Provider experience and order selection in the sharing economy

The sharing economy, enabled by digital platforms, which connect providers and consumers for peer‐to‐peer exchanges, experienced rapid growth in recent years. Although researchers attempted to explore the societal or business impact of the sharing economy market, little is known about how individual providers operate their businesses, given that providers are capacity‐constrained, self‐scheduled and unprofessional. In this study, we are interested in the relationship between experience and providers' order selection behaviours. Leveraging a rich and proprietary dataset from a large sharing economy platform—which facilitates the exchanges of home‐cooked meals in China—and employing multiple identification strategies and estimation methods, we find that the number of orders declined by a provider first increases with their experience, but later decreases. However, their sales revenue keeps increasing with experience. Our investigation further reveals that this happens because providers adjust their order selection strategies at different experience levels to achieve higher revenue in the sharing economy. Our study is among the pioneering studies to empirically understand providers' market behaviours in the sharing economy and offers important practical implications.


| INTRODUCTION
The sharing economy (also known as collaborative economy, collaborative consumption or gig economy), is a new type of peer-to-peer business enabled by digital platforms and mobile technologies (Bonina et al., 2021;Sundararajan, 2016).Typically, the sharing economy has a two-sided market that connects both providers (who provide under-utilized resources, e.g., goods and services and more broadly, skills and time) and consumers (who utilize assets offered by providers) (Gansky, 2010).The peer-to-peer exchanges or transactions between these two sides are commonly facilitated by a third-party mobile platform.Popular examples of the sharing economy include ridesharing (e.g., Uber), room-sharing (e.g., Airbnb), workspace-sharing (e.g., WeWork), meal-sharing (e.g., EatWith), resource-sharing (e.g., NeighborGoods) and peer-to-peer lending (e.g., Lending Club).The sharing economy recently enjoyed a remarkably rapid growth and attracted significant attention.For instance, there were approximately 760 million sharing economy users in China in 2019, and the average annual growth rate exceeded 30% over the next few years (AP News, 2020).Moreover, the total value of the global sharing economy is predicted to increase to about $335 billion by 2025, from only $15 billion in 2014 (Statista, 2019).
Compared with traditional e-commerce settings, the sharing economy has an important and unique feature regarding providers.Specifically, the sharing economy market has low entry barriers, and providers are mainly unprofessional individuals (Cohen & Sundararajan, 2015). 1 For instance, Uber drivers are not professional full-time drivers, and Airbnb hosts are not professional hoteliers (Li et al., 2015).Typically, these providers already have fulltime employment and tend to work on the digital platforms as a hobby, only investing their limited spare time and energy to make extra cash (Sundararajan, 2016).Although existing studies explore how the sharing economy affects the society (e.g., Brazil & Kirk, 2016;Greenwood & Wattal, 2017), or some business and economic outcomes (e.g., Burtch et al., 2018;Cramer & Krueger, 2016;Fraiberger & Sundararajan, 2015;Zervas et al., 2017), there is a striking shortage of works that examine the unique features of providers.
Given the limited spare time and energy, unprofessional providers on sharing economy platforms always face capacity constraints and are unlikely to accept and handle a large number of sales orders (Wirtz et al., 2019).This is substantially different from a traditional business, where providers have relatively "unlimited" capacity and always expect, and are able to take, more orders.Since sharing economy providers are self-scheduled, they can freely choose which consumers to serve, according to their own evaluations and judgements.Faced with the dilemma between capacity constraints and earning maximization, they usually make choices among orders from consumers.There is indeed evidence that sharing economy providers will selectively accept/decline orders (Edelman et al., 2017).For instance, Uber drivers may accept orders based on trip features, and Airbnb hosts may decline orders according to guest characteristics.
What adds complexity to this dilemma is the influence of provider experience on providers' capacity constraints, their expertise, and, subsequently, their order selection strategy-which is an important, but understudied, aspect related to the unique feature in the sharing economy market.As unprofessional individuals, sharing economy providers lack not only expertise of running their business, but also self-perception of competence when they start their business.Without knowing how to better serve consumers and cater to their needs, they are likely to attract a small number of sales orders (Kalnins & Mayer, 2004).Unlike traditional businesses with well-established routines and onthe-job training, they must explore how to operate their businesses on their own.It is imperative for sharing economy providers to engage in experiential learning, the idea of "learning-by-doing" (Anand et al., 2016), because it helps them understand their capacity limits and better operate their business under capacity constraints.Nevertheless, the processes of developing expertise and coping with capacity limits are complex and can vary with different levels of experience.The dynamics and tensions between expertise gained and capacity constraints complicate the role of provider's experience in sharing economy providers' order selection strategies and decisions.Essentially, provider experience could be an important determinant of order selection, and the eventual sales revenue in the sharing economy, which, unfortunately, remains unexamined.
