Ethnic Foodscapes: Foreign Cuisines in the United States

Abstract

Existing scholarship on consumption has focused on how purchasing patterns vary by age, race, and socioeconomic status. This article opens a new perspective by asking how consumption is distributed geographically and how these patterns are associated with neighborhood characteristics. This paper departs from previous studies of consumption by concentrating on not just what is consumed but where. Merging data from Google Places with demographic data from the American Consumer Survey produced a data-set with occupations, ethnic composition, and incomes of neighborhood residents as well as the number and type of ethnic restaurants. Census tracts with high percentages of professionals and highly educated residents have a greater diversity of restaurant genres, whereas wealthier census tracts have less diversity. The share of residents in knowledge-intensive jobs impacts the number of ethnic cuisines in a census tract. So does the immigrant population, but this association takes on different forms for three types of cuisines: destination, popular, and enclave. Traditionally sophisticated destination cuisines, like French and Japanese, thrive in elite neighborhoods independent of ethnic composition and income. Mexican and Chinese restaurants are popular in all sorts of neighborhoods, almost independent of ethnic composition, while more exotic cuisines, like Vietnamese and Cuban, remain largely confined to immigrant enclaves.

Keywords:

• food practices
• ethnic food
• migration
• social class
• taste

Acknowledgments

I gratefully acknowledge the support of Paul DiMaggio and Andreas Wimmer, whose comments and advice greatly improved the research summarized here. I would also like to thank the reviewers for their helpful suggestions and additions.

Notes

1. Terms like “ethnic” and “exotic” are relational. That is, Ethiopian food is neither ethnic nor exotic to Ethiopians. It is ethnic only to outside groups, like white Americans. Krishnendu Ray (2011, 2016) and Lucy Long (2004) both discuss this relational aspect. In this paper, I use the terms ethnic and exotic self-consciously to describe the perspective of most Americans.

2. Of course, elites serve as “early adopters” only for some types of innovations. Things that people would purchase out of necessity or due to financial concerns would not apply here. For example, high-status people are unlikely to be early adopters of innovations in fast food like happy meals.

3. We can, of course, distinguish ethnic eateries that cook for ethnics, from those that cook for elites.  For example, a Mexican restaurant that cooks for Mexicans depends on immigrants, but it may also benefit from adventurous foodies.  In contrast, a Mexican restaurant that cooks for elites depends on wealthy diners, but does not depend on immigrants.  In this study, there is no way to differentiate between the two kinds of ethnic restaurants, but it is important to keep this variation in mind.

4. See Appendix A.

5. Typically, restaurant owners select the appropriate genres for their Google Places profile. Users can also update, amend, and add genres to an establishment by clicking on “suggest an edit” on its Google Places profile. The coders and programmers at Google Places will then verify the amendments before making the changes to the Google Pages profile.

6. Many restaurants have more than one genre associated with them. For example, a Bahn Mi truck would be labeled “Food Truck,” “Sandwich Shop,” and “Vietnamese.”

7. See Appendix C for a complete list and description of census variables.

8. I also constructed a measure of the creative class (based on the proportion in the census tract employed in arts, entertainment, recreation, and food services), but it proved insignificant.

9. I ran all my models with and without each category of job in the knowledge-based occupations variable to check for robustness. The effect of knowledge-based occupations was not driven by any individual occupation.

10. There is a separate ethnic density measure for all ethnicities corresponding to the restaurant genres I collected.

11. I ran all analyses with a restaurant count variable instead of ENG, and it produced similar results. I also constructed a Gibbs–Martin index. The correlation between ENG and Gibbs–Martin is 0.912, indicating that the indices measure diversity in the same way.

12. Even after removing all census tracts with no residents or no restaurants from the data-set, the data was overdispersed and the variance of restaurant counts (the dependent variable) was greater than the mean. Poisson models perform better when the data are not over-dispersed, so a negative binomial was the most appropriate given the structure of the outcome variable. Negative binomial models rationalize overdispersion with a model that assumes that, for every neighborhood, restaurants are indeed Poisson distributed, but that all individuals do not have the same mean rate. Instead, neighborhood rates are assumed to be gamma distributed in the population. This improves upon the Poisson model by allowing for another level of variability in the data, the cross-neighborhood variability in the mean rate of occurrences of restaurants (specified by the gamma distribution) (Land, McCall, and Nagin 1996). After running several models—a Poisson, a negative binomial, a zero-inflated Poisson (ZIP), and a zero-inflated negative binomial (ZINB)—I evaluated the negative binomial model against the others by testing whether theta = 1 (or log(theta) = 0). I could clearly reject the null hypothesis that the better performance of the negative binomial model over the Poisson model is due to random sampling alone. The output from the negative binomial regression is available upon request. Essentially, when the variance is larger than the mean, the negative binomial performs better than the Poisson, and additional tests indicated that the negative binomial model was the best fit for the data.

13. I ran similar regressions with percent of residents over the age of 25 with an HS diploma and percent with a BA degree instead of median years of education for the population aged 25 and over. Each iteration of the model produced similar results, although the effect was not as strong as median years of education.

14. The diversity measure includes all the genres included in Appendix B, not just the twelve listed in Table 2.

15. Of course, income and education are highly correlated, and in most cases, neighborhoods high in education will also have a high average income. This analysis shows us that neighborhoods with the same levels of income but different levels of education are likely to have different levels of culinary diversity (and vice versa).

16. Taking the log of income helps deal with the skewedness of the income variable and makes the variable more normally distributed.

17. I looked at employment in many occupations individually. Artists, college students, and graduate students had no statistically significant relationship with any genre other than Mexican.

18. This is true even when financial and healthcare jobs are included separately in the regression table. They remain statistically significantly associated with most genres of restaurant.

19. The assimilation indices were created using probit regression to identify the characteristics that are most strongly associated with immigrant status. Once that model was constructed, all information about real immigrant status was eliminated from the data-set so that the model could assign a probability to each individual that he or she is foreign born. Finally, the information concerning individuals’ immigration status was taken into account, again to determine how accurate the probabilities were. An immigrant group would be given a score of 100, or complete assimilation, if it was impossible to distinguish immigrants from natives.