Workers in the Crowd: The Labour Market Impact of the Online Platform Economy

In this paper, we compare wages and labor market conditions of individuals engaged in online platform work and in traditional occupations by exploiting individual-level survey data on crowdworkers belonging to the largest micro-task marketplaces, focusing on evidence from the United States and Europe. To match similar individuals, survey responses of crowdworkers from the US and EU have been harmonised with the American Working Conditions Survey (AWCS) and the European Working Conditions Survey (EWCS). Our findings indicate that traditional workers retain a significant premium in their earnings with respect to online platform workers, and that those differences are not affected by the observed and unobserved ability of individuals. This holds true also taking into account similar levels of routine intensity and abstractness in their jobs, as well as the time spent working. Moreover, labour force in crowdworking arrangements appears to suffer from high levels of under-utilisation, with crowdworkers being more likely to be left wanting for more work than comparable individuals.


Introduction
Among the "mega-trends" which characterise the future of work, the growth of the online platform economy has been steady and fast in the recent years and has been contributing to the changing nature of work (OECD, 2016, Harris & Krueger, 2015. 1 Technological progress and digitalisation are at the basis of its current development. Due to the overall exponential growth of internet facilities, indeed, recent years have shown an increasing number of workers participating in what is described as the gig, on-demand, or platform-based economy (Degryse, 2016, Prassl & Risak, 2015. These workers are usually called crowdworkers, where crowdwork is defined as an "employment form that uses an online platform to enable organisations or individuals to access an indefinite and unknown group of other organisations or individuals to solve specific problems or to provide specific services or products in exchange for payment" (Eurofound, 2015).
The economic conditions of crowdworkers have been analysed in a number of recent descriptive studies (e.g. Berg et al., 2018, Berg, 2015, Difallah et al., 2018, Hara et al., 2018, Pesole et al. (2018), De Groen et al., 2017 showing how these workers suffer from the erosion of fundamental labour rights, the loss of social protections and difficulties in exercising collective action. It would be a mistake, however, to assume that platform work has a causal effect on working conditions solely based on the evidence of these descriptive studies, as it could be argued that the characteristics of crowdwork are intrinsically different from traditional salar- Given the likelihood that the online platform economy will further expand in the coming years, it is crucial for governments and social partners to take an active role in designing labour market institutions (e.g. minimum wages, employment protection, health and safety regulations) that can ensure labour and social rights for this type of workers. This is especially urgent for platform workers involved in the so-called micro-tasks (a series of small tasks which together comprise a large unified project and can be performed independently over the Internet in a short period of time), which are more exposed to risks concerning low pay, precariousness 1 According to the OECD (2016), the online platform economy is the economic activity which enables transactions -partly or fully online -of goods, services and information.
2 During the UN General Assembly in September 2015, the four pillars of the Decent Work Agendaemployment creation, social protection, rights at work, and social dialogue -became part of the new UN 2030 Agenda for Sustainable Development (United Nations, 2015, Transforming our world: the 2030 Agenda for Sustainable Development). At the same time, the European Parliament resolution of 19 January 2017 recognised the need to set a European Pillar of Social Rights also for 'atypical or non-standard forms of employment, such as temporary work, involuntary part-time work, casual work, seasonal work, on-demand work, dependent self-employment or work intermediated by digital platforms' (European Parliament, 2017, European Parliament resolution of 19 January 2017 on a European Pillar of Social Rights).
2 and poor working conditions (Prpić et al., 2015). 3 In light of these critical issues, in this paper we analyse a large fraction of the available evidence on earning and working conditions of crowdworkers involved in micro-tasks. Our focus is on the evidence from the United States and Europe and our main goal is to answer to the following questions: Are individuals involved in micro-task crowdsourcing intrinsically different from traditional salaried workers involved in comparable occupations, and are there difference between micro-task crowdworkers from the US and from Europe? Is it possible to estimate the real impact of micro-task crowdwork on wages and working conditions of platform workers? Is incidence of labour market slack in crowdwork higher than that in traditional forms of salaried employment? We focus on the supply side of these labour markets and intend to measure how much individual characteristics influences those differences in pay.
Our contribution is based on an empirical analysis of cross-sectional data collected from three different surveys and harmonised in order to obtain the greatest degree of comparability.
We adopt a quasi-experimental approach: the aim is to provide an unbiased comparison of Our findings indicate that crowdworkers earn about 70% less than traditional workers with comparable ability, while working only a few hours less per week. Also, platform workers appear to be uninterested in looking for other forms of occupation, while still expressing the desire to work more than what they currently do. These results suggest that most crowdworkers are similar to a form of idle workforce, which is excluded from traditional employment and is still under-utilised.
To the best of our knowledge, this is one of the first attempts to provide an unbiased comparison of platform and traditional workers in terms of earnings and working conditions in a quasi-experimental design. Moreover, contrarily to most other studies on the online platform economy, which aim their attention at US crowdworkers, we intend to focus on both United States and Europe.
The rest of the paper is organised as follows. Section 2 outlines the online micro-task labour market, Section 3 is dedicated to a review of the literature, Section 4 describes the data used for our empirical analysis, Section 5 outlines our empirical specification and Sections 6 and 7 show our results and robustness checks. Finally, in Section 8 we discuss our conclusions. The Appendix is dedicated to additional descriptive statistics and regressions.
2 The online micro-task labour market Phenomena such as crowdwork do not exist in a vacuum, but are fostered and facilitated by wider socio-economic trends, and the development of "virtual work" can surely be identified as one of these. The term virtual work has been used by many authors to describe all of the various forms of work characterised by the execution of work through the Internet, computers, or other IT-based tools. However, not all digital jobs are necessarily a novelty per se, and not all new jobs are digital. While new forms of employment have surfaced, pre-existing ones have acquired a new role and relevance, thanks to the influence of new technologies. 4 Crowd employment is one of these new forms of work and transcends traditional arrangements by de facto requiring a tripartite relationship in which an intermediary agent -the platform -manages workers -or, rather, service providers -not only by matching them with clients but also controlling pay levels, providing ratings and generally exercising many other functions that affect workers directly. Within the platform, through an open call, client companies can offer online tasks, which are performed by contractors in exchange for remuneration (see, e.g., Eurofound, 2015). Because the majority of online platforms explicitly deny the existence of any employment relationship between the parties, individuals in crowdwork are generally characterised as independent contractors, performing their work in a discontinuous or intermittent basis.
Crowd employment can then be identified as a phenomenon that essentially entails a new, and substantially cheaper, way of outsourcing tasks to a large pool of workers through IT-based platforms (Prassl & Risak, 2015) and, because of this, it has also been defined as "crowdsourcing". 5 By requiring platforms as intermediate actors, crowdwork manages to virtually nullify most transaction costs, thus allowing for a flexible and 'extremely scalable' workforce (De Stefano, 2015) to enter the labour market and maximise the use of under-utilised assets such as human capital. 6 Crowdwork arrangements may vary greatly: skill requirements for outsourced jobs may 4 Eurofound (2015) has identified nine distinct new forms of employment: employee sharing, job sharing, interim management, casual work, ICT-based mobile work, voucher-based work, portfolio work, crowd employment and collaborative employment. 5 This term which was first used by Jeff Howe in his article "The Rise of Crowdsourcing", Wired Magazine, 14, 2006. 6 The ability to provide services online significantly enlarges the scope of crowdwork markets, thus enabling services to be provided globally, as opposed to the local focus of the services offered by work-on-demand platforms (such as Uber, Foodora, or Taskrabbit), which are characterised by the physical and tangible nature of the tasks being offered. range from high to low and, while tasks with high routine intensity and low abstract content are prevalent -as, for example, most tasks in Amazon Mechanical Turk (AMT), Clickworker and Figure-Eight -complex and even creative activities are also present. Amazon Mechanical Turk easily stands as a prime example of a crowdwork platform, being widely recognised as one of the most popular ones (see Harris & Krueger, 2015). The short and repetitive tasks offered in the platform often include: image/video processing, translation, data verification, information gathering and processing, audio and visual editing, amongst many others. 7

