Back to Vol. 2, No. 3 Table of Contents

Gender Language Style and Group Composition in Internet Discussion Groups

Victor Savicki,
Dawn Lingenfelter
Merle Kelley

Department of Psychology
Western Oregon State College

Full Screen

Table of Contents

Abstract

This study focuses on group gender composition and the seeming relatedness between gender roles and group process functions described as task and maintenance, as found on the Internet. The sample was drawn from randomly selected set of 27 online discussion groups from both the Internet and from commercial information services (e.g. Compuserv) using the ProjectH dataset. The 2692 valid messages were coded for language content (fact, apology, first person flaming, status, etc.) that has been related to gender role in other research. Each message was also coded regarding the gender of its author. Results held with the conventional impression that men far outnumber women as participants in online discussion groups. However, results were mixed in regard to the relation of language patterns and group gender composition. Gender composition was related to patterns of computer mediated communication in this context. However, there were an unexpectedly high proportion of participants of indeterminate gender in this dataset, it is difficult to test the hypotheses with precision. However, the sample is comprised of "real-life" groups, so what is lost in experimental control is compensated for in generalization to other uncontrolled settings.

Introduction

Both the popular press ( Tannen, 1994) and research studies (Herring, 1994) have noted both the increased use of networked computers and the relationship between gender and language in computer mediated communication (CMC). Such examination is timely since more and more emphasis is being placed on CMC in education and in the workplace. Additionally, the Internet, the current version of the "Information Superhighway", dominates predictions of a truly "networked nation" (Hiltz & Turoff, 1993) in which far flung participants can interact with information and with each other overcoming barriers of time and geography. Historically, CMC has been a male dominated domain (Herring, 1994). However, women are taking advantage of this mode of communication in ever growing numbers. Previous studies cite situations in which CMC is a hindrance to full participation for both genders (Herring, 1994) and situations in which CMC enhances communication for both genders (Herschel, 1994). The current study will build on work of Herring (1993,1994) and of Savicki, Kelley, and Lingenfelter (1996, in press), to test the predictions that online communication in groups is related to the gender composition of the groups.

The overall framework of this study focuses on group gender composition and group process functions. We presume, consistent with previous studies (Dindia & Allen, 1992; Mulac, Wiemann, Widenmann, & Gibson, 1988), that groups composed of all men or all women will represent extreme positions on several gender related variables, while mixed groups of both men and women will fall between the extremes. Also we presume, consistent with previous studies (Anderson & Blanchard, 1982; Carli, 1989; Eagly & Karau, 1991), that women will behave consistently with maintenance or socio-emotional group process roles and men will behave consistently with task oriented roles.

Specifically in a CMC context, Herring (1994) describes different gender styles consistent with the task versus socio-emotional distinction: "The male style is characterized by adversariality: put-downs, strong often contentious assertions, lengthy and/or frequent postings, self-promotion, and sarcasm." "The female-gendered style, in contrast, has two aspects which typically co- occur: supportiveness and attenuation" (1994 pp 3-4). She indicates that the gender composition of an online group is related to the CMC of the whole group; "Entire lists can become gendered in their style" (1994 p 6). However, individual members find that the communication style of their gender becomes "mixed" with the style of the group, since gender styles are "more resistant to conscious reflection and modification" (1994 p 6). These observations are consistent with those of Kanter (1977) who predicts that effects of gender will be proportional to the gender composition of the group.

Studies by Savicki, Kelley and Lingenfelter (1996, in press) have found specific communication differences related to gender composition in small (4 to 6 member) groups using asynchronous CMC over 3 to 4 week periods to complete specific tasks. These results are consistent with both task versus socio-emotional contrasts and with Herring's (1994) assertions. First, women in female only (FO) groups were more satisfied with the group process and had more advanced levels of group development than did either male only (MO) or evenly mixed gender (MIX) groups. FO groups used the greatest proportion of "I" (I, me, my, myself) pronouns, the most self-disclosure, and issued the most opinions.. Although MO groups used more factual assertions in one study (1996), that result did not appear in the following study (in press). Use of "we" pronouns (we, us, our, ourselves) was not different across groups with different gender ratios. However, FO groups did send more messages explicitly referring to other members of the group than did either MO or MIX groups. Finally, consistent with Herring's (1994) hypothesis, MO groups used more coarse and abusive language and changed their opinions less than did the FO groups. Some of the differences between the Herring (1994) predictions and the actual results may be explained by the CMC context of the groups' activities.

