17 October 2016 Christine URBANO Comments Off
Imagine you’ve just changed the recipe for one of your products;

you’ve added a flavor and you want to make sure your customers can tell the difference between the old and the new recipe.

You might want to try the so-called triangle test. This test is one of the basics of sensory analysis: panelists are presented with three coded samples – one different and two alike – and are simply instructed to identify the odd sample (their choice is forced).

If none of the panelists perceive the difference between the products, they will answer randomly, and the success rate will be around one-third (which means that the panelists have a one in three chance of getting the right answer at random). It is said that, assuming nobody can tell the difference between the two products, the number of successes follows a binomial distribution rule with a sequence of n tests (the number of panelists) and a one-third chance of success.
The higher the number of successes you obtain (with results deviating from one-third proportionately), the more likely it will be that some of your panelists have in fact perceived a difference and that you ought to reject your assumption, as you were not in line with that binomial distribution rule.

For example, if you administer this test to a panel of 20 tasters and obtain 10 right answers, your risk of being mistaken in saying that some of your panelists can tell the difference between your products (i.e. rejecting your assumption) is α = 9.2% (i.e. less than a 1 in 10 chance). Would you be satisfied with this? It isn’t too bad, but if your risk threshold is 5% maximum, which is the usual threshold in sensory analysis, your attempt to show a substantial difference will prove unsuccessful (your initial assumption cannot be rejected).

But does this mean you proved that your old and new recipes were similar? Not at all! In fact, even if you didn’t reject your initial assumption, you didn’t prove it was right either. That you didn’t succeed in proving the recipes were different does not mean you succeeded in proving they were similar either! Frustrating, isn’t it?

In order to prove that two samples are similar, one has to take into account the so-called β risk and a “discriminator percentage”. But that’s another story… to be continued. 🙂