Which Simulation Method Should I Use?

Last Updated: 10 Aug 2017Hits: 5079
I conducted an experimental CBC study and I would like to use the simulator. However, I do not feel very confident to test which of the methods would be the best in terms of real choice data for example. Is there any recommendation for one of the simulation methods when using CBC/HB data?

Generally, we recommend Randomized First Choice (RFC).  But, so much depends on the nature of your data and your goals.

For example, if you have just two attributes (brand and price), then RFC doesn't make much sense (just takes up extra computing time without doing anything for you in terms of correction for product similarity).  That's because we recommend the Price attribute be "turned off" from receiving correction for product similarity (Price, then, get's unique error perturbations).  Then, if all the brands are formatted as separate levels (as they usually are, leading to unique error perturbations on them as well), then RFC could provide no correction for product similarity on the brand attribute (since products that have different levels for an attribute are viewed as entirely distinct with respect to that attribute).  You'd be left to what's pretty darned close to the faster share of preference (logit) method.

For conjoint studies involving more attributes (than brand and price) where you think that similarity between attributes (in terms of shared levels on attributes) should lead to enhanced competition among such product concepts, RFC offers a practical and pretty fast way of doing it.

That said, once you have estimated utilities at the individual level via something like HB, the differences between using RFC and the standard logit (Share of Preference) rule are usually fairly minor...except in the situations involving simulations involving pairs of extremely similar products (in competition with other products).

Another solid approach (which you can trick our software into doing with a bit of data manipulation of the .HBU file, and which makes formal statisticians & academics happier) is to use 50 to 200 draws per respondent.  Then, you simulate across those draws using the Share of Preference (logit) rule.  It messes up your standard errors in our simulator (you'd need to ignore them), but the shares of preference are just fine.

Here is a document from the Technical Papers section on our website which explains more information about how to properly use market simulators with conjoint data:

  Introduction to Market Simulators for Conjoint Analysis (2009)