Appendix A: Methods

The following question guided this study: Is there a significant difference in service quality between providers in states with no or lighter license requirements and those in neighboring states with more burdensome requirements?

Data

To measure quality, we relied on Yelp business ratings, which others have found to accurately reflect service quality. 1 These ratings spanned October 2004 through October 2020 for locksmiths and October 2004 through June/July 2019 for all other occupations. The first year in our study is 2004 because Yelp started collecting ratings on October 12 of that year. 2 All ratings are on a five-star scale, with one star being the worst rating a business can receive and five stars the best. Each rating also has a written review associated with it, but our analysis used only the numerical rating. For each business, we averaged the ratings across the entire time span.

Control variables included population, percentage of the population with a bachelor’s degree or higher, and median household income. Prior studies similar in design to ours have shown these to be important control variables. 3 We collected these data from the Census Bureau’s 2019 five-year American Community Survey estimates at the block group level—the lowest geographic level for which data were available. 

Sample

The sampling unit (and the unit of analysis) was businesses within specific occupations. To determine the sample of occupations and states in our study, we used data from the second edition of the Institute for Justice’s License to Work report to identify occupations with stark licensing differences across neighboring states. 4 Because IJ released the second edition of License to Work in 2017, we checked relevant licensing requirements to ensure nothing significant had changed in the intervening years. 

Licensing is not limited to state laws. Counties and cities can and often do enact their own licensing laws. 5 Failing to account for such laws, where they exist, would produce spurious analytical results. We therefore also examined a small sample of cities in border counties in our states of interest to determine whether local licensing was present (it was not). We also examined salon/shop licensing at the state and local levels for barbering- and cosmetology-related occupations as this, too, can sometimes affect licensing requirements for workers. However, such requirements had no impact on our analyses.

We limited the businesses in our sample to those within narrow bandwidths on either side of state borders. For each occupation-state comparison, we used a bandwidth that would result in a sample size sufficiently large for analysis. Consequently, bandwidths differ by occupation-state comparison (see Table A1). We had to modify two sets of comparisons that used the CA-NV border (interior designer and tree trimmer) due to a lack of businesses along the states’ shared border. Instead of using bandwidths, we selected businesses located in border counties in the Lake Tahoe region. 6 These counties provided a sufficient number of firms for our analyses, while other regions along the CA-NV border were too rural to do so. Final sample sizes for businesses by occupation-state comparisons are presented in Table A2. 

Table A1: Final Comparisons and Bandwidths

OccupationStatesBandwidth
BarberNJ to PA15 miles
CosmetologistNY to CT5 miles
CosmetologistNY to NJ5 miles
Interior Designer∗CA to NVCounties in the Lake Tahoe region
LocksmithPA to NJ5 miles
Manicurist†CT to MA15 miles
ManicuristCT to NY15 miles
Tree TrimmerNV to CACounties in the Lake Tahoe region
Tree TrimmerVA to MD10 miles
∗ Though California does not license interior designers, it does offer title protection to those who hold certification with the California Council for Interior Design Certification. Cal. Bus. & Prof. Code §§ 5800–12.
† As of January 1, 2021, Connecticut licenses manicurists. H.B. 7424, 2019 Gen. Assemb., Reg. Sess. (Conn. 2019); Connecticut State Department of Public Health. (n.d.). Nail technician. https://portal.ct.gov/DPH/Practitioner-Licensing–Investigations/Nailtechs/Nail-Technician. However, it did not do so during our study period.

Table A2: Number of Businesses for Each Occupation-State Comparison


Lower Burden/UnlicensedHigher Burden/LicensedTotal
OccupationState# of FirmsState# of Firms# of Firms
BarberNJ206PA420626
CosmetologistNY45CT4994
CosmetologistNY940NJ3191,259
Interior DesignerCA63NV3699
LocksmithPA100NJ94194
ManicuristCT45MA4489
ManicuristCT144NY262406
Tree TrimmerNV32CA92124
Tree TrimmerVA58MD3391

Analysis

We analyzed these data using regression models, treating individual businesses as the unit of analysis. In the analyses, we employed a geographic regression discontinuity design to isolate the potential effect of high licensing burdens in counties bordering either unlicensed or less burdensomely licensed states. The areas in which these businesses are located should be similar except for variation in state licensing requirements. Nevertheless, we included community characteristics as covariates to improve the estimates’ precision.

We ran separate regressions for each occupation-state comparison. This allowed for a clear interpretation of a specific license’s effect rather than the general effect of all the licenses captured in our models, which vary in their burden. 7

In geographic discontinuity, if a license were an important determinant of service quality, we would expect to find a measurable difference in service quality between states when comparing businesses across borders, but—for reasons described above—we would not expect to find differences in community characteristics such as population, percentage of the population with a bachelor’s degree or higher, and median household income. We tested this in a series of regressions in which the community characteristics, instead of the Yelp ratings, were the dependent variables. 

Results showed almost no significant differences in community characteristics across states, but because there were a few, we included the community characteristics as covariates in the regressions. Table A3 provides the descriptive statistics for these covariates in each of our comparisons.

Table A3: Descriptive Statistics for Each Comparison’s Covariates


Population per Block Group Education (Percentage BA or Higher)Household Income (Median Dollars)
Occupation/StateMeanSDMeanSDMeanSD
Barber





NJ (less burdensome)1,612 759 0.2520.14280,253 36,323 
PA (more burdensome)1,408 711 0.3070.20168,258 33,926 
Cosmetologist





NY (less burdensome)1,266 501 0.4190.158105,941 64,626 
CT (more burdensome)1,405 752 0.4860.214111,226 43,677 
Cosmetologist





NY (less burdensome)1,449 843 0.5730.181109,303 56,695 
NJ (more burdensome)1,701 989 0.4240.19995,657 43,179 
Interior Designer





CA (unlicensed)2,113 2,630 0.2450.10080,545 33,899 
NV (licensed)1,428 977 0.1830.14054,983 27,838 
Locksmith





PA (unlicensed)1,650 821 0.2800.20066,795 38,874 
NJ (licensed)1,344 561 0.3010.15382,518 41,414 
Manicurist





CT (unlicensed)1,612 677 0.3400.09778,797 30,448 
MA (licensed)1,728 769 0.2270.10176,204 32,104 
Manicurist





CT (unlicensed)1,590 701 0.3980.137108,573 49,221 
NY (licensed)1,390 580 0.3260.15092,781 50,110 
Tree Trimmer





NV (unlicensed)1,186 441 0.2530.12459,947 26,459 
CA (licensed)1,837 2,141 0.2430.10568,037 30,698 
Tree Trimmer





VA (unlicensed)2,120 1,008 0.4000.146133,486 57,237 
MD (licensed)1,720 606 0.3460.142109,468 39,188 
Note: We report all covariates at the block group level.

Our general model took the form:

          Y = β0 + β1(state)+ β2(distance) + β3(Θ) + β4(X) + ε

Where Y refers to a business’s average Yelp rating; state represents whether a business is in a more burdensomely licensed, less burdensomely licensed or unlicensed state; distance represents the number of miles a business is from the state border; Θ represents a series of dummy variables that identify counties directly opposite the state border from each other; and X represents a vector of control variables, as defined above.

A researcher who completed a similar analysis filtered out businesses with fewer than 10 Yelp reviews to avoid skew. 8 Had we done so, there would have been too few businesses for the analyses. As a contingency, however, we ran a second set of all models with the number of reviews per business included as a covariate. Results of those models indicated the number of reviews made only trivial differences in the results. Consequently, we present the results without the number of reviews.