Thus, in this study, we are interested in two research questions in the sharing economy context.First, we seek to understand how provider experience affects providers' order selection behaviours by exploring whether a nonlinear effect between provider experience and declined sales orders exists.Second, we attempt to understand how providers operate their business for better selection, which eventually leads to higher revenue.These important questions were not answered in prior research, due to theoretical and empirical reasons.Theoretically, although the relationship between experience and increased sales performance may already be observed in traditional business where there is relatively no capacity constraint (Perkins & Rao, 1990), experience allowing capacity-constrained, selfscheduled, unprofessional providers to selectively accept/decline orders is a new and unique phenomenon in the sharing economy.More importantly, the effect of provider experience on order selection is likely to vary at different levels of experience, given the changing dynamics of provider expertise and capacity limits.Empirically, though prior studies could analyse the sharing economy at the aggregate level, they did not have access to transaction data, and more importantly, providers' market behaviour data.
With a rare opportunity, we can examine these questions using a rich and proprietary dataset from a large sharing economy platform, which facilitates the exchange of home-cooked meals in China.Based on multiple identification strategies and estimation methods, we find evidence that providers will decline more orders during the early stage of business when they have less experience and will decline less orders once they become more experienced.
Interestingly, providers' sales revenues keep increasing with experience, even when they decline more orders.Our investigation further reveals that this is because providers adjust their order selection strategies at different experience levels to reap the maximum benefits of the sharing economy to achieve higher revenue.
Our research has a few notable contributions.First, our research is among the pioneering studies to understand how capacity-constrained, self-scheduled, unprofessional providers operate their business in the sharing economy market.Second, we uncover the role experience plays in providers' sales and market behaviours in the sharing economy.Third, our paper is unique in that it provides empirical research based on rich and granular-level data in the context of the sharing economy.Lastly, the notable findings from this research also offer important and practical implications to stakeholders.

| Peer-to-peer businesses
Our work focusing on the sharing economy is generally related to the literature on peer-to-peer (P2P) networks or businesses whose services are provided by peer individuals, but not business professionals.P2P networks are distributed networks in which users can share and consume their own resources separate from others (Hosanagar et al., 2010).P2P networks are used in many applications, but were popularized in the sharing of files (mostly information goods, such as music) (Johar et al., 2011).Relevant research studied several important aspects of P2P networks, including the impact of both positive and negative network externalities on the optimal network size (Asvanund et al., 2004), the pricing mechanism for content distribution (Lang & Vragov, 2005), the supply-side factors (i.e., file search and redistribution incentives) (Hosanagar et al., 2010), the effects of network congestion on user incentives for sharing (Johar et al., 2011), and the antecedents of users' continued-sharing behaviours (Xia et al., 2012).P2P networks have recently evolved and now play an important role in the online lending market, where individual lenders make unsecured loans to individual borrowers.P2P lending has also attracted significant academic interest (e.g., Herzenstein et al., 2011;Liu et al., 2015).For instance, Herzenstein et al. (2011) examined how identity claims constructed in narratives by borrowers influence lender decisions in P2P online lending.Using data from Prosper.com, they found that unverifiable information affects lending decisions above and beyond the influence of objective, verifiable information.Moreover, Liu et al. (2015) investigated how friendships act as pipes, prisms and herding signals in online P2P lending.They reported that friends of the borrower act as financial pipes by lending money to the borrower; the prism effect of friends' endorsements via bidding on a loan negatively affects subsequent bids by third parties; and when offline friends of a potential lender place a bid, a relational herding effect occurs as potential lenders are likely to follow their offline friends with a bid.
Despite the growing literature on P2P businesses, individuals' resource limits or capacity constraints have been largely overlooked in the literature.While an individual's resource limit might not be a major concern in some P2P services for information goods sharing, the case could be very different when a P2P service involves sharing tangible resources, such as financial resources and physical assets.

| The sharing economy
In addition to P2P networks for information goods sharing, recent P2P networks also facilitate the short-term sharing, or rental, of physical goods, such as cars and houses (Fraiberger & Sundararajan, 2015)-colloquially called the 'sharing economy' (Sundararajan, 2013).Although such markets have become increasingly popular (Hu, 2021), relevant research has been surprisingly insufficient.Apart from research on the sharing economy's structures (e.g., governance structure Bai & Velamuri, 2021) and business models (Abhishek et al., 2021), there exist a small number of studies that focus on the impact of sharing economy.Generally, these studies can be classified into two streams.
The first stream of literature studied the societal impact of the sharing economy (e.g., Brazil & Kirk, 2016;Edelman et al., 2017;Greenwood & Wattal, 2017;Huang et al., 2020).For instance, Edelman et al. (2017) conducted an experiment on Airbnb and found that applications from guests with distinctively African American names are 16% less likely to be accepted compared to identical guests with distinctively white names.This discrimination occurs among landlords of all sizes, and it is most salient among hosts who have never had an African American guest.More importantly, Greenwood and Wattal (2017) investigated how the entry of ride-sharing services influences the rate of alcohol-related motor vehicle fatalities.They found a significant drop in the rate of fatalities after the introduction of Uber X.Furthermore, the effect of the Uber Black car service is intermittent and manifests only in large cities.However, Brazil and Kirk (2016) exploited differences in the timing of the deployment of Uber in US metropolitan counties from 2005 to 2014 to examine the relationship between the availability of Uber's rideshare services and total traffic fatalities.They also found that the deployment of Uber services in a given metropolitan county had no association with subsequent traffic fatalities.Benjaafar et al. (2022) found that the introduction of ride-sharing can lead to increased traffic, even though car ownership is reduced.Differently, Huang et al. (2020) analysed the relationship between online labour supply and offline unemployment.They demonstrated a positive and significant association between local (US county) unemployment in the traditional offline labour market and the supply of online workers residing in the same county, as well as significantly larger volumes of online project bidding activity from workers in the same county.