Literature review
Tackling the issues related to micro-task crowdsourcing has proven to be a multifaceted effort which, so far, has seen the intervention of different disciplines such as law 8 , information technology and economics. Until recently, the body of research on the economics of crowdsourcing has been, so far, remarkably thin, compared to other areas of study: a glaring lacuna, considering the growing size of the platform economy. As suggested by Hara et al. (2018), this scarcity of literature is mostly attributable to the absence of publicly available data on crowdwork platforms and their workers, in addition to a variety of methodological issues concerning the type of data to analyse and the empirical approach to be used. Nonetheless, as discussed by Horton et al. (2011), Paolacci et al. (2010 and Berinsky et al. (2012), crowdwork platforms potentially present themselves as an ideal environment for empirical studies, in particular those based on experimental research. In this regard, Horton & Chilton (2010) offer one the first attempts to obtain empirical evidence on reservation wages in crowd employment from an experimental framework. Other contributions focus instead on the demand side of these markets, concentrating on task pricing and worker productivity optimisation (e.g. Mason & Watts, 2009, Singer & Mittal, 2013. Several additional descriptive studies have been provided. Harris & Krueger (2015) document the development of the platform economy and call for the recognition of an independent worker status, while other studies, receiving support from international institutions such as ILO (Berg, 2015 andBerg et al., 2018) and FEPS (Huws et al., 2017), have contributed to the literature with a thorough overview of the demographics of crowdsourcing. Hara et al. (2018) document wage and working time amongst AMT crowdworkers, discussing the necessity of including the time spent searching for tasks in working time indicators, while a recent paper from Difallah et al. (2018) summarises the main take-aways from a longitudinal survey on AMT whose data has been published in the mturk tracker website, curated by Ipeirotis (2010). 9 7 As described in AMT website: https://www.mturk.com/ (last accessed: 19th September 2018). 8 See, for example, Prassl & Risak, 2015, De Stefano, 2015 The survey contains data on country, gender, age, income from AMT, time spent on AMT, marital 5 Another important contribution on the analysis of the platform economy in US comes from Katz & Krueger (2018), where the two economists, in the context of studying the evolution of all alternative work arrangements from 2005 to 2015, estimate that, out of all occupations, 0.5% involve the direct selling of activities and services mediated by an online intermediarya figure that can proxy the size of the so called gig-economy (see Harris & Krueger, 2015).
Crowdwork can be considered as another form of service outsourcing, and as such other contributions should be taken into consideration. There is an ample body of literature on service outsourcing and its labour market effects, mostly dedicated to analysing whether aggregate labour demand is affected by complementary or substitution effects. Amiti et al. (2005) and Amiti & Wei (2009) offer evidence on the impact of service offshoring in the UK and US, predicting no significant effects on aggregate employment. In contrast, Görg & Hanley (2005) find negative employment effects for both material and service outsourcing. Other scholars -such as Degryse (2016) -suggest that crowd employment could be equated to a form of digital migration and, in this regard, Ottaviano et al. (2013) offer a valuable study of the labour market effects of migration and task offshoring. Proxying substitutability through routine intensity of tasks -a concept originally introduced by Autor & Dorn (2013) which spurred a novel body of literature focusing on the task-based approach to labour markets - Ottaviano et al. (2013) find that service outsourcing, while having no effect on employment, has changed the task composition of native workers.
A few recent works, however, have focused on a number of supply and demand factors which contribute to the deterioration of earnings in online labour markets. Dube et al. (2018) address monopsony in online labour markets, finding that their peculiar structure allows platforms to impose a considerable markup on workers' productivity, leading up to a 20% contraption in their earnings. Looking at the supply of online workforce, the relationship between unemployment and micro-task labour markets was further explored in Borchert et al. (2018), where labour demand shocks have been found to affect temporary participation in online labour markets. Negative spill-over effects from crowdwork markets may be less obvious, but cannot be excluded. Focusing on on-demand labour platforms, Berger et al. (2018) explore the effect of introduction of Uber across taxi drivers, finding a negative association with their hourly earnings. Finally, the effects of digital labour markets on high skilled service flows are investigated in Horton et al. (2017), who focus on the UpWork freelancing platform.
While these studies all improve our understanding of important factors contributing to wage deterioration of online platform workers, none of these contributions focuses on the issue of self-selection into crowdsourcing, leaving the effect of individual ability unmeasured.
We believe that a complete picture on working conditions in online crowdsourcing can only be achieved by measuring how much individual characteristics of online platform workers status, household income and household size of Mechanical Turk workers, and can be accessed at the address: http://demographics.mturk-tracker.com/ 6 contribute to these conditions, a task which we intend to pursue with this paper.

