Wishing for bad weather: demand shocks and the labor supply behavior of Bologna Pizza Vendors
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Wishing for bad weather: demand shocks and the labor supply behavior of Bologna Pizza Vendors Alessandro Saia∗ University of Bologna and OECD Job Market Paper This version: October 2015 Download the most recent version here Abstract A number of recent influential papers have analyzed the labor supply behavior of New York taxi drivers (Camerer et al., 1997; Farber, 2008; and Crawford and Meng, 2011), the findings of which are in sharp contrast with the prediction of the standard model of labor supply, but are consistent with alternative theories of labor supply decisions with Reference-Dependent Preferences (Kőszegi and Rabin, 2006). In the spirit of previous works, I study the labor supply behavior of Home Delivery Pizza Vendors in Bologna who, like taxi drivers, experience temporary variations in their earning opportunities, and are free to work more or less by accepting (or not accepting) orders. Unlike previous papers, however, I can use weather shocks as instrument for earning opportunities. By utilizing the variation in demand due to adverse weather conditions, I find that the behavior of pizza vendors is perfectly in line with the predictions of the standard model of labor supply. When truly exogenous demand shifters are used, the target component plays no role in vendors’ labor supply decisions. Moreover, I show that using the identification strategies of previous papers (in particular, Crawford and Meng, 2011), pizza vendors behave like taxi drivers in terms of their labor supply decisions. This shows that the findings of previous studies are driven by the (incorrect) way in which demand shocks are identified. Keywords: Labor Supply, Reference-Dependent Preferences, Elasticity of Intertemporal Substitution JEL Classification: D01, D03, J22 ∗ Department of Economics, University of Bologna & Economics Department, OECD. E-mail: alessandro.saia@unibo.it. I am particularly grateful to Andrea Ichino. I would like to thank Maria Bigoni, Matteo Cervellati, Juan J. Dolado, Irene Ferrari, Riccardo Ghidoni, Giovanni Mastrobuoni, David Reiley, Alfonso Rosolia, Vincenzo Scoppa, Giulio Zanella and seminar participants at Bologna, EUI, Aix-Marseille School of Economics, the BOMOPAV 2014 Meeting, the XIII Brucchi Luchino Workshop and the XIV TIBER Symposium on Psychology and Economics. The views expressed in this paper are those of the author and do not necessarily reflect those of the OECD and its member countries.
1 Introduction There is a growing list of influential papers documenting a series of behavioral anomalies that question the validity of standard economic models. Several studies have provided robust evidence of the fact that individuals tend to overweigh the present compared to the future (Laibson, 1997), that they make decisions based purely on a subset of the available information (Dellavigna and Pollet, 2009), and that the utility of the people in question is also a function of the utility of other people (Charness and Rabin, 2002). An assessment of the extent to which the core assumptions and the predictions of standard economic models are correct, is of crucial importance to academics and policymakers alike. Incorporating behavioral insights into the design of public policies not only could improve the effectiveness of existing policies, but could also offer new tools for the future (Chetty, 2015). The implications of behavioral biases may be particularly important within the context of labor economics. For example, the current design of income taxes or unemployment insurance policies is entirely based on the prediction of standard models of labor sup- ply: labor supply should increase in response to transitory, positive changes in earnings opportunities (i.e. positive Frisch elasticity). Is this prediction true? Until recently, empirical research into the relationship between workers’ efforts and their earning opportunities was conducted using individual panel data available for standard workers. However, using standard workers to test the predictions of labor supply models is problematic for two main reasons. First, these workers are faced with variations in wages that are of a permanent nature, and are thus correlated with significant income effects. Second, they usually operate within settings in which (at the intensive margin) it is not possible to establish their own efforts. Therefore, while the external validity of these results is beyond any doubt, what is difficult to identify is the underlying labor supply behavior. Due to the serious limitations of available data, scholars had gradually reduced their interest in studying workers’ labor supply decisions. However, in recent years the extensive body of literature within economics that focuses on the behavior of the underlying labor supply, has once again become central to academic debate. Instead of investigating the labor supply decisions of standard workers, recent papers have analyzed the labor supply behavior of piece-rate workers such as taxi drivers (Camerer et al., 1997; Farber, 2008; and Crawford and Meng, 2011) and bike couriers (Fehr and 2
Goette, 2007). These specific categories of worker have proven to be ideal when investigat- ing labor supply decisions. On the one hand, they operate in settings where they are faced with constant changes in their earning opportunities (in other words, the income effects on the labor supply are likely to be negligible). On the other hand, they are free to decide whether to work, and to what degree, and this is also easy to measure. However, despite the fact that their efforts can be set, these papers have provided evidence that is in sharp contrast with the prediction of the standard model, but is consistent with alternative theo- ries of labor supply decisions with reference-dependent preferences. Reference-dependent preferences (henceforth, RDP ) labor-supply models predict that workers’ labor supply depends not only on earning opportunities, as standard neoclassical theory claims, but also on a target (Reference Point). The idea is that a taxi driver, for example, decides to stop driving after having reached his/her target, for example a given level of income, and not on the basis of whether future potential earnings are high or low. Crawford and Meng (2011) is the most recent work analysing the daily labor supply of New York City taxi drivers, and it suggests that the behavioral component is an important part of labor-supply decisions. In their paper, the aforesaid authors derive a model of daily labor supply as a survival model whereby at the end of each journey, a driver decides whether to continue or to stop driving. They estimate a reduced-form model of the probability of stopping as a linear function of cumulative hours and cumulative income, and in order to investigate the relevance of target behavior they include two dummy variables indicating whether hours or income exceed the targets. Building on Kőszegi and Rabin’s (2006) theory of RDP, targets reflect the driver’ rational expectations in terms of hours and income, and represent the driver’s belief concerning possible outcomes. However, due to the fact that targets are not observed, they operationalize them as functions of endogenous variables. Given the lack of suitable demand-shifting instruments, they compute sample proxies for drivers’ rational expectations using driver-specific sample average income and hours prior to the current shift, allowing them to vary across days of the week in order to take account of variations in demand throughout the course of the week. Their results show a strong deviation from the standard neoclassical model of worker behavior, and suggest that behavioral biases are a key component of drivers’ labor supply decisions. However, Crawford and Meng are not able to identify an exogenous demand shifter and therefore why taxi drivers’ efforts vary throughout the day is not clear. In the spirit of previous works, I analyze the labor supply behavior of a specific group of 3
