The following question guided this study: Is there a significant relationship between the number of food trucks and restaurants from 2005 through 2016?
Annual food truck and restaurant data came from the 2005 to 2016 annual Census County Business Patterns database. Food trucks (the independent variable) were represented by mobile food establishments, NAICS code 722330. Restaurants (the dependent variable) were represented by full service, quick service and cafeteria style establishments, NAICS codes 722110, 722211 and 722212 (pre-2012) and 722511, 722513 and 722514 (post-2012).
Control variables included population estimates and unemployment rate. It is self-evident that counties with greater populations will generally have more food establishments of any kind, making it important to control for population. We drew county population estimates from annual Census Bureau data. The number of food establishments may also depend on the economic health in a county. Thus, we controlled for that using unemployment rates—a common proxy for economic health 1
—at the county level drawn from the Bureau of Labor Statistics Local Area Unemployment Statistics. We collected and used one instrumental variable found to be a predictor of the presence of food trucks: ethnic diversity. 2
We estimated this as a Herfindahl-Hirschman Index using racial percentages from the Census Annual County Resident Population Estimates.
We analyzed the data for all counties (n = 3,133) and then for non-rural counties (n = 1,165). We coded the urbanicity of counties using data from the 2013 USDA Rural-Urban Continuum Codes. 3
Rural was defined as non-metro counties with populations of fewer than 2,500 people adjacent to and not adjacent to metro counties (codes 8 and 9 on the continuum). Non-rural counties included all other counties (codes 1 through 7 on the continuum).
We analyzed these data using dynamic panel data analysis, specifically an Arellano-Bond model in a one-step estimation. 4
In the analysis, we treated unemployment rate and population as exogenous, food trucks as a predetermined variable, and HHI as an instrument. The model also included year fixed effects and used robust standard errors. Year fixed effect controlled for any idiosyncratic year effects. Robust standard errors accounted for heterogeneity present in the data (total sample: Greene LR test = 149,000, p = .000; non-rural sample: Greene LR test = 43,610, p = .000). For robustness checks, we also analyzed the data using traditional ordinary least squares fixed effects, an Arellano-Bond model in a two-step estimation 5
and an Arellano-Bond-Blundell model in a two-step estimation. 6
Results were substantively similar to the Arellano-Bond model in a one-step estimation. This means the results reported below are not an artifact of a particular estimation method but, instead, reflect a substantive relationship between food trucks and restaurants in the manner we describe.
In all of the preceding analyses, the food truck variable was modeled to be contemporaneous with restaurants and with a one-year lag. Our use of a one-year lag, rather than a longer lag (e.g., two years, three years, five years), was informed by media reporting 7
and academic literature 8
suggesting that food trucks’ potential effects on restaurants—if any—would be observed sooner rather than later. In media reporting, for example, restaurant owners opposed to food trucks have described how their businesses suffered shortly after the arrival of food trucks and predicted their firms would shutter. 9
One restaurant manager said, “When our count is down, we can just go outside and count the trucks right in front of us,” while another said, “[W]hen there’s a whole bunch (of trucks), we see a significant drop in sales.” 10
Moreover, research on factors that contribute to restaurant failure indicates competition is a consistent and significant predictor. 11
Independent restaurants, in particular, struggle to remain operational in areas with greater competitive density. In media reporting, it is owners and managers of just those types of restaurants who are quoted as objecting to competition from food trucks.
Added to this are still other media articles that suggest exogenous factors have near-immediate effects on restaurants. An article about the 2019 federal government closure, for instance, described how four weeks into the closure restaurants were already reporting 20% to 60% losses in sales and significant reductions in employee work hours. Restaurants were described as just “trying to sustain their business.” 12
In the absence of prior systematic evidence about the effects of food trucks on restaurants (specifically, a survival analysis on the relationship between trucks and restaurants), taken together, the literature described above acts as a guiding theory and suggests a one-year lag is appropriate. However, some might argue that one year is too short—that restaurants may be able to hold out for more than a year when faced with competition from food trucks. We allowed for that possibility and ran all models described herein with a second lag for food trucks. The results were not robust across all models. Specifically, inconsistencies appeared in statistical significance, signs on coefficients and magnitudes of coefficients. Such results, plus the theory guiding our use of a one-year lag, compelled us to report results for only the one-year lag model.
Table A1 includes results for all counties and the non-rural sample. Findings for restaurant and food truck variables are, of course, the same as reported above. Both unemployment and population are significantly related to number of restaurants and in an expected direction. Greater unemployment (a sign of a comparatively weaker economy) is associated with fewer restaurants. More populous counties have more restaurants. Notably, results are quite similar when comparing all counties and the non-rural sample. This suggests the dynamics between food trucks and restaurants are largely a suburban and urban phenomenon, which is entirely logical given the paucity of both restaurants or food trucks in rural areas. Finally, Table A1 includes autocorrelation results. As is desirable, results confirm first differences in the Arellano-Bond model are significantly correlated in the first order, indicating dynamic effects; no significant second-order autocorrelation appears in the first differences of errors.