Indicators and Individuals are Both Variables: Simulating Alternative
Parameterizations of Micro-to-Macro Models in Small Groups and Teams
Research

ABSTRACT

Applied researchers working on small groups and teams have been slow to
adopt multilevel latent variable measurement models for predictor variables
in micro-to-macro models—where individuals responses to multiple indicators
reflecting predictor variables are clustered within teams and the relations
of interest are between teams. We extend previous technical work to show
that researchers can apply latent variable modeling procedures to datasets
that have both small numbers of teams and small numbers of members in those
teams. We conducted a Monte Carlo simulation study in Mplus to evaluate
three approaches to parameterizing predictor variables where measures
satisfy conventional psychometric criteria in micro-to-macro models: (1)
aggregating responses across both indicators and individuals, (2) two level
latent variable modeling of the same data, and finally (3) single level
latent variable modeling. We found for latent variable modeling procedures
compared to classical procedures across all conditions that we produced (1)
over 80% convergence rates, (2) dramatically reduced bias, (3) surprisingly
high coverage, despite relatively inefficient standard errors, and (4)
dramatically increased power. We also found that Bayesian estimation with
mildly informative priors dramatically increase the efficiency of the SEs
across all conditions. Based on our findings, we urge small groups and
teams researchers to adopt latent variable modeling procedures, and
consider moving to Bayesian estimation, to produce more trustworthy, if not
perfect, results from their data.

Keywords:

Small Groups and Teams research; Sample size and trustworthy results;
Multilevel structural equation modeling (MSEM); maximum likelihood (ML) and
Bayes estimation;