In fact, more academic efforts have been devoted to the second stream of research that focuses on the economic and business impact of the sharing economy (e.g., Benjaafar et al., 2019;Burtch et al., 2018;Cohen & Sundararajan, 2015;Cramer & Krueger, 2016;Fraiberger & Sundararajan, 2015;Jang et al., 2021;Jiang & Tian, 2018;Li et al., 2015;Li & Srinivasan, 2019;Luo et al., 2020;Santi et al., 2014;Skiti et al., 2022;Weber, 2014;Zervas et al., 2017).From the consumers' perspectives, while Santi et al. (2014) analysed a large dataset to show that taxi sharing may lead to passenger inconvenience but generate collective benefits, Fraiberger and Sundararajan (2015) reported that peer-to-peer rental markets may change consumption mixes significantly, substituting rental for ownership and lowering used-good prices while increasing consumer surplus.From the perspective of the lender or owner of goods, some researchers were interested in the sales impact of the sharing economy.For instance, Li et al. (2015) found substantial differences in the operational and financial performances between professional and nonprofessional hosts on Airbnb due to pricing inefficiencies, which led to a 16.9% difference in daily revenue.Zervas et al. (2017) explored the impact of Airbnb's entry on hotel room revenue and reported that the causal impact on hotel revenue is in the 8%-10% range, and the impact is more salient for lowerpriced hotels and hotels that do not cater to business travellers.Similarly, Li and Srinivasan (2019) studied how the sharing economy fundamentally changes the way the industry accommodates demand fluctuations and how incumbent firms should strategically respond.They found that Airbnb's flexible supply helps to recover the underlying demand lost due to seasonal hotel pricing (i.e., higher prices during high-demand seasons), and even stimulates more demand in some cities.
Lastly, prior studies have also shown that the sharing economy may have the moral hazard problem (Weber, 2014), impede innovation (Cohen & Sundararajan, 2015), and reduce lower quality entrepreneurial activity by offering viable employment for the un-and under-employed (Burtch et al., 2018).
We summarize the details of the above literature in Table A1 in the appendix.In summary, past empirical studies on economic impact mainly focus on some aggregate level outcomes but overlook market behaviours and outcomes at the individual provider level.Moreover, extant literature has not yet explored the role of provider experience in the sharing economy context.

| Provider experience
Experience has been a fundamental research topic in many disciplines for decades.Typically, experience comes from the repeated conduction of a particular activity (Anand et al., 2016).Repetition is helpful because practice enhances efficiency through experiential learning processes (BCG, 1970).Experiential learning or 'learning-by-doing', describes how people leverage and reflect on past experiences to create new knowledge or develop new skills (Kolb, 2015).
Past research has examined applications of experiential learning in higher education and various workplace contexts (Lewis & Williams, 1994), and demonstrated that experience is a key driver of performance in a wide range of operational processes.Generally, existing literature in marketing, information systems and management has shown that experience can help business professionals better collect, evaluate and utilise information (Perkins & Rao, 1990); manage knowledge (Ko & Dennis, 2011); increase individual proficiency (Reagans et al., 2005); develop dynamic capabilities (King & Tucci, 2002) and compete more effectively in the market (King & Tucci, 2002).Ultimately, experience may lead to better earnings/business performance (Altu g & Miller, 1998;Campbell, 2013;Kalnins & Mayer, 2004;Ko & Dennis, 2011;Medoff & Abraham, 1980) or career success (Gong et al., 2018;Igbaria et al., 1994;Igbaria & Wormley, 1992).For instance, in the provider-consumer relationship, providers with more sales experience have been found to know more customer traits (i.e., have more facts memorized) and more selling strategies (i.e., more rules) to achieve better sales performance (Sujan et al., 1988).
Although the above studies have explored the role of experience in work, none of them have been conducted in the context of the sharing economy.Prior research in diverse traditional work settings mainly focus on workers' experiential learning through routine activities and operations in the workplace (Perkins & Rao, 1990) and professionals' accumulation of experience in established businesses (Anand et al., 2016).On the contrary, providers in the sharing economy context are capacity-constrained, self-scheduled and unprofessional individuals.Therefore, in such a highly flexible and unstructured working environment, the effect of experience on individual providers' selling strategies and sales performances could exhibit different patterns.Thus, while it has been observed that traditional businesses with relatively few capacity constraints experience increases in sales performances (Perkins & Rao, 1990), there is a need to revisit experiential learning in the context of the sharing economy.Specifically, as providers strategize for revenue maximization under capacity constraints, we conjecture that there might exist a nonlinear relationship between providers' experience and sales performance in the sharing economy.