piece-rate workers: Home Delivery Pizza Vendors in Bologna. At the intensive margin (i.e. during a day) vendors, like taxi drivers, can decide how much they want to work by accepting (or not accepting) orders. A key feature of this paper is that unlike previ- ous works, I am able to identify an exogenous, transitory labor demand shifter. Using exogenous, transitory changes in weather conditions as instrument for vendors’ earning opportunities, I invariably find that vendor’s behavior is consistent with the prediction of the standard model of labor supply. This result contrasts sharply with the results of previous studies, and I show that this difference in findings is due to differences in the analytical approach adopted rather than in the data used. The remainder of this paper is organized as follows. The next section describes the labor market for Home Delivery Pizza Vendors and the data utilized in this paper. Section 3 presents empirical evidence of the relevance of adverse weather condition for pizza demand using Web search data. In Section 4, I present the empirical approach and my findings. Section 4.1 sets out initial evidence of the relationship between demand shocks and the labor supply behavior of food vendors. In Section 4.2, I replicate the analysis of Crawford and Meng (2011) using Pizza Vendors’ data, and the results I obtain are consistent with those found for taxi drivers. In Section 4.3, I present the empirical results obtained by utilizing the exogenous variation in rainfall as instrument for earning opportunities, and I show that the behavior of home delivery pizza vendors is perfectly consistent with the predictions of the standard labor supply model. Section 6 offers some conclusions. 2 Data and Background The data used in this paper consists of high-frequency data regarding orders carried out through an on-line takeaway pizza delivery service (henceforth, the Website). The Website acts as a web-based intermediary between privately-owned restaurants and their customers (a modern variant of the Yellows Pages, if you like). The Website allows customers to select a restaurant delivering pizzas and to place their orders with that restaurant. When the customer places an order, a message is sent to the restaurant through a 2-way till receipt/order machine. Pizza vendors receive the order through the machine box, and can then accept or refuse the order. If they accept it, a confirmation message is sent to the customer with an indication of the delivery time (set by the vendors). Payments are collected by the vendors on delivery of the pizzas, and the Website charges a commission for each order. 4
The pizza delivery restaurants in question are located in Bologna, and are privately owned family businesses. Unlike normal restaurants, their distinguishing feature is that they do not provide table service, and the majority of customer orders are carried out by telephone or on the Web. Therefore, I only observe one part of total demand (i.e. those orders received through the Website); however, it is reasonable to assume that the two main channels (telephone and Web) are positively related. Unlike previous works, this paper investigates the labor supply behavior of a group of workers. In the rest of this paper I consider each single restaurant as a group of workers, and I refer to all of the workers from one restaurant as its vendors. Moreover, I assume that the group (in terms of size and composition) remains constant throughout the observation period. Given the labor market characteristics in question (i.e. the restaurants are all family businesses), it is fair to assume this to be true. 2.1 Data on Pizza Sales The Website provided me with all orders placed with 59 pizza delivery restaurants during the period February 2010-July 2011. As previously explained, vendors receive orders through a machine box, whereupon they can choose whether to accept or reject them. Rejected orders are vital if a credible estimate of the intertemporal substitution elasticity for labor supply (Frisch elasticity) is to be provided. Unfortunately, as in previous works, in this paper I rely only on the equilibrium outcomes (i.e. accepted orders), since rejected orders were not recorded. From February 2010 to July 2011 a total of 147,090 orders were observed. Figure 1 plots the location of each pizza delivery restaurant in Bologna (the red dots in the left panel) and the spatial distribution of orders observed in the sample (the blue dots in the right panel). I have detailed information about the date and time each order was placed, the total amount of the delivery, the distance between the restaurant and the place of delivery, and the expected delivery time, all of which is communicated to customers when orders are accepted by the vendors. [Figure 1] Summary statistics are shown in Table 1. Restaurants’ daily average earnings across the whole sample amount to 68.15 e, while the average number of orders accepted in a day is around 8. Not surprisingly, daily earning opportunities vary across days of the week and across months. Earning opportunities are higher during the winter and at weekends (for 5
