USING THE CHOICE SEQUENCE IN TIME TRADE-OFF AS DISCRETE CHOICES - DO THE TWO STORIES MATCH?
Time trade-off (TTO) is the de facto gold-standard for eliciting health state utilities (HSU) for EQ-5D health-states. However, there is a growing interest in the use of discrete choice experiments (DCE), as DCE is cheaper to administer. Importantly, the standard way of eliciting TTO values is through a series of discrete choices. The aim of this study was to compare TTO-derived HSUs and HSUs estimated by inferring DCE values from TTO responses.
TTO-values for 10 EQ-5D-5L health states were collected from 202 study participants. Mean TTO-HSU were calculated for each of the 10 health states. For comparison, the same data was treated as a series of discrete-choices. We used a “non-stopping TTO” procedure, meaning that all respondents answered all possible TTO questions for all 10 health states. Each strict expression of preference was recorded as a DCE. For indifference a random DCE-draw was used. After removing 69 incomplete responses (3.4%), we generated 42922 discrete choices. Each DCE-pair consisted of (A) 10 years in an impaired health-state, preceded by 10 years of lead-time without impairment, and (B) a shorter life without impairment.
We estimated simple binary logit/probit models to estimate latent utility, assuming that respondents are utility-maximizers subject to an IID-error term. The latent utility for each choice-situation was modeled as
Y = (t*DA – (b*HSV)*DA) – t*DB*DD+ error
where t is a parameter for the value of time without impairment (1), DA and DB are the durations of choice A and B respectively, HSV is a dummy vector indicating the impaired state in choice A, and b is a vector of parameters to be estimated and interpreted as the disutility of spending time in the reduced health state. A is chosen whenever Y>0, corresponding to A having greater utility than B. Each health-state was also characterized worse-than-death or not included in the model by the dummy DD.
The logit-model achieved the highest observed maximum likelihood. The estimated DCE-utilities agree well with the TTO-utilities for health states with low utility (close to 0). However, states with high TTO-utilities are not matched by their DCE counterparts; and even exceed 1 – inconsistent with the QALY-model.
The values estimated using the inferred discrete choices did not match the scale of the corresponding TTO values.