Beyond Labor Market Outcomes: The Impact of the Minimum Wage on Nondurable Consumption

Transcription

1 Beyond Labor Market Outcomes: The Impact of the Minimum Wage on Nondurable Consumption Cristian Alonso First Version: October 2015 This Version: December 2015 Abstract How effective is the minimum wage at raising nondurable household consumption through the redistribution of income towards poor workers? Using novel data on retail sales by county, I exploit variation in the minimum wage rates across states and over time to answer this question. I find that a 10% increase in the minimum wage increases sales by 1.1%, a surprisingly large number considering that only a tiny fraction of the American labor force earns the federal minimum wage rate. I argue that such a large effect is explained by positive spillovers benefiting the bottom quarter of the labor income distribution. As expected, the expenditure response to minimum wage hikes is stronger in counties where the policy binds more. JEL codes: J38, J20, D31, E21, E64 1 Introduction The minimum wage is a controversial policy in the United States. A large literature has studied the effects of the policy on labor market outcomes such as employment and wages. However, if the goal of this controversial policy is to raise welfare, its evaluation should not be limited to labor market outcomes. Instead, it should also include the effect on consumption. The measurement of this effect is challenging due to limited data. But at least among nondurables, where adjustment costs are low, households response is likely to be immediate and so, coordinated and easier to measure in the data. So, my research question is does a minimum wage hike increase nondurable consumption? To overcome data limitations, I use a novel dataset on retail sales of groceries to produce a measure of nondurable consumption at the county level. While the evolution of this variable appears I thank Greg Kaplan for the valuable advice he has provided at each stage of this project. I am grateful for the comments and suggestions from Mark Aguiar, Will Dobbie, Alan Krueger, Alexandre Mas, Atif Mian, Benjamin Moll, Ezra Oberfield, Richard Rogerson, Mark Watson, Juan Pablo Xandri, and participants at the Macro Student Workshop at Princeton. I thank the Kilts-Nielsen Data Center at the University of Chicago Booth School of Business for providing access to the Kilts-Nielsen Retail Scanner data. All errors are my own. Princeton University, calonso@princeton.edu 1

2 to be similar to that of nondurable consumption as reported by the national accounts at the state level, the use of retail sales has several advantages for this project. Retail sales data is available at any frequency. I set the frequency to be a quarter, which allows me to study the dynamics of the response in more detail. Also, retail sales are available at a very disaggregated geographic level, so that I can effectively control for local economic conditions. Finally, this geographic detail allows to take advantage of the heterogeneity across US counties to assess the differential effect given how much the minimum wage binds in each area, strengthening the identification strategy. I exploit the wide dispersion of state minimum wage rates across the country and over time during the period to identify the expenditure response. At any point in the sample at least a quarter of the states had set minimum wage rates above the federal level; and in some periods half of the states had done so. I use all the increments in the state legislated minimum wage rate. In addition, the federal wage had a three-step increment starting in July Since the prevailing rates at the state level were different, the federal raise was not uniform across the country and so, it allows for the identification of the effect by comparing areas where the rate went up significantly with those where the new legislation did not have an impact because the local rate was already very high. A panel data research design with two-way fixed effects is a suitable framework for this project. By complementing time and county fixed effects with local area variables to proxy for economic conditions, my identification strategy rests on the assumption that counties where the minimum wage rate went up would have experienced the same increase in sales that was observed in counties where the rate was unchanged, conditional on observable economic conditions. In this setting, county fixed effects capture any static unobserved heterogeneity at the local level, whereas time fixed effects absorb the common time trend. In addition, variables such as employment and house prices at the county level enable me to control for local economic conditions, which is particularly important in the context of the Great Recession since not all the counties were hit with the same severity by the crisis and that severity could be correlated with policy changes. Then, my preferred specification controls for gross state product, and for employment, population, and house prices at the county level, in addition to county and time fixed effects. I find that the minimum wage is effective at raising aggregate expenditure. An increase of 10% in the minimum wage rate increases nominal sales by 1.1% and real sales by 0.7%. The magnitude of the estimated consumption response appears large when compared to the mean consumption growth of 1% in the post-war period. Furthermore, the elasticity is greater in counties where the minimum wage binds relatively more. I argue that this magnitude is consistent with the redistributive implications of a minimum wage hike beyond the very bottom of the earnings distribution. Using CPS data I show that the minimum wage induces modest but significant increases in labor income for households up to the 25 th percentile of the distribution. A 10% minimum wage hike would increase sales by 2 billion dollars, a quarter of the predicted extra labor income. My results are then consistent with high marginal propensities to spend for poor workers. 2

