The conventional wisdom about Black Friday (the day after Thanksgiving) Shopping in the US is that it is the primary driver of retail sales for the year! Economic reasoning would tell us that retailers would not be opening earlier and earlier, creating labor problems, and giving deeper and deeper loss-leader discounts if something economically important wasn't happening. So when Paul Dales of Capital Economics published the chart above, which contradicts the convenitonal wisdom, the story was picked up by many news outlets (for example, in the Washington Post, Black Friday is a bunch of meaningless hype, in one chart).
What the chart shows is that when percentage change in sales during Thanksgiving week (the x-axis) is use to predict overall retail sales (the y-axis), the relationship is very weak or nonexistent (the regression line is almost flat). Readers familiar with the "wrong-level problem" (see my earlier post here) should immediately be skeptical and suspect we are seeing another example of Simpson's Paradox.
Simpson's Paradox results when data from different groups are combined and correlations between independent and dependent variables are reversed. My guess (and I'm only guessing here) is that data displayed above are from stores of different sizes. Assume for the moment that Big Box stores (Walmart and Target) have smaller percentage increases than boutique stores. Assume that the red lines imposed on Chart 2 above connect stores of similar size. Each store can have an increase in Black Friday sales while the overall macro-relationship is negative.
In order to resolve this paradox, yearly sales in each store have to be analyzed as a function of Thanksgiving week sales and their same-store sales coefficients analyzed to determine the effect of Black Friday sales. It's hard to believe that individual stores aren't doing this kind of analysis and would continue with Black Friday Blowouts if their results were not positive. If Mr. Dales had access to store-level data, he could have done the analysis with the hlmmc package.
No comments:
Post a Comment