Information Frictions and Observable Experience:
The New Employer Price Premium in an Online Market
Christopher T. Stanton
Harvard, NBER, and CEPR
Catherine Thomas
LSE, Centre for Economic Performance, and CEPR
November 2015∗
Abstract
Unfamiliar markets are likely to have aspects or features that are learned about via experience. In the global online labor market oDesk, new employers appear uncertain about the value
of hiring a worker. In addition, engaging in transactions appears to require high effort for new
employers that tapers with usage. As employer experience is publicly observable and workers
are differentiated, workers take expected employer uncertainty and effort costs into account
and submit wage bids that are 10-percent higher to new employers. A structural model of
worker supply and employer demand for inexperienced and experienced employers shows that
two-thirds of the 10-percent wage bid premium can be attributed to the higher costs incurred
when working for inexperienced employers. The remaining one-third is due to inexperienced
employers’ uncertainty about market value, which makes their hiring decisions less elastic with
respect to wage bids and alters optimal markups. Because employment relationships tend to
be short-term, workers bids do not account for the longer-term benefits in the market from
employer learning. A counterfactual analysis shows that offsetting some of the initial wage
bid premium by offering lower platform fees to inexperienced employers creates spillovers to
future demand and would increase platform profits.
∗
We thank seminar participants at the AEA Meetings, the CEPR Workshop on Incentives, Management and
Organisation, Harvard, LMU, LSE, Mannheim, NBER Summer Institute, and Yale, along with Ajay Agrawal, Nava
Ashraf, Ricard Gil, Ed Lazear, Arnaud Maurel, Luis Rayo, Yona Rubinstein, Scott Schaefer, Kathryn Shaw, and
Nathan Seegert for helpful comments.
1
1
Introduction
In many markets, first time buyers are offered prices that differ from those available to experienced
buyers. In the online global labor market oDesk, where the buyers are employers of remote labor
services, those that are new to the market face higher wage bids than employers that have made
prior hires on the site. That is, there is a new buyer wage premium, rather than a first-time buyer
price discount. The premium discourages transactions and many first-time buyers leave the market
without hiring. This paper uncovers the reason for the new buyer price premium in this market,
and investigates when it is likely to generalize to other market contexts.
The new employer wage premium is shown to arise because new employers face incomplete
information about the market. Their unfamiliarity with the marketplace raises the costs of their
transactions, and their uncertainty about the distribution of market value affects the elasticity of
their hiring demand. Workers are able to submit wage bids that reflect the greater costs imposed
by inexperienced employers because employer experience is publicly observed, and, in addition,
because workers are also differentiated in the market, bids to new employers include higher optimal
markups reflecting these employers’ relatively inelastic demand.
Both these forms of incomplete information act as market entry barriers that reduce the number of transactions in the market relative to a benchmark where new employers have the same
information set as experienced employers. A counterfactual exercise shows that setting experiencedependent platform fees, rather than the current fixed fee, can offset some of these frictions and
result in increased profits for the platform.
On oDesk, employers post projects and then workers make wage bids, and the data allow observation of the set of offered and contracted wages over the entire history of employers’ use of
the market. To facilitate a comparison of employers posting similar jobs over time, the analysis
in this paper concentrates on relatively homogeneous administrative support jobs where the modal
job posting involves data entry work. Even when posting relatively simple administrative support jobs, inexperienced employers receive wage offers that are, on average, more than 10 percent
higher than employers with four or more prior hires. A large and positive gap is present in OLS
regressions and in models with fixed effects for workers, employers, and for workers and employers.
Among employers who hire in the market, the wages paid to the hired workers exhibit similarly
large differences.
A structural model of labor demand and supply, estimated using the data, permits a decomposition of the new buyer wage bid premium into a part that is due to higher costs and a part that
2
comes from higher worker markups. Because wage bids are likely correlated with unobserved worker
and job-match quality that is observed by the employer, estimating demand requires an instrument
that shifts workers’ costs but not employer demand. We use worker-country exchange rates as the
instrument. Workers on oDesk come from different countries, and they are exposed to differential
cross-country macroeconomic fluctuations that are plausibly orthogonal to individual worker-level
unobserved characteristics and employer demand after accounting for workers’ participation on the
platform. Because the worker’s outside wage is typically paid in their local currency while the oDesk
contract is paid in US dollars, we use the standardized log of the local currency to dollar exchange
rate to instrument for the log of workers’ wage offers. The residuals from this first stage regression
serve as control functions, allowing us to estimate demand using mixed logit models.1
Workers’ bids are assumed to be equilibrium price offers in a differentiated products Bertrand
game. When applying to an employer’s posted job, the worker anticipates the costs he will face and
marks up this cost by an amount that reflects the employer’s expected wage elasticity of demand
given observable experience. Segmenting employers into those that are new to the market and those
that have four or more prior hires allows for estimation of the market equilibrium for each group.
Both worker costs and wage elasticities of demand differ across the two employer groups.