3 A number of additional checks confirm the robustness of my results. First, I show that there are not significant pretrends in sales nor in observable variables that could jeopardize my identification strategy. Second, I perform a diff-in-diff-in-diff exercise and I find that counties where the minimum wage binds more experience a greater expenditure response when compared to other counties in the same state. Third, I show that smaller counties present a larger, but noisier, elasticity, which is consistent with smaller counties being less productive. But the expenditure response is still positive and significant for the largest counties. Fourth, I find that the elasticity is not statistically different when I break the sample period into and , but the point estimates are greater in the latest period. Fifth, I obtain similar elasticities when running the model for sales at the store and at the state level and for nondurable consumption in NIPA, which suggests the baseline Nielsen dataset is reasonably representative. Fifth, I show that my results are not driven by the housing bubble, nor by the subset of states who annually index their minimum wage rates to the federal inflation. To the empirical literature studying the effects of the minimum wage, this paper offers a new take by focusing on the implications for consumption rather than labor market outcomes, an arguably better proxy for welfare. The measurement of the effects of the minimum wage has interested labor economists for decades with significant contributions made by Neumark and Wascher (1992), Card and Krueger (1994), Lee (1999), and Autor, Manning, and Smith (2015), among others. In this line of work, this paper draws on the methodology from recent work on the identification of the effects on labor market outcomes, most notably from the ongoing discussion between Dube, Lester, and Reich (2010) and Neumark, Salas, and Wascher (2014b) about the need to control for unobservable spatial employment growth heterogeneity. I argue that this concern does not emerge during my sample period and so, my baseline specification is a two-way fixed effects panel design. Nevertheless, I show that my results are robust to the presence of unobservable spatial time trends. Only one other paper explores the effect of the minimum wage on consumption. Using CEX data, Aaronson, Agarwal, and French (2012) measure the response of vehicle purchases to minimum wage raises and find a positive effect, although not statistically significant and very unequally distributed among low-income households, which they interpret as evidence of borrowing constraints and adjustment costs. Aaronson, Agarwal, and French (2012) also report an insignificant effect of the minimum wage on consumption of a combined category of nondurables and services. However, the measurement of grocery purchases in the CEX-Interview is noisy 1 and its combination with services may further confound the effect. In this paper, I complement the analysis in Aaronson, Agarwal, and French (2012) by focusing on nondurables rather than vehicles and by employing a better measure of nondurable expenditures at the county level. The rest of the paper is organized as follows. In section 2, I describe the data and show that the store sales data appears to be a reasonable approximation to consumption aggregates from the national accounts. In section 3, I discuss the identification strategy. In section 4, I present the 1 Households are asked the usual amount of weekly expense for grocery shopping and not the actual amount spent in a given and recent week, where recall could be higher. 3

4 main results. In section 5, I include robustness checks. In section 6, I explore the quantitative link between the estimated expenditure response and the impact of the minimum wage on the labor income distribution. Finally, section 7 concludes. 2 Data Description My empirical analysis exploits the significant policy heterogeneity across the country and a new measure of consumption by county. In this section, I first describe the significant dispersion in the prevailing minimum wage rates at the state level in the period Then, I argue that retail sales data are a good proxy for consumption at the county level, which allows me to control for the evolution of local area observable variables. 2.1 Minimum Wage In this subsection, I discuss the two sources of variation in the minimum wage that I exploit: federal and state level changes. The rich variation observed in minimum wage rates (Figure 1) emerges from the institutional setting determining the policy. Federal, state, and local governments in the US can set their own minimum wage. The effective minimum wage rate is the maximum of the three. Since only a few counties have recently enacted their own legislation, I abstract from this variation and focus instead on the federal and state minimum wage rates. This data was retrieved from the federal and state Departments of Labor. The Fair Minimum Wage Act of 2007 increased the federal rate in three steps during my sample period. The federal minimum wage rate rose from $5.15 to $5.85 on July 24, 2007, to $6.55 on July 24, 2008, and to $7.25 on July 24, The law induced different changes in the effective rate across states because some states had already legislated rates higher than the federal one. I exploit this variation in my identification strategy. A number of states enacted changes in the state minimum wage rates during my sample period. Ten states 2 adopted an annual automatic adjustment of their minimum wage rate to compensate for inflation. In nine of these states 3, the adjustment follows the federal rate of inflation and it is then arguably exogenous to local economic conditions. Seventeen other states have legislated higher rates as well, usually with increments taking place sequentially over a few years. The interaction of these state and federal increments produces valuable dispersion in the effective minimum wage rate across the country (Figure 2). For half of the county, mostly the south and the center, the change in the minimum wage between 2006 and 2013 was given by the federal increment, i.e. 41%. For 16 states, the effective minimum wage rose by less than 30% in these eight years. 2 Arizona (2006), Colorado (2006), Florida (2004), Missouri (2006), Montana (2007), Nevada (2004), Ohio (2006), Oregon (2002), Vermont (2005), Washington (1998). The number in brackets is the year the automatic adjustment was approved. 3 The exception is Colorado where the minimum wage is adjusted using the Consumer Price Index for the Denver- Boulder-Greeley combined metropolitan statistical area. 4

5 Figure 1: State-Level Effective Minimum Wage Evolution Min Max 5.5 Mean Median P Significant dispersion in state level minimum wages and over time The quarterly rate is computed as the average of the minimum wage rates prevalent in each month of the quarter. The effective minimum wage is defined as the maximum of the state and federal minimum wage rates. New York had the smallest increment in the period, only 7%. Finally, 8 states, mainly in the west, experienced increases above the federal change. Nevada had the largest increment in the period, 60%. 2.2 Nielsen Sales In this subsection, I argue that grocery sales as computed from a Nielsen dataset provide a reasonable measure of nondurable consumption at the county level. To approximate nondurable consumption, I obtain sales by county from a Nielsen dataset mainly covering groceries and drugs. The Nielsen Retail Scanner Data contains weekly information on pricing, volume, and merchandising conditions generated by participating retail store point-of-sale systems in all 48 contiguous states and DC, from 2006 to It covers approximately 40,000 stores of very diverse size (Figure 3). The dataset contains all the products in Nielsen-tracked categories (mainly food, non-food grocery items, health and beauty aids, and select general merchandise) that are labeled with a UPC. According to Nielsen s own estimates, the information included represented in 2011: 53% of the total sales in food, 55% in drugs, 32% in mass, 1% in liquor, and 2% in C-store 4. To distinguish real from nominal effects, I compute two measures of sales. Nominal sales aggregate all the transactions by store and quarter using current prices. In contrast, real sales aggregate transactions using the 2012 average price for each product. Real sales are then meant to abstract from price changes and price differences across stores and give a sense of the change in quantities sold. 4 Kilts Center for Marketing (2014) 5

