Avoid overliance on average partworth and importance charts
Running scenarios at the individual levels is an important sanity check on average partworths, and another advantage of metric conjoint analysis over approaches that ignore individual differences.
Metrisim includes these charts to give you a sense of the overall average importance of each attribute and the impact of each level on purchase likelihood ratings.
However, while it is traditional to examine these to give you an initial sense of direction, you should keep in mind that it is better to run simulations to confirm your understanding.
The reason for this is that the simulations are run at the individual level and translated into shares via decision rules and at times some other calculations such as external effects calculations.
Correlations
If there is a correlation between partworth utilities, this can result in a quite different share change from what you might have expected looking at the averages.
One way in which you can visualize this is by running a cluster analysis on the segment tab.
In the UHT milk example, you will notice that in a ‘3 cluster’ analysis, certain clusters exist which are brand senstitive, but not so price sensitive; while others are price sensitive but seem less interested in other attributes. These results indicate some correlation may exist between partworths, and so it is quite possible that for certain respondents who value a certain brand, that they might be less price sensitive. So when you select this brand in the simulator, and run simulations you may notice that this brand show less change in market share when changing price than for other brands.
The same might be true of other attribute levels such as packaging – where if you select an environmentally friendly option such as recycled packaging, other attributes such as price may then seem less influential.
Weights
If you have volume weights or a compound weight that represents both volume and respondent representation, you may also notice different outcomes from expected given the average partworths. This is because their may be a correlation between volume purchased (by volume I mean both actual volume measures such as ‘litres of milk per month’ or the number of units bought – such as the number of chocolate bars purchased each month) and partworths.
For example, if a group of heavy buyers preferred brand A to brand B then switching to A in a simulation may have a larger positive impact on share than expected. If they prefer B, then of course the reverse is true.