Workers charging higher markups to inexperienced employers are equilibrium outcomes in a
model of employer demand where experience reduces uncertainty about the expected value of hiring a worker. Employers’ willingness to hire an individual applicant for a job at a given wage bid is
the probability that the applicant is the highest perceived order statistic of wage-adjusted productivity among available workers. Workers’ payoffs are discontinuous around the hiring threshold, and
differences in the spread between the highest and second-highest order statistic creates local market
power. When employers’ beliefs about the distribution of match quality are less precise, the spread
between the highest expected order statistic and the next-highest order statistic is greater. Therefore, in the event that the worker has the highest expected wage-adjusted quality among available
workers, an inexperienced employer with greater uncertainty about the distribution of worker match
quality has a higher expected relative willingness to pay for that worker compared to alternatives.
In other words, the inexperienced employer has less wage bid elastic demand. This intuition is
general beyond labor or matching environments and applies to other contexts with unit demand:
noise in the environment when evaluating products in goods markets or applicants in labor markets
1
We do not estimate a dynamic model of employer demand. An existing literature estimates empirical models
of optimal search with learning, but they are challenging because the parameters of the distribution of beliefs are
unobserved, serially correlated state variables. Allowing for experience-dependent prices adds an additional layer of
complexity that requires more restrictive modeling assumptions.
3
has the potential to change price elasticities, altering equilibrium markups.
Estimates of the wage bid elasticity of demand for inexperienced and experienced employers,
together with the observed wage bids and other information about worker characteristics and job
attributes, allows the decomposition of the inexperienced employer wage bid premium. Estimates
of the sum of the workers’ outside wages and hassle costs are plausible; they decrease with employer
experience and are correlated with country-level gdp per capita. Hassle costs are responsible for
about 66 percent of the difference in bids, and experience-dependent employer demand differences
are responsible for the remaining 34 percent.
The distribution of net expected match quality for experienced employers also has a higher mean
than the distribution for inexperienced employers. This suggests drift in valuations from using the
market as employers gain experience. If workers’ hassle costs are a result of employers’ learning to
use the market platform, then it is likely that employers also bear some of this cost and experience
reduces this net cost over time. When combined with evidence that part of higher wage bids reflects
hassle costs to workers, employers appear to bear some share of learning costs through higher wage
bids.
The results suggest that employers learn about market value with experience, and several additional features of the data are consistent with the predictions of a search model with learning
about an unknown distribution. In fact, this paper offers novel evidence from the actual behavioral
taken on sequences employer hiring that learning is important.
2
With an unknown distribution of
match quality, or employer-specific value of the market, as employers become more certain about
the distribution, the hiring reservation value declines. This is because the option value of waiting
falls as the precision of beliefs increases, so the value of the marginal interview declines with cumulative experience.3 When inexperienced employers hire after only a small number of interviews, they
forgo further learning opportunities. An employer would choose to forgo these opportunities only if
an early interview yielded a match that exceeds a high reservation value. Employers who conduct
more interviews are likely to have worse job outcomes because the reservation value declines with
cumulative search and they are revealed-preferred to have not found a match that exceeds the high
2
For theory, see Lars Ljungqvist and Thomas Sargent’s textbook treatment of the job search model with an
unknown offer distribution. The labor economics literature has extensively considered learning models over the
career. See Gibbons and Farber (1996), Altonji and Pierret (2001), Lange (2007), Kahn (2013), and Kahn and Lange
(2014). For empirical work in the context of learning one’s own type, see Miller (1984) and Arcidiacono, Aucejo,
Maurel, and Ransom (2013). This paper is also related to the literatures in industrial organization and marketing
about how demand changes as consumers learn about product quality (see Erdem and Keane, 1996; Ackerberg,
2003).
3
In support of this prediction of a learning model, the data show that employers tend to interview fewer workers
on successive job postings.
4
initial reservation value. In the data, an inexperienced employer who hires after having interviewed
5 to 8 workers is 10 percentage points less likely to report a successful job outcome than an employer
that interviews a single employee.
The oDesk revenue model currently imposes a tax of 10 percent on all transactions on the site.
Using estimated differences in demand according to employer experience and estimated differences
in workers’ costs of interacting with employers in each group, we examine whether a tax rate that
varies with employer experience would increase transactions and generate additional profits for the
platform. Although inexperienced employers have less elastic demand with respect to wage bids,
oDesk could, nonetheless, generate higher platform profits by charging lower fees to inexperienced
employers. This is because of the large spillovers to future demand of increasing the probability
of making a first hire on the site. The short-term nature of job contracts on the site means that
no individual worker has an incentive to subsidize employer learning because the benefits of doing
so would accrue to the employer and the employer’s future hires. In contrast, the platform itself
does have some incentive to subsidize employer learning, and offset part of the premium charged
by workers to new employers, since it captures a share of all future revenues that the employer
generates in the market. This means that optimal platform pricing partially internalizes the effect
on wage bids that is facilitated by the platform by publicly revealing employer experience.
