Keeping Teams Continuous: A Dimensional Test of Group Personality Composition

Models that predict how well a small team will perform from its members' personalities make three simplifications. They sort teams into "high" and "low" performers at an arbitrary cutoff. They reduce each team's mix of personalities to a few averages and extremes. And they drop the members who barely participate. Psychiatry spent two decades learning what these choices cost. When a continuous phenomenon is forced into categories, cases near the boundary become unclassifiable. Shared mechanisms get hidden across the bins. The people just below a threshold look almost identical to those just above it. This note proposes to import that lesson into team-composition research, then test whether the field's headline results survive it.

The core idea fits in two sentences. Treat team outcome, member traits, and member engagement as continuous quantities, and model the team with a small, interpretable function made of per-member effects plus pairwise interaction effects. Then ask two questions. How much of the reported benefit of "personality composition" is real signal and how much is an artifact of the cutoff? And do the interaction terms show whether traits combine by complementarity or by synergy? The first test reanalyzes an existing three-person collaboration corpus. The second is a modest prospective study that adds a behavioral measure of how much mental effort each member is willing to spend, taken from transdiagnostic psychiatry.

Where the two literatures meet

The closest existing system is the Personality Composite-Network of Predicting Group Performances Using a Personality Composite-Network Architecture During Collaborative Task. It lets a team's personality profile steer which stretches of its recorded speech a recurrent network attends to. On 97 three-person sessions from the NTHULP corpus it reports 70.0% unweighted average recall (UAR, the mean of per-class recall) at classifying teams as high or low performing. The same paper reports 63.1% for speech alone and 63.6% for personality alone. Three of its design choices are exactly what the psychiatric literature warns against. Group scores are cut at 50 into two classes. Composition is summarized as the min, max, mean, and median of each Big-Five trait. Passive "sleeper" members are excluded. This work motivates the proposal and supplies its main baseline.

The algorithmic team-formation work already moves partway toward continuous description, and it shows why the move matters. Synergistic Team Composition replaces yes/no skill checklists with graded proficiencies and continuous post-Jungian personality scores. It still hard-codes a categorical "leader" type. Optimising team dynamics finds that the weighting that maximizes project scores differs from the one that maximizes members' relational well-being. That result is a warning that "team quality" is not a single axis. We use this second paper as a constraint: continuous outcomes are reported separately rather than merged.

The theoretical frame comes from Team Performance with Test Scores. It models team performance as a set function and distinguishes two cases. Under complementarity (submodularity), each added member helps less when the team already covers similar ground. Under synergy (supermodularity), members make each other more productive. The paper proves that no score computed on individuals in isolation can guarantee a near-optimal team in these regimes. This constrains our claims: we build predictors, not selection guarantees. It also motivates the pairwise terms, because their signs are an empirical estimate of which regime a given trait pair lives in.

From the diagnostic side, the method and cautionary evidence come from three papers. Linked dimensions of psychopathology and connectivity in functional brain networks finds four dimensions of psychopathology that cut across diagnoses. These dimensions track brain connectivity at r = 0.68–0.71, and all four share a common network signature that the categories obscured. Disorders of compulsivity: a common bias towards learning habits shows that three distinct diagnoses share one computational shift toward habit, while a superficially similar group (obese individuals without binge eating) does not. The transdiagnostic structure of mental effort avoidance measures, in 811 people, how often they choose a low-demand task. Compulsivity tracks less avoidance, and lack of perseverance tracks more. All three studies measure individuals in isolation. None asks how such dimensions combine inside a working group, which is the gap we target.

Core Claim

We claim that binarizing team outcomes and summarizing teams by trait extremes misattributes predictive signal. We test two hypotheses. First, many errors of current models fall on teams near the cutoff, where the categorical label is least meaningful. Second, a continuous model with explicit pairwise interactions will rank held-out teams at least as well as the attention network, using far fewer parameters. The signs of its interaction terms will be stable enough across folds to say whether, for example, high agreeableness complements or amplifies low neuroticism. Separately, we claim that treating participation as a continuous member dimension, rather than excluding low participators, carries information about the weakest teams. A behavioral effort-avoidance measure may add signal that self-report personality misses. This last claim is the most speculative and is tested prospectively.

The team model is deliberately simple. For a team $T$ with members $i$ described by continuous profile vectors $x_i$ (Big-Five scores, plus an engagement score in later stages), predicted outcome is

$$\hat{y}(T) = b + \sum_{i \in T} u^\top x_i + \sum_{i<j \in T} x_i^\top W x_j + v^\top s(T),$$

Here $b$ is an intercept, $u$ holds per-member effects, $W$ is a sparse symmetric matrix of pairwise trait interactions, and $s(T)$ is an optional vector of speech features from the session. With $W = 0$ the model is purely additive. Positive entries of $W$ on a trait pair indicate synergy for that pair, negative entries indicate diminishing returns, and the trained model is fit to the continuous score $y$ rather than its binarized version.

Research Program

Stage one is a boundary audit of NTHULP. We will request the raw group scores that the benchmark cuts at 50. We will reproduce the personality network, the speech network, and the combined network under the original protocol, then plot each model's errors against the distance of a team's true score from the cutoff. If errors concentrate within a narrow band around 50, the 70.0% figure is partly measuring label noise near an arbitrary line. This is the team-level counterpart of the subthreshold problem in diagnostic classification. We will also recompute UAR at alternative cutoffs to show how sensitive rankings of methods are to that choice.