example, daily earnings and daily orders on Sundays in the month of January are almost twice those of the overall sample). [Table 1] Figure 2 shows the distribution of orders throughout the day. The number of orders received varies greatly: orders are concentrated around lunch breaks (generally between 12 noon and 3 p.m.) and dinner hours (generally 6-9 p.m.). [Figure 2] 2.2 Weather Data Data for weather conditions in the city of Bologna over the entire period are provided by the Emilia Romagna Regional Agency for the Protection of the Environment (ARPA). Hourly rainfall data are provided by a weather station located in Bologna’s city center (Bologna Urbana)1 . [Figure 3] Figure 3 plots the number of days per month with measurable rainfall. From February 2010 to July 2011 there were 206 rainy days (i.e. days on which there was at least 0.1 mm of rain). As expected, the average monthly number of wet days varies considerable from one season to the next: figures are much higher in winter, and much lower in summer. Figure 4 shows changes in rainfall intensity,measured in terms of the number of rainy hours observed during the course of the day2 . The considerable variation in the rainfall observed during the period in question would seem to suggest that the intensity of rainfall during the day needs to be taken into account. [Figure 4] 3 The Demand for Pizza Delivery in Bologna Demand for pizza delivery fluctuates daily due to demand shocks caused by day-of-the- week effects, sport events, holidays, etc. Among others, weather conditions are one of the 1 Note that the city of Bologna, with a population of about 384,000 and a total area of about 140.7 km2 , is relatively small. New York City, for example, has a population of over 8.4 million, living in an area of 1,214 km2 . 2 Wet hours are defined as those in which more than 0.00 cm of rain was observed. 6
key shifters of demand for pizza delivery services. Workers in the food delivery industry recognize that rainfall is strongly correlated with positive short-term fluctuations in the demand for pizza delivery services. For example, in 2012, while announcing a rise in his firms’ profits, the CEO of Domino’s Pizza3 , Lance Batchelor said ”When the weather turns nasty we always see sales turn upwards”. Why does rainy weather positively affect demand for pizza delivery? Tony McNaulty, the owner of Papa Murphy, another chain of pizza restaurants, suggests that ”if it’s going to take you 40 minutes to get home on an average day with no rain...it’s going to take you an hour to get home. When it rains, everybody’s running a little bit later and maybe the family doesn’t have enough time to prepare dinner now, so they use us” 4 . In this section I use Web search data to provide evidence of the positive effect of adverse weather conditions on the demand for pizzas. Several studies have recently demonstrated that Web search behavior correlates in near real time with individual activity, and that search counts are predictive of economic outcomes, including unemployment levels (Et- tredge et al., 2005), home sales (Wu and Brynjolfsson, 2013) and consumers activities such as cinema attendance and music sales (Goel et al., 2010). Likewise, I use daily Web searches to derive information about the elasticity of pizza demand to rainy conditions. Pizza-related search queries are provided by Google Trends. Google Trends provides me with an index5 that indicates the daily volume of queries submitted to Google that include the Italian words for ”pizza restaurant Bologna” in the area of Bologna. Figure 5 plots Google Trends’ daily index of searches for the words ”pizza restaurant Bologna” from February 2010 to July 2011. The pattern of Web searches is consistent with the behavior of pizza demand: searches follow a seasonal trend that is compatible with pizza consumption and the line peaks during the cold months. [Figure 5] I examine to what extent adverse weather conditions affect daily Web search volume, 3 Domino’s Pizza is an American restaurant chain and international franchise pizza delivery corporation with 145,000 employees and more than 1 billion USD of revenue. 4 The full article is available here. 5 Google Trends data does not report the raw level of queries for a given search term; rather, it reports a query index that indicates the volume of searches for the specific search term divided by the total number of queries in a given geographic location at a certain point in time. 7
using the following model: log(Searchest ) = β + β1 Rainy Dayt + Zt β + t (1) where Searches t represents the Google Index on day t, Rainy Day is a dummy variable that takes value 1 if it has rained on the corresponding day, and Z is a vector of variables that includes day-of-the-week, monthly and yearly fixed effects. Column (1) of Table 2 reports the results of the OLS estimation using the model that also includes day of the week, monthly and yearly fixed effects. As expected, the coefficient of interest is positive and statistically significant: adverse weather conditions during the day lead to approximately 6% of search queries being related to pizza consumption. [Table 1] Not surprisingly, results are robust when investigating the effect of rainfall at the intensive margin. Results in Columns (2) and (3) of Table 2 present estimates that alternatively take, as their variables, the number of rainy hours in a day and the total amount of rainfall (in millimeters), respectively. The corresponding estimate is sizable. For example, five hours of rain are associated with an increase of around 4% in the daily volume of pizza- related queries submitted to Google6 . Finally, it should be pointed out that bad weather might also directly affect the supply of pizza delivery services. For example, driving in the rain is harder, more difficult and less enjoyable. Therefore, it is possible that adverse weather conditions negatively affect the supply of pizza delivery services. In order to explore this possibility, I use the infor- mation about delivery times contained in pizza sales data. The idea is that the expected delivery time should provide information regarding the difficulty of delivering pizzas in 6 Previous works have suggested that ”rain may have a number of effects on the market for taxi rides. First, it likely increases” (Farber, 2015). However, rain might also decrease demand for taxis. For example, I usually try to avoid taking taxis when it rains because rainy weather affects traffic congestion, it reduces average car speed and increases the average cost of the ride (see also Footnote 13 in Farber (2005)). Using Google searches, I investigate the relationship between demand for taxis and the quantity of precipitation. I collect data for the quantity of daily Web searches for the word ”Taxi” in New York City from January 2010 to June 2011 (in order to have similar sample size, I have used an equivalent time period to the one employed in the main text). Results are shown in Appendix A6. OLS estimation results reported in Table 18 suggest that weather conditions do not affect Google searches for the word “Taxi” in NYC. It is interesting to note that Taxi-related searches are correlated with other demand shifters. The results reported in Table 19 show that the volume of searches is higher at the weekend and during public holidays. These estimates suggest that Google searches are a good proxy for taxi demand, and the lack of correlation that I observe between web searches for the word ”Taxi” and adverse weather is probably due to the fact that adverse weather does not positively affect demand for taxis (if anything, the corresponding point estimates are always negative). 8
