Additive and multiplicative probabilistic models of infant looking times

Area of Science:

• Developmental Psychology
• Cognitive Science
• Statistical Modeling

Background:

• Habituation, a fundamental learning process, is studied in infants using looking time (LT) measures.
• Previous research has utilized additive and multiplicative regression models to analyze habituation, with varying conclusions on their fit and interpretation.
• The choice of probability distribution and modeling approach (e.g., pooling vs. partial pooling) can influence the analysis of LT data.

Purpose of the Study:

• To compare the goodness-of-fit of additive and multiplicative regression models for infant habituation looking times.
• To evaluate multiple probability distributions (Weibull, gamma, lognormal, normal) within these models.
• To assess the impact of experimental factors (contrast type, number of trials, age) and modeling choices (pooling methods, model complexity) on habituation analysis.

Main Methods:

• An infant habituation experiment was conducted with participants aged 3-11 months, varying contrast types, number of habituation trials, and age cohorts.
• Additive and multiplicative regression models were fitted to looking time data, incorporating various probability distributions.
• Model performance was evaluated using log-likelihood, mean absolute error, predictive accuracy, and parameter correlations. Both non-pooled and partial-pooled (hierarchical) models were compared.

Main Results:

• The gamma distribution demonstrated the best fit and predictive performance among the tested distributions.
• Partial-pooling models yielded more precise predictions and reduced parameter correlations compared to non-pooled models, despite a slightly worse data fit.
• Infants showed stronger dishabituation to luminance contrast changes (dark to bright) than orientation contrast. Dishabituation effects were more pronounced after more habituation trials, though not always statistically robust.

Conclusions:

• The gamma distribution is a suitable choice for modeling infant habituation looking times.
• Partial-pooling (hierarchical) models are recommended for their superior predictive accuracy in developmental studies.
• While multiplicative models are generally preferred for habituation, the choice between additive and multiplicative models may have less impact on statistical inference than previously thought.