In addition, because valuations drift over time with experience and employers learn through
hiring, marketplace operators may have ex-post monopoly profits because there are substantial
costs associated with learning about an alternative market. Encouraging employers to learn the
market and to learn the distribution of worker productivity by charging them lower prices provides
justification for many of the ”freemium” pricing structures that are currently observed in online
markets and elsewhere, where some aspect of product or service is offered to customers free of
charge, and a fee is charged for some further features or functionality. An implication of the results
presented in this paper is that it is likely to be optimal for anything that allows or accelerates
learning to be part of the ”free” or reduced-price part of the ”freemium” package.4
The paper proceeds as follows: Section 2 provides stand-alone details about oDesk and how
employers search in this market. Section 3 introduces evidence on differential pricing by observable
experience. Section 4 estimates demand by employer experience levels and decomposes wage bids
into costs and markups. Section 5 provides additional evidence that a search model with learning
can fit important moments of the data, lending additional support to the hypothesized mechanism
4
We thank Ajay Agrawal for this connection to the design of ”freemium” product offerings.
5
driving the result. Section 6 evaluates how changes in the market affect hiring rates and profitability.
Section 7 presents some robustness tests that rule out alternative reasons for higher wage bids.
Section 8 concludes.
2
The Search Process on oDesk.com
oDesk.com is an online platform that allows employers to contract with remote workers.5 It facilitates both search and matching for workers as well as remote task and project management and
payments. Work includes a range of jobs where output can be delivered electronically, and the most
frequently observed job categories are Web Development and Administrative Support. Jobs tend
to be short-term spot transactions, with the majority of postings requiring under three months of
work. Around 85 percent of the transactions in the market span international borders and, therefore, constitute international labor services trade. This paper focuses on oDesk’s matching role,
which is as follows:
An employer with a job opening to fill creates an account, free of charge. To post a job opening,
the employer is asked to select the job category, give the job a title, and describe the work to be
done as well as the skills needed and the expected duration. Once the posting is in the system,
potential employees learn about the job by searching on the site or through automatic notification.
Interested workers then submit applications for the job posting and bid an hourly wage. Employers
also have the option of searching worker profiles directly and inviting candidacies from individual
workers. Workers’ profiles, visible to potential employers, contain information about their skills,
education, prior offline work experience, and experience on oDesk. The country where they are
located is also displayed prominently on the profile. For workers that have experience on the site,
their profile shows a summary feedback score out of five, and all worker profiles include a requested
hourly wage rate that the worker is free to change at any time.
After receiving applications or initiating candidacies, employers can opt to interview any number
of workers for the job. If applicants agree to be interviewed, the interview usually takes place via
Skype. Whether an interview actually occurs is not recorded in the oDesk database, so the remainder
of the paper will refer to the intention to interview as an interview.6 An employer can choose to
hire an applicant with or without interviewing them first. Upon hiring, the employer can monitor
5
Other prominent platforms included elance and Guru, the first of which merged with oDesk in 2014. The merged
company has recently changed its name to Upwork.
6
The data record that an interview takes place whenever an employer sends an interview request to a worker,
whether that worker was invited to apply for the job or initiated the candidacy himself.
6
workers via software provided by oDesk, and oDesk manages all payments. When a job is complete,
the employer is asked for feedback about the worker and vice versa. The employer is also asked
whether or not the job was completed successfully.
The employers’ search sequence on a single job, as described above, is observed on each of
their successive job postings. For each posting, the data contain information about the complete
applicant pool; which candidates, if any, are interviewed; and which candidate, if any, is hired.
This paper does not use information about on-the-job monitoring, but does exploit information on
feedback and indicators of job success. While several recent papers use oDesk data (Agrawal et
al., 2013a and 2013b; Horton, 2013; Lyons, 2013; Pallais, 2013; Ghani, Kerr, and Stanton, 2014;
and Stanton and Thomas, 2012), this is the first paper to use this empirical setting to study how
employers’ hiring processes evolve over time and how workers respond to employer experience in
the market.
3
Data and Summary Measures
The employers studied in this paper are those employers who posted Administrative Support jobs
on oDesk between January 2006 and June 2010. The paper analyzes Administrative Support jobs
because, among the most frequently observed job categories, these jobs are the least likely to require
on-the-job coordination among workers hired by the same employer. This means that the identity
of previously hired workers is less likely to affect subsequent hiring behavior through a productionrelated channel, relative to jobs posted in programming-intensive categories. This allows for a
focus on search-process-related dependencies between past and future hiring behavior. Around one
third of the Administrative Support jobs in the data are for Data Entry, around 20 percent are for
Personal Assistants, and another 20 percent entail Web Research.7 In an effort to ensure that jobs
are similar, analyses that estimate demand restrict the sample to Data Entry jobs.
Employers are mostly located in the United States include private individuals as well as those
hiring on behalf of established firms. Of the 16, 903 employers who post jobs without having hired
previously on the site, 10, 664 received all the applications for this job before going on to post
another job. For these jobs, the hiring periods can be segmented by time and, throughout the
paper, this set of jobs is referred to as the non-overlapping or sequential sample. This restriction
reduces the total number of postings from 55, 207 to 38, 905. Figure 1A presents the histogram of
the distribution of total number of jobs posted per employer. 99 percent of employers in the data
7
The remainder are: Email Response Handling, Transcription, and Administrative Support - Other.
7
post fewer than 30 jobs throughout the period. Figure 1B presents the same histogram, this time
looking at the distribution of the number of job openings per employer where the posting results in
a hire.