6 Figure 2: Change in the Minimum Wage > 45% (42%, 45%] 41% (Federal Change) (15%, 40%] [0%, 15%] Large geographic variation in the change in the minimum wage Percentage change in the effective minimum wage between January 2006 and December The effective minimum wage is defined as the maximum of the state and federal minimum wage rates. I collapse sales to the county level, my preferred unit of analysis. For consistency, I drop all the stores that are not present during the 32 quarters of my sample period. I show in section 5 that this restriction does not affect my results, so the potential bias introduced by the lack of entry and exit of stores is not a concern. Finally, I drop the county of Ringgold in Iowa because it lacks employment data in the fourth quarter of My sample then includes 2,263 counties across the 48 contiguous states and DC. Retail sales appear to be a good proxy for nondurable consumption at the local level. First, the data provides good geographic coverage as evidenced by Figure 4, which shows the number of stores in each county. Second, the evolution of retail sales in the period is fairly consistent with the changes measured by the national accounts at the state level (Figure 5). 2.3 Other Sources of Data Additional variables are used to characterize counties and local economic dynamics. I complement the Nielsen and minimum wage data with public sources of data at the county level: annual population by county from the Census Bureau s Population Estimates Program, quarterly data on employment and average wages from the Quarterly Census of Employment and Wages (QCEW), the number of tax returns filed by county from the Internal Revenue Services, house prices from the Federal Housing Finance Agency, and statistics on income distribution from the American Community Survey. I also use data at the state level: Employment from the Bureau of Labor Statistics and Regional Accounts from the Bureau of Economic Analysis. Table 1 displays summary statistics for the resulting county-level dataset. 5 Personal Consumption Expenditures statistics by state for 2013 have not been released yet. 6

7 Density Figure 3: 2013 Annual Sales Distribution by Store Sales (USD millions) The dataset contains both small and large stores Annual sales for stores that are present during the 32 quarters of the sample. Figure 4: Nielsen Data: Number of Stores by County (9,555] (3,9] (2,3] [1,2] Nielsen data offers a good geographic coverage Number of stores that are present during the 32 quarters of the sample in each county. 7

8 Figure 5: Nielsen Nominal Sales vs. NIPA Consumption State Level Growth Panel A. PCE Panel B. Goods PCE ID RI IL NV DE AZ DC LA NC MA MTWA VT SC NY VA PA ME MD MO NH OR CT WI GA MNAL CA NJ FL MI CO TN OH TX WY NM NE IN ND WV AR KY MS SD OK KSIA UT Goods ID RI IL NV DE AZ NC DC MT ME MO MA NH VT PA MNAL VA WI SC FL OR CA MD CT MI NJ LA WA NY GA CO TN OH TX NE WY NM IN WV AR KY MS ND SD OK IA KS UT Nielsen Nielsen Panel C. nondurable Panel D. Food Nondurable ID RI IL NV DE AZ NC DC MT LA WA FL MOOR NY ME MA VA GA NHPA MN CA VT WI SC AL MD CT MI NJ CO TN OH TX NM NE WY IN ARWV Nielsen MS KY ND SD OK IA KS UT Food RI ID DE IL NV AZ NC VA ME FL CA MD MNAL NJ DC MO MT LA NY OR WA MA GA CT PA VT NH WI SC MI CO TN OH TX NM WY NE IN AR ND KY WV MS Nielsen SD IA KS OK UT Nielsen sales data appears as a reasonable approximation to consumption Change in log values between 2006 and Nielsen refers to the nominal sales data. NIPA corresponds to different items of the Personal Consumer Expenditures by State. 8

9 Mean Std. Dev. P25 Median P75 Minimum Wage Sales Employment Population House Price Index Average Weekly Earnings Number of Stores Table 1: Summary Statistics in 2012 There is significant heterogeneity across counties Mean, Standard Deviation, 25 th percentile, median, and 75 th percentile of each variable by county in Minimum wage and average weekly earnings are in nominal dollars. Sales are in millions of nominal dollars. Employment and population are in thousands. Finally, to reconcile my expenditure results with previous work regarding employment and wages, I use the Outgoing Rotation Group from the Current Population Survey as discussed in section 6. 3 Identification Strategy In this section, I argue that a panel data framework is the appropriate choice to study the consumption response to minimum wage changes. County fixed effects absorb static unobservable local characteristics, whereas time fixed effects capture global time trends. In addition, variables such as employment, population, and house prices allow me to control for potentially heterogeneous observable time trends on local conditions. Let Sales c,t be the amount of total sales (nominal or real) in quarter t in county c, located in state s. Then, my baseline model is: log(sales c,t ) = κ c + τ t + βlog(minimum Wage s,t ) + γx c,t + ɛ c,t where β is the coefficient of interest, the elasticity of sales (nominal or real) to the minimum wage. Minimum Wage s,t is the effective minimum wage rate in state s in quarter t, κ c is the county fixed effect, τ t is the time fixed effect, and X c,t is a set of observables at the county or state level (employment, population, gross state product, house prices, and regional price deflator). My identification strategy rests on the assumption of parallel trends across counties. I assume that conditional differences between those counties where the minimum wage increased and those where it did not would have remained constant in the absence of a minimum wage hike. This 9