Data
Finding appropriate sources of information on micro-task crowdworkers has proved to be a rather demanding task. The first difficulty has been the identification of crowdworkers in existing large-scale survey data on workers and working conditions. The European Working Conditions Survey (EWCS) (Parent-Thirion et al., 2016) and the American Working Conditions Survey (AWCS) (Maestas et al., 2017) both contain comparable data on wages, job quality and skills but, in these cases, it is arduous to disentangle platform workers from any freelancer working from home. As micro-task crowdsourcers tend to perform specific, routine intensive activities, we expect that equating them to any freelancer working from home will likely pose a serious source of bias. While growing, the size of the platform economy is still minor, so platform workers will naturally be under-represented in general surveys.
Dedicated surveys on crowdworkers have been very useful in this regard. However, while there is currently plenty of information on work on digital platforms -acquired either through online questionnaires (e.g. Berg, 2015, Berg et al., 2018, Huws et al., 2017, Ipeirotis, 2010 or web plug-ins (e.g. Hara et al., 2018) -the methodologies behind the collection of this data often differ significantly, with the resulting surveys varying not only in their sample sizes but also in terms of item comparability.
With the aim to provide a reliable empirical analysis of the effects of crowdwork on labour market conditions in United States and in Europe, our initial efforts have focused on ascertaining which datasets would have allowed us to maximise the comparability of our results while retaining a satisfactory pool of observations and key variables.

Treatment and control groups
Our treatment sample uses information on European and US crowdworkers from the two rounds of the ILO Survey on Crowdworkers (Berg, 2015 andBerg et al., 2018). Thanks to the similarities in terms of the relevant variables of analysis, our control groups were constructed using data from the American Working Conditions Survey and from the European Working Conditions Survey. We harmonise the ILO Survey on Crowdworkers with these general working conditions surveys in our attempt to put these new forms of work into a comparative and global perspective.
The dataset from Berg (2015) and Berg et al. (2018) consists of two consecutive surveys  12 Further details on the sampling methodology followed in the ILO surveys are available in Berg et al. (2018). 13 For most our estimates, we decided not to narrow our control group based on profession. While an analysis of earnings and outcomes across comparable tasks (for example, in terms of routine intensity, as suggested in Dorn, 2013 andOttaviano et al., 2013) will not be disregarded, our causal estimates focus on comparing workers while controlling for their ability, disregarding any bias-inducing factor -such in the case of occupations -that could affect our estimates. For similar reasons, a small number of individuals, which have been reporting to do freelancing work from home as their main occupation, has been omitted from the estimations. This being considered, we restrict our control group to workers in occupations with comparable routine and abstract task-intensity in Table III, so to provide a more complete picture of the crowdworking phenomenon: the results included in said table, for all the aforementioned reasons, are included for descriptive purposes and should be intended void of any causal interpretation.

Figure I Participation in crowdsourcing versus traditional occupations by age
N otes: The figure shows the probability density functions of age by type of work across the US and European samples. Control sample is restricted to employed and self-employed individuals in working age, excluding freelancers working from home. nificant differences with both control groups, with the typical crowdworker being more likely to be younger, single and more educated overall. These differences are likely explained by the younger relative age of platform workers, being years of schooling and marital status obviously correlated with age. Notably, Figure I pictures participation in crowdsourcing conditional on age for both forms of work, showing how platform workers tend to occupy those younger age cohorts where individuals are more likely to be excluded from traditional forms of employment.
This age differential affects the likeliness of not being married or having children, explaining the higher propensity of being the main earner in the household and the smaller household size amongst crowdworkers. The condition of caring for children or disabled relatives, as will be discussed later, also appears more common to platform workers.
Looking at each region, differences in earnings also appear to be much more pronounced in the United States than in Europe, where the differential with traditional occupations increases from 10.47 USD in EU to 23.56 USD in US. Similarly, European crowdworkers, on average, appear to work fairly less than their US counterparts. Other disparities emerge in terms of gender (where a male majority is statistically significant in EU), health status and education.

Selected labour market indicators and controls
In order to compare crowdworkers and traditional salaried workers, we selected a number of key labour market indicators. With our data being extracted from different sources, a number of variables have been subjected to recoding, for the sake of harmonisation. Keeping the changes at a minimal level, the final coding sometimes differs across the US and European samples. In many cases, the changes have been negligible, but will nonetheless be reported when needed. 14 Hourly and weekly nominal earnings have been selected as our key variables of interest.
We were also able to control whether crowdsourcing is the main source of income for the respondent: we expect weekly earnings to be altered by this condition, while hourly earnings should remain unaffected.
Another crucial dimension of interest is weekly working hours. Thanks to the ILO survey, we were able to estimate how much time crowdworkers spend on the platform between paid and unpaid tasks. This allowed us to investigate the differential in our earnings estimates between crowdworkers and traditional workers when accounting for unpaid working hours.
In all instances, availability of weekly working hours proved essential for computing hourly earnings, as all surveys do not report the hourly rate of pay, but rather weekly, monthly or yearly absolute earnings. 15 We were also able to conduct our analysis on a different set of dependent variables, allowing us to paint a more nuanced picture of the crowdworking phenomenon. Along with indicators of skill use and job satisfaction, the EWCS, AWCS and ILO surveys contain items for identifying if the surveyed individuals would like to work more than what they currently do or whether they are currently looking for another occupation, 16 serving as proxies for labour use in the platform. This enabled us to identify involuntary crowdwork as a dimension that goes beyond standard employment statistics.
In our analysis we consider a number of controls. We first control for age, gender and education and, from there, we add other predictors. In the literature, returns to education on earnings have been widely documented, 17 while gender pay gaps have also been studied thoroughly. 18 We can also expect marital status and the number of people living in the household to affect earnings and working conditions in general. Finally, we control for state specific effects and for whether the respondent is the main earner of his household. Another fundamental variable in our analysis is caregiving, indicating whether the respondent has been involved in full-time caring for children or disabled/elderly relatives. The implications of this variable for our 2SLS model will be discussed later.