3.1
Employer Behavior
Tables 1A to 1D present summary statistics about job postings, grouped by the number of previous
hires made (across all job categories) for the relevant employer. Table 1A includes all job posts in
the sample. Table 1B includes only the subset of these jobs for which the job search was complete
by the time the prior or next job search was complete—the sequential sample. Tables 1C and 1D
present summary statistics for the sample restricted to Data Entry jobs, which includes 21, 407 job
postings all together and 15, 144 sequential job postings. These tables provide initial evidence that
both employer and worker behaviour differs according to the number of previous hires the employer
has made on the site.
Employers in Table 1A without prior hiring experience receive an average of 34.51 applications
prior to closing the job or making a hire. In the sequential sample, inexperienced employers receive
an average of 34.7 applications per job. As employers gain experience, the number of applications
per job tends to fall, with the largest decline taking place between having hired zero and one prior
worker. The average hourly wage bid for inexperienced employers in the overall sample is 5.01
USD, and in the sequential jobs sample it is 5.19 USD. Column 4 shows that the average hourly bid
declines with employer experience. For those with one prior hire, the average bids fall to 4.41 USD
(4.49 USD in the sequential sample). After four or more previous hires, employers receive bids that
average 3.56 USD in both the complete and sequential samples.
Employer learning about the value of the market is one potential mechanism explaining why employer behavior changes with experience, and some evidence of learning is present in these summary
measures. Employers plan to conduct an average of around 3.8 interviews per job posting when
inexperienced, both for all jobs and in the sample of sequential jobs. Column 5 of Table 1 shows
how this statistic evolves with employer experience. Employers tend to interview more applicants
on earlier job postings, again with the largest decline taking place after the first hire. The overall
probability that a job opening results in an eventual hire also differs substantially with employer
experience. Column 6 shows that inexperienced employers are least likely to hire. The probability
of hiring doubles for employers with at least one prior hire.
The data also provide some evidence that workers appear differentiated to employers. When a
8
hire is made, and controlling for the mean value attributed to workers’ observable characteristics,
employers are unlikely to hire a candidate whose hourly bid is in the lowest decile of the bids
received. Figure 2A shows that the employee hired made a bid in the lowest decile of received bids
for only 25 percent of hires. Figure 2B examines the bid decile of the hired worker having controlled
for job category and workers’ country, feedback score, education and english skill level. The hired
worker’s residual bid is in the lowest decile for less than 20 percent of hires.
To motivate the hypothesis explored in the paper that differences in demand may play a role
in the observed differences in bids documented in Table 1, Figure 3 provides a summary sketch of
the relationship between wage bid and the probability of being hired for workers applying to Data
Entry jobs. To attempt to remove unobserved heterogeneity, the sample is restricted to workers
from the Philippines, which is the most frequent country of origin for job applicants. Bids are also
residualized based on observable worker characteristics and time fixed effects. The two connected
lines come from separate local polynomial regressions evaluated at the plotted nodes. This figure
establishes that the nature of employer demand is associated with the extent of employer experience
rather than a causal relationship at this stage. However, with this caveat in mind, the relationship
between bids and the probability of hire appears quite different for employers with and without
experience. The figure shows that over the most frequently observed wage bid range, experienced
employers’ demand is shifted out and less steep compared to inexperienced employers’ demand.
3.2
Workers’ Hourly Bids and Hourly Wages
Table 2 investigates how wage bids vary with employer experience. The omitted category is employers prior to making their first hire. The coefficients show the percentage change in the average
wage bid received when workers observe that the employer has hired one, two, three, four or five
or more prior workers. Worker-level controls include the applicants’ feedback and an indicator for
missing feedback, an indicator for having done a previous job without receiving feedback, and the
number of previous hires. Job-level controls include job category fixed effects and expected job
duration fixed effects. Monthly time fixed effects are also included, and, to isolate changes in the
information that workers are able to observe when bidding, the sample includes only those jobs that
are sequential.
Column 1 shows that, in the cross section, employers with one observable prior hire receive bids
that are eight percent lower than employers who have no observable experience hiring. This falls to
nine percent lower and then to eleven percent lower for employers with two and three prior hiring
9
spells, respectively. Employers with five or more hiring spells receive bids that are 16.5 percent lower
than employers with no experience. Column 2 adds controls for changes in qualitative features of
the job post, including a third-order polynomial in the number of characters in the description and
fixed effects for expected duration. These controls are intended to allow for the possibility that the
employer is learning about how to write an effective job posting. The observed decline in average
wage bids received changes very little.
This pattern of examining changes in the qualitative feature of job posts is repeated throughout
the table, where subsequent columns add various further controls to account for other plausible
explanations for the decline in wages bid with employer experience. Columns 3 and 4 add employer
fixed effects. A given employer receives bids that are about four percent lower with one prior
hire and about ten percent lower with five or more prior hires compared to bids when he has no
experience. Columns 5 and 6 are estimated with worker fixed effects. The result is robust within
worker, as a given worker submits lower bids to experienced employers. Within worker, bids to
employers on their fifth or more hiring spell are, on average, about nine percent lower than bids to
employers with no observable experience. The last two columns are estimated with both employer
and worker fixed effects, and the parameter estimates show the same pattern of declining wage bids
with experience. These columns show that the observed wage bid declines with employer experience
are not due to workers segmenting employers based on employer experience when choosing whether
or not to apply to a job.