| The platform
To empirically address the above questions, we obtained access to the rich and proprietary data from a large sharing economy platform in China.It is a location-based, third-party mobile platform/application that connects both providers and consumers in order to facilitate their peer-to-peer exchanges.The providers are individuals who have spare time and are willing to cook some dishes or meals in their home kitchens to share.The consumers can be any registered users besides restaurant owners, operators or cooks in the local city.The platform verifies the kitchens' identities (providers) and occupation when they first join the platform, which makes the platform a suitable research context to investigate the capacity-constrained, self-scheduled and unprofessional nature of service providers in the sharing economy.The platform started its business in Beijing in October 2014, and then expanded to six more cities, including Shanghai, Hangzhou, Wuhan, Changsha, Guangzhou and Shenzhen.It is the largest company facilitating such home-cooked, food sharing business in China and has received five rounds of funding.Figure 1 presents a list of providers/kitchens nearby.Figure 2 presents the dishes from a chosen provider/kitchen.Figure 3 presents some basic information of the provider/kitchen.
A typical exchange via this platform consists of the following steps.First, a provider (or cook) uploads dish details (e.g., name, description, image, price, etc.) to their kitchen.Second, a consumer chooses any available dishes from a kitchen to place an order.Third, the provider accepts or declines the order.If the order is accepted, the provider prepares the ordered dishes.Fourth, the consumer receives and enjoys the food.There are three ways that the F I G U R E 1 Providers/kitchens nearby consumer can receive the food: (1) the provider can deliver the food to the consumer; (2) the consumer can pick up the dish from the provider or even dine at the cook's home (with permission); (3) the food is delivered to the consumer by some third-party delivery team.
The platform provided us with rich data on its business in Beijing from October 2014 (when the business started) to December 2016, including information on all the providers, dishes, and orders.

| Competition
Regardless of whether providers' market entries are driven by hobby or economic benefits, after entry, providers will face competition from other providers in the same market.Competition will certainly affect sales performance.Thus, a key issue in our empirical analysis is to identify a provider's competitors to account for competition effects.
When a provider registers on the platform, they need to decide and configure their business radius (i.e., the distance in which the provider is available to consumers).Consequently, the radius determines the provider's market F I G U R E 2 Dishes offered by a provider/kitchen size (i.e., the region within the business' radius).As the platform is location-based, a consumer can only observe and place orders from a provider if the consumer is within the provider's business region.
Because providers enter the market at different times, the competition each provider encounters also varies across time.Thus, for a focal provider i in a time period, we check each and every other provider j in our dataset to identify whether j is a competitor to i.As we have the latitude and longitude information of each provider's kitchen, we consider that i and j are competitors if their distance (calculated based on latitudes and longitudes) is shorter than or equal to the sum of their business radiuses.In other words, if the business regions of providers i and j have an overlap, they are competitors.
Therefore, in the subsequent empirical analysis, we employ this approach to identify all the competitors for a focal provider during a time period and to construct all the competitor-related factors. 2 F I G U R E 3 Provider/kitchen information 4 | MODEL AND ANALYSIS

| Empirical model
Based on the above discussion on our motivations, we first seek to examine how provider experience affects their declined sales order by focusing on the potential nonlinear (quadratic) relationship.We conduct our analysis at the provider-week level to construct all our model variables.Let subscript i denote each individual provider in our dataset, and subscript t denote each week.To investigate the impact of provider experience on sales order, our dependent variable, declined sales orders (ODR_DC it ), indicates the number of sales orders provider i rejected in week t.Our independent variable, provider experience EXP it , is measured by the number of active days provider i worked on the platform up to week t. 3 Lastly, we have a comprehensive set of factors at various levels as our control variables, including: (1) number of dishes offered by provider i in week t (DSH_CNT it ), (2) average price of dishes offered by provider i in week t (DSH_PRC it ), (3) value of redeemed coupons offered by provider i in week t (CPN it ), (4) number of consumer reviews provider i received up to week t (RVW_CNT it ), ( 5) average number of stars (from 0 to 5) of consumer reviews (i.e., review valence) provider i has received up to week t (RVW_STR it ), ( 6) provider i's business radius (in kilometres) (RDS i ), ( 7) gender of provider i (MAL i , 1: male, 0: female), ( 8) age of provider i (AGE i ), ( 9) number of provider i's competitors in week t (CP_CNT it ), (10) average number of dishes from provider i's competitors in week t (CP_DSH_CNT it ), ( 11) average dish price from provider i's competitors in week t (CP_DSH_PRC it ), ( 12) average value of redeemed coupons offered by provider i's competitors in week t (CP_CPN it ), ( 13) average number of consumer reviews provider i's competitors have received up to week t (CP_RVW_CNT it ), ( 14) average number of stars of consumer reviews provider i's competitors received up to week t (CP_RVW_STR it ), (15) accepted sales order (ODR_AC it ), ( 16) a set of time dummies at the weekly level (T t ).
Based on our motivation, we construct and include the squared term of EXP to model the quadratic effect of EXP on ODR_DC.The panel-level model is specified in Equation ( 1): where βs and ω are the model coefficients, α i captures unobserved provider-specific and kitchen-specific effects, and ε it indicates the residual random error term.The coefficients β 1 and β 2 captures the quadratic effect of provider experience on declined sales order.We take the logarithmic form on our dependent variable (Hendricks & Sorensen, 2009;Lin et al., 2017) to account for the right-skewness of the dependent variable.Tables 1 and 2 present the descriptive statistics and the correlation matrix of our model variables.