10 assumption is untestable, but I will show in subsection 5.2 that trends in observable variables before the policy changes were not significantly different between treatment and control. The use of panel data is important to control for fixed unobservable heterogeneity. Counties where the minimum wage is high could be counties where the average wage is also high and so, a cross sectional estimate would not yield the impact of the policy but rather it would just reflect the fact that richer counties consume more. County fixed effects solve this problem by capturing all the static county characteristics and using the change rather than the level of the minimum wage to identify the elasticity. The panel data framework also allows for a flexible specification of time trends. Time fixed effects capture any shocks that affect all the counties in the same way, including seasonality. However, since my sample period includes the Great Recession, a concern emerges from the differential severity of the crisis across the country. As long as the heterogeneous shocks experienced by the counties are uncorrelated with the changes in the minimum wage, the estimated elasticity will be consistent. But if the shocks are correlated with the changes in the minimum wage, the estimation will be biased. To attenuate this concern, in some specifications I control for local observable economic conditions, allowing different groups of counties to experience different shocks. The inclusion of these controls comes at the cost of additional endogeneity concerns because sales, employment, and house prices could be jointly determined. My preferred specification controls for employment, population, state product, and house prices, and it also includes time and county fixed effects. However, in the following two sections, I present my results with and without these local controls and find that the estimates for the elasticity of sales to the minimum wage are fairly stable across specifications. My results are robust to the presence of heterogeneous unobservable time trends, a source of ongoing debate in the empirical literature on the minimum wage 6. I discuss this concern in detail in subsection 5.3, where I perform a diff-in-diff-in-diff exercise and find that even after allowing for state-specific time trends, counties where the minimum wage binds relatively more experience larger increments in sales. In addition, subsections 5.1 and 5.2 show that heterogeneous time trends do not seem to be a concern during my sample period. Since my goal is to obtain a measure of the aggregate response of sales to the minimum wage, I weight counties using their population level in the 2010 Census. Given that Nielsen tracks more stores in the more heavily populated areas, this approach also gives more weight to the counties where sales are measured better. Nevertheless, I show in subsection 5.10 that the estimated elasticity is not significantly different when weights are not used. Finally, following Dube, Lester, and Reich (2010), I cluster standard errors at the state level to account for any possible serial correlation and the bias introduced by the minimum wage being the same within the state (Bertrand, Duflo, and Mullainathan (2004)). 6 Dube, Lester, and Reich (2010) 10

11 4 Results In this section, I present the main results of the paper. In my baseline specification, retail sales increase by more than 1% after a 10% increment in the minimum wage. The response is stronger in counties where the minimum wage binds more. In terms of composition, the expenditure response is quite homogeneous across product groups. 4.1 Baseline Results I present my baseline results in this subsection. Given the concern about heterogeneous shocks at the county level during the Great Recession, my preferred specification includes county and time fixed effects and controls for employment, population, and house prices at the county level, and state level output. I include alternative specifications as well. Nominal sales increase by more than 1% after a 10% minimum wage hike, a response that is both economically and statistically highly significant. My preferred specification, Model (5), yields the most conservative estimate of the elasticity of nominal sales to the minimum wage: But the point estimate is fairly stable across specifications, oscillating between and (Panel A of Table 2). While the addition of controls does not change the points estimates significantly, it does improve their precision. The exception is Model (6) where the inclusion of the regional price deflator comes at the cost of the reducing the sample size by two years 7. The coefficients on the controls have, for the most part, the expected sign. The coefficient on employment is positive, and the one on population is negative. The similar absolute value of these estimates suggests that the employment to population ratio could be a sufficient control. The coefficient on current Gross State Product is positive and significant only when house prices are not included. The positive and highly significant coefficient on house prices indicates that this variable is a better proxy for local economic conditions than the state level of output. The price deflator is not statistically significant from zero. Real sales rise significantly after a minimum wage increment too. The point estimate for the elasticity ranges from in my preferred specification to In each of the six specifications, the real sales elasticity is around 0.04 points lower than the nominal sales elasticity. Even if the difference between the nominal and real sales elasticities is not statistically significant, it suggests that some of the increase in sales emerges from consumers paying higher prices. It could be that stores raise prices in response to higher demand or that consumers change their shopping behavior after an increase in the minimum wage, replacing cheap stores with expensive ones, and so, increasing nominal expenditures more than real quantities. The latter appears to receive support from the data as discussed in subsection 5.7, where I find that nominal and real sales experience the same change when measured at the store level. 7 Regional price deflator is only available since