Panel B of Table 2 shows that it is employer experience, rather than public employer feedback,
that is associated with lower bids. In these columns, the sample is limited to employers’ with no
observable hiring experience and to employers with five or more prior hires. Indicators for the
employer having no observable feedback and for the employer having observable feedback of 4.5 or
higher are interacted with the indicator for having five or more prior hires. Including these controls
does not change the main result that experienced employers receive lower bids. In fact, the main
point estimate measures the effect for experienced employers with bad feedback: these employers
still receive bids that are significantly lower than inexperienced employers. The interactions of employer experience and good employer feedback are, in most cases, a fraction of the main effect. In
specifications comparing how bids change within employer, the estimates are statistically indistinguishable from zero. With worker fixed effects, the feedback interactions have positive parameter
estimates. However, if a lack of feedback was the primary driver of bid differences, one would instead expect that the interaction of good feedback and experience would result in lower bids. When
10
adding both worker and employer fixed effects, these interactions are both small. The main message
from these regressions is that experience tends to swamp any differences in how employer feedback
evolves over time in explaining differences in wage bids.
Panel C of Table 2 shows that the relationship between employer experience and lower bids
is robust to including controls for feedback that the employer has given to workers on prior jobs.
As in Panel B, the sample is employers with no prior hiring experience and employers with five
or more prior hires. Indicators for the employer having given no observable prior feedback and
for having given good observable prior feedback are interacted with an indicator for having five or
more prior hires. The significant negative coefficients on having experience remain when including
these controls. The estimated coefficients on these controls are positive in sign, which is consistent
with the interpretation that workers attribute having left no feedback or giving good feedback as
positively correlated with employers’ willingness to pay.
To further visualize these bid differences, Figure 4 presents a binned scatterplot of within workermonth mean bids to data entry jobs posted by inexperienced employers against the within-worker
month mean bid to jobs posted by experienced employers. The 45-degree line is also plotted in red.
Workers’ who bid very low wages for experienced employers appear more likely to increase their
bid for inexperienced employers; the difference between bids appears to decline for workers who bid
higher wages for experienced employers. Although it is not causal, the bid differences appear to be
greatest in regions of bid levels where the demand curves plotted in Figure 3 differ most. We now
turn to estimating demand more credibly.
4
A Model of Worker Supply and Employer Demand
This section presents a model that allows for two channels whereby employer experience can affect
workers’ equilibrium bids. The first channel is that experienced employers may impose different
costs on job applicants and also incur different costs themselves. The second is that experienced
employers’ relative familiarity with the market affects their wage elasticity of demand. Estimating
the model reveals the role that each of these aspects of employer experience has on observed wage
bid differences. In the model, the worker’s decision is the hourly bid to make for a job, and the
employer’s decision is which, if any, worker to hire.
11
4.1
A Worker’s Optimal Bid
When choosing the wage to bid for a job posted by employer i, worker j’s objective function is:
max pij (log wij ) × exp (log wij − log(1 + τ )) + [1 − pij (log wij )] × cj .
log wij
where log wij is the log of j’s wage bid and pij , is the probability employer i hires worker j, which is
a decreasing function of wij and (left implicit) an increasing function of all wage bids received from
the other workers in the applicant set Ji . Because oDesk has an ad-valorem fee τ, the wage received
by the worker if hired is wij (1 + τ )−1 and the wage the employer pays is wij . If worker j is not
hired, he receives the flow outside opportunity cost cij , which captures the opportunity cost of his
outside option plus any hassle costs of work.8 This objective is static, but the first order condition
for the workers’ problem is the same as in the dynamic case as long as no other jobs offers arise
over the duration of work on the job posted by employer i. By increasing his bid wij , the worker is
trading off a lower probability of being hired with a higher wage if hired.
The system of Ji first order conditions from each worker j ∈ Ji ’s objective function determines
equilibrium bids in a Bertrand Nash game in prices (here, in wages). Each worker’s first order
condition is given by:
∂pij
wij (1 + τ )−1 − cij + pij wij (1 + τ )−1 = 0.
∂ log wij
(1)
which implies:
cij = w (1 + τ )−1 + pij wj (1 + τ )−1 /
∂pij
.
∂ log wij
(2)
∗
gives:
and solving for worker j’s optimal bid wij
∗
wij
∂pij
= cij (1 + τ ) 1 + pij /
∂ log wij
−1
.
(3)
Worker j’s equilibrium wage bid, hence, depends on the employer-specific hassle costs and the
worker-specific outside option, which, together, sum to cij , the employer’s demand function, pij (log wij ),
and the tax rate τ , which is fixed across different employers at present in the market. We turn now
8
The objective could alternatively be written:
max pij (log wij ) × exp (log wij − log(1 + τ ) − ciH ) + [1 − pij (log wij )] × cjO
log wij
where ciH is a hassle cost associated with being hired by employer i and cjO is the outside wage for worker j. The
first order condition in this case makes clear that only cij = ciH + cjO can be identified.
