Sir Richard Blundell, a very well renowned British economist (as you might guess from the Sir in front of his name), is visiting UBC this week. During his time here is giving a series of advanced lectures; today’s lecture was on the theory and applications of revealed preference theory. Here I am going to give an outline of one piece of his work that he discussed: a paper that looks at the changes in demand for cigarettes in Britain during the 1980s and 1990s. This paper is particularly important because it simultaneously advances economic theory and contributes to a very important policy debate. The paper in question is (somewhat blandly) called Prices versus Preference: Taste Change and Revealed Preferences, and is co-authored with Abi Adams, Martin Browning and Ian Crawford.

This post is reasonably technical, but should (I hope) be readable for most. If you manage to stick through the economics at the beginning then you will see some remarkable insights at the end regarding the effects of taxes and health information on the behaviour of smokers. To summarize, health information has very little effects on lower educated, heavy smokers but a much greater effect on higher educated, lighter smokers. However, increases in cigarette taxes do have an effect on lower educated, heavy smokers.  [Aside: what do you think the welfare implications of these results are?]

Economists usually work on consumption data within a paradigm called the revealed preference paradigm. Imagine an economist who observes people making consumption choices as they go about their lives. The economist considers these choices to be indicative of, and driven by, the preferences that the people hold. If the economist sees a person purchase bananas when they could also have purchased apples (for the same price), then the economist reasons that the person must like bananas at least as much as they like apples.

If an economist views a person’s purchases on a number of different occasions then they can start to build a complete picture of the person’s preferences. It is then worthwhile to check whether the preferences of the person are sensible. If someone purchased lots of bananas when bananas were expensive (and apples were cheap), and then purchased lots of apples when apples were expensive (and bananas were cheap) then we would say that this person behaved irrationally, and point out that they may be better off if instead they purchased more apples and bananas when each were cheap. If a person behaves in fashion consistent with a model of behaviour then we say that the model rationalizes their behaviour (and often claim that this is good for both the model and the person).

Now, it is entirely possible that preferences can change over time. You might like bananas today, but over time come to prefer apples. Changes in preferences of this sort can create problems for standard economic models. If, on one hand, we assume that preference are fixed over time (which is an assumption that is often made) then we risk accidentally mis-classifying people as making irrational choices when really all that has happened is a simple shift in preferences. On the other hand, if we allow for a model in which preferences can change then the model will be able to rationalize any observed data (a point that Blundell et al. prove in their paper). That is, once we allow for preferences to change then the model loses its power to tell us anything interesting.

They first key innovation in the Blundell et al. paper is that they introduce a methodology for finding the smallest possible preference change required to rationalize a data set. This is really useful, because it enables us to identify when the changes in preferences required to rationalize the data are too large to be plausible. For example, if you are willing to pay at most $1 for a banana today, but are willing to pay $100 next month (and your income hasn’t risen dramatically), then we would feel pretty comfortable in claiming that this change in behaviour cannot be explained by a reasonable change in preferences.

The second, more technical, key innovation in the Blundell et al. paper is that they generalize their procedure up to the population level. Ideally, you would observe the same people over and over again so that you can measure individual preferences. But unfortunately most data sets contain different people’s data at different time periods. To get around this, Blundell et al. use quantile regression techniques to estimate the minimal changes in the distribution of preferences across the population that are required to rationalize the data. This requires some reasonably strong assumptions about the structure of population preferences, but is nevertheless quite remarkable.

Blundell et al. then test their model in a domain where they have fairly good reason to think that preferences have changed over time: preferences on smoking between 1980 and 2000. This provides a good diagnostic test for the model (and its associated assumptions) as this was the period when the harmful effects of smoking became widely known, and we should expect this to reduce the preference for smoking. However, current evidence cannot tell us how much of the decline in smoking over this period was due to price increases in cigarettes (i.e. not caused by preference changes) and how much was caused by changes in preferences.

As might be expected, on aggregate there is a reasonably large decrease in the preference for smoking. Blundell et al. split their data into a few different groups based on educational level and amount of smoking, and the separated data is also very interesting. Highly educated, light smokers experience the biggest decline in preferences for smoking, while heavy smokers (and particularly low education, heavy smokers) experience almost no change in preferences for smoking. This indicates that almost all of the reduction in smoking for heavy smokers over the period 1980-2000 was caused by rising prices (and not health campaigns).[1]

The model appears to give results that are plausible, whilst also providing some new insights into the data. All in all, a pretty good result. It would be very interesting to see what insights this type of modelling approach could produce in other domains. Certainly, there are a couple of heavy technical assumptions that may or may not be reasonable in other domains, but testing them and searching for relaxations seems like a useful avenue for future research.

Now, what are the implications for welfare of the above results [note: this is me being somewhat speculative, and is not in the paper]. One key concern of raising taxes on products that are primarily consumed by poorer individuals is that they can be very regressive. For smoking, this concern might be eased if we think that health education campaigns are reducing the preferences for smoking. If people are smoking less because they wish to smoke less, rather than because higher prices are forcing them to smoke less, then the regressive effects of the tax are ameliorated.

The data presented by Blundell et al suggests, however, that amongst the cohort for which heavy smoking taxes are the most regressive (low education, heavy smokers) there is no evidence of any shift in preferences. The share of income spent on tobacco for this cohort is roughly constant over the period 1980-2000 which, given the 80% rise in prices, implies about a 44% decrease in tobacco consumption over that period. However, this decline in consumption is not attributable to a decline in desire for smoking — the data suggests that the desire for smoking has remained roughly constant and the change in consumption is driven purely by the price increases. Smoking taxes may actually be reducing the welfare of this cohort.

[1] It is worth noting at this point that the study here focused a cohort of people that were between the ages of 25 and 35 in 1980, so the data does not tell anything about the effects of health campaigns on the number of young people who might start smoking.