12 (1) (2) (3) (4) (5) (6) Panel A. Nominal Sales Minimum Wage 0.166** 0.148** 0.189*** 0.168*** 0.111*** 0.190*** (0.075) (0.057) (0.046) (0.042) (0.037) (0.059) Employment 0.414*** 0.565*** 0.496*** 0.412*** 0.457*** (0.068) (0.072) (0.079) (0.090) (0.096) Population 0.566*** 0.569*** 0.446*** 0.459*** (0.108) (0.114) (0.123) (0.158) GSP 0.216** (0.093) (0.091) (0.073) House Prices 0.140*** 0.115* (0.034) (0.065) Deflator (0.575) Panel B. Real Sales Minimum Wage 0.124* 0.106** 0.142*** 0.123*** 0.072** 0.147** (0.068) (0.048) (0.040) (0.035) (0.032) (0.056) Employment 0.421*** 0.556*** 0.490*** 0.416*** 0.437*** (0.063) (0.078) (0.083) (0.094) (0.099) Population 0.507*** 0.510*** 0.401*** 0.375** (0.102) (0.107) (0.118) (0.146) GSP 0.205** (0.092) (0.097) (0.075) House Prices 0.124*** 0.109* (0.036) (0.060) Deflator (0.547) Observations Counties Quarters Clusters Table 2: Baseline Results A 10% increase in the minimum wage increases nominal sales by 1.1% * p < 0.1, ** p < 0.05, *** p < Standard errors (in parenthesis) clustered at state level. Time and county fixed effects are included in every specification. Sample period for Specifications (1) to (5) is Sample period for Specifications (6) is

13 4.2 Sales Elasticity by Minimum Wage Bindingness Counties where the minimum wage binds more should exhibit a stronger expenditure response following a minimum wage hike. In this subsection, I show that there is indeed a stronger response in counties where the average wage is relatively lower with respect to the minimum. To assess the differential impact of an increase in the minimum wage depending on how binding the rate is in a given county, I use the ratio of average to minimum wage. Previous work 8 has employed the ratio of median to minimum wage to measure the bindingness of the policy. Unfortunately, median wages are not available at the county level and so, I use average wages (from the QCEW) instead. With the ratio of average to minimum wage in 2005, i.e. the year before the beginning of my sample, I sort counties into four different groups and I run the baseline regression for each group separately. Results are summarized in Figure 6. Counties where the average wage relative to the prevailing minimum rate is lower, and so where the policy binds more, show a stronger expenditure response. Following a minimum wage hike of 10%, nominal sales increase by 2.4% for the quarter of the counties where the policy binds the most, but only by 0.7% for the group of counties where it binds the least. Results are very similar for real sales, with an estimated response of 2.7% for the lowest and 0.3% for the highest quartile. While we can reject the elasticity of both nominal and real sales being equal to zero for the half of counties where average wages are low relative to the minimum, the estimates are not statistically different from zero for the other half. The precision of the estimates across these four groups varies significantly and the confidence intervals become narrower for higher-wage counties. The standard error of the estimated elasticity is more than twice as large in the lowest quartile as it is in the highest one. This difference in precision mainly emerges from the use of weights. Since high average wage counties tend to be more populated and better measured by Nielsen, the elasticity is estimated with more precision in the highest quartiles. Nevertheless, the use of weights does not affect the qualitative results of this subsection. In the appendix, I replicate Figure 6, but without using weights and I find that the estimated expenditure response is again decreasing in the level of average wages. 4.3 Sales Elasticity by Fraction of Workers Affected by the Minimum Wage Counties where a larger fraction of the population are likely to receive the minimum wage should exhibit a stronger expenditure response following a minimum wage hike. In this subsection, I use a measure of income distribution at the county level and show that the increase in consumption is indeed significantly larger when most workers make low wages. The fraction of tax returns filed by bracket in each county provides a rough measure of income distribution at the local level. In 2012, a full-time worker making the minimum wage would be earning between $15,131 (7.25, federal rate) and $18,866 (9.04, rate in Washington). Then, I use the fraction of tax returns filed with a labor income below $25,000, the smallest bracket that contains 8 Lee (1999), Autor, Manning, and Smith (2015) 13

14 0.6 Figure 6: Minimum Wage Bindingness Panel A. Nominal Sales Panel B. Real Sales β n 0.2 β n Quartile Quartile The expenditure response is stronger in counties where the minimum wage binds more 95% Confidence Intervals included. Standard errors clustered at state level. β n is the elasticity of sales to the minimum wage from Specification 5 when the sample is limited to those counties that in 2005 belonged to the n th quartile of the Average Wage to Minimum Wage distribution. The minimum wage binds more in states in the lowest quartiles. Specification (5) includes employment, population, gross state product, and house prices, and time and county fixed effects. these amounts. I exploit the cross-county heterogeneity by interacting the (log) minimum wage with the fraction of workers making less than $ 25,000 a year. More precisely, I run the regression: log(sales c,t ) = κ c + τ t + β 1 log(mw s,t ) + β 2 F raction c log(mw s,t ) + γx c,t + ɛ c,t where X c,t is a set of local controls as before. In this context, the elasticity of sales to the minimum wage in county c is given by: β 1 + β 2 F raction c. In Table 3 I report the sales elasticities evaluated at four different levels of the fraction of workers making less than $25,000 a year. Two of those are out-of-sample estimates: when everyone in the county makes the minimum wage (Fraction = 1) and when no one does (Fraction = 0). The remaining two are in-sample estimates: when 56.7% and 26.7% of the workers make the minimum wage, which are respectively the 99 th and 1 st percentiles of the distribution of the fraction. Counties with more people in the lowest brackets exhibit a greater elasticity of sales to the minimum wage, as expected. In counties where more than half of the workers report a low labor income, sales rise by more than 2% in response to a 10% minimum wage hike, and the elasticity is highly significant. On the other hand, when only a quarter of the workers in the county report low income, the estimated elasticity is basically zero and insignificant. The out-of-sample estimates provide further illustration of the positive connection between the fraction of workers directly 14