Stage two fits the continuous interaction model above under nested cross-validation, with sparsity on $W$ tuned in the inner loop. We report rank correlation with the true score, together with UAR after thresholding predictions at the original cutoff so the result is directly comparable. We then bootstrap the fitted $W$ and report which interaction signs hold in at least 90% of resamples. With 97 teams, only a handful of pairwise terms can be estimated. The sparsity penalty and the stability requirement are how we avoid overclaiming.

Stage three addresses passive members. If NTHULP's excluded sessions can be obtained, each member receives a continuous engagement score computed from speaking-time share and turn count. That score enters $x_i$, and all sessions are modeled together. If the excluded sessions are not available, this stage moves entirely to stage four.

Stage four is a prospective study of about 150 three-person teams solving a collaborative problem with a continuous expert-rated outcome. Everyone is retained. Before teams meet, each member completes Big-Five and brief transdiagnostic symptom scales plus the Demand Selection Task from the effort-avoidance study. We preregister one question: does adding each member's low-demand choice rate to $x_i$ improve held-out rank correlation beyond Big-Five plus engagement? Relational well-being is collected as a second, separately reported outcome, following the dissociation in the Edu2Com paper.

Baselines

The naive reference point comes in two parts. The first predicts the majority class, or the training-set mean for continuous metrics. The second is an additive linear model on per-member Big-Five scores with $W = 0$ and no speech. Beating the naive predictor shows there is signal at all. Beating the additive model is the minimum evidence that composition, rather than a simple sum of member traits, matters.

The strongest existing method is the Personality Composite-Network from Predicting Group Performances Using a Personality Composite-Network Architecture During Collaborative Task. It reports 70.0% UAR on NTHULP high/low classification, against 63.1% for its speech-only network and 63.6% for its personality-only network. It is the strongest competitor because it is the only published model that combines composition and in-session behavior on a real collaborative task. We will retrain it on our exact cross-validation folds with the same eGeMAPS speech features, the acoustic feature set it uses. Its hyperparameter search gets the same number of configurations as our model, and it is scored on both UAR and rank correlation. For the rank-correlation comparison we also train a regression-output variant of it. For NTHULP rank correlation with the continuous score, no reported result was found on Lacuna, so the continuous comparison has no published number to beat and rests on our retraining. The speech-only network, reproduced under the same folds, stands in for "just observe the team working," which is the kind of team-level signal that the Kleinberg–Raghu impossibility result points toward.

Two ablations isolate the proposal's ingredients. Setting $W = 0$ tests the pairwise-interaction claim. Swapping per-member profiles for the attention network's 20-number min/max/mean/median summary tests whether keeping members as a set rather than as extremes matters.

On the Barcelona classroom data, if it is accessible, the reference points are SynTeam's standardized Kendall tau distance of 0.15, teachers' 0.28, and random assignment's 0.42 from Synergistic Team Composition. With only 48 students this is a secondary check, not a primary benchmark.

Beating the strongest baseline means one of two outcomes. The first is equal or higher UAR at the original cutoff together with significantly higher rank correlation under a paired bootstrap over teams. The second is equal performance on both metrics with a stable, interpretable $W$ and roughly two orders of magnitude fewer parameters. Only the first outcome would support the claim that the continuous model is better rather than merely simpler.

Why not just…?

Why not just keep the binary benchmark and report UAR, as the field does? Because the boundary audit in stage one is exactly the test of whether that is safe. The analogous problem is well documented in the diagnostic literature. Linked dimensions of psychopathology and connectivity in functional brain networks shows that categories hide structure shared across them, and Disorders of compulsivity: a common bias towards learning habits shows that labels can group together cases that differ on the underlying mechanism. If errors turn out to be spread evenly across the score range, this objection wins for NTHULP, and we will say so.

Why not just train the existing attention network with a regression output? This is the most serious threat, because it may capture most of the benefit of continuous outcomes with no new model. We run it as a baseline. If it matches our model on rank correlation, the contribution narrows to the boundary audit and the interpretable interaction estimates. We will not claim a predictive gain in that case.

Why not just use in-session behavior, since communication and ties dominate traits? The Strength of the Strongest Ties in Collaborative Problem Solving and The Effect of Team Communication Behaviors and Processes on Interdisciplinary Teams' Research Productivity and Team Satisfaction do find that relational and process variables outweigh individual attributes. On NTHULP, however, speech alone reaches 63.1% UAR against 70.0% when composition modulates it, which suggests the two sources carry partly separate information. Behavior is also unavailable at the moment teams are formed, which is when composition models are meant to be used. Our model includes speech as an optional term, so the fitted weights will show directly how much each source contributes.

Why not just use Big-Five neuroticism and conscientiousness instead of a behavioral effort task? They may well suffice. The preregistered stage-four test is whether the Demand Selection Task adds held-out predictive value over them. A null result there is informative, because it would tell practitioners they can skip the behavioral task.

Decisive Checks

The proposal lives or dies on four results. The first is whether errors cluster near the cutoff in the stage-one audit. The second is whether the continuous interaction model beats the regression-trained attention network on rank correlation without losing UAR. The third is whether at least some pairwise terms keep their sign across 90% of bootstrap resamples, which is the minimum for any claim about complementarity versus synergy. The fourth is whether engagement and effort avoidance add held-out signal in the prospective sample. A negative answer to the first two still yields a useful correction: evidence that the categorical benchmark is not distorting conclusions in this setting. A negative answer to the third means 97 teams cannot resolve composition effects at the level of trait pairs, and that result sizes the study the field actually needs.