Model specification
Inspired by the Treatment Effect literature, we estimate the effect of working in online platforms on labour market outcomes in a quasi-experimental framework, where we compare earnings and working conditions of platform and 'traditional' workers across control and treatment groups. From this point of view, our approach has drawn inspiration from LaLonde (1986) and is not dissimilar from previous studies on part-time employment which instrument hours of work through household size and fertility (Ermisch & Wright, 1993, Hotchkiss, 1991, and Blank, 1998. In our case, the treatment group is composed by crowdworkers interviewed in the ILO survey, while the control group includes workers from the AWCS and EWCS surveys. As platform workers are usually paid by task, and not by hour, hourly earnings are determined first by the demand for those specific skills and characteristics over which clients can discriminate upon (factors which we can mostly control for with our set of observable covariates) and, on the supply side, by the ability of each individual worker to complete these tasks efficiently (which is mostly unobserved).
A simple comparison of the average outcomes between control and treatment groups is then not sufficient for answering our research question. Descriptive analyses or ordinarily least squares may produce biased results, potentially overestimating the effect of the platform weekly nominal earnings by weekly working hours. 16 This last item was however only recorded in the AWCS and ILO. 17 See, e.g., Angrist & Krueger (1991) and Card & Krueger (1992). 18 E.g., Blau & Kahn (2003) and, Altonji & Blank (1999).
economy on wages and working conditions. Indeed, it could be argued that individuals in crowdsourcing arrangements possess unobserved characteristics which make them qualitatively different from more traditional salaried workers, thus leading to a problem of self-selection into online labour markets. To account for this potential selection bias and offer a more appropriate comparison between the different outcomes, we adopt an instrumental variable approach. We choose the following instrumental variable two stages least squares specification: where i refers to each individual, Y is the set of our outcome variables (natural logarithm of hourly earnings and of hourly earnings adjusted for unpaid activities), while X is a vector of k-2 controls, and F is a dummy which indicates whether the respondent is female. 19 The full set of controls in the X vector are age, age squared, number of people in household, main earner (i.e. if an individual is the main earner in the household), main source of income (i.e. if the reported earnings refers to the individual's main source of income), education level, marital status and state/country of residence.
In the first stage regression (2), the treatment T (a dummy which equals 1 when crowdwork is the individual's main paid activity) is regressed on our chosen instrument Z plus the same controls we use in the second stage regression (1). Using the predicted value of T (the estimated linear probability of working in the platform) in (1), we obtain the impact of crowdwork on our desired outcome through the coefficient λ. In case the treatment T is really assigned exogenously conditionally on Z, then the coefficient on λ will not suffer from selection bias (Angrist, 2006).

Instrumental variable identification
Drawing from the demographical evidence from the studies mentioned above (see Section 3), a number of instruments have been considered for our analysis. Not all candidates for instrumentation, however, could be used, due to differences between the surveys we used.
Nonetheless, we considered and tested different types of potential instruments.
Initially, we looked at exogenous variables such as age or having a debilitating health condition, which are both significantly correlated with crowdwork (age: -0.1643***; poor health: 0.0193***). 20 We discarded those variables as we believe their adoption would lead to a violation of the exclusion restriction, biasing our estimates downwardly: younger workers typically earn less than older individuals, while workers in poor health may take longer times to complete their work activities, leading to a reduction in hourly earnings.
We then considered an alternative instrument: time spent in caregiving at home. This variable is potentially a good instrument since it is exogenous and potentially highly correlated with crowdwork. The underlying reasoning is that people may be more involved in crowdwork if they are compelled to stay at home to look after children or elderly relatives: this type of work, indeed, can be a reasonable source of income for them, given their circumstances. Caregiving appears to be highly correlated with crowdwork in our US sample (estimated correlations: caregiving 15h=0.0521***; caregiving 40h=0.1698***). This relationship is similar in Europe where caregiving also reveals itself as a significant predictor of platform work, but only at higher thresholds (caregiving 40h=0.0933***). These differences hint at the possibility of welfare-biased differential effects of caregiving, as caregivers may have access to more labour law safeguards in Europe than in US, reducing the need for auxiliary earnings from crowdwork. Evidence from Germany (Bick, 2016), indicates that a large fraction of working mothers in part-time would work full-time if they had greater access to subsidised child care.
It is then not unreasonable to expect labour market policies to similarly influence participation in crowdwork.
At the same time, caregiving also appears to be consistently correlated with the gender of the respondent: females are over-represented among crowdworkers who are caregivers, with the correlation between being in caregiving (40h) and crowdwork raising from a full sample (US+EU) correlation coefficient of 0.1920*** to 0.2502*** for the female population. This finding supports previous evidence that men's caregiving is a variable phenomenon mainly layered by societal roles, putting its exogeneity into question, 22 and uncovering a serious source of bias in the instrumental variables estimates, where the exclusion restriction is violated if gender is found to be correlated with the dependent variable.
Nonetheless, we trust that these issues can be mitigated by assuming that platform work has no intrinsic effect on gender-dependant outcomes, arguing that, after controlling for individual's characteristics and ability, crowdwork arrangements do not tend to reinforce discriminations based on the sex of the worker, due to the relative anonymity that service providers enjoy on the platform (as found in Adams, Abi and Berg, Janine, (2017) "When Home Affects Pay: An Analysis of the Gender Pay Gap Among Crowdworkers"): clients are, indeed, usually unable to ascertain the gender of online service providers. Should this assumption hold, all differences between genders will then be linked to common structural trends across control and treatment groups which can be identified linearly, and the interaction between gender and the selected instrument can be added to the instrument pool in the first stage of the estimation process. As a final check, the 2SLS estimates that can be drawn from the pool of female workers can be also said to hold for the rest of the sample, after adjusting for structural linear effects. This adjustment can be done following this simple formula: where λ f will be the effect of platform work on the female population as predicted by our instrument, ϕ 2 the linear common gender effect predicted in the full sample model, and exp(φ)−1 will indicate the baseline effect of platform work on the selected dependent variable. As our 2SLS estimation will be based on the full US-EU sample, 23 region-specific differential gender effects can also be isolated by the coefficient of the interaction between gender and the regional dummy, and then applied to the final estimates using a similar procedure, if needed.
In first part of our empirical analysis we will show that the coefficient of the interaction term between crowdwork and gender is not statistically different from zero when controlling for other observables, allowing us to generalise the common structural term predicted with ϕ 2 .
Split sample instrumental variable models -or TS2SLS -have already been explored in the past by Angrist & Krueger (1995) and Inoue & Solon (2010), who address those events when the instrument and the outcome are not measured in the same sample. In our case, however, the two subsamples -male and female -are not homogeneous. It is vital, then, to assume the differences between the two subsamples to be linear and, most importantly, to assume the structural relations within them to remain the same.
As long as these assumptions are reasonable, we only need to worry about the causal channel between caregiving and our outcome variables on the female population. Evidence from the literature on female caregiving finds that working hours -and, by extension -total earnings are affected by this condition (see Wakabayashi &Donato, 2005, andEarle &Heymann, 2012) but, to the best of our knowledge, there is no mention of hourly earnings. 24 We believe that, with the inclusion of our observable controls, we are able to filter out the endogenous effects of caregiving -due to its relationships with household size and marital status especially. In any case, we do not believe caregiving to be able to influence ability in any way: it is reasonable to assume that caregiving affects the opportunity to work more, not the relative skills of an individual -and how much the labour market rewards these skills.
A similar reasoning prevents the use of this instrument for the estimation of the effect of crowdwork on some of the other working dimensions, such as working hours. As the crowdwork 'complier' group 25 is made out of individuals spending a significant amount of time in caregiving, we can expect 2SLS estimates of working hours and weekly earnings to inevitably suffer from a downward bias, as will be discussed in the next section. In such a case, we will then mostly rely on OLS to obtain more reasonable -yet still biased -estimates of these dimensions.

