So in January I asked for the $3 million-not-put-out-to-tender survey data that was collected by Melanie Wakefield to track the introduction of plain packaging. It eventually arrived. The survey and supporting documents can be found here.
The survey itself consists of 26 waves of surveys beginning 9 April 2012 and concluding 30 March
2014. Then there are follow up surveys that begin 7 May 2012 and conclude 4 May 2014. So the data consists of 26 stacked cross-sectional surveys. So I’ve been playing around with the data. So far not all the results already reported from the data set can be exactly replicated – but most are close enough. Then there is is cherry-picking – not all results are reported. This data set is going to keep me busy for some time.
What I want to show this morning is some regression analysis I did exploiting the cross-sectional features of the data.
So for each wave I estimated a probit regression with “Daily Smoker” as the dependent variable. I included a number of demographic control variables – age, sex, education and income, then a health concern variable, and whether they approved of plain packaging as a policy. Then variables comparing how smokers thought their tobacco compared to a year previously on “Quality”, “satisfaction”, “Value for Money” and “harmfulness”. Finally a variable indicating whether they “Like the pack” and an interaction between “Like the pack” and whether the current pack they are consuming is a “Plain Pack”.
So I ran 26 regressions – one for each wave – and collected the coefficients and then graph them. The graph below shows the coefficients for “Like the pack” (LIKEPACK) and the interaction term (LIKEPACK*CURR_PLAIN). If the plain packaging policy worked as advertised we would expect to observe a statistically significant negative coefficient on LIKEPACK and a positive and statistically significant coefficient on the interaction term.
The Like the pack coefficient is only ever negative and statistically significantly different from zero in one instance – that big negative spike that you see at Wave 20. Wave 20 begins on 11 September 2013 –
11 days after the massive 12.5% increase in tobacco excise. The interaction term is positive and statistically significantly different from zero (p < 0.05) when plain packaging is first introduced and then when the excise is increased. I interpret those results to suggest that the only time the plain packaging policy "worked" was when it was first introduced and when the excise increase came into effect.So there is a Hawthorn effect in the data and a price effect.
Update: A good friend emails to point out an error on my part. While the excise increases have normally being introduced in September, the first 12.5% increase was in December. So not coinciding with wave 20 but rather with wave 23. This is a good error to have made because it strengthens the argument. Plain packaging “worked” when it was introduced and at one other random point in the survey period.