15 (1) (2) (3) (4) (5) (6) Panel A. Nominal Sales Fraction = ** 0.797*** 0.670** 0.664** 0.521* 1.067*** (0.229) (0.270) (0.284) (0.317) (0.303) (0.396) Fraction = p(99) 0.267*** 0.333*** 0.323*** 0.307*** 0.227** 0.441*** (0.087) (0.093) (0.085) (0.095) (0.095) (0.135) Fraction = p(1) (0.098) (0.078) (0.082) (0.081) (0.069) (0.084) Fraction = (0.185) (0.180) (0.201) (0.214) (0.193) (0.234) Panel B. Real Sales Fraction = ** 0.707*** 0.594*** 0.589*** 0.461** 0.796*** (0.182) (0.186) (0.199) (0.222) (0.209) (0.297) Fraction = p(99) 0.211*** 0.277*** 0.269*** 0.253*** 0.182*** 0.333*** (0.073) (0.063) (0.060) (0.066) (0.065) (0.104) Fraction = p(1) (0.088) (0.065) (0.066) (0.063) (0.054) (0.074) Fraction = ** (0.156) (0.138) (0.150) (0.157) (0.140) (0.184) Table 3: Sales Elasticity by Fraction of Workers Affected by the Minimum Wage The expenditure response is stronger in counties where more workers make the minimum wage β 1 + β 2 F raction c evaluated at four different levels of F raction c : 100%, 56.7%, 26.7%, and 0%. * p < 0.1, ** p < 0.05, *** p < Standard errors (in parenthesis) clustered at state level. Time and county fixed effects are included in every specification. Specifications differ on the controls included as indicated in Table 2 Sample period for Specifications (1) to (5) is Sample period for Specifications (6) is benefiting from the minimum wage increment and the expenditure response. An alternative source of data for this exercise is the American Community Survey (ACS), which reports the fraction of households with total income lower than certain thresholds by county. I prefer the IRS data because it provides not only total but also labor income, which is a more relevant variable for my application. Additionally, the version of the ACS covering all the counties reports 5-year averages rather than annual estimates. In any case, I repeat the exercise in the Appendix using the ACS data and find very similar results. 4.4 Sales Elasticities by Product Group In this subsection, I show that the expenditure response is quite homogeneous across product groups. Decomposing the sales response by product category may be interesting for at least two reasons. 15

16 First, it provides evidence on how consumers allocate their additional income, which is instructive to infer the nature of different goods. Second, the decomposition may be of relevance to a paternalistic planner who worries that poor workers could increase their consumption of alcohol instead of healthy food. The sales response is not statistically different across products. In my preferred specification, Model (5), an increase of the minimum wage induces a greater nominal expenditure in most product groups, the exception being Alcohol for which the point estimate is basically zero. The strongest nominal responses are in the categories Food (0.123) and General Merchandise (0.137). Interestingly, the real response is less than a third of the nominal one for General Merchandise suggesting that higher prices may play a more important role in this category. On the other hand, the point estimates for the nominal and real elasticities are very similar for Health and Beauty Aids and Non-Food Grocery. There is no evidence that the minimum wage increases alcohol consumption disproportionately. In real terms, sales of Alcohol have a positive but very small coefficient, which is estimated with very poor precision, so the elasticity is not statistically significant from zero. In Model (1), not my preferred specification, the point estimates for Alcohol are slightly larger, around 0.09, but still lower than the elasticities of the remaining product groups. Since Nielsen tracks only a small fraction of liquor sales at participating stores 9, the poor precision of the estimates is not surprising and prevents further inference. 5 Robustness In this section, I show that my results are robust to a number of alternative specifications and provide evidence favoring my identification strategy. First, I show the absence of significant pretrends in sales by exploring the dynamic response to the policy. Second, I show that no differential pretrends show up in observable variables measuring local economic conditions. Third, I check that my results are robust to unobservable time-changing spatial heterogeneity by running a diff-in-diffin-diff analysis. Fourth, I show that the expenditure response is stronger in smaller counties, but even in the largest counties it is positive and significant. Fifth, I find that the point estimate of the expenditure elasticity is larger in the second half of my sample, suggesting that recent minimum wage hikes at the state level have affected relatively more workers. Sixth, I show that my results are not driven by indexing states. Seventh, I argue that limiting the sample to stores always present in the Nielsen dataset does not introduce a significant bias for lack of entry and exit by running the regressions at the store level. I also find that nominal and real sales experience a very similar change at the store level, suggesting that consumers move to more expensive stores, rather than stores increasing prices. Eighth, I find a positive effect on nondurable expenditure as measured in the preliminary PCE data by state, providing further evidence on the quality of the Nielsen data. Ninth, I show that my results are not driven by the bust of the housing bubble by excluding the 9 Kilts Center for Marketing (2014). 16

