Monday, July 09, 2012
Thoughts on the accurate prediction of respondent-level behavior
Let me start this post with a little story. A couple of weeks ago some friends were coming round for dinner. After leafing through a few cookbooks to find suitable dishes, I carefully noted down the required ingredients and headed off to the store. But the store was out of two key ingredients: halibut and raisins. Instead, I bought pork loin and dried cranberries.
And so it goes for many of our purchases in life. Despite our best intention, something intervenes to deflect us from carrying through on that intent. The same applies when we shop for branded goods and services. We might start out thinking we will buy one brand and end up buying another. That does not make those intentions wrong. If an interviewer had called me up before I set out on my trip to buy dinner ingredients, I would faithfully have told them what I intended to buy. I just did not get the opportunity to make good on my intent. In a survey, people can only tell you about what they might do given their current understanding of the purchase context.
So why am I telling you this? Because there is an ongoing debate about whether you should predict respondent behavior based on individual people’s responses or at the aggregate level (after summing up responses of all the interviewees). To my mind, this is a completely irrelevant debate. As with all research, methodological validity depends on the objective. What are you trying to predict? Both approaches are valid but they address different needs. And neither approach is guaranteed to produce a completely accurate prediction of behavior, either in aggregate or at the respondent level. Intent is an important influence on behavior but not the only one.
In the typical tracking study, we measure a respondent’s predisposition to buy a brand. Note that I prefer to use the word “predisposition” rather than purchase intent, precisely because it reflects the feeling that people have about brands under most circumstances. Predisposition implies an inclination to do something. Intent implies determination to do something and that usually only gels close to the point of purchase.
And as people approach the point of purchase, various barriers and facilitators will have an influence on whether or not a specific individual’s behavior reflects their initial attitude. These barriers and facilitators usually only become apparent to the individual as they move through the purchase process, e.g. sales and discounts, lack of availability, recommendation from friends, poor review ratings, etc. They are usually unknown at the point of interview.
So is it worth measuring predisposition? Of course it is. If someone is predisposed to buy one brand more than others, then that brand is in pole position at the start of the purchase race. But few people are strongly predisposed to one brand alone. Depending on how well brands compete to capitalize on people’s predisposition they may buy another brand in their consideration set, not the one they were most predisposed to buy. Or, increasingly in these days of digital search and advocacy, people may come across a completely new brand that meets their needs while shopping.
So what can we do to improve our prediction of behavior? That all depends on your objective.
If you want to understand the probability that a brand will increase or decrease in market share across a longer time frame, then the wisdom of crowds works in your favor. En masse the mismatch between individual attitude and behavior tends to cancel out, particularly if you take macro level influences on behavior into account, like brand size (people are far more likely to fulfill their inclination to purchase big brands than small ones).
But if you want to predict the behavior of an individual, then you have to allow for the influence of their specific barriers and facilitators from the time of the survey up until purchase. And to do that properly, you need to collect data from the same individual over time, not just ask questions at one point in time. We have done this for a media company, contacting people again the week after the interview to see if they followed through on their viewing intent. Initially the fit between intent and behavior was very close, but as the season progressed, various factors intervened to mean that fewer people followed through on their intent.
At Millward Brown, we deploy respondent-level models and have developed them to accurately predict a brand’s current power, its ability to command a price premium, and its potential to grow. The first two are validated against respondent level and aggregate level data from sources like Shopcom. The latter, an aggregate metric validated against market share change over the course of a year, as it should be.
So my viewpoint is that we should cut the needless debate and make sure our tools match the task. What’s yours? Please share your thoughts.
This entry was posted on Monday, July 09, 2012
and is filed under Research.
You can leave a response.