Many variables may moderate the relationship between gender and CMC. Dramatic differences in context certainly may obscure or speciously highlight results. The context of Internet discussion groups is one in which membership is usually large, members probably do not know all others in the group (especially if there are a large number of "lurkers", i.e. members who read messages but do not write responses and therefore are not visible in the text- based discussion), and the task is not to produce a specific result, but rather to generate ideas and discuss them. In contrast, the small task group often has a limited number of members (fewer than 12) who know of all other members (no lurkers), and whose task is to produce a specific cognitive product (e.g. the evaluation of a product or the ranking of priorities). The current study tests hypotheses generated both from observations of the Internet (Herring, 1994) and observations of small, online task groups (Savicki, Kelley & Lingenfelter, 1996, in press). In both contexts gender composition of the group is predicted to relate to language and CMC. Results from the small task group, Savicki, Kelley and Lingenfelter (1996, in press) studies will be tested in the context of the Internet to evaluate their generalization to that context.

Current Hypotheses

Hypothesis 1: The larger the proportion of men in discussion groups, the more the members will use language that a) states facts without personal ownership, b) challenges group members, c) calls for explicit action, d) is argumentative, e) uses coarse and abusive language, and f) indicates the members status. Hypothesis 2: The larger the proportion of women (i.e. smaller proportion of men) in discussion groups, the more the members will use language that a) self-discloses, b) states personal ownership of opinion, c) apologizes, d) asks questions, e) uses "we" pronouns, f) responds directly to others in the group, and g) seeks to prevent or alleviate tension or arguments.

Methods

The methods build on the work of ProjectH (Rafaeli & Sudweeks, 1993b), an international research group investigating computer mediated communication.

Subjects

The sample was drawn from a randomly selected set of 30 online discussion groups from both the Internet and from commercial information services (e.g. Compuserve). For each discussion group 100 sequential messages were selected for a total of 3000 messages so that themes and threads as well as individual messages could be examined. During the 100 message sample period, the number of contributors to the group discussion ranged from 13 to 81 with an average of 46.2 contributors per group. There was no way to determine the actual size of each group because "lurkers," subscribers to the lists who did not contribute, did not make their presence known.

Two groups were discarded from the current research sample because their focus was not on discussion: one group only posted jokes, another only posted advertisements to sell or swap electronic equipment. A third group was discarded as an outlier since it accounted for a dramatically disproportionate level of conflict (e.g. 90% of all extreme conflict messages came from this group) and thus was not representative of the general sample. Eight of the remaining 2700 messages were deleted as uncodable for a total of 2692 messages sent by 1208 different individuals in 27 different online discussion groups. The complete ProjectH selection methodology is documented in the ProjectH Technical Report ( Rafaeli & Sudweeks, 1993c).

Instruments

Content analysis was done on the messages using the ProjectH Codebook (Rafaeli & Sudweeks, 1993a). Some adaptations of original codes were made to enhance reliability and to assure linearity in the coding categories. Table 1 shows the scales of interest for this research. The complete codebook is available electronically (see Rafaeli & Sudweeks, 1993a).