wet weather. Using information on delivery times, I estimate: log(Delivery T imeih ) = α + β1 log(Ordersih ) + β2 Rainy Hourh + +β3 Rainy Hourh ∗ log(Ordersih ) + ih where Delivery T imeih represents the average delivery time observed for vendors i dur- ing hour h, Ordersih is the number of orders accepted by vendors i during hour h and Rainy Hour takes a value of 1 if it has rained during the corresponding hour. The co- efficient of interest is β3 , and indicates the combined effect of the number of orders and adverse weather conditions. Column (1) of Table 3 contains the corresponding OLS es- timates. The coefficient of interaction term is positive and significant: expected delivery times are thus longer when it is raining. This finding is in keeping with the idea that rain complicates the delivery of pizzas. Similar results are reported in Column (2), where millimeters of rain are used as a measure of bad weather. Taken as a whole, the findings presented in this section indicate that while adverse weather conditions positively affect demand for pizza delivery services, rainy weather has a negative impact on pizza vendors’ working conditions. 4 Estimations of Labor Supply Models 4.1 Demand shocks and labor supply I begin the analysis of vendors’ labor supply behavior by firstly providing preliminary evidence of the relationship between demand shocks and labor supply behavior. Pizza vendors are piece-rate workers, and their labor supply decisions concern how much effort they want to make by accepting orders. Then, as a natural proxy of vendors’ efforts, I use the number of orders accepted during the course of the day7 . The results presented in the previous section suggest that adverse weather conditions significantly increase pizza delivery demand. By utilizing the exogenous increase in pizza delivery demand, I can investigate the relationship between labor supply and labor de- 7 In principle I could use the number of pizzas produced or total earnings during the day as alternative measures of effort. The adoption of alternative measures of labor supply does not change results very much. Appendices A3 and A4 show the results obtained with the dependent variable consisting of total revenue collected, or the number of pizzas produced during the day, respectively. 9
mand shocks using the following model: log(Ordersit ) = α + β1 Rainy Hoursit + Zit β + it (2) where the dependent variable is the number of orders accepted by vendors i in day t, the variable Rainy Hours indicates the number of hours of adverse weather experienced during the course of the day by vendors i and Z is a vector of variables including vendors- day-of-the-week fixed effects and the full set of monthly and annual fixed effects. For each pizza vendor, the number of rainy hours observed in a day is computed taking a given hour as the starting point and the time at which the last order was placed as the ending point. This starting hour corresponds to the earliest hour at which an order was observed over the entire sample. Thus, while the starting hours vary among vendors, it is assumed to be constant over the period in question. An alternative approach is to consider the time at which the first order was placed during the day as the starting point. This approach, which is implicitly adopted by Crawford and Meng (2011), allows starting hours to vary among vendors and across days, but relies on the strong assumption that the waiting time for the first order is always equal to zero. This approach seems inconsistent with the daily activity of pizza vendors, since as with taxi drivers, a series of preliminary steps are required (in this case, for example, a certain stable oven temperature is required for pizzas to be made)8 . [Table 4] Column (1) of Table 4 contains OLS estimates obtained using the number of wet hours observed in the day as the variable of interest. The corresponding coefficient is positive and statistically significant. The demand for delivered pizzas is positively influenced by adverse weather conditions, and consequently vendors react by making more effort. Simi- lar results are obtained when the variable in question is the total rainfall (in millimeters) during the day (Column 2). Overall, the estimates shown in Table 4 provide evidence of the positive relationship between labor demand and labor supply of food delivery vendors: vendors work more when earning opportunities are greater. The size of the coefficients suggests that 5 rainy hours (corresponding to the average number of wet hours observed on rainy days) result in a 5% increase in the number of orders accepted by food delivery vendors. Moreover, it 8 It should be pointed out that when the alternative measure is adopted, all the results are qualitatively similar. 10
is interesting to note that the coefficients in Table 4 are fairly similar to the coefficients obtained in Section 3 using Google Trend data. However, it should also be noted that these results, while providing initial evidence of the positive sign of labor supply elasticity, do not provide very much information. Positive elasticity of labor supply is not inconsistent with the presence of target behavior. Vendors fail to respond to a positive change in labor demand if, and only if, labor supply decisions are exclusively driven by reference-dependent preferences9 . Since labor supply decisions cannot plausibly be driven solely by the target behavior, it seems reasonable to investigate vendors’ behavior using a more flexible empirical strategy permitting the co-existence of the neoclassical component of labor supply with the reference-dependent effect. 4.2 Estimations using Crawford and Meng’s Target In the next section I replicate the analysis of Crawford and Meng (2011) by using pizza sales figures. Obtaining a similar finding is of crucial importance since it implies that despite the intrinsic differences between taxi drivers and pizza vendors (e.g. single vs group of workers, different settings and cities), any differences in the results are due to the different econometric approach rather than to the differences between the two settings10 . Using pizza-delivery data, I estimate a probit model of the probability that vendors i will stop delivering on day t after order x has been accepted, as a function of cumulative orders accepted during the day and a dummy variable indicating whether orders accepted in a day exceed the target: P (stopitx = 1) = Φ(δI[Yitx > Tit ] + γYitx + Zitx β) (3) where Φ(·) is the standard normal cumulative distribution function, I[·] is a function that takes value 1 if the number of orders accepted at that point in the day is more than the target level of orders Tit and Z is a vector of variables that includes controls for weather conditions and vendors’ day of the week fixed effects and the full set of hour of the day, monthly and yearly fixed effects. 9 In a model of labor supply decisions where total utility is defined as a weighted average of consumption utility and gain-loss utility, workers will not respond to a demand shock only in the case where the gain- loss utility’s weight is equal to one. See Appendix A1. 10 I depart from their analysis by modeling pizza vendors’ labor supply decision solely as a function of their effort. However, it is important to note that the fundamental results of this section are qualitatively the same if I adopt a labor supply model with two targets, namely specific sample average income and hours worked prior to the current day, like the one adopted by Crawford and Meng (2011). 11