Results
Columns (1) to (3) from Table II  Additional key controls are: gender, age (and its squared term), number of people in the household, marital status (whether the respondent is married or lives with a partner), and two dummies indicating whether the respondent is the main contributor to the household's income and whether crowdwork is her main source of income. We also take into account a set of controls for the different US and EU28 states of residence (with a total of 79 states) and for the level of education (distinguishing among six different education levels).
Standard errors are also robust to clustering on the level of the federal or member state.
As shown in the table, the effect of crowdwork on earnings is always negative and significant, confirming the results from the descriptive analyses of previous studies where earnings from micro-tasks are well below national averages (as found in Berg, 2015 andHara et al., 2018). The effect of the female dummy is also always negative and significant, confirming a gender pay gap in the labour market. In the third column we present our full specification: 24 Our findings -see Table VI below -similarly suggest that hourly earnings are not significantly affected by caregiving after controlling for other observables. 25 We here define as "compliers" all individuals in caregiving who participate in crowdwork arrangements and all individuals not in caregiving who stay in traditional forms of work. 26 Observations with missing values are excluded from the estimation. while both dummies for being the main earner in the family and for the surveyed occupation being the respondent's main job are positive and significant. Controlling for all other observables, the interaction term between gender and crowdwork is not statistically different from zero, while, most notably, the coefficient on gender alone retains its magnitude and significance, showing a negative linear effect on earnings (-16.5%) and no notable variation between specification (2) and (3), where the interaction is introduced. This finding provides support to our hypothesis that crowdwork platforms do not generate any intrinsic gender discriminatory effect other than reaffirming common structural gaps.
Columns (4) to (6) (6), the effect is now much closer to our estimate for the US sample, equalling to a 65.5% reduction in hourly earnings. The effect of the gender dummy is also negative and significant, this time indicating a smaller reduction in earnings (-6.8%).
A negative gender effect can also be found across European crowdworkers, albeit with a 5% statistical significance.
A significative improvement in our estimates is offered in column (7), where a full sample (US+EU) specification is presented. The difference in general region-specific gender effects is isolated by the coefficient of the EU × F emale interaction term, whose positive effect counteracts the negative sign of the F emale term, now referring to the baseline US sample. 28 Most importantly, the Crowdwork × F emale interaction turns not significant again, as its effect seems to be recaptured by the regional gender effects, confirming that crowdwork platforms do not generate any intrinsic gender discrimination on earnings. Finally, the effect of crowdwork on PPP-adjusted net hourly earnings is estimated up to a 63.5% reduction. Also, in all instances, the negative effect of working in digital labour market is slightly reduced when crowdwork is the main source of income. it is however true that an analysis which focuses only on the individuals who perform similar Table III OLS estimates of the effect of online platform work on earnings in the US and EU (1) (3) (8)  Control sample restricted to occupations whose routine and abstract task content is comparable to the 5th and 95th percentile of crowdwork occupations by their routine and abstract task content. *p<.05; **p<.01; ***p<.001

Figure II Estimated OLS coefficients from varying task-intensity splits (US+EU)
N otes: OLS coefficients for the 'Working in crowdwork' dummy after restricting the control sample (US+EU) by increasing routine task-intensity and decreasing abstract task-intensity. Sample sizes from each estimation are reported as a percentage of the full control sample. Control sample restricted to employed and self-employed individuals in working age, excluding freelancers working from home.
occupations can enhance our ability to explore the actual wage premium of traditional workers with respect to platform workers. To this aim, we split our treatment and control samples for both United States and Europe according to the degree of routine task intensity, abstractness, and a combination of the two indicators. We assign routine and abstract task intensity score to individuals in the control group using the indicators from Autor & Dorn (2013), where each occupation is given a score based on O*NET task measures. We then compute, using a similar methodology, the same scores from the ILO sample, disaggregating each observation into the five most common tasks, and assigning each task a score based on the routine and non-routine cognitive O*NET measures, as reported in Acemoglu & Autor (2011), and then averaging the scores after re-weighting each task by its relative frequency. Finally, we restrict the control groups to those observations whose routine and non-routine task intensity falls within the range of scores obtained in the treatment sample. Our results show that the coefficients do not diverge excessively from our initial results, displaying a negative -and slightly strongereffect on earnings for platform workers, in all the regressions considered (US, EU, US+EU), indicating that the routine and abstract content of micro-task jobs might not capture the reduction in earnings from traditional professions in any way.
As we cannot ascertain the full comparability of the routine and abstract task-intensity scores between controls and treatment, we provide a further robustness check in Figure II, where we restrict the control sample by decreasing abstract and increasing routine taskintensity scores, and estimate the 'Working in crowdwork' coefficient (y-axis) using the same least squares specifications from Table III (columns 7 and 8). The x-axis indicates the minimum abstract task-intensity and the maximum routine task-intensity score used for the sample split.
The figure suggests that, the more the abstract intensity of control occupations is lowered, the more the effect of crowdwork on earnings is reduced. A similar decrease is made evident when we set a higher routine content for control occupations. Nevertheless, our previous interpretation is not invalidated: these contractions in the effect of crowdwork on earnings are very minimal, as we consider that the coefficient fully maintains its sign and significancy, and that the estimated effect ranges from 57.8 to 65.5% only when performing splits on abstract intensity, and from 63.5 to 61.5% when increasing the minimum routine content. The great majority of the earnings differential between platform and traditional work remains then unexplained by the abstract and routine task-intensity of crowdsourcing. Table IV. When investigating time spent on the platform, the estimates appear particularly sensitive to the way working hours are computed. In particular, in columns (1), (4) and (7) we find that, on average, when only paid activities are considered, working in crowdwork reduces the number of weekly working hours by 16 hours, also indicating a 7 hours differential between US and the EU platform workers.