12
∗
to how employer experience in the market can affect wij
in equation (3) through cij and pij (log wij ).
4.2
Employer Demand
Employers posting jobs have a publicly observed experience level, e, which is the number of previous
interactions with workers. For inexperienced employers, e = 0 and, for experienced workers, e ∈
[1, 2, .....]. After posting a job, employer i with experience e evaluates the employer-specific wageadjusted quality of each worker j ∈ Jie . In addition to the wage bid, this is a function of three
variables that are observed by the employer: worker-specific quality that is commonly valued by
all employers, qj , a worker-employer Type-1 extreme value distributed error, εij , and an observed
signal of that worker’s employer-specific quality, mij .9
Employer experience, e, determines the employer i’s evaluation of worker j’s employer-specificwage-adjusted quality conditional on the signal mij . The signal mij is modeled:
log mij = ηij (µi + de ) + ξij ,
(4)
where ηij (µi + de ) is a random effect with employer-specific mean (µi + de ) that is orthogonal to
common quality qj by construction, and ξij is another idiosyncratic error component. The term
µi is a permanent, employer-specific term that captures his mean valuation for the market, and is
not known with certainty but can be estimated more precisely by employer i having learned from
experience. de is drift that is common to all employers that allows employer valuations to change
in a common way with hiring experience.
Although it is not necessary for the general intuition of the model, a convenient way to parameterize employer uncertainty and learning about µi is a normal-normal learning model (Farber and
Gibbons, 1996; Altonji and Pierret, 2001; Kahn and Lange, 2014). Suppose employers believe the
2
distribution of µi is normal, with prior mean and variance µie and σµe
. The random effect ηij is
also normally distributed with mean (µi + de ) and variance ση2 .
As a Bayesian employer gains experience, his beliefs about µi change, and his interpretation of
the signal becomes more precise. To see this, if an employer with experience e interacts with one
9
We assume any interaction with a worker, whether an interview or a hire, provides the same signal. A detailed
analysis of forward-looking strategic search and stopping where an interview provides a less precise signal than a hire
is available on request.
13
more worker j and observes log mij , the distribution of the posterior mean of µi(e+1) is:
"
2
µi(e+1) |µie , σµe
˜N
#
log mij − de
µie
+ 2 /
2
2
ση + σξ
σµe
−2
ση−2 + σξ
+ σµe ,
!
1
−2
ση2
+
−1
σξ2
−2
+ σµe
,
(5)
The conditional distribution of the individual random effect given the signal mij and employer
experience e is:
"
2
ηij | log mij , µie , σµe
∼N
#
!
log mij
µie
1
−2
,
de +
+ 2
/ σξ−2 + σµe
+ ση−2 , −2
−2 + σ −2
σξ2
σµe + ση2
σξ + σµe
η
(6)
which has conditional expectation:
#
µie
log mij
−2
−2
−2
/
σ
+ 2
+
σ
+
σ
.
= de +
µe
η
ξ
2
σξ
σµe + ση2
"
2
η̂ij = E ηij | log mij , µie , σµe
(7)
With more experience, the prior belief about µie receives more weight in the conditional expectation
of worker-employer specific quality, and the noisy signal mij receives less weight. This is because
the total variance around µie declines. This variance is the sum of the actual variance around the
true mean, ση2 , and uncertainty about the mean through the variance of the prior at experience level
2
2
2
/∂e < 0, ση̂e
, also declines with experience. That is,
. Because ∂σµe
e, σµe
2
∂ση̂e
∂e
< 0.
2
The variance of the conditional expectation of worker-employer specific match quality, ση̂e
,
determines the expected order statistics in the employer’s choice set Ji . The employer’s demand is
a corner solution and is equivalent to choosing the worker j ∈ Ji that provides the highest value of:
J
2
log qj + E ηij | log mij , µie , σµe
+ εij − α log wij j=1 ,
where the parameter α scales the sensitivity of log wages relative to the variance of the idiosyncratic
error. Alternatively, if the expected wage-adjusted quality for this worker is less than the value of
the outside option of not hiring on the site, the employer does not make any job offers. With this
n
o
outside option, employer i’s choice set for the job is Ci = {qj , mij , εij , wij }Jj=1 , {0, 0, εi0 , 0} .10
10
Workers are assumed to be available when they initiate an application, which is reasonable given aggregate
statistics on potential unobserved availability. This requires that the probability that a worker receives two offers
over a time interval ∆ is sufficiently small. Similar justifications are used when formulating many stochastic processes.