Following Crawford and Meng (2011), I compute the target using vendors’ monthly and day-of-the-week specific sample average orders accepted in one day, prior to day t (i.e. the target for vendors i, for day-of-the-week d in month m is the average number of orders accepted by vendors i in previous days-of-the-week d in month m up to but not including the day in question)11 . Despite the presence of endogenous regressors, the empirical strategy proposed by Crawford and Meng (2011) has the advantage of isolating the neoclassical component of labor supply from the reference-dependent effect. While the estimated coefficient of cumulative orders provides evidence of that part of labor- supply behavior related to standard preferences, the coefficient of the dummy indicates the variation in the probability of the vendor continuing work after the target has been reached. If positive and statistically significant, the coefficient of the target will provide evidence of a significant deviation from the standard labor supply model. Table 5 contains the marginal probability effects obtained using the reduced form model of stop-working probability. Following Crawford and Meng (2011), the model in the first column contains weighted estimates, where the weights are equal to the number of observations used to compute the target. Since sampling error tends be larger in the first part of the sample period, weighted estimates reduce the bias due to sampling variation. Results from unweighted regression, where sampling variation is ignored, are shown in Column (2). Results obtained using unweighted and weighted regressions are essentially the same. In both models, the target’s coefficients are positive and statistically significant: vendors are more likely to stop working when they hit their target. At the same time, the level of orders accepted has a significant and negative effect, providing evidence of the presence of a neoclassical component in the labor supply decisions of pizza vendors. [Table 5] Despite the differences between the two settings, the results presented in this section are substantially consistent with those of Crawford and Meng (2011). Pizza vendors appear to act like taxi drivers: being above the set target increases the probability that will stop working. Overall these results suggest that reference-dependent preferences play a key role in the labor supply behavior of pizza vendors: when demand is higher than expected (i.e. 11 An alternative approach would have been to construct the target as the expected rainfall for the current day. The idea is that vendors may form their expectations regarding labor demand on the basis of weather forecast data. I have collected hourly weather forecasts for the corresponding period from a major weather forecast provider. The corresponding results are presented in Appendix A3. Results are not consistent with the predictions of the reference-dependence model. However, the poor performance of weather forecasts raises certain doubts about the fact that vendors rely on such forecasts to form their expectations of pizza demand for the day in question. 12
vendors have already reached their target for the day but pizzas are still being ordered), vendors are more likely to stop working. However, it is not clear why vendors reach their target (i.e. what are the factors that make pizza demand higher on that specific day), and it also doubtful whether (and difficult to understand how) target is correlated to the factors underlying demand for pizza delivery services. Not surprisingly, the coefficient of the dummy variable indicating whether it was raining during the corresponding hour12 is negative and statistically significant: workers’ efforts are positively related to any positive variation in demand for pizzas. In other words, when earning opportunities are temporarily higher, vendors’ behavior is consistent with the prediction of the standard model of labor supply. It should be pointed out that the role played by reference-dependent preferences in labor supply decisions would indicate that workers will respond differently to demand shocks if they have already reached their target13 . Unlike Crawford and Meng (2011), I can investigate this theoretical prediction by combining the dummy indicating whether orders accepted at that point in the day exceed the target, and the variable indicating whether it was raining during the corresponding hour. If reference-dependent preferences play an important role in determining vendors’ labor supply decisions, then the coefficient of the interaction term is expected to be positive and statistically significant (i.e. if vendors have reached their target they are less likely to respond to an increase in demand for pizza delivery). [Table 6] Table 6 reports the coefficient of the interaction model. The coefficients of the inter- action terms are not statistically significant at conventional levels. Results from both the weighted and the unweighted models show that the coefficient in question is actually positive, as theory predicts, but is far from significant: when earning opportunities are temporarily higher, the fact of surpassing the target plays no role in vendors’ labor supply decisions14 . The results presented in this section indicate that i) adverse weather conditions are pos- itively related to workers’ efforts, and ii) the role of target is unclear when demand is actually temporarily higher (i.e. during rainy hours). Overall, these results raise cer- 12 Similar results are obtained when rain is measured in millimeters, see Appendix A5. 13 In Appendix A1 I formulate two alternative models of labor supply for piece-rate workers operating in a market characterized by menu costs, designed to produce theoretical predictions 14 The coefficient of the interaction term is negative when adverse weather conditions are defined in terms of the amount of rain (in millimeters) that falls within an hour. 13