| Result
We first estimated a fixed effects (FE) model of declined sales order (ODR_DC) on all the control variables.As reported in Table 3, Column (1), various control variables have significant relationships with ODR_DC, implying that our control variables can help explain the variation in ODR_DC.We also estimate a random effects (RE) model of ODR_DC on these control variables and generally find similar results in Table 3, Column (2).
Beyond these control variables, we then estimated a full FE model by further including the independent variable of provider experience (EXP Â EXP, EXP), measured by the number of a provider's active days on the platform.We summarize the results in Table 3, Column (3).As indicated, the significant coefficients of EXP Â EXP (À0.009) and EXP (0.064) shows that EXP has an inverted U-shaped relationship with ODR_DC, and the extreme point of the inverted U shape is at EXP = 3.42.Figure 4 shows the predictive margins of declined sales order at different values of EXP.This implies that as experience increases, a provider may first decline more orders.However, when the provider gains more experience (i.e., active for at least 342 days ≈ 49 weeks on the platform), they may be less likely to decline orders.We then follow Lind and Mehlum (2010) and Haans et al. (2016) to formally and rigorously test for the inverted U-shaped relationship.After conducting the three-step testing procedure of the U-shaped relationship, we obtain a p-value of 0.038.Hence, we reject the null hypothesis on the monotone or U-shape relationship and then conclude the inverted U-shape relationship.Therefore, we can fully observe the inverted U-shaped relationship within the 114-week range of our current dataset.Specifically, the number of orders declined by the provider first increases as the provider's experience grows and will start to decrease once they become more experienced.
T A B L E 1 Descriptive statistics In addition to the full FE model, we further estimate a full RE model of ODR_DC on all the independent and control variables and summarize the results in Table 3, Column (4).Consistently, the estimated coefficients of EXP Â EXP (À0.009) and EXP (0.061) are also statistically significant.The Hausman test result (χ 2 = 336.20,p = 0.000) shows that FE should be chosen over RE.Thus, we consider the results of the full FE model in Column (3) as our baseline results.We also employ the FE model estimation for subsequent analysis.

| Identification
The above analysis might be subject to a potential endogeneity issue as provider experience could be endogenous due to the omission of relevant factors.For instance, we were not able to taste the dishes offered by each provider.
How a dish tastes is likely to be correlated with providers' order selection behaviours, as well as experience, which might lead to endogeneity.Next, we address the potential endogeneity concern.
We use the instrumental variable (IV) approach.The chosen IV is the number of weeks since provider i's registration on the platform until week t (WKS_REG it ).A larger (smaller) WKS_REG may indicate a longer (shorter) period a provider has been with the platform, which is related to the maximum number of active days that the provider could use to serve customers on the platform.Nevertheless, as providers could only experientially learn and develop their order selection strategies by actually serving customers, the time at which the provider joins the platform should have no direct relationship with providers' order declining strategies.By using the IV and its quadratic terms (Wooldridge, 2010), we perform the two-stage least squares (2SLS) and summarize the results in Table 4, Column (2).For ease of reference, Table 4, Column (1), presents the baseline results from Table 3, Column (3), whereas Table 4, Columns (3) and (4) display the results of IV estimation in the first-stage equation.For brevity, from this point onward, we only present estimates of our focal variables but include all the control variables during estimations.We obtain largely consistent results as our baseline results.In addition, to eliminate the concern that our independent variable and IV are measured in different units, we re-perform 2SLS by using an alternative operationalization of our independent variable EXP_WK it , which is measured by the number of active weeks provider i worked on the platform up to week t.We obtain significant and consistent results, which are summarized in Table 4, Column (5).
In addition to the endogeneity issues, another possible concern is the unobserved heterogeneity of providers.To address this, we apply a random coefficient model (Boudreau & Jeppesen, 2015).This model can Predictive margins of ODR_DC capture the possibility that order selection behaviours vary due to any latent, unobserved provider-specific heterogeneity, for example, each provider may have an idiosyncratic, unique cooking style and skill.We summarize the estimates in Table 4, Column (6).As indicated, the estimates of EXP Â EXP and EXP remain significant and consistent.
Therefore, after accounting for the potential endogeneity and heterogeneity issues based on the above various strategies, we consistently find that the provider experience exhibits a clear inverted U-shaped relationship with the declined orders.This suggests that as the provider gains more experience, they are likely to first decline more orders and later decline less.

| Robustness
We further corroborate our findings by checking the robustness and consistency in multiple ways.For ease of reference, Table 5, Column (1); Table 6, Column (1) and Table 7, Column (1) present the baseline results from Table 3, Column (3).
First, a possible concern is the potential collinearity among our model variables.The Variance Inflation Factor (VIF) of EXP Â EXP and EXP are 10.15 and 12.12, respectively, and are slightly larger than the rule of thumb of 10.
We perform mean-subtracted centralization to the focal explanatory variables.We re-estimate Equation (1) based on these centralized variables and summarize the results in

T A B L E 5 Robustness (1)
Variable Second, one may be concerned about potential serial correlation.Thus, we estimate a model with a first-order autoregressive (AR1) disturbance structure.As indicated in Table 5, Column (3), the estimate of EXP Â EXP under an AR1 structure remains similar.This implies that our results are robust to serial correlation.
