“ Shopping cost and brand exploration in online grocery ” by Andrea Pozzi Web

The very definition of brand exploration requires the identifications of the set of brands that are already known to a consumer at t0. I construct this set using the first few months of purchase data which are not then used for the estimation. Choosing the length of the period used to calibrate initial conditions imposes a trade-off. On the one hand, the longer the amount of data allocated to recovering the set of known brands the more accurate this will be. On the other hand, extending the length of the calibration window reduces the number of observations left for the estimation. For the estimates reported in Table 7 of the main text I implicitly assumed that three months was a long enough period to credibly recover the set of cereal brands explored by the household in the past. In reality, this assumptions surely generates spurious instances of brand exploration, although it is not obvious that this bias would be at all correlated with the shopping channel chosen for the trip. Nevertheless, I present below evidence that results are not overly sensitive to different assumptions. A first concern relates the fact that I retrieve the set of brands already known to a household using a limited span of observation and only using purchases occurring at one specific grocery chain. Ideally, this set should be constructed using the entire shopping history of the household at all the stores visited. In Panel A of Table A-1 I asses the practical relevance of this shortcoming using an additional piece of scanner data: a subset of observation from the HomeScan panel, collected by AC Nielsen. Each household in this sample is equipped at home with a scanning device through which they can record all their purchases at every grocery chain visited. The sample I obtained covers over 89,000 households in the period between January 1998 and December 2007, although not all the households are in the sample for the entire span. 85,827 households shopped for breakfast cereals at least once for a total of 2,752,802 trips. The number of households engaging in online shopping of grocery is limited, which explains why this data source could not be used for the core analysis. The HomeScan panel allows to recover the set of known brands using a much longer shopping history, potentially leading to accuracy gains. For purchases generated by households active between June 2004 and June 2006, the difference between the number of brand trials counted using a three months calibration span and the number resulting from using the full previous history for the calibration is 71,194 cases. This implies that mislabeled brand exploration instances make up 3% of the overall number of trips. Some of these households, however, may not have been in the sample very long before June 2004 which would reduce the gap between using three months or the entire history to define brand exploration.