17 Nominal Sales Real Sales (1) (5) (1) (5) Health and Beauty Aids 0.128* 0.055* 0.117* 0.051* (0.076) (0.031) (0.067) (0.030) Food * * (0.114) (0.069) (0.085) (0.049) Non-Food Grocery (0.111) (0.056) (0.094) (0.053) Alcohol (0.214) (0.203) (0.215) (0.237) General Merchandise 0.231** 0.137*** (0.094) (0.038) (0.074) (0.033) Figure 7: Sales Elasticities by Product Group The expenditure response is quite homogeneous across product groups * p < 0.1, ** p < 0.05, *** p < Standard errors (in parenthesis) clustered at state level. The product group Food includes Nielsen s departments of Dry Grocery, Frozen Foods, Dairy, Deli, Packaged Meat, and Fresh Produce. County and time fixed effects are included in both specifications. Employment, population, gross state product, and house prices are also included in Specification (5). quarter of counties most affected by the drop in house prices. Tenth, I show that the estimated expenditure response is even larger when observations are not weighted by population. 5.1 Dynamic Response to the Minimum Wage Significantly different pretrends in sales between treatment and control could raise concerns on the validity of my identification strategy, which rests on the assumption of conditional parallel trends. In this subsection, I explore the dynamic response to the minimum wage and I find that there are no statistically significant pretrends. To explore the dynamic response to the minimum wage I follow Dube, Lester, and Reich (2010) and estimate my preferred specification with distributed lags spanning 20 quarters, 12 leads 10 and 8 lags: log(sales c,t ) = κ c + τ t + 7 j= 12 β j log(mw s,t+j ) + β 8 log(mw s,t 8 ) + γx c,t + ɛ c,t The use of changes rather than values produces coefficients representing the cumulative response to the minimum wage. Figure 8 summarizes the results. 10 At the time of this draft, information on minimum wage rates is available until the end of 2015, so the use of 12 leads reduces the sample period to for this exercise. 17

18 Figure 8: Dynamic Response to the Minimum Wage Nominal Sales Real Sales β j 0 β j Quarters Quarters Pretrends are not significantly different from zero 95% Confidence Intervals included. Standard errors clustered at state level. β j is the cumulative response to a change in the minimum wage after j quarters. Specification (5) is used, which includes employment, population, gross state product, and house prices, and time and county fixed effects. Pretrends are not significantly different from zero prior to the change in the minimum wage. Between one and three years before the policy change, the point estimates do not exhibit a trend and are approximately zero. Four quarters prior to the minimum wage hike, the estimate rises for both nominal and real sales, although it is still not statistically different from zero even with 90% confidence. This increase is consistent with consumption smoothing if agents are allowed to borrow against future higher incomes, but it could also be picking up on the fact that most increments during my sample period take place at one-year intervals and are of similar sizes. After the minimum wage hike, point estimates exhibit a clear upward trend, which is statistically significant for nominal sales, but not for real sales. Two years after a 10% minimum wage increment, nominal sales increase by 2.3% and real sales by 1.9%, around 75 basic points more than the effect at the time of the hike. 5.2 Pretrends in Observables My identification strategy rests on the assumption of conditional parallel trends, an assumption that is untestable by definition. In this subsection, I show that treatment and control do not exhibit different trends in observable variables measuring economic conditions To assess the existence of pretrends on observables, I regress them on twelve leads of the minimum wage: log(y c,t ) = κ c + τ t + γ + 0 j= 12 β j log(mw s,t+j ) + ɛ c,t 18

19 0.2 Figure 9: Pretrends in Observable Variables Employment 0.2 Current GSP β j 0 β j β j Quarters House Prices Quarters β j Quarters Regional Deflator Quarters Observables do not exhibit pretrends 95% Confidence Intervals included. Standard errors clustered at state level. β j is the elasticity of the observable variable to the level of the minimum wage within j quarters. County and time fixed effects are included. where Y c,t is employment, house prices, current gross state product, and regional price deflator. In this case, I choose to regress on levels and not on changes of the minimum wage because the interpretation of cumulative response does not make sense for these variables. Figure 9 shows that economic conditions were not changing in a systematically different fashion prior to the minimum wage hike for treatment and control counties. 5.3 State-Specific Time Fixed Effects In this subsection, I show that my results are robust to the presence of unobservable spatial heterogeneity in time trends through a diff-in-diff-in-diff exercise. The desirability of local area controls to account for unobservable and heterogeneous time trends is at the center of an ongoing debate in the literature measuring the impact of the minimum wage on 19