Third, we check our findings across differences in variable measures.In addition to the number of weeks on the platform as the measure for EXP, we use the number of past consumers a provider served as a new measure (PST_CSM it ).We re-estimate our model using the new measure and summarize the results in Table 5, Column (4).As indicated, the focal estimates remain consistent.
Fourth, we check the robustness of our findings across different model specifications.We first estimate a population-averaged (PA) model that allows for an exchangeable correlation structure of a generalized linear model and then a random effects model estimated via maximum likelihood (RE-ML).The corresponding results for the PA and RE-ML models are shown in Table 5, Columns ( 5) and ( 6) respectively.Moreover, we construct an alternative independent variable on order selection IS_ODR_DC it , which is a binary indicator that takes one if there are declined sales orders and takes zero otherwise.We then estimate a panel logit model to investigate providers' likelihood of declining orders and report the estimation results in Table 5, Column (7).Furthermore, since the number of declined sales orders (ODR_DC) is a count variable that can only take non-negative integers, we use ODR_DC as the dependent variable and estimate the model using a panel Poisson regression.To further account for the excess zero counts in ODR_DC, we estimate a zero-inflated Poisson (ZIP) model that presents the data generating process for count values and zeros separately by including a logit inflation model to predict excess zeros.Table 5, Columns ( 8) and ( 9) summarize the results for the panel Poisson and ZIP respectively.In general, the model parameter estimates across all these models remain consistent with that in Column (1).
Fifth, one may still be concerned that the inverted U-shaped relationship, which indicates a provider's order selection behaviour, might only occur when the provider receives a large number of orders, which are beyond capacity constraint.Consequently, the findings might not be applied to providers with a small number of orders.To address this, we re-estimate our model based on various sub-samples with different levels of orders (i.e., the number of orders is smaller than 10, 20, 30, 40, 50 and 60, respectively).As indicated in Table 6, Columns (2) through ( 7), the estimates remain largely consistent.Moreover, the mean of sales orders accepted (ODR_AC it ) in the full sample is 36.382.Our estimates, based on various sub-samples, indicate that our findings will hold in different order levels both below and above the average level.This also suggests that, even when the amount of orders is still within capacity, an experienced provider may selectively decline orders.
T A B L E 6 Robustness (2) Variable (1) Baseline (2) Order <10 (3) Order <20 (4) Order <30 ( Sixth, there could exist an alternative explanation for the inverted relationship between EXP and ODR_DC. That is, a reputable and popular provider (as reflected by RVW_CNT) may attract more consumers and, thus, have the opportunity to decline more orders.Therefore, when RVW_CNT increases, ODR_DC may increase.
However, a decreasing marginal return of reputation can occur once a provider has attracted all the consumers they can.In other words, the inverted U-shaped pattern of ODR_DC might be driven by RVW_CNT but not EXP.
To dismiss this alternative explanation, we construct and include the squared term of RVW_CNT for estimation.
As shown in Table 7, Column (2), RVW_CNT does not have any quadratic effect on ODR_DC, which rules out the alternative explanation.
Lastly, we include additional measures on competition to account for cases in which providers in the business radius might experience different levels of competition based on the similarity of their offered products.We first construct a series of variables of product similarity (i.e., CP_SIM_NAM it , CP_SIM_WAY it , CP_SIM_MEA it , CP_SIM_VEG it ) to indicate the similarity in dishes between provider i and their competitors in week t. 4 We find consistent results as our baseline outcome.
In conclusion, we are confident that all the various checks confirm the robustness of our findings regarding the inverted U-shaped relationship between provider experience and declined sales order.
T A B L E 7 Robustness (3)  7 has been corrected in this current version.On column 3, row 4, "0.007" should be "(0.023)".On column 3, row 5, "(0.007)" should be "0.007".On column 3, row 6, "(0.009)" should be "(0.007)".Lastly on column 1, row 5, "RVW_CNT * RVW_CNT" should be "RVW_CNT Â RVW_CNT".] The above results show robust evidence that provider experience has a quadratic effect on the number of declined sales orders.This suggests that a decline in sales orders will first increase with experience, but then decrease after the provider has reached a certain experience level.It is interesting and surprising that the effect of provider experience on order declining behaviour varies at different levels of experience.Specifically, experience increases (decreases) the number of orders declined when providers are inexperienced (experienced).Hence, we explore the underlying mechanism.
As discussed above, we postulate that providers with more experience have a better understanding of their own competency levels and capacity constraints, as well as more expertise to better cater to consumers' needs.In the early stage, unprofessional providers have limited ability to handle more sales orders (i.e., a lower bound of capacity limit).Once they gain experience and begin to learn more about their competency and capacity constraints, they become better able to judge the value of sales orders so they can more efficiently self-schedule given their capacity constraints.Therefore, when they do not have a lot of experience, they tend to decline more low-value orders.As they gain more experience, they only keep the orders with high values.This is consistent with the information shown in Figure 5, which depicts a model-free plot of the average number of declined orders, accepted orders, product variety, product quality and product promotion 5 over time.Specifically, we can observe a slight increase in OrderValueAvg in the early stage.In the later stage, however, providers become more "professional" and capable of processing more orders (i.e., increased capacity limit) even with the same level of energy.As shown in Figure 5, we can observe that providers offer a wider variety of products (ProductVarietyAvg) without compromising the quality (ProductQualityAvg) over time.Therefore, they can process and accept more orders (decline fewer orders) when they gain more experience.Consequently, although experience could increase providers' sales performances, in general, the effect of provider experience on their order selection behaviours varies at different experience levels.