tain doubts about Crawford and Meng’s estimates and about the relevance of reference- dependent preferences to vendors’ labor supply decisions. 4.3 Estimations using Rainfall Variation Unlike Crawford and Meng, I can use changes in weather conditions to isolate exogenous variations in vendors’ earning opportunities. In this section I investigate the relevance of the target component, using demand shocks, defined in terms of rainfall variation, as instrument for pizza demand during the day. Using adverse weather conditions on a given day, I create two instruments for estimation purposes. I construct the variable Cumulative Rainy Hours as the number of rainy hours observed up to that point in the day15 . I also formulate a target variable labeled Target Rainy Hours, using vendors’ day of the week/ monthly sample average daily number of rainy hours observed up to, but not including, the day in question. As in the previous section, the first variable isolates the neoclassical component of labor supply, while the second variable is a proxy for the vendors’ rational point expectations of demand for pizza during the day. [Table 7] Due to the non-trivial solution of a standard two-stage regression in a probit model with multiple endogenous variables, I first present the reduced-form estimates16 . Table 7 reports the marginal probability effect obtained using the reduced-form model, where the probability that vendors end their working day is a function of the number of rainy hours observed during the day, and a dummy variable indicating whether rainfall reaches the estimated target. The left-hand panel shows the estimates of the weighted model. The coefficient of cumulative rain is negative and statistically significant, and in keeping with the results presented in Section 4.1 it suggests the presence of a neoclassical component in the labor supply decisions of pizza vendors. However, the estimated coef- ficient of dummy variables indicating whether daily rainfall exceeds the target is close to zero, and is far from being statistically significant. Similar findings are reported in the 15 As explained in Section 4.1, cumulative rainfall is calculated for each pizza vendor using a given hour as a starting point. The starting hour corresponds to the earliest hour at which an order was observed over the entire sample. Thus, while the starting hour varies from one vendor to the next, it is taken to be constant over the period in question. 16 The information underlying the relationship between rainfall-based and orders-based variables is provided in Table 10. 14
right-hand panel containing estimates of the unweighted regression. Table 8 shows the conventional two-stage least-square estimates17 . Columns (1) and (2) contain the 2SLS results from the weighted and unweighted models respectively. In both cases, the Angrist-Pischke first-stage F statistics (Angrist and Pischke (2008) are suitably high, and confirm that initial stage relationships are strong enough. Reported estimates are consistent with the reduced-form results, and provide no support for the relevance of reference-dependent preferences to the labor-supply records of pizza vendors. The effect of cumulative orders during the day is negative and highly significant, while the coefficient indicating whether orders accepted exceed the target, is devoid of statistical significance in both models. [Table 8] Overall, these results contrast with the estimates obtained in Section 4.2. When truly exogenous demand shifters are used, the behavior of pizza vendors is perfectly consistent with the predictions of the standard model of labor supply, and reference-dependent preferences play no role in vendors’ labor supply decisions. 5 Conclusions It is vitally important to understand the role played by behavioral components in workers’ labor supply decisions. If behavioral factors play an important role in labor supply deci- sions then policies that neglect such anomalies are going to be sub-optimal. In this paper I investigate the role of behavioral anomalies in the context of labor supply decisions. Using high frequency data for pizza sales, I study the labor supply behavior of Home Delivery Pizza Vendors in Bologna. Building on Crawford and Meng’s (2011) analysis, I show that when targets are defined as average sample realizations of previous labor supply decisions, I find evidence of reference-dependent behavior. As Crawford and Meng (2011) also dis- covered, the probability of vendors stopping work is higher when they reach their targets. However, unlike with previous works, I am able to identify an exogenous, transitory de- mand shifter. Unlike Crawford and Meng (2011), I can investigate the relevance of the target behavior using rainfall figures as an instrument for earning opportunities. When truly exogenous demand shifters are used, the target component plays no role in vendors’ labor supply decisions, and the observed behavior is consistent with the prediction of the standard labor supply model. 17 The comparable linear probability model’s estimates using Target Orders and Cumulative Orders are given in Table 9. 15
References Angrist, J. D. and Pischke, J.-S. (2008). Mostly harmless econometrics: An empiricist’s companion. Princeton university press. Camerer, C., Babcock, L., Loewenstein, G., and Thaler, R. (1997). Labor supply of new york city cabdrivers: One day at a time. The Quarterly Journal of Economics, 112(2):407–441. Charness, G. and Rabin, M. (2002). Understanding social preferences with simple tests. Quarterly journal of Economics, pages 817–869. Chetty, R. (2015). Behavioral economics and public policy: A pragmatic perspective. NBER Working Paper Series, No. 20928. Crawford, V. P. and Meng, J. (2011). New york city cab drivers’ labor supply revis- ited: Reference-dependent preferences with rational expectations targets for hours and income. The American Economic Review, 101(5):1912–1932. DellaVigna, S. and Pollet, J. M. (2009). Investor inattention and friday earnings an- nouncements. The Journal of Finance, 64(2):709–749. Ettredge, M., Gerdes, J., and Karuga, G. (2005). Using web-based search data to predict macroeconomic statistics. Communications of the ACM, 48(11):87–92. Farber, H. S. (2005). Is tomorrow another day? the labor supply of new york city cabdrivers. Journal of Political Economy, 113(1):46–82. Farber, H. S. (2008). Reference-dependent preferences and labor supply: The case of new york city taxi drivers. The American Economic Review, 98(3):1069–1082. Farber, H. S. (2015). Why you can’t find a taxi in the rain and other labor supply lessons from cab drivers. The Quarterly Journal of Economics, page forthcoming. Fehr, E. and Goette, L. (2007). Do workers work more if wages are high? evidence from a randomized field experiment. The American Economic Review, 97(1):298–317. Goel, S., Hofman, J. M., Lahaie, S., Pennock, D. M., and Watts, D. J. (2010). Predicting consumer behavior with web search. Proceedings of the National Academy of Sciences, 107(41):17486–17490. Kőszegi, B. and Rabin, M. (2006). A model of reference-dependent preferences. The Quarterly Journal of Economics, 121(4):1133–1165. 16
Laibson, D. (1997). Golden eggs and hyperbolic discounting. The Quarterly Journal of Economics, pages 443–477. Wu, L. and Brynjolfsson, E. (2013). The future of prediction: How google searches foreshadow housing prices and sales. NBER Chapters. 17
Tables and Figures Figure 1: Pizza Delivery Restaurants and Orders Across the City of Bologna. Figure 2: Distribution of Orders over the Day. Orders by Clock hour .2 .15 Frequency .1 .05 0 9 11 13 15 17 19 21 23 01 03 Clock hour 18
Table 1: Descriptive Statistics Full Sample Sundays January Obs. Mean St. Dev. Obs Mean St. Dev. Obs. Mean St. Dev. Number of Orders 19,746 7.44 7.54 2,916 10.89 9.80 1,319 8.26 7.93 19 Daily Revenues (in e) 19,746 68.15 66.71 2,916 96.25 89.43 1,3198 80.03 76.41 Number of Pizza Produced 19,746 14.91 17.45 2,916 20.91 22.07 1,319 16.77 16.26 Daily Precipitation (in mm) 19,746 1.04 3.24 2,916 1.66 3.84 1,3198 0.66 2.23 Number of Rainy Hour in a Day 19,746 0.87 2.18 2,916 1.55 3.33 1,319 0.83 2.66 Notes: Notes. The table reports the summary statistics for all the variables used throughout the empirical analysis. The unit of observation is the pizza delivery restaurant in each day.
