OBJECTIVES:

Challenges arise when discrete choice experiments (DCEs) involve many attributes. Two methods to deal with many attributes are Partial Profiling and Hierarchical Information Integration, but they require additional assumptions and pose practical difficulties. We present a new approach, fold-in-fold-out (FiFo), which makes it possible to retain all necessary attributes while reducing the cognitive burden.

METHODS:

A DCE was conducted to quantify contributions of various aspects of COPD to its burden of disease. Choice sets consisted of 15 attributes with three levels. In order to make the choice task feasible, 13 attributes were grouped into three dimensions. In some choice sets, all dimensions were folded-in: all attributes in that dimension were set at the same level. In most sets, one dimension was folded-out, so attribute levels varied.

In the Bayesian mixed logit analysis extra parameters were added: (1) ϕ for additional complexity in folded-out choices, and (2) λ, λ, λ for differences in choice behaviour due to each dimension being folded-out or in. Raw regression coefficients represented attributes’ folded-in states. Folded-out could be represented by coefficient*(1+ λ). A burden of disease Index on a 0-100 scale was developed by extrapolating coefficients, without and with partial adjustment to folded-out status. The latter was considered the best representation of the choice context.

RESULTS:

The ϕ parameter indicated that folding out led to more complexity and to less consistency. The λ parameters showed that attributes got more weight when folded-out. When no λ adjustment was made in the extrapolation, the fatigue attribute made the largest ighest contribution to the Index (maximally 14 points). With adjustment, limitations at daily activities became equally important (11pt for both), while these limitations seemed considerably less important without adjustment (8pt).

CONCLUSIONS:

FiFo makes DCEs with many attributes feasible, but requires more complex statistical models with additional parameters.