Projects Funded for Hal Giuliani Gordon

2020-2021

Predicting Demand for Plant-Based Meat

Meredith Fowlie and Hal Giuliani Gordon

Abstract

Specifics of the Project:
This project aimed to better understand consumer demand for plant-based meat. In contrast to other meat substitutes (tofu, veggie burgers), plant-based meat products are being marketed as indistinguishable in taste and appearance to meat. Producers of plant-based meats (PBM) aim to compete more directly with, and eventually replace, meat.

We received a panel of purchases from well over 200,000 grocery store shoppers who have bought plant based meat at least once, as well as a similarly sized control panel in a proprietary dataset from a nationwide grocery chain. We found that buying and more importantly, rebuying PBM is associated with having previously bought less meat and more meat substitutes. In addition, the people entering the PBM market are no more likely to have bought meat than those who first started buying it, suggesting PBM is struggling to expand its reach to those who could most easily switch away from real meat. In addition, because of how promotional pricing is determined at this nationwide chain, we were able to run event study regressions to test the theory that PBM has is a robust substitute for beef in grocery stores. In these regressions, we find little evidence for switching between meat and PBM.

Summary of Results:
After receiving the large sample of purchases, we worked to create covariates from the purchases. While PBM was only first introduced in limited stores in mid 2017, all purchases since the beginning of 2016 are included in the data. This allowed us to characterize households by their purchases in prior periods, as well as flexibly examine what types of purchases could ultimately predict who will buy PBM and who will rebuy it, with the ultimate aim of trying to better understand if PBM are attracting the types of customers who are likely to substitute it for real meat.

Covariates created from this pre-period, as well as those matched from a credit company who provides estimates of the head of household’s age and race, and the size and income of the household, were used to see how well we could predict three things: buying plant based meat, rebuying plant based meat, and the date of first buying plant based meat.

Those who bought PBM were more likely to be younger, to have a higher income, and to be shopping in more liberal voting areas. While households who bought PBM were more likely to be "low meat" households that spent less than 5% of their pre-period budget on meat, the difference was only a half percentage point. This supports the often repeated fact that PBM customers also buy meat. However, they still bought less meat, with a larger gap in their average market basket raw percent devoted to meat in the pre period. In addition, buyers of PBM bought about 3 times as much veggie burgers and tofu.

Rebuyers were more likely to be "low meat" purchasers and had spent more on traditional replacements and a little less on meat in the pre-period. That seems to infer that PBM was more likely to catch on with customers who were already interested in replacing meat in their diets. Early buyers were younger and less wealthy (although still wealthier than those who never bought at all). Earlier buyers were more likely to be ’low meat’, bought less meat and ground beef, and bought more meat replacements.

In order to examine whether the large number of covariates coming from pre period purchases could more accurately predict buying PBM, rebuying it, and buying it later or earlier, we implemented regressions using the least absolute shrinkage and selection operator (Lasso), a common machine learning algorithm that automatically selects a model so that many coefficients go to zero. While not causal, these models are used to better understand who the buyers were, and who was most likely to rebuy or buy later.

The OLS and Lasso results have less predictive power than hoped. The highest R^2s are for the buy/don’t buy decision. There, the OLS models (with nearly 500 variables) have an R^2 of just 0.16. The Lassos have slightly lower R^2s. Unfortunately, the Lassos only collapse the coefficients to zero of about half the categorical variables. This means we are not able to single out a handful of traits that are much more important in predicting the buy/don’t buy decision.

The models on early and late adoption and rebuying within 3 months have even lower R^2s, but are at least able to reduce the number of non-zero coefficients a bit better. When the sample is reduced even further to just California based stores, the Lasso was able to collapse many more coefficients to zero. Table 1.8 and 1.9 from our paper are reproduced below, showing the covariates related to rebuying and related to the week of first buy.

From a climate and animal welfare perspective, we would have liked to see rebuying popular with meat eaters, but instead, our Lasso regressions found the opposite (Table 1.8). Milk, bacon, beef, meat, pork, and ground beef are all predictive of not rebuying PBM, while meat and dairy alternatives, seafood, tofu, kombucha and expensive produce are related to rebuying. This table tells a clear story that customers who are more likely to incorporate plant based meat into their diets are much less likely to have had a diet high in the ground beef that PBM cheerleaders hope to be replacing.

While those households who bought more meat in the pre period were less likely to buy PBM, we were hoping that those types of households would be more likely to have bought PBM in the later period (as information about PBM started reaching more households due to increased media stories), but in table 1.9, we did not find that to be the case. Early adopters were more likely to have bought meat and dairy alternatives, tofu, and organic vegetables. This fits the story that early adopters of PBM were already interested in meat alternatives. However, none of the coefficients positively related to week of first purchase (meaning they bought later) were related to meat purchases, which suggests PBM was not spreading to meat eaters during the end of 2019. From these data, I conclude that PBM has a lot longer to go on making itself more attractive to the kinds of customers who are most likely to be substituting away from beef.

To study if PBM was crowding out beef, we used the as good as random variation of when PBM went on promotion (sale price). After interviewing officials with the corporate office of the grocery chain and running our own tests, we found that the temporary price cuts of PBM were imposed at the region level, and were not coordinated with each other, nor were they correlated with prior sales. We then used these events in event study difference in difference regressions.

While these regressions did show that PBM promotions greatly increase sales of PBM (figure 1.3 from our paper reproduced below), they have little to no measurable effect on beef (figure 1.4). There are a number of reasons to think these regression results are not perfect. First, the number of clusters, 12, was quite low, but that was the level at which prices were set. Second, the predicted demand change for PBM in the first week of the promotion (0.003 lbs) was probably not large enough to overcome the normal noise in the much, much larger ground beef markets. Yet, the main finding remains that we are unable to detect any effect of PBM on the demand for meat.