For example, if, from the worker’s perspective, job offer arrivals follow a Poisson process, then the probability of
receiving a job offer in the interval (t, t + ∆) is λ∆ + o (∆) , and the probability of two offer arrivals in (t, t + ∆) is
o (∆). Although declined job offers are not observed, this assumption on arrival of offers seems to be reasonable given
data from two other measures. First, the observable arrival rate of interview requests fits this assumption. When the
worker-day is the interval of analysis, only 1.6% of the worker-days in the sample have more than 1 interview request
14
Each candidate anticipates that they will be hired only in the event that they are the high2
falls with experience afest order statistic after adjusting for wage bids, but the fact that ση̂e
fects employer i’s willingness to pay for the most preferred worker and, hence, affects the optimal
markup for all workers. To illustrate, we hold fixed log qj , α, εij , and log wij , while concentrating on
2
E ηij | log mij , µie , σµe
. In the normal-normal problem, the expected difference between the highest
2 11
. Conditional on being the highest
and second highest order statistic is inversely related to ση̂e
order statistic, the expected spread between oneself and the second best alternative is increasing
2
2
with ση̂e
. That is, an inexperienced employer, with a larger ση̂e
, evaluates the second preferred can-
didate to be a less good substitute for the preferred candidate. The expected preferred candidate
faces less competition from the second-most preferred worker and, hence, a more inelastic demand.
Equation (3) shows that when workers anticipate that employer demand is less elastic, they will
submit a higher wage bid.
To uncover how the components of the optimal bid function given in equation (3) differ between
jobs that are posted by inexperienced and experienced employers, we segment employers in the
data by experience level. Inexperienced employers, with e = 0 are grouped and labeled N , and
experienced employers, with e > 2 are grouped and labeled E. We estimate demand separately
2
for the subsamples of N and E employers. Splitting the sample allows us to recover ση̂N
for
2
inexperienced and ση̂E
for experienced employers, and to identify (dE −dN ), or the drift in valuation
2
with experience. However, we do not separately identify σξ2 and dynamics of σµe
. With this
approach, all that is required for identification is that the conditional expectation in equation (7)
becomes more precise and centered around the true (µi + de ) as employers gain experience.12 When
this is the case, experienced employers are less likely to hire based on noise, ξij , and are less likely
to search and experiment because of uncertainty about µi .
To complete the model, we introduce a time subscript, t, that is intended to make clear which
worker characteristics (or wage bids) may change with calendar time. We assume that the commonly
valued component of productivity has functional form log qj = Xjt β +υjt , where observable common
arriving; 0.15% of the worker days have more than 2 interview requests arriving. Second, a post-candidacy survey
asks employers for reasons why particular workers were not hired or workers for reasons why they exited the active
candidate set. In 0.02 percent of responses, employers or workers explicitly report a realized scheduling conflict. In
some cases, employers invite workers to apply for a job. In 8.7% of the instances when an employer initiates contact
with a worker, the worker reports that they are too busy to follow through on the invitation to apply or are not
interested in the job. Explicit rejections of application invitations are dropped when estimating demand.
11
This is found to be true in simulations with normally distributed random variables, however, we are not aware
of a rigorous proof.
12
A cost of this approach is that we are unable to credibly estimate the speed of employer learning. But the
robustness of this approach to any assumed specification—or mis-specification—of the functional form of the learning
model was judged to outweigh this drawback.
15
factors Xjt , including feedback, the workers’ past experience, and observable resume characteristics,
are ”priced” in the market. Market participants likely observe more information about common
quality than the econometrician, and thus υjt is likely to be correlated with the wage offered,
introducing an endogeneity problem to be addressed in demand estimation.
4.3
Demand Estimation
First Stage
To address concerns about the endogeneity of wage bids due to correlation with unobserved common
value (υjt ) or worker-employer match quality (that portion of log mij that is observed to workers
when bidding), we use an instrumental variables strategy that exploits changes in the local currency
to dollar exchange rates for individual workers and for their competitors from other countries.
Workers’ local labor market opportunities are indexed in the local currency. With frictions in
the offline labor market due to sticky offline wages or imperfect financial markets, local offline
opportunities are likely to adjust more slowly to exchange rate changes than online transactions.
When the local currency appreciates relative to the dollar (one dollar provides more local currency
units), workers’ wage offers when bidding for jobs falls, reflecting the reduction in dollar value of
their alternative labor market opportunities, and hence in their opportunity coss of online work
(part of cij in equation (3)). This change in the attractiveness of outside labor markets has a
substantive and statistically significant effect on workers’ wage bids.13
This exchange rate instrument induces substantial variation in wages, illustrated in Figure 5.
The vertical axis is the ratio of the monthly mean of Indian workers’ log bids and the monthly mean
of Filipino worker’s log bids, both in US dollars; the horizontal axis is the log of the Filipino to
Indian dollar exchange rate. The ratio of bids tracks the exchange rate ratio closely. The Philippines
and India have the largest shares of oDesk workers, so this source of variation is most important in
the data. Similar patterns are found across different countries. To make the instrument comparable
across countries, we take z-scores using the aggregate time series for each country over the sample
period.
The ideal instrument holds fixed the distribution of unobserved quality for applicant workers
and shifts relative prices among workers. This enables estimation using control function methods.14
13
Calendar time fixed effects remove country-invariant aggregate market changes.
Results from control function estimation may be sensitive to estimation error in the control function. The precision
of the first-stage estimates suggests that this problem is unlikely to be severe. Petrin and Train (2009) discuss the
properties of the estimator.