Table 1

Content codes used with definitions and examples

Content Scale Definition Example
FIRSTPER measured verbal self-disclosure, a statement by the author of the message about the author of the message: 1 = yes, 2 = no. "I like opera", "I'm an email junkie", "My hair is black" but not "My mother's hair is black" or "My cat is black"
OPINION measured statements of the personal opinion of the message author; it had to indicate the first person directly or indirectly. This scale was adapted into a 2 point scale; 1 = no opinion was present, 2 = opinion was present. "I think chocolate is the best flavor ice cream", "Chocolate is a favorite flavor of mine".
FACT measured statement of fact whether or not the fact was correct, without first person reference to the message sender: 1 = no statement of fact, 2 = one or more statements of fact. "The world is really flat." "The government is loaded with freeloaders."
APOLOGY measured any form of apology (implied or direct): 1 = no apology present to 2 = clear apology. "I am sorry I said what I said", "I take my words back"
QUESTION measured the presence of a question: 1 = no, 2 = yes. "How can I operate this software?"
ACTION measured any call for action on the part of the reader: 1 = no, 2 = main content of the message. "Write your congressman," "Go see this movie."
CHALLENGE measured the presence of a challenge, dare, or bet: 1 = no, 2 = yes. "I challenge you to support that statement"
COALIT1 measured degree of agreement or disagreement with another person or statement previously appearing in the group discussion. This scale was adapted so that only response to previous communicators and its intensity was scaled regardless of its direction; 0 = no reference to another person's message, 1 = mild response to another person on the list, 2 = strong response to another person on the list. "I really agree with Ralph." "I think Ralph and Inge's idea sucks."
COALIT2 measured the use of the first person plural pronouns (we, us) to refer to others on the discussion list: 1 = no, 2 = yes. "We seem to be able to consider these ideas well." "Good for us!"
FLAME1 measured 6 levels of argumentativeness of a message: 1 = positive, neutral or no opinion to 6 = hostile: profanity, tirades, to the point of ignoring original issue. "I have to take issue with you on that one." "Only a real dork would take such a stupid position."
FLAME2 measured 5 levels of the use of coarse or abusive language in a message: 1 = no abusive language to 5 = abusive language about content and persons in and out of the group. "I can only say that you must be a real asshole."
FLAME3 measured efforts to prevent or alleviate tensions or arguments in the discussion: 1 = no such efforts, 2 = tries to calm ongoing tension. "I think things are getting out of hand here. Let's cool the tirades and get back to the point."
STATUS measured whether or not the message body or header identified the personal status of the author: 1 = no, 2 = yes. "Dr. H. I. N. Mighty, M.D., Ph.D. Founder and President."

Interobserver reliability for these scales was determined by having two independent observers completely code 8 of the 27 lists (800 messages); thus reliability is based on 29% of the whole sample. Prior to coding the lists in the research dataset, each coder completed a common pretest dataset to an acceptable standard. After coding the research messages, if the two coders could not reach a satisfactory level of agreement, a third coder was enlisted to also code the entire list (see Rafaeli & Sudweeks, 1993c). Percentage of interobserver agreement (agreements/agreements + disagreements) yielded the following results: FIRSTPER = 71%, OPINION = 66%, FACT = 61%, APOLOGY = 95%, QUESTION = 83%, ACTION = 87%, CHALLENGE = 97%, COALIT1 = 57%, COALIT2 = 88%, FLAME1 = 84%, FLAME2 = 97%, FLAME3 = 96%, STATUS = 90%.

Procedures

Each message was coded regarding the gender of its author. Original coding indicated some unreliability in gender coding, and all individuals coded "can't tell" were re-examined. Two independent coders re-coded the appropriate messages. When re-coders disagreed, they conferred with a third coder to reach a conclusive gender code for each individual contributor to each list ( 0= indeterminate gender, 1= Female, 2= Male). Every effort was made to categorize each author on the list into either male or female, however roughly 13% of messages fell into the indeterminate gender category. The group composition index for this study was developed by calculating the proportion of men in the group as a percentage of all specifically gender categorized members: PropM = N of men/(N of men + N of women). Because those members categorized as indeterminate gender could have been either male or female, they were not used to determine the gender composition of the groups. However, they were included in the data analysis, since they were exposed to the overt gender composition of the group in the same way as was any other group member.

