Description
A maximum likelihood model, fitted like a regression, that tests whether a behavioural outcome is better described by a cusp with two stable states controlled by a 'normal' and a 'splitting' variable than by a linear model.
The cusp catastrophe describes a behaviour that changes smoothly when a splitting variable is low but jumps between two stable states, with hysteresis, when it is high. Cobb made this deterministic model statistical by adding stochastic noise to the dynamics, which yields a probability distribution that can be unimodal or bimodal depending on the control variables. As in regression, the behavioural state is a weighted sum of measured outcome items, and the normal (alpha) and splitting (beta) controls are each linear functions of measured predictors. Estimation gives coefficients, standard errors and fit indices, and the cusp is compared with linear and logistic alternatives using AIC and BIC. A classic example in the book is attitude data where intensity of feeling loads only on the splitting axis: low-involvement respondents are spread around the middle, high-involvement respondents polarise. The book's rules are to simulate data from the model first to understand power, to use independent cross-sectional observations (the method is not valid for time series), and to remember that a better cusp fit is evidence, not proof. For UX this turns claims about polarised attitudes or all-or-nothing adoption into testable models.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Siloed
- Stage
- Discovery, Evaluation
- Data
- Quantitative
- Effort
- Days
Steps
- Specify the behavioural outcome and candidate predictors for the normal axis (pushes the outcome up or down) and the splitting axis (amplifies polarisation, such as involvement).
- Simulate data from a cusp with plausible parameters and fit it, to check that your sample size can recover the effects.
- Inspect the data descriptively: histograms overall and at high levels of the splitting predictor, and pairwise plots.
- Fit the full cusp model with each predictor allowed on both axes, then reduced models that drop non-significant paths.
- Compare the cusp with linear and logistic models using AIC and BIC, and check how many observations lie inside the bifurcation set.
- Interpret which predictors act on which axis and report the result alongside other catastrophe flags.
Sources
- van der Maas, H., Complex Systems Research in Psychology (2024)
Also known as
- Cobb's cusp method
- Stochastic cusp model
- Cusp regression
- cusp R package
Categories
- Analytics & Behaviour
- Statistics & Modelling