, and D. Mahr (2014), “Take their word for it: The symbolic role of linguistic style matches in user communities,” Management Information Systems Quarterly, 38 (3), 1201–17.
Pennebaker (2007), “LIWC: Linguistic Inquiry and Word Count,” LIWC2007 Manual, 1-22.
Schillewaert, Niels and Tom De Ruyck (2012), “Give ’em all something to talk about. Tips for effective community management,” April, Quirk’s Marketing Research Review.
Schillewaert, Niels, De Ruyck Tom, Ludwig Stephan and Manfred Mann (2011). “The dark side to crowdsourcing in online research communities.” CASRO Journal, pp. 5-9, http://issuu.com/casro/docs/casro-2011_journal.
Von Hippel, E. and Ralph Katz (2002), “Shifting innovation to users via toolkits,” Management Science, 48 (7), 821–33.