OLS estimates for working hours indicators are shown in
When crowdwork is also the main source of income, these figures are further reduced, and all crowdworkers appear to be working circa 7 hours less than traditional workers, all else being equal.
If, however, the indicator is adjusted for the time spent in unpaid tasks -as in columns (2), (5) and (8) -the magnitude of the coefficient changes again, showing a 9 hours increase in working hours across the US and the EU. For individuals whose main occupation is crowdwork, the differential with the control is reduced even more, to the point that, on average, US crowdworkers appear to be working even more than comparable workers. Significant disparities with the European sample remain, indicating that, for EU workers, there is no discernible difference in working hours between platform and traditional workers when crowdwork consists in the main source of income of an individual.
Moving to factor utilisation, we are presented with some intriguing figures. In (3), (6) and (9), our OLS model suggest that most platform workers would like to work more than they currently do in either crowdwork or in other forms of employment, suggesting a degree of factor under-utilisation. While not shown in the table, we also found out that these figures are halved when respondents are asked whether they would prefer to work in non-crowdwork occupations (even when crowdwork is the main source of income). These findings partially confute the perception of platform work as a temporary form of occupation for the underemployed, configuring it as a rather stable condition with unremarkable mobility towards other forms of employment -for many, at least. However, even if not actively looking for a job, Table IV OLS estimates of the effect of online platform work on job quality in the US and EU (1) (3)

Figure III Estimated 2SLS coefficients from varying full-time caregiving thresholds (US+EU)
N otes: Second-stage coefficients for the "Working in crowdwork" dummy instrumented through a caregiving instrument with increasing weekly hours threshold. Control sample restricted to employed and self-employed individuals in working age, excluding freelancers working from home.
Our IV estimates for hourly earnings are displayed in Table V, together with the OLS estimates from both the full sample and a female-only sample. 29 In the 2SLS regressions the estimates for the full sample and the female sample show both weak predictive power when instrumenting caregiving with a 15 hours weekly threshold (columns 3 and 4): while the first stage displays a high R-squared, the crowdwork coefficient is never statistically different from zero and the instrument always fails to pass the F score test for excluded instruments.
The 40 hours threshold generates instead much more reasonable coefficients for working in crowdwork (columns 5 and 6), predicting a general and statistically significant reduction (-63.46%; coeff.: -1.007) in hourly earnings. While very close to our OLS estimates, it could be argued that these estimates still suffer from bias due to the interaction between caregiving and gender (even if this interaction is included in the instrument pool). Restricting our study to the female population, working on crowdwork platforms reduces earnings by 60.07% (column 6, coeff.: -0.918) over working age women, all else being equal. This is well below the -1.05 (-65.18%) log points that the least squares model would predict over the female sample (column 2). In both cases, anyway, all instruments pass the F score tests for excluded instruments, with the first-stage partial R 2 also yielding remarkable results (see Bound et al., 1995). Complete first stage regressions are shown in Appendix B.
As argued earlier, while we cannot confidently attest the exogeneity of the instrument on implying that, if randomisation is achieved though its channel, the -0.918 coefficient could be considered close to an unbiased parameter of the effect of crowdwork on the earnings of the whole population after accounting for the linear gender estimates, with women still earning circa 20% less than men in both control and treatment groups. After adjusting for the rest of the sample using equation (3), we obtain a baseline reduction in earnings of 67.7%, raising our confidence in the results from the previous full sample specification. This interpretation holds even if we assume that there is some sort of gender based self-selection into the crowdworker population: should this hypothesis be true, then only our interacted OLS estimates would be biased. Since, however, we are now interested in the effect of earnings, irrespective of gender, this estimate could be considered appropriate for both men and women if the sample conforms to the target population.
In order to achieve a better understanding of the variability of the 2SLS estimates as the instrument changes its threshold, and to reduce the conceptual differences between the definitions of full time caregiving between the control and treatment groups, Figure III plots the selected threshold against the estimated effect of working in crowdwork, together with their significance level. It is evident from the figure that, with caregiving becoming a significant predictor of crowdwork at its 36 hours per week threshold, the estimated coefficients also follow a more reliable pattern with little variation in their sign and statistical significance.
Most importantly, full and split sample estimates conform to very similar trends, providing evidence that our instrument choice adequately controls for gendered bias in caregiving.