20 labor market outcomes. Dube, Lester, and Reich (2010) and Allegretto, Dube, Reich, and Zipperer (2015) argue that the standard two-way fixed-effects panel model fails to account for preexisting decreasing time trends in employment during the 1990s and 2000s. Since these decreasing time trends appear to be spatial in nature, the authors propose to account for them by including region or division-specific time fixed effects, or by comparing contiguous border counties. The assumption is that counties within the same region or division, or neighboring counties are better control groups. On the other end of the debate, Neumark, Salas, and Wascher (2014a) and Neumark, Salas, and Wascher (2014b) argue that such an approach throws away a lot of valid information and mechanically produces insignificant results. In their view, it is not clear that within region or division counties, or neighboring counties are better control groups. Unobservable spatially heterogeneous trends do not appear to be a concern in my sample period. The previous two subsections have shown that there are not significant pretrends in sales or in economic conditions. Nevertheless, I explore the robustness of my results through a diff-in-diff-indiff exercise. I interact the (log) minimum wage by a measure of its bindingness at the county level and I incorporate state-specific fixed effects, τ t,s. Formally, I run the following regression: log(sales c,t ) = κ c + τ t,s + βlog(minimum Wage s,t ) Bindingness c + γx c,t + ɛ c,t In this new specification, β measures the differential impact of the minimum wage in counties where the policy binds more tightly by comparing them with other counties within the same state but where the policy binds less tightly. Table 4 presents the results for three alternative measures of bindingness: the ratio of median income to minimum wage, the fraction of workers filing tax returns by less than $25,000, and the fraction of households with income lower than $25,000. I find that the expenditure response is indeed stronger in counties where the minimum wage binds more tightly even after allowing for state-specific unobservable and heterogeneous time trends, despite the strong requirements in the data imposed by this identification strategy. In Panel A, counties where the median income is high compared to the minimum and so, where the policy binds less tightly, have a lower response to increments in the rate as expected. The estimate is not statistically significant in Specification (5), but it is fairly stable across models. In Panel B, the use of IRS fraction of tax returns does not yield statistically significant results, but the points estimates are still positive. Finally, in Panel C, counties where more households have low incomes experience a larger expenditure response to increments in the minimum wage. Thus, my results are robust to heterogeneous time trends. 5.4 Minimum Wage Effect by County Size In this subsection, I explore the differential response of sales to the minimum wage by county size. The exercise is reveals whether the identification of the elasticity is coming from small or large counties. The former are more likely to be affected by the minimum wage as they are relatively less productive, whereas the latter are better measured by Nielsen. 20

21 (1) (2) (3) (5) Panel A. Median income Nominal Sales 0.044*** 0.063*** 0.038* (0.014) (0.015) (0.020) (0.019) Real Sales 0.038** 0.057*** 0.037** (0.015) (0.015) (0.017) (0.016) Panel B. Fraction of Tax Returns Below $25,000 Nominal Sales (0.432) (0.494) (0.624) (0.592) Real Sales * (0.294) (0.335) (0.425) (0.392) Panel C. Fraction of Households with Income Below $25,000 Nominal Sales 0.714*** 0.950*** (0.235) (0.253) (0.360) (0.336) Real Sales 0.662*** 0.903*** 0.609** 0.448* (0.192) (0.194) (0.255) (0.226) Observations Counties Quarters Clusters Table 4: State-Specific Time Fixed Effects The expenditure response is stronger in counties where the minimum wage binds more tightly Coefficient is the elasticity of sales to the interaction between the minimum wage and a measure of the bindingness of the policy. In Panel A, Median income refers to the median household income in the county from the ACS divided by the minimum wage and adjusted by the number of working hours in a year to facilitate interpretation. * p < 0.1, ** p < 0.05, *** p < Standard errors (in parenthesis) clustered at state level. County and state-specific time fixed effects are included in every specification. Specifications differ on the controls included: (1) does not include any controls, (2) controls for employment, (3) controls for employment and population, (4) controls for employment, population, and house prices. 21

22 I use the 2010 Census population to sort counties into four groups and run my baseline specification, Model (5), in each group separately. For this exercise, I do not use weights to highlight how sales in larger counties are measured more precisely. Figure 10 summarizes the results. The expenditure response is stronger in smaller counties, but the estimates are not very precise. After an increase in the minimum wage of 10%, the group of smallest counties experiences a growth of 2.8% in nominal sales and 2.3% in real sales. The large expenditure elasticity is consistent with small counties being less productive 11 and thereby more affected by minimum wage hikes. However, since Nielsen tracks fewer stores in these counties, the measure of sales is noisier and the confidence intervals for the estimates are wide. In the case of real sales, the elasticity to the minimum wage is not statistically distinguishable from zero at the 95% level for the two groups of counties with population lower than the median. In larger counties, the expenditure response is smaller but statistically distinguishable from zero. After an increase in the minimum wage of 10%, the group of largest counties experiences a growth of 9.6% in nominal sales and 7.5% in real sales. The use of weights in the baseline specifications in section 4 explains the similarity between these estimates and those presented earlier. elasticities are more precisely estimated for this group of counties as Nielsen tracks more stores and the idiosyncratic noise washes out. The nominal and real sales elasticity are significantly different from zero with a level of confidence of 99% and 98% respectively. 5.5 Breaking Sample Period In this subsection, I explore whether different patterns emerge for different subperiods of the sample. I find that the expenditure response is larger in the second half of the sample, but also is estimated with more noise. I break the sample period into two subperiods of 16 quarters each. The first subperiod covers , it contains the three federal increments in the minimum wage and all of the Great Recession as defined by the NBER 12. The second subperiod goes from 2010 to 2013 and it only includes increments of the minimum wage rates at the state level. The expenditure response is larger in the last four years of the sample, but is also estimated with less precision. The In my preferred specification, Model (5), the elasticity of nominal sales to the minimum wage is in the period , but only in the previous four years. Similarly, the point estimate for real sales is four times larger in the more recent subperiod (0.106 versus 0.025). The difference could emerge from state level increments in the second half of the sample applying over already larger minimum wage rates and so, potentially affecting a larger share of the labor force. However, the standard errors are much larger in the second half of the sample because the variations of the minimum wage exploited are much smaller. Thus, the difference in elasticities is not statistically significant. 11 Rosenthal and Strange (2004) 12 According to the NBER, the Great Recession started in the last quarter of 2007 and ended in the second quarter of Source: 22