A valence-based account of group interaction and decision making

ABSTRACT

Group participants often develop a range of problem solutions before discussion. We addressed whether, and at what level of analysis, initial opinions influence discussion and perceptions of decision outcomes. The Group Valence Model (GVM) presents a dual-process approach to interaction and decision making as a function of the distribution of supportive and oppositional comments. GVM predicts that discussion reflects individual-level opinions until a group solution emerges, whereupon discussion is influenced by group-level factors. Data from four previous studies were machine-coded for supportive and oppositional statements. Results indicated that the model holds in some degree at the group level but not at the individual level. Discussion focuses on mechanisms that drive interaction prior to the emergence of a group-level solution.

KEYWORDS:

• Decision making
• valence
• interaction
• disagreement
• convergence

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 It is worth noting that the association between comments and cognitive resources seems weak for both group interaction (Meyers, 1989; Meyers & Seibold, 1990) and interpersonal compliance-gaining discussions (Applegate, 1982; Dillard, 1988). See Hewes (1996, 2009) for a contrasting viewpoint.

2 The private descriptions of the group’s decision have not been used in prior analyses and only very rudimentary analyses of the private profiles (number of distinct impressions) and discussion data were used.

3 In most cases, we relied on participants’ punctuation to identify sentences, though in some cases we modified a response if it did not have sentence structure (e.g., a list).

4 Data from 10 groups were excluded because of recording failures—though profile and decision task data were available for those groups we chose not to use them in the analysis.

5 We computed sentiment scores for other combinations of the number of words preceding and following the target words. Correlations among the different methods were all above .90 so we chose to use the default setting of five preceding and two following words.

6 Because opinion scores can be negative, we used a modified version of the Gini index to compute disagreement scores (Chen et al., 1982).

7 Chi-squared based model fit statistics (e.g., CFI, RMSEA) for MSEM reflect misspecification from both levels, though the estimates are weighted more heavily in favor of the within level by virtue of it having substantially more observations than the between level. We examined level-specific fit statistics to evaluate the model (see Hox et al., 2018; Ryu, 2014; Ryu & West, 2009).

8 Mplus issued a convergence warning for our initial models and indicated that the problem was negative residual variance estimates for both pre-AT discussion and the opinion profiles. Constraining those variances to zero fixed the problem and all model estimates presented are based on models those constraints (see Jak, 2018).

9 In this case, the Gini index indicates profile similarity but not direction. Thus, a low Gini score reveals that profiles are similar but says nothing about opinion polarity (i.e., whether opinions are mostly negative or positive within a group). Including mean opinion scores for groups controls for opinion polarity.