Course · 1. Background and motivation

From choices to value

Utility

The models behind conjoint assume each person gives each product a score, its utility, and picks the highest. The utility is a sum over the product's attributes:

utility(Surge Berry 500ml at $2.49) = value of Surge + value of Berry + value of 500ml − price sensitivity × $2.49 + a random part

Each "value of a level" is a part-worth. The random part is everything the attributes do not capture: mood, what is in the fridge, which product was at eye level. It is why the same person does not always choose the same way, and why the model gives probabilities of choice, not certainties.

What the model must assume

Two assumptions decide almost everything a conjoint model can tell you.

1. Do people differ in their tastes? A model where everyone has the same part-worths (an aggregate model) is quick but blurs segments together. Modern methods let part-worths vary across people. The industry standard, hierarchical Bayes (HB) logit, does this, and so does our model.
2. How is the random part shaped? The standard logit assumes the random parts of different products are independent of each other. That one assumption fixes how the model thinks people switch between products, before it has seen any data. The next page shows why that matters.

Reading the numbers

• Part-worths are relative within an attribute: what matters is the gap between Surge and Volt, not either number alone. Chosŏn shows them centred on zero.
• Importance is an attribute's range of part-worths as a share of the total: how much it can move a choice.
• Willingness to pay divides a part-worth gap by the price sensitivity, turning it into money.
• Every estimate comes with a 95% range. See Uncertainty.

Next: How shoppers switch