First, we examine how a provider's experience affects their sales revenue and order value.We construct a DV, REV it , to indicate the total sales revenue of provider i in week t.We estimate the impact of EXP on REV and find a positive and significant estimate in Table 8, Column (1).This suggests that a provider will earn more revenue when their experience increases.We further estimate whether there exists a quadratic effect by including EXP Â EXP.
While the estimate of the squared term in Column (2) is significant, the further U-shape test (Lind & Mehlum, 2010) indicates that the extreme point is far outside of the interval, suggesting a monotone positive relationship between provider experience and sales revenue.This implies that a provider's revenue increases almost linearly with their experience, even when they decline more orders.To test the underlying mechanism of this increased revenue with fewer orders, we construct a DV, OV it , to indicate the average order value (i.e., REV divided by ODR_AC) of provider i in week t.We estimate the impact of EXP on OV and find a positive and significant estimate in Column (3).We include EXP Â EXP to check whether a quadratic effect exists by including the squared term and summarizing the results in Column (4).The U-shape test (Lind & Mehlum, 2010) shows an insignificant inverted U-shaped relationship (p = 0.138) between provider experience and average order value.Therefore, during our observation period, the average order value increases when the provider is less experienced and highly capacity-constrained, but may start to decrease when they become more experienced and their capacity limit increases.This is consistent with our baseline results, as well as our conjecture above that providers' order selection strategies vary at different levels of experience.Specifically, providers tend to select orders with higher value when the influence of capacity constraint dominates at their inexperienced stage, whereas they will be less 'picky' on high-value orders when they are more experienced and have gained sufficient expertise to efficiently run their business.Nevertheless, one might argue that the increase of average order value is not due to providers' order selection behaviours, but their receipts of higher value orders as they become more experienced.To tease out this alternative explanation, we further constructed and included OV_ALL it , the average order value of all accepted and declined orders as a control variable in the estimation.As shown in Table 8, Column (5), we obtain significant and largely similar estimates for EXP and its quadratic term as those in Column (4).This indicates an insignificant inverted U-shaped relationship (p = 0.330) between provider experience and average order value.Therefore, despite the value of all sales orders increasing, in general, we are still able to observe providers' order selections in regards to their experience levels.
Furthermore, we are interested in analysing a provider's business strategy (i.e., how they operate in the market) across different experience levels.In other words, we attempt to understand an experienced provider's strategies to achieve higher revenue.We test the impact of EXP on various strategy-related factors, including (1) product quality, which is proxied by review valence, a widely recognized indicator of product quality (Forman et al., 2008) (i.e., RVW_STR), (2) product promotion (i.e., redeemed coupon value, CPN) and (3) product variety (i.e., number of dishes, DHS_CNT).To account for the endogeneity and simultaneity between EXP and these factors, we estimate a system of equations using three-stage-least squares and seemingly unrelated regressions separately and report the results in Table 9 Columns (1) to (4) and Columns ( 5) to ( 8) respectively.We obtain largely consistent estimates on F I G U R E 5 Model-free plot of average declined sales orders, accepted sales orders, order value, product variety, product quality and product promotion over time  the effect of provider experience on declined sales orders in Columns ( 1) and ( 5).Moreover, the estimation results in Columns (2) to (4) (( 6) to ( 8)) reveal some interesting relationships between provider experience and the three strategy-related factors.First, the estimate in Table 9, Columns (2) and ( 6) show that when experience increases, a provider will improve product quality.In addition, the estimated impacts of EXP on product promotion in Columns (3) and ( 7) suggest that a provider will offer a lower promotion as they gain more experience.Furthermore, the results in Columns ( 4) and ( 8) show that a provider tends to maintain a larger set of product choices for consumers when they gain more experience.
Lastly, we demonstrate that these business strategies will indeed lead to more sales orders (i.e., all orders placed, including orders accepted and declined).In addition, providers will become better equipped at handling more orders and improve their order selection strategy as their experience level increases.We construct three DVs: (1) ODR_AC it (the sum of accepted orders), (2) ODR_ALL it (the total number of all orders received by provider i in week t, that is, the sum of accepted orders (ODR_AC) and declined orders (ODR_DC)), and (3) ODR_AC_R it (the ratio of accepted orders to total orders).for a provider to choose from and enable the provider to accept and handle more orders.

