by Michiel Bliemer » Thu May 10, 2018 9:10 am
Pivot designs in Ngene are not optimised for each individual respondent, but rather are optimised based on a reference level, and relative (or absolute) pivots are used to tailor (but not optimise) the design to the specific respondent.
In the simplest case, the reference level represents the average attribute level across the population.
In the more advanced case, the population is split into groups and each group has separate reference levels.
For example, consider route choice and assume that the average travel time in the population is 30 minutes. Then setting traveltime.ref[30] means that Ngene is optimising designs for a reference respondent, but within the survey the actual levels are replaced by respondent specific data; so if a respondent states that his/her travel time is 50 minutes, then levels of 50 +/- 25% are presented. Clearly, the further away the travel time of the respondent is from the average, the more the design will lose efficiency.
More efficiency can be maintained by splitting the population for example in short and long distance travel time with for example average travel times of 20 and 50 minutes. Then you can create two pivot designs in Ngene, one with reference level 20 and one with reference level 50. If a respondent has a travel time of more than (say) 40 minutes you can take the 50-optimised design, while for all other respondents you take the 20-optimised design.
The more segments you consider, the more efficient your data collection will be. This will in the end generate a database of designs and for each respondent you pick a design from this database. Segments can be combinations of different attributes, e.g. (low travel time, train user), (high travel time, car user), etc. For a study in the UK we created a large library of designs (more than 1000) off-line, such that within the survey we could simply pick an efficient design.
If you already have information available from the respondent (that means, you are doing your survey in a two-stage process), then you could optimise for each individual respondent. This would simply mean generating many designs individually within Ngene.
What we found is that using average values for the whole population or segments of the population is generally good enough for efficiency, it is generally not necessary to optimise for each individual separately (which is a lot of work). For more information, please refer to:
Rose, J.M., M.C.J. Bliemer, D.A. Hensher, and A. Collins (2008) Designing Efficient Stated Choice Experiments in the Presence of Reference Alternatives. Transportation Research Part B, Vol. 42, pp. 395-406.
Michiel