Results

As table 2 indicates, the conventional impression that online discussion groups are primarily populated by men holds for this sample. Fully 73% of the subjects were men and almost 75% of the messages were authored by men. However, these percentages are far below the 95% male figure often cited (Chronicles, 1994, September 5).

Table 2

Gender composition of the 27 groups was indicated by the following results: two groups were all male, 25 groups had more men than women, 10 groups had more members with indeterminate gender than women, only one group had roughly equal distribution of men and women, no group had a majority of women. Number of participants in the 100 message sample for each group (range 13 to 81) was not related to the PropM index (r = -.029). Figure 1 shows the distribution of group composition using the proportion of males index (PropM).

Figure 1

The two hypothesis were tested using multiple regression of language usage in the message dataset as related to gender composition of groups using the PropM index. Multiple regression was used because the hypothesized patterns of language are seen as co-occurring, not independent. Therefore, multiple regression will treat the hypothesized relationships as a set of possibly interrelated variables.

Hypothesis 1:

The first multiple regression tests whether a larger proportion of men in discussion groups, is related to members using more language that a) states facts without personal ownership, b) challenges to group members, c) calls for explicit action, d) is argumentative, e) is coarse and abusive , and f) indicates member status.

As table 3 indicates, the overall relation of group composition to hypothesized male language patterns was significant (R2 = .025, F(6) = 11.525, p < .0001). However, only 2 of the predicted 6 language use variables added significantly to the overall result. As predicted, subjects did use more fact oriented language and more calls for action in groups with higher proportions of males. However, there was no significant contribution from language which was challenging, argumentative, contained coarse and abusive words or status indicators.

Table 3

Hypothesis 2:

The second multiple regression tests whether a larger proportion of women (i.e. smaller proportion of men) in discussion groups, is related to members using more language that a) self-discloses, b) states personal ownership of opinion, c) apologizes, d) asks questions, e) uses "we" pronouns, f) responds directly to others in the group, and g) seeks to prevent or alleviate tension or arguments.

Table 4 indicates the overall relation of group composition to hypothesized female language patterns was significant (R2 = .03, F(7) = 12.020, p < .0001). Two of the 7 language use variables added significantly to the overall result in the predicted direction; subjects in groups with lower proportions of males did use more self-disclosure and more attempts at tension prevention and reduction. Four of the language use variables did not add to the overall significance; there was no relation of use of opinion, apology, questions, or "we" pronouns to group gender composition. Finally, 1 of the 7 variables added significantly to the overall result in the opposite of the predicted direction; subjects responded less to others in the group when the proportion of males was low.

Discussion

The first consideration in discussion of these results is the call for caution in interpretation. Caution is necessary for at least two reasons: the proportion of indeterminate gender subjects and the smallness of the overall effect. Because fully 13% of subjects could not be categorized as either male or female, and because 38% of all groups had more indeterminate gender members than they had female members, there is a large ambiguity in the dataset concerning gender. This surprising number of messages were not able to be categorized as being sent by either a male or female for several reasons. Some authors used pseudonyms, others used only initials, or did not indicate who the author was at all other than the electronic address in the message header. Some of the headers did reveal reliably coded gender related names, but others only listed a cryptic aggregate of numbers and/or letters with no clue as to the gender of the sender. More interesting was the number of senders with names of indeterminate gender; for example, Chris, Leslie, or Merle. Even though the PropM index attempted to control for this ambiguity, the results must be taken with caution. Another caution is that with this large a dataset, large significance levels can be generated by small effects. Only 2.5% of the variance was accounted for the analysis for hypothesis 1 and only 3% in the analysis for hypothesis 2. Clearly, there are many other factors at work in the dataset; group gender composition has only modest explanatory power.

Given those cautions, it is interesting to note the mixed results of language choice as related to gender within this CMC context. The expectation was supported that groups with higher proportions of women would be conducive to group members self-disclosing and seeking prevention and reduction of tension. However, other predictions were not confirmed. It may be that when language choice is considered using multiple regression rather than individual, one-way analyses, self-disclosure and tension reduction account for the lion's share of the variance with other predicted language patterns subsumed by these two.