| DISCUSSION AND CONTRIBUTION
Analysing a rich and proprietary dataset from a large sharing economy platform, which facilitates the exchanges of home-cooked meals in China, we discover several notable findings.First, in the sharing economy market, a provider will deal with sales orders based on their experience.In the early stage, when a provider is inexperienced, they tend to decline more sales orders as experience increases.However, in the later stage when they gain experience, they are less likely to decline orders and are capable of handling more accepted orders for better revenue.Thus, providers adjust their order selection strategies at different experience levels to reap the maximum benefits of the sharing economy.Second, an experienced provider will adopt strategies of improving product quality, offering more promotions, and increasing product variety.Through these strategies, the provider could attract more placed orders from consumers and have the ability to handle more orders.Consequently, they have better judgement on which orders achieve higher revenue as they gain more experience.
Our research findings have the following contributions.First, our research is among the pioneering studies to understand providers' market behaviours (order selection and business strategies) in the sharing economy.Existing studies explore how the sharing economy affects the society (e.g., Brazil & Kirk, 2016;Greenwood & Wattal, 2017) and some business and economic outcomes (e.g., Burtch et al., 2018;Cramer & Krueger, 2016;Fraiberger & Sundararajan, 2015;Zervas et al., 2017).However, there is still a striking shortage of research that focuses on providers.Thus, we contribute to the literature by empirically examining individual providers' market behaviours to offer relevant insights.
Second, our study serves as an innovative work to discuss and explore the unique features of providers (i.e., capacity-constrained, self-scheduled, unprofessional individuals) and the related unique role of provider's experience in the sharing economy.We add to the experiential learning literature (Anand et al., 2016) by documenting the important mechanisms of why and how providers can adjust their order selection strategy to earn more in the sharing economy.These findings also shed light on the difference between the sharing economy and traditional business, which helps people understand the nature of the sharing economy.
Third, our paper is unique in that it provides empirical research based on rich data in the sharing economy settings.The sharing economy has become extremely popular in recent years.However, there is still a lack of relevant granular-level data for academic research.Existing studies (mostly still ongoing) are either mainly commentary (e.g., Cohen & Sundararajan, 2015;Eckhardt & Bardhi, 2015;Sundararajan, 2013;Sundararajan, 2014) or analytical (e.g., Fraiberger & Sundararajan, 2015;Weber, 2014) or exploit limited data about the entry decision of a sharing economy platform (e.g., Uber and Airbnb) combined with other sources of data at the aggregate level (e.g., Brazil & Kirk, 2016;Burtch et al., 2018;Greenwood & Wattal, 2017).In contrast, our research has obtained detailed information about providers' transactions and market behaviours directly from a large sharing economy platform.Thus, we contribute to the literature on the sharing economy by providing more detailed empirical investigations and insights on a much more granular level.
On a practical level, we also provide guidance to platform operators in the sharing economy.Our research shows that inexperienced providers sometimes decline less valuable orders because they are still incapable of handling more volume at the same level of energy or effort given the expertise that is gained under capacity constraints.This might hurt consumers' platform experience, discourage the low-consuming population, or impede the platform growth.
Therefore, platform operators could consider enhancing providers' proficiency and capacity.For instance, platform operators can arrange some training activities to inexperienced providers to improve their efficiency in taking more orders when they first join the platform.Moreover, platform operators could offer incentives to or implement regulations on these inexperienced "amateurs" to reduce the number of orders that are declined.Moreover, our findings provide advice to consumers.In traditional e-commerce settings, consumers always follow word-of-mouth and consider star providers the "safer" choices.We suggest that such information might be inadequate for making informed purchase decisions in the sharing economy context where their orders could be declined.For instance, consumers who have urgent requests for orders should consider both their order value and provider experience.If their order value is not very high, they should seek experienced providers who are more capable of serving consumers and running their business in order to mitigate the possibility of their orders being declined.On the other hand, consumers with large orders can seek more variety by transacting with less-experienced providers who are more responsive in attending to high-value orders.
Our paper has several limitations, which may shed light on possible future research.First, while providers' selfmanaged capacities are reflected in many aspects, our paper investigates only one capacity construct, that is, providers' order selections.Due to the limitations of our data and context, we are unable to examine other aspects, such as providers' decisions on business radius (which remains largely invariant in our data context).Thus, future research could replicate our research in other sharing economy contexts (e.g., car-sharing), where providers tend to frequently change their service/business radius.Second, we only focus on one specific sector of the sharing economy, that is, meal-sharing platforms, on which providers can serve more than one customer simultaneously.Thus, our findings on providers' order selection strategies might not be directly generalized to other sharing economy platforms where one provider can only serve one customer at one time (e.g., accommodation and car-sharing) and may need to be revisited in future research.
Number of observations = 135 855, number of providers = 8358, number of weeks = 114.All variables are at the provider-week level.Variables used in the identification, robustness and mechanism checks are also reported.The provider will receive orders as long as they are active on the platform.During estimation, variables EXP, CPN, CP_CNT, CP_CPN and CP_RVW_CNT are scaled by dividing by 100; variables RVW_CNT and PST_CSM are scaled by dividing by 1000.
T A B L E 8 Mechanism: Order accepted, revenue, order value Table 10's Column (1), (5) and Table 11's Column (1) present the estimated impacts of these strategy-related factors on ODR_ALL, ODR_AC and ODR_AC_R, respectively.As expected, higher product quality (RVW_STR), more product promotion (CPN) and larger product variety (DSH_CNT) will attract more orders (ODR_ALL)