Is students’ teamwork a dreamwork? A new DCE-based multidimensional approach to preferences towards group work

The study investigates how lower secondary students in Poland prefer different learning modes (group work, independent study, working with a tutor or parent) across subjects (Polish/Native language, mathematics, English, geography). Motivated by the importance of social and emotional skills for later success, and by concerns that prevalent Likert-type survey measures (e.g., PISA indices) may misrepresent true preferences, the authors ask whether students prefer group work and how such preferences vary by subject and achievement. They posit that aligning instructional modes with student preferences can increase satisfaction, effort, and academic outcomes, drawing on efficiency wage theory and research linking interest, motivation, and achievement. The introduction highlights potential biases in traditional preference measurement and the need for a methodology that captures trade-offs and heterogeneity.

The review situates group work within global competence and documents both benefits (enhanced metacognition, critical thinking, interactive engagement) and drawbacks (unequal contributions, dominance issues). Preferences toward group work are shaped by prior experiences, beliefs, social norms, personality (introversion/extraversion), gendered patterns of sociability, and achievement levels, with mixed findings on whether high achievers prefer collaboration. Age and educational stage also influence perceptions and preferences. The literature notes substantial heterogeneity and context dependence (by subject and learning situation) and critiques Likert-type measures for bias and lack of trade-off information. The authors argue for stated-preference approaches (DCE) to capture multidimensional, context-sensitive preferences, moving beyond a simple alone-vs-together dichotomy and considering subject-specific self-efficacy and gender differences.

Design: A Discrete Choice Experiment (DCE) grounded in Random Utility Theory and Lancaster’s characteristics theory was used to elicit preferences. Alternatives described weekly school projects varying on three attributes: study mode (independently; in a group of five peers; with a tutor; with a parent), subject (mathematics, Polish/Native, English, geography), and required time (continuous). Each respondent completed six choice tasks with three alternatives each (no opt-out), choosing the preferred option. Sampling and data: Data come from the Evidence Institute’s Competences 2018 online assessment in Poland targeting final grades of lower secondary and primary (ages 14–15). Of 34,560 assessed students, 21,381 answered this DCE module; after exclusions, the final DCE sample is 20,942 (≈50% female), yielding 125,652 choices (376,956 observations in stacked format). The overall survey also included online standardized assessments in mathematics, reading comprehension, and foreign language, linked to the PISA scale via common items, and a background questionnaire. Achievement measurement: Competencies were assessed using Item Response Theory (IRT), estimating ability parameters comparable to PISA scales; these were used to explore relationships between preferences and achievement. Estimation strategy: Baseline Multinomial Logit (MNL) estimated main effects of study mode, subject, and time (price). To capture heterogeneity, Random Parameters Logit (RPL/mixed logit) with normally distributed random coefficients was estimated (Stata mixlogit, 500 replications). Willingness-to-pay (WTP) in time units (hours) was derived as marginal rates of substitution between attribute levels and time. To address issues with ratios of random coefficients, a mixed logit in WTP space (Stata mixlogitwtp) was also estimated, with time modeled log-normally. Heterogeneity by achievement (top/bottom 20% in math) and gender was examined via subgroup MNL WTP estimates, and subject-by-mode interactions were included to test context dependence. Individual-level parameters were imputed (Revelt & Train approach via mixlbeta). A multilevel model decomposed variance at student and school levels using individual RPL coefficients, estimating intraclass correlations (ICCs).