Another question as yet to be answered is whether or not there is a "threshold" gender composition proportion beyond which women may be able to have an even greater effect on group language choice. Clearly, the range of group gender composition is restricted in this sample with the lowest PropM number being 53%; no group had more women than men. Both Herring (1994) and Kanter (1977) indicate that language choice by women may be different in groups that are all or mostly women. Additionally, Kanter posits that women in "skewed" groups (85% to 99% male) may play out a more stereotyped "token" role; whereas women in "tilted" groups (65% to 84% male) take on "minority" status and may have a more visible effect on the culture of their groups (1977 pp 208-209). In the present sample, women in most groups found themselves in "token" proportions. Different proportions of men and women may lead to different actions by the non-dominant members. Savicki, Kelley, and Lingenfelter (1996, in press) note significantly different patterns in female only groups than in equally mixed or male only groups. Future research could examine the threshold question in this context by looking at extreme groups. With the large ambiguity introduced by the indeterminate gender proportions of this dataset, it is difficult to test these hypotheses with exactness. However, the sample is comprised of "real life" groups, so what is lost in experimental control is compensated for in generalization to other uncontrolled settings.

The expectation held that groups with higher proportions of men would be conducive to more impersonal, fact oriented language and more calls for action. However, other predictions were not confirmed. An interesting result was the lack of significance for "flaming". Both Herring (1994) and Savicki, Kelley, and Lingenfelter (1996, in press) found more argumentativeness and coarse language in male dominated groups. The failure to replicate such findings in this sample, again, may be a result of restricted range: i.e. having no groups with a majority of women. Most groups did not use extreme degrees of flaming, although most did show diverging or disagreeing opinions and ideas. As predicted by Herring (1994), the value of full discussion of ideas was supported in this predominantly male sample; tension reducing efforts were mostly related to use of coarse and abusive language (FLAME1-FLAME3: r = .19), not to argumentativeness or disagreement (FLAME2-FLAME3: r = .04), The most optimistic interpretation of these results is that online communication may be becoming more civilized.

One variable not included in the current study which could moderate the language choices of group members regardless of the gender composition is the group task. McGrath and his associates (McGrath, 1984; McGrath & Hollingshead, 1993; McGrath, Arrow, Gruenfeld, Hollingshead, & O'Connor, 1993) indicate that cognitive tasks such as generation of ideas and discussion not resulting in group decision require less cooperation and collaboration and thus are more likely to call forth individually oriented responses or what Hewes (1986) has dubbed a "collective monologue" rather than group oriented responses from group members. The online discussion groups studied in this sample (and those on the Internet generally) are engaged in such tasks. In such a situation observers might expect to see more argumentativeness, more impersonal statements of fact, more challenge, and more marshaling of influence such as status. These predictions bear a striking similarity to the male style of language choice elaborated earlier. To the extent that the Internet is populated mostly by males engaged in tasks that do not require group orientation, the question becomes is the resulting language choice pattern a "male" pattern or a "task" pattern.

Finally, as the CMC field develops, it becomes clearer that findings for one context of CMC do not necessarily generalize to other CMC contexts. Distinctions must be made between findings from asynchronous versus synchronous interaction, along dimensions of media richness (e.g. text only versus multimedia), along dimensions of task, and, of course, in relation to group factors such as size and composition. For example, to what degree do the differences in the findings in the present study versus the findings in Savicki, Kelley, and Lingenfelter (1996, in press) reflect differences in gender effects or differences in the context of CMC. In the current study large groups (up to 81 members) who did not know all others in the group (because of size and lurkers) generated ideas and discussed them with no specific assigned outcome. Whereas, in Savicki, Kelley, and Lingenfelter (in press) small groups (4 to 6 members) who knew all others in the group completed assigned tasks that required collaboration and group agreement. Different contexts will require different research, and, generalizations across contexts must be tested; as this study attempted.

References