Table VII
2SLS estimates of the effect of online platform work on net hourly earnings, adjusted for unpaid tasks, in the US and EU (1) (3)  We do not report 2SLS estimates for working hours. The reason is that the condition of caregiving may prevent crowdworkers from working more or from pursuing other sources of income, whereas the desire to work more may be biased by the complications associated with the transition to caregiving. In this case, our interpretations from Table IV should then be understood as neither final or conclusive, and alternative instruments should be considered for this specific analysis.
Caregiving certainly influences weekly earnings though two distinct channels: first, as more time is allocated to caregiving, the total number of maximum weekly working hours is reduced; secondly, this activity may also generate costs for the caregiver which influence how much he or she will necessitate to earn each week. As discussed earlier, our focus on hourly earnings allows us to filter most of these issues out under the assumption that, in a static setting such as in our cross-section, platform workers are unable to individually influence their hourly salary, which is only determined by how efficiently they work. However, a less obvious implication stemming from this reasoning is that these caregiving costs may lead to a lowering of the reservation wage, which in turn could also affect participation in online labour markets and raise concerns with regards to the exogeneity of the instrument.
While this mechanism is expected and motivates our identification strategy by providing a theoretical justification for transition into crowdwork for individuals in caregiving, whether ability is linked to the level of prior and posterior reservation wages is, instead, a source of concern. In other terms, if individuals previously outside of the workforce are enter the labour force because of caregiving and are only able to join crowdworking arrangements because of their ability, then self-selection into online labour markets cannot be excluded, and estimated will suffer from bias. At the same time, caregiving may not affect may not affect participation in crowdwork for individuals who already have access to other forms of income.
Following from this reasoning, some final checks for our instrument are provided in Table   VI, where hourly earnings are regressed over the instrument and the full set of control covariates across partitions of our sample. 30 If heterogenous ability factors which we cannot already control for affect reservation wages and, in turn, participation in crowdwork, then we should see differential effects of caregiving in our reduced form estimates across treatment and control groups: caregivers who crowdwork should earn less, on average, than other crowdworkers not in caregiving (due to the participation of individuals who would have not worked before), while non-caregivers in traditional occupations, instead, should also earn more than counterparts in caregiving.
Our results, however, tell us a different story, indicating that our covariate selection already controls for these effects relatively well. While caregiving appears to have a negative and slightly significant effect on earnings in our full sample of female workers, these effects are rendered insignificant when performing the same regressions over the control and treatment 30 This analysis is similar to the one presented in Madestam et al. (2013).
groups, indicating that the negative sign of that initial coefficient is entirely linked to the first-stage relationship between caregiving and crowdwork. Notably, in no case the caregiving coefficient reaches any level of statistical significance once modelling the same regressions on the full sample (men and women). In any case, while these checks and the first-stage tests give us a good confidence in our results, it could be argued that our model may still suffer from some form of bias, due to the inability to distinguish between different forms of caregivinga problem which will be addressed in the next section. 31 Last but not least, we model hourly earnings again while accounting for time spent in unpaid activities in Table VII. As a consequence, hourly earnings -columns (1), (3) and (5) -fall well below our previous estimates, displaying a coefficient of -1.323 (-73.3%), with the prediction moving to -70.6% when instrumenting our treatment in column (7). Comparable results also apply to the female population (columns 2, 4, 6, and 8).

Instrumental variable specification
In this section we perform robustness checks for our 2SLS model. The choice of caregiving in the female population as an instrument for participation in crowdwork calls indeed for a number of robustness checks, as it could be argued that the effect of caregiving on participation in crowdwork may change with time, or that caregiving affects the participation in crowdwork but not the duration of crowdwork arrangements. Differences in control and treatment survey items may then cause issues with identification of caregivers when these individuals have been working on the platform for a long time.
While the EWCS and AWCS surveys inquire how much time does the respondent currently spent in caregiving, the ILO survey records whether the respondent was engaged in full-time caregiving right before starting to work on the platform. The design of the ILO survey then allows us to maintain the causal channel between caregiving and platform work (back when they started working online), while the controls enable us to identify whether comparable individuals in the complier group are still employed in traditional forms of work. This approach, however, imposes that, if caregiving is an exogenous determinant of crowdworking, we should reasonably assume that crowdworkers who entered this form of employment due to caregiving are still engaged in this activity.
To account for these issues, we control in Table VIII for time spent in the current occupation, a control that was previously excluded from the final model due to its -obviouscollinear relationship with our outcome and treatment variables.
In the final model of Table VII, we made the assumption that most crowdworkers have not 31 We here refer to our inability to disentangle caring for children from caring for elderly or disabled relatives.    (2) to (4) and (7) to (9), the sample is restricted to individuals who have been working in their current occupation for less than 4 years, 2 years and finally 1 year. *p<.05; **p<.01; ***p<.001 29 been engaged in this form of employment for a long time and the ones acting as caregivers when starting platform work are still engaged as such, based on the finding that 75.51% of crowdworkers have not been engaged in this form of employment for more than two years. We now relax this assumption in Table VIII, where we run the same final IV specification from  (1) and (5). 32 In the subsequent specifications -columns (2) to (4) and (5) to (8) -we perform a similar analysis by restricting the sample to people who have been working for less than 4 years, 2 years and finally 1 year. By comparing workers that have been working in their current occupation for similar time, the more we reduce the years they have been spending in their current occupation, the more our assumption that these workers are still in caregiving is made reasonable: in this way, we believe to be able to filter out the effects of time spent in a given occupation through the first stage of the 2SLS model. The trade-off is that, the more we reduce our sample size, the more our estimates lose in precision. Nevertheless, the interpretation of our results stays relatively unchanged, with the coefficients retaining their signs and significance. The magnitude of our coefficient for platform work, however, seems somewhat sensible to the sample reduction: in any case, it never overestimates the coefficient of the OLS model, while remaining relatively stable after individuals with more than 5 years of employment have been accounted for. After adjusting for gender specific linear trends, as in equation (3), we can reasonably argue that working in crowdwork generates a negative effect on earnings sitting between 67.2 and 58.35% less than for comparable workers after controlling for time spent in current occupation.
Finally, as mentioned earlier, our econometric model may also raise a number of concerns with regards to the exogeneity of our instrument. Evidence from studies such as Kremer & Chen (2002) suggests that fertility may be influenced by a number of social drivers. While we believe that our controls are sufficiently apt at filtering these influences out, 33 we here intend to relax this assumption and treat fertility as endogenous. Even if, as discussed, conflicting survey designs prevent us from fully separating individuals caring for children from the ones caring for disabled or elderly relatives, we can nonetheless identify individuals in caregiving who, at the same time, do not have kids -and, therefore, are most surely not caring for children. We then switch our instrument with the new one ("Caring for elderly or disabled relatives only" and present our results in Table IX, adopting the same approach used for the robustness checks in Table VIII. The reductions in the "complier" treatment group leave to an increase in the variability of our estimates which appear particularly sensible to the reduction in sample size. Since this time we are only able to compare individuals with no children, some kind of bias can still be expected: in fact, while our estimates maintain their sign and  (2) to (4) and (7) to (9), the sample is restricted to individuals who have been working in their current occupation for less than 4 years, 2 years and finally 1 year. *p<.05; **p<.01; ***p<.001 do not diverge too much from our results in Table VIII, they surely suffer from some level of overestimation. In any case, these results do not contradict our previous findings.