- Time cost: More required time significantly reduces utility (RPL time coef −0.257, SD 0.258). - Preferred study mode: Relative to group work (reference), other modes are on average less preferred, but with strong heterogeneity: • With tutor: RPL mean −0.597 (SD 1.299); 32.3% have positive preference. • Independent: RPL mean −0.181 (SD 1.258); 44.3% positive. • With parents: RPL mean −0.437 (SD 0.931); 31.9% positive. Overall, group work is most preferred. - Subject preferences (relative to Polish/Native): • English: RPL mean 0.515 (SD 0.952); 70.6% positive. • Mathematics: RPL mean 0.245 (SD 0.885); 60.9% positive. • Geography: RPL mean −0.170 (SD 0.765); 41.2% positive. - WTP in hours (MNL; consistent with WTP-space model): • To switch from working with a tutor to group work: −2.34 h (i.e., willing to give up >2 h of leisure to be in group). RPL WTP mean ≈ −2.44 h. • From parents to group: −1.69 h (RPL ≈ −1.62 h). • Independent vs group: −0.22 h (RPL ≈ −0.61 h); on average, small dispreference to independent relative to group. - Heterogeneity by preference sign (average WTP among those with positive vs negative coefficients): e.g., tutor +1.42 h vs −3.00 h; independent +1.42 h vs −1.68 h; English +2.06 h vs −0.69 h; mathematics +0.86 h vs −0.44 h. - By achievement (top vs bottom 20% in math; MNL WTP): • Top performers more strongly prefer group over adult-supervised modes: with tutor −3.129 h; with parents −2.226 h; bottom performers −1.959 h and −0.979 h, respectively. • Independent vs group: top −0.072 h (slight preference for group); bottom +0.398 h (≈24 minutes preference for independent work over group). • Subjects: top value mathematics +2.350 h and English +2.631 h; bottom value English +2.116 h, mathematics +0.208 h, and strongly dislike geography (−2.356 h). - By gender (MNL WTP): • Both genders prefer group over tutor/parents; males especially dislike tutor (−3.007 h). • Independent vs group: males +0.304 h (prefer independent); females −0.802 h (prefer group, willing to give up ≈48 minutes to avoid independent work). • Subjects: males value mathematics more (1.704 h) than females (0.620 h); English valued by both (≈2.2–2.6 h); females dislike geography (−1.408 h) more than males (−0.219 h). - Subject–mode interactions (MNL with interactions): preferences depend on subject. Students prefer mathematics with a tutor/parents over Polish in a group; they prefer Polish in a group over studying mathematics independently. English is preferred to Polish across modes; geography with tutor/parents is disliked relative to Polish in a group. - Multilevel variance: School-level ICCs are small (<1% across attributes; e.g., tutor mode ICC 0.78%), indicating most preference variation is at the individual level, not between schools. - Sample/estimation: 20,942 respondents; 125,652 choice tasks; 376,956 observations; MNL and RPL parameters significant at 5% with substantial SDs indicating heterogeneity.

The findings show that when preferences are elicited via DCEs with explicit trade-offs, Polish 14–15-year-olds generally prefer group work, contradicting interpretations from Likert-scale indices (e.g., PISA) that suggested aversion to teamwork. Preferences are highly heterogeneous and context dependent: they vary by subject, achievement, and gender. Aligning instructional modes with student preferences can enhance satisfaction and potentially increase effort and performance. Notably, students value English and mathematics projects, and top performers especially prefer group work over adult-supervised modes, while low-performing males prefer independent work. The subject–mode interactions indicate that support (e.g., tutor) can increase acceptability for demanding subjects like mathematics, whereas geography is less favored under adult supervision. Methodologically, DCEs overcome biases in rating scales by forcing trade-offs and enabling estimation of WTP in a common metric (time). The small school-level ICCs suggest that interventions should be individualized rather than school-wide one-size-fits-all policies. The results inform policy and practice on structuring collaborative learning to maximize engagement, reduce free-riding, and accommodate diverse student needs, including in digital learning contexts where social interaction may be limited.

This study introduces a DCE-based, multidimensional approach to measuring students’ preferences across study modes and subjects, linked to IRT-based achievement. It demonstrates that group work is, on average, the most preferred mode, with substantial heterogeneity by student characteristics and subject. Preferences measured via DCEs differ from those inferred from Likert-type indices, underscoring the value of stated-preference methods in education. Policy implications include designing collaborative learning that is well-structured, equitable, and sensitive to subject demands and learner profiles, potentially mitigating math anxiety through supportive modes and enhancing motivation and outcomes. Future research should replicate in diverse socio-cultural contexts and grade levels; incorporate additional attributes (e.g., grading stakes, frequency of modes, group size, varieties of group work), and measure prior experiences and expectations to disentangle preferences from beliefs. Extending DCEs to other non-cognitive educational constructs could improve evidence for policy and classroom practice.

- Context specificity: Conducted in Poland; generalizability requires replication elsewhere and across grades. - Experience not measured: No data on students’ prior frequency/quality of group work or whether experiences were collaborative vs cooperative; cannot separate preferences from expectations. - Attribute scope: Group size not varied; varieties of group work not differentiated; grading stakes and frequency of operation in different modes not included. - Measurement limits: While DCE reduces some biases, unobserved factors and hypothetical bias may remain.