Figure 3: Number of Rainy Days per Month. Figure 4: Number of Rainy Hours per Day. 20
Figure 5: Daily Search Volume for the Words ”Pizza Restaurant Bologna” Table 2: Google searches for the words ”pizza restaurant Bologna” and weather conditions (log) Searches (log) Searches (log) Searches Rainy Day 0.0635** (0.026) Number of Rainy Hours in the Day 0.0072*** (0.002) Daily Precipitation (in mm) 0.0069*** (0.002) Controls Yes Yes Yes R-squared 0.7030 0.7030 0.7046 Observations 546 546 546 Notes: *** Significant at the 1% level, ** at the 5% level, * at the 10%. Robust standard error are in parentheses. Estimates based on data on daily search Google Index provided by Google Trends. I used the GI of the Italian word for ”pizza restaurant Bologna” for the period February 2010 - July 2011. All columns include dummies for day of the week, month, year. 21
Table 3: Adverse Weather and Delivery Time (1) (2) (log) Delivery (log) Delivery Time Time Orders * Rainy Hour 0.0063** (0.003) Orders * Hourly Precipitation (in mm) 0.0027** (0.001) Controls Yes Yes R-squared 0.7401 0.7400 Observations 75,944 75,944 Notes: *** Significant at the 1% level, ** at the 5% level, * at the 10%. Standard error by vendors and day are assumed. Columns include vendors-day of the week FEs, month FEs, year FEs and hour of the day FEs and the average distance of delivered orders in the corresponding hour. The variable Rainy Hour takes value 1 if in the corresponding hour it has rained. 22
Table 4: Adverse Weather Conditions and Vendors’ Labor Supply (log) Orders (log) Orders Number of Rainy Hours in the Day 0.018*** (0.002) Daily Precipitation (in mm) 0.010*** (0.001) Controls Yes Yes R-squared 0.6782 0.6778 Observations 19,746 19,746 Notes: *** Significant at the 1% level, ** at the 5% level, * at the 10%. Standard error clustered by vendors are assumed. Columns include vendors-day of the week FEs, month FEs, year FEs and hour of the day FEs. Table 5: Marginal Effects on the Probability of Stopping - Crawford and Meng’ Target Weighted Unweighted Cum. Orders > Target Orders 0.034*** 0.040*** (0.003) (0.003) Cumulative Orders /10 -0.047*** -0.054*** (0.004) (0.004) Rainy Hour -0.024*** -0.021*** (0.005) (0.005) Controls Yes Yes Log likelihood -114,775.29 -28,521.724 Pseudo R2 0.3847 0.3866 Observations 493,517 120,209 Notes: *** Significant at the 1% level, ** at the 5% level, * at the 10%. Standard error clustered by vendors and day are assumed. Columns include vendors-day of the week FEs, month FEs, year FEs and hour of the day FEs. The model in the first column contains weighted estimates, where the weights are equal to the number of observations used to compute the target. Results from unweighted regression, where sampling variation is ignored, are shown in the second column. The variable Rainy Hour takes a value of 1 if it has rained during the corresponding hour. Evaluation point: after 10 orders on a rainy hour at 9:00 pm. 23
Table 6: Target Behavior and Heterogeneous Response to Demand Shocks Weighted Unweighted Cum. Orders > Target Orders 0.144*** 0.164*** (0.015) (0.014) Cumulative Orders /10 -0.200** -0.225*** (0.017) (0.016) Rainy Hour -0.125*** -0.099*** (0.035) (0.032) (Cum. Orders > Target Orders) * Rainy Hour 0.042 0.024 (0.046) (0.043) Controls Yes Yes Log likelihood -114,773.62 -28,521.592 Pseudo R2 0.3847 0.3866 Observations 493,517 120,209 Notes: *** Significant at the 1% level, ** at the 5% level, * at the 10%. Standard error clustered by vendors and day are assumed. Columns include vendors-day of the week FEs, month FEs, year FEs and hour of the day FEs. The model in the first column contains weighted estimates, where the weights are equal to the number of observations used to compute the target. Results from unweighted regression, where sampling variation is ignored, are shown in the second column. The variable Rainy Hour takes a value of 1 if it has rained during the corresponding hour. Table 7: Marginal Effects on the Probability of Stopping - Crawford and Meng’s Target Reduced-Form Estimates using Rainfall Variations. Weighted Unweighted Cum. Rainy Hours> Target Rainy Hours 0.003 0.008 (0.005) (0.005) Cumulative Rainy Hours /5 -0.015*** -0.018*** (0.004) (0.004) Controls Yes Yes Log likelihood -115,103.21 -28,630.311 Pseudo R2 0.3829 0.3842 Observations 493,517 120,209 Notes: *** Significant at the 1% level, ** at the 5% level, * at the 10%. Standard error clustered by vendors and day are assumed. Columns include vendors-day of the week FEs, month FEs, year FEs and hour of the day FEs. The model in the first column contains weighted estimates, where the weights are equal to the number of observations used to compute the target. Results from unweighted regression, where sampling variation is ignored, are shown in the second column. Evaluation point: after 5 wet hours at 9:00 pm. 24
Table 8: Probability of Stopping - 2SLS Results Weighted Unweighted Cum. Orders> Target Orders 0.067 0.118 (0.088) (0.078) Cumulative Orders /10 -0.120** -0.144*** (0.050) (0.047) AP 1st Stage F-stat: Cum. Orders > Target Orders 16.95 22.70 AP 1st Stage F-stat: Cumulative Orders /10 8.32 13.70 Controls Yes Yes Observations 493,519 120,210 Notes: *** Significant at the 1% level, ** at the 5% level, * at the 10%. Standard error clustered by vendors and day are assumed. Columns include vendors-day of the week FEs, month FEs, year FEs and hour of the day FEs. The model in the first column contains weighted estimates, where the weights are equal to the number of observations used to compute the target. Results from unweighted regression, where sampling variation is ignored, are shown in the second column. Table 9: Probability of Stopping - OLS estimates - Crawford and Meng’s Target Weighted Unweighted Cum. Orders> Target Orders 0.070*** 0.070*** (0.002) (0.002) Cumulative Orders /10 -0.097** -0.095*** (0.002) (0.002) Controls Yes Yes R-squared 0.2826 0.2888 Observations 493,519 120,210 Notes: *** Significant at the 1% level, ** at the 5% level, * at the 10%. Standard error clustered by vendors and day are assumed. Columns include vendors-day of the week FEs, month FEs, year FEs and hour of the day FEs. The model in the first column contains weighted estimates, where the weights are equal to the number of observations used to compute the target. Results from unweighted regression, where sampling variation is ignored, are shown in the second column. 25
Table 10: Relationship between precipitation-based and orders-based variables Weighted Unweighted Cum. Orders Cumulative Cum. Orders Cumulative > Target Orders Orders /10 > Target Orders Orders /10 Cum. Rainy Hours > 0.113*** -0.026 0.124*** -0.022 Target Rainy Hours (0.043) (0.049) (0.041) (0.042) Cumulative Rainy Hours /5 0.054 0.211*** 0.023 0.191*** (0.042) (0.058) (0.038) (0.049) Controls Yes Yes Yes Yes Log likelihood -209,824.46 n.a -52,032.884 n.a R-squared 0.2888 0.7131 0.2579 0.7034 Observations 493,517 493,519 120,209 120,210 Notes: *** Significant at the 1% level, ** at the 5% level, * at the 10%. Standard error clustered by vendors and day are assumed. Columns include vendors-day of the week FEs, month FEs, year FEs and hour of the day FEs. The model in Column (1) and (2) contains weighted estimates, where the weights are equal to the number of observations used to compute the target. Results from unweighted regression, where sampling variation is ignored, are shown in Column (3) and (4). 26