Model specification
All robustness checks we previously presented rely on the correct specification of the IV estimator. In this section, instead, we address the concerns related to this approach by relying on an alternative specification for the estimation of the causal effects of crowdwork.
An interesting result from our IV estimates is that first-stage regressions produce comfortably high R-squared statistics, meaning that our set of observables is rather competent at predicting assignment into treatment. If we have a correct specification for the probability to work in online labour platforms, then a binomial model can be used to compute propensity scores, which can be used to re-weight observations across control and treatment groups. Re-weighting can be achieved through inverse probability weighting (first proposed by Rosenbaum, 1987; see Austin, 2011, for a methodological review of uses of propensity scores in quasi-experimental settings), where new weights are produced by assigning each observation the inverse of the conditional probability of its treatment status. This means that, in our case, individuals in crowdwork will receive a weight equal to 1/p i , while traditional workers will be weighted 1/(1 − p i ), where p i indicates the propensity score; in other words, it indicates each individual probability P (T = 1|X) to be working in crowdsourcing, given a set of covariates X.
The differences of inverse propensity scores weighted averages will yield the effect of platform work, under the caveat that the underlying propensity score model is correct.
To overcome this issue, an inverse-probability-weighted regression-adjustment (IPWRA), first covered by Robins et al. (1994) and further developed in Wooldridge (2007) is proposed, where inverse probability weighting is combined with regression adjustments in order to produce a doubly robust estimator. In IPWRAs, regression models are fit on inverse probability weighted observations according to their treatment status (meaning that the model is fit on two separate treatment and control samples), and the parameters from these models are used to predict counter-factual outcomes on an individual level, for all observations. The difference in means between treatment and control predicted outcomes will then yield the ATE.
The IPWRA estimator ensures that, as long as one of the two models, one for predicting assignment, and the other one for modelling outcome, is correct, then the results will not suffer from bias. We then use a binomial logistic model to calculate propensity scores, 34 and then assign the new weights to control and treatment observations so that a linear model for earnings (hourly, natural log) can be fit across treatment and control groups, using the same covariate specification from Table II, column 7 (omitting, for obvious reasons, the "Working in crowdwork" dummy). The results from our test can be found in Table X. Our estimates show that, after controlling for these different models, working in crowdwork still produces a statistically significant -69.62% reduction on earnings (adjusted for unpaid tasks). Extending our double robust approach to the estimation of the ATET, we find the effect on the treated to be close to -70% as well (with comparable statistical significance).
These results are remarkably similar to the ones obtained by our previous instrumental variable approach (and OLS, by extension), and reinforce our finding that working conditions in crowdwork are generally unaffected by the characteristics of individuals working in these arrangements.

Conclusions
In this paper we have provided an empirical analysis of the effect of crowdwork on working conditions in both the United States and Europe. We use a quasi-experimental design and we assemble data from different sources, coming from online surveys on crowdworkers, web plugins and commonly used extensive surveys on workers' conditions in the US and Europe. To the best of our knowledge, this is one of the first attempts to provide an unbiased comparison of platform and traditional workers in terms of earnings and working conditions.
In our contribution, we focused on the effects of individual ability on earnings in the plat-form economy, finding that most of the differences between platform workers and traditional workers are unexplained by individual characteristics. As we show that the effect of crowdsourcing on earnings is even larger as it could be expected from simple differences in means, our estimates cast a dark shadow over platform work: crowdsourcers earn 70.6% less than comparable workers in terms of ability, while spending nearly as much time working in the platform as their counterparts do in traditional occupations. Most importantly, labour force in crowdworking arrangements appears to be highly under-utilised, with all crowdworkers being more likely to be left wanting for more work than comparable individuals. All these findings, along with the fact that these individuals do not appear to be looking for other jobs more than 'traditional' workers, relegate crowdworkers into a new category of idle workers whose human capital is not being fully utilised nor adequately compensated.
It should be noted that while these results hold for US and EU platform workers, the external validity of our estimates is threatened by the nature of crowdwork platforms themselves and, while our conclusions may be extended to routine-task intensive platforms such as Crowdflower or Clickworker, our analysis may not hold in other contexts where more diversified tasks, requiring more creative input from service providers, are offered, such as in the case of UpWork. 35 The observed disparities should then be attributes to factors other than individual ability.
We were able to rule out the possibility that most of these differences were caused by the routine and abstract content of online platform jobs, as workers with comparable routine and abstract tasks still retain most of their salary premium, indicating that the relative simplicity and repetitiveness of these tasks does not necessarily lead to a sizeable decrease in earnings.
This leads us to believe that this effect could be better explained by the following factors: 1. competition from equally skilled but cheaper labour from other countries within the same platform; 2. scarcity and heterogeneity in demand for these kind of activities; 3. lack of labour rights and minimum standards stemming from the status of independent contractors.
In the first case, the earnings effect of platform work can be attributed to excess supply: indeed, the influx of "digital immigrants" may lead to an increase in labour supply and infratask competition, lowering remunerations due to the low complementarity of these workers.
Indeed, Borchert et al. (2018) have found that unemployment shocks, leading to increased participation in online markets, can have a positive effect on wage elasticities in crowdwork.
In the second case, it could be argued either that firms and clients are mostly uninformed about the possibility of outsourcing through online platforms, or that the sample of clients which employs online labour is intrinsically different in its nature from other firms, generating scarcity in demand. While panel data is necessary to study these effects, the lack of particular differences between crowdworkers in 2015 and 2017 -in the ILO quasi-panel -indicates that, so far, the demand for these services has seen little growth. Also, while Katz & Krueger (2018) estimate a general rise in participation platform economy between 2005 and 2015 (from 10.7 to 15.8% in the US), Farrell & Greig (2017) find evidence in support of the claim that these markets have, overall, reached their peak in 2016. Still, persistence of slack and factor underutilisation in these markets is indicative of the presence of a mismatch between supply and demand which, if not found to change over the next years, could be described as a structural condition of crowdsourcing as a consequence of the nature of its clients.
In the third and final case, the lack of standards, linked with the monopsonistic nature of platforms, enables the imposition of a heavy markup over online workers, allowing clients to operate at prices well below the market's marginal costs. These considerations are consistent with the results of Dube et al. (2018). As our results refer to year 2015, the influence of these factors could change in the future, in parallel with the evolution of the platform economy. In any case, we believe that the poor working conditions crowdsourcers have to live with are the result of an interplay between these elements, and it is up to future research to test each of these hypotheses individually, disentangling the effect of each of these factors from the others.   2SLS First and Second Stage coefficients of the effect of online platform work on earnings in the US and EU (Table V) (1) (3)