choice-metrics.com

Dear all,

I have an attribute in my design which can get both negative and positive direction, depending on respondents preferences. The attribute is "life expectancy before treatment" and in the pilot study I coded the attribute as dummy and with priors very close to zero (b3.dummy [0.0004|0.0002] * LE [8,5,2]).

I conducted the pilot study and obtained coefficient of 0.37 (sd: 0.11, p=0.001) for life expectancy at 8 years and 0.18 (sd: 0.12, p=0.154) for life expectancy at 5 years compared to 2 years.

I have been discussing the results and some expect that people prefer to treat patients with shorter life expectancy which is not the case with the pilot results. The litterature shows mixed results depending on the level of the life expectancy (if it is some months or years?)

Does the fact that the prior in my pilot study have positive direction affect the findings? I think it was better to use zero priors for all levels!!!

Do you recommend using the optained priors in a Baysian efficient design?

Thank you,
Nasrin

Prior values of 0.37 and 0.18 are still relatively small and will not have a major impact on the design. A Bayesian efficient design would indeed be preferred and for example a prior like (n,0.18,0.12) will also draw some negative values.

There may be a lot of heterogeneity in the population, with people attaching negative or positive values to certain attribute levels. You will pick this up when you estimate a mixed logit or latent class model. When you estimate a multinomial logit model you can only pick up population averages but the true parameter values may be both on the positive and negative side.

Michiel

Thank you Michiel.

I estimated a latent class model and as you mentioned people attached both positive and negative values to this attribute.

Best, Nasrin