Appendix A1: RDP and heterogenous responses to demand shocks In the following sections I develop two models of labor supply for piece-rate workers that operate in a market with menu costs. Reference-dependent preferences model predicts that workers should respond differently to demand shock if they are above the target. Evidence of the reference-dependent preferences in labor supply decision can be obtained by interacting the dummy variable indicating whether effort exceeds the target with the rainfall variable. The results reported in Tables 6 and 17 do not provide such support. A standard model of labor supply Pizza vendors operate in a monopolistically competitive framework. The market is char- acterized by the following demand function: q = (A − p) (4) where p and q represent price and quantity of pizza and A>0. The vendors’ utility is U = p · q − c(q) (5) where p and q represent price and quantity of pizza produced and c(q) = θq represents linear effort cost. Using (1) and (2) the vendors’ maximization problem can be written as max U = (A − q) · q − θq (6) q The first order condition of (3) w.r.t. q is ∂U =A−2·q−θ =0 (7) ∂q utility maximization yields ? A−θ q = (8) 2 and a price of ? A+θ p = (9) 2 Consider a demand shock (AR >A) and suppose that the food vendors are able to change the price but only at a menu cost of z (i.e. the cost of literally printing new menus). 27
In the presence of menu cost the food vendors might find it optimal to not adjust pizza’s price to the level AR + θ p?R = (10) 2 rather, they might choose to keep the price at p? and adjust its output to18 AR − θ qR? = (11) 2 In a market with the presence of menu costs a change in aggregate demand affects quantity but not prices. A model with reference-dependent preferences Following Kőszegi and Rabin (2006), I write the vendors’ total utility V (q|q T ), as a weighted average of standard utility U (q) and gain-loss utility R(q|q T ), with weights 1 − η and η, as follows: V (q|q T ) = (1 − η)U (q) + ηR(q|q T ) (12) where the gain-loss utility R(q|q T ) = µ(q − q T ) (13) where q T is the effort target value19 ,and µ(·) is a ”universal gain-loss function” that satisfies the following properties: • µ(x) is strictly increasing • µ00 (x) ≥ 0 for x0 The vendor’s maximization problem is max V (q|q T ) = (1 − η) ((A − q)q − θq) + η µ(q − q T ) q (14) 18 I observe such behaviour whenever z > ∆Ũ where ∆Ũ (∆Ũ =UR − U˜R = pR · qR − C(qR ) − (p · q̃R − C(qR )). 19 Following Kőszegi and Rabin (2006), I model target as set equals to the agents’ rational expectation. Building on Crawford and Meng (2011), I treat them as point expectations rather than distributions E(A)−θ q T = E(q) = 2 . 28
The first order condition of (11) w.r.t. q is ∂V = (1 − η)(A − 2q − θ) + η µ0q (q − q T ) = 0 (15) ∂q 20 I can identify the relation between A and q through ∂λ ∂q 1−η = − ∂A ∂λ =− 00 (q − q T ) (16) ∂A ∂q −(1 − η)2 + η µ qq if η = 1 (i.e. the standard utility weight is equal to 0): ∂q =0 (17) ∂A if (q − q T ) < 0 (i.e. the vendor is below his target): ∂q − ∂λ ∂A 1−η = − ∂λ = − 00 (·) (18) ∂A ∂q −(1 − η)2 − η µ qq if (q − q T ) > 0 (i.e. the vendor is above his target): ∂λ ∂q + 1−η = − ∂A ∂λ =− 00 (·) (19) ∂A ∂q −(1 − η)2 + η µ qq and ∂q + ∂q − < (20) ∂A ∂A 20 λ indicates equation (12). 29
Appendix A2: Target using Forecast Data An alternative approach would have been to construct the target as the expected rainfall for the current day. The idea is that vendors could form their expectations regarding po- tential labor demand for the day using weather forecast data. Since adverse weather con- ditions positively affect demand for pizza delivery, when vendors observe adverse weather forecasted for following day, they expect demand to be higher. I have collected hourly weather forecasts for the period between February 2010 and July 2011 provided by one of Italy’s main weather forecast providers (http://www.3bmeteo.com/). Forecast data are provided the day before the day in question, and refer to the Bologna urban area. Despite the short-term nature of forecasts, they tend to perform rather poorly. During the period from February 2010 to July 2011 there were 212 days for which weather conditions were wrongly predicted (i.e. when the actual number of rainy hours in the day was different from the corresponding forecast). Figure 6 shows the days for which the weather was wrongly forecasted (dark lines). Figure 6: Days With Wrongly-Forecasted Weather. The frequency of dark lines suggests the poor accuracy of weather forecasts. Figure 7 plots the number of days per month for which the weather was wrongly forecasted, and provides additional information about the low level of forecast accuracy. On average, the weather was wrongly forecast 11 days per month, and in two cases (February 2010 and December 2010), the number of days when the weather was wrongly forecasted exceeded the number of days for which it was correctly forecasted. 30
Figure 7: Number Of Days per Month on which the Weather was Wrongly Forecast. Bearing in mind the unreliability of weather forecasts, I replicated the analysis of Section 4.1 using the following model: log(Ordersit ) = β1 + β2 (N o. of Rainy Hours during the Dayit > F or. N o. of Rainy Hours during the Dayit )+ +β3 N o. of Rainy Hours during the Dayit + Zit β + it where the dependent variable is the (log) number of orders accepted by vendors i on day t, the variable (N o. of Rainy Hours during the Dayit > F or. N o. of Rainy Hours during the Day) is a dummy indicating that the day’s rainfall hit the forecasted number of rainy hours for that day. Z is a vector that includes vendors’ day of the week fixed effects and the full set of monthly and yearly fixed effects. The results are shown in Column (1) of Table 11. The coefficient of the variable indicating the actual number of rainy hours is positive and statistically significant. This result confirms the findings presented in Section 4.1: vendors respond positively to changes in earning opportunities. Moreover, the coefficient of the variable indicating whether rainy hours in a day exceed the forecasted number of wet hours, is positive and statistically significant. Similar results are obtained when adverse weather conditions are defined in terms of milliliters of rain (Column 2), the only difference being that the target coefficient is no longer statistically significant. While the results of this section do not provide any precise general predictions, and are only meant to offer supplementary evidence, the sign and significance of the estimated coefficients provide no evidence of the presence of target behavior. 31
You can also read