14
16
However, as is often the case with any cost-shifting instrument, market participation decisions are
likely to vary with the instruments as well as the decision of what wage to bid. Appendix 1 illustrates
the potential endogeneity of the applicant pool when there is sorting into market participation with
variation in the cost-shifting instrumental variable. To account for how differences in participation
change the distribution of unobserved worker quality, we use a second instrument that is a measure
of competition among workers. Because workers can see aggregate statistics about the prior bids
made for each job opening before making their bid, changes in observed bids by other workers
are likely to affect participation decisions. To capture this, we use other workers’ average bids
(aggregate) excluding own country. We then remove a monthly time fixed effect to ensure that only
cross-sectional variation is used for identification. This instrument picks up cross-country differences
in the intensity of competition from other workers. The instrument is correlated with participation
behavior, but other workers’ bids are plausibly uncorrelated with unobserved quality from the
perspective of employers after controlling for aggregate sources of time effects. With this additional
instrument, we estimate a participation equation at the monthly level using a probit model to correct
for any potential sorting based on the instrument. We include the inverse mills ratio estimated from
this probit model into the first stage when estimating the control function.
Table 3 provides details about the participation probit and the first stage regression. The results
make clear that the instruments are strong, however, the parameter estimate on the local exchange
rate is sensitive to having the inverse mills ratio included, suggesting that there is sorting on the
instrument, as described in Appendix 1.15
Choice Model
The probability that employer i with experience e ∈ {N, E} chooses worker j is the probability that
worker j has the highest expected wage-adjusted quality in the choice set Jie . This probability is:
2
log (qj ) + E ηij | log mij , µie , σµe
+ εij − α log (wij ) >
2
+ εik − α log (wik )
log (qk ) + E ηik | log mik , µie , σµe
(8)
for all k ∈ Ci . With the employer’s problem defined and a strategy to account for demand endogeneity, the estimated model must still account for heterogeneity across employers through (µi + de ) and
15
We have also estimated models that do not include the inverse mills ratio in the first-stage regression used to
obtain the control function. Both experienced and inexperienced employers are estimated to be less own-bid elastic
than in richer specifications with the inverse mills ratio included.
17
within-employer changing perceptions about η̂ij with experience. A modified random-parameters
logit model, estimated separately for N and E employers accomplishes both of these objectives.
The probability that worker j is hired for a job posted by employer i with experience e, conditional on η̂ij and µie the wage bid wij , and the worker-time level control function CFjt ψ is:
2
+ Xjt β − α log wij + CFjt ψ /
pj|i,η̂,µ = exp η̂ µi + de , σµe
2
1 + ΣJj i exp η̂ µi + de , σµe
+ Xjt β − α log wij + CFjt ψ .
(9)
Marginal choice probabilities for employer i and each potential choice j come from integrating over
the conditional distribution of ηije (µi ) and the distribution of µi :
Z Z
pij =
2
2
pj|i,η̂,µ f η̂ij |µi + de , ση̂e
f µi |µ̄e , σµe
dη̂ije dµi ,
(10)
2
2
where η̂ije ˜N µi , ση̂e
and µi ˜N µ̄e , σµe
, and:
"
2
η̂ije = E ηije | log mij , µi , σµe
#
µi
log mij
−2
=
+ 2
/ σξ−2 + σµe
+ ση−2 .
2
2
σξ
σµe + ση
The distribution of random parameters is identified because the outside option does not contain
ηije (µi + de ) or µi . The distribution of µi is identified in the data by variation in the attractiveness
of employers’ candidate pools as follows: When an employer with a very favorable candidate pool
chooses not to hire, part of this decision is due to the idiosyncratic Type-1 extreme value error and
part may be due to a low value of µi ; similarly, when an employer with an unfavorable candidate
pool chooses to hire, he may have a high µi or may have received a good draw of εij . The extent to
which employers with unfavorable pools hire and the extent to which employers with favorable pools
choose the outside option, hence, identify µ̄ for the inexperienced, the deviation from µ̄, captured as
2
drift de for the experienced, and σµ2 . Identification of ση̂e
comes from the extent to which employers
make choices that differ from the distribution of ranks implied by the extreme value error and the
parameters α, β, and ψ given log (wij ) , Xjt , and CFjt ψ.16
The parameters in equation (10) are estimated using just-identified simulated method of moments where the moment conditions exploit orthogonality between the control functions, character16
2
Note that ση̂e
is identified relative to the variance of the extreme value error, which is normalized for all al2
ternatives. Because the value of the outside option does not include η̂ij , ση̂e
is identified based on variation in
inside-alternative choices relative to the choice of the outside option.The identification argument is similar to the
identification of branch-wise variances in nested logit models.
18
istics, wage bids, and the difference between the actual choice and the simulated choice probabilities
2
2
. In the simulation procedure, MLHS draws for each employer
f µi |µ̄, σµe
under f η̂ije |µi + de , ση̂e
are generated for µi . These sequences are designed to provide good coverage. For each of these initial draws, a larger sequence of standard normal random variables are drawn for each applicant j,
allowing for the estimation of the distribution of η̂ije conditional on each draw of µi .
2
has an e subscript to reflect that the unbalanced
It is important to note that the variance σµe
nature of the panel may result in different moments of the latent distribution of valuations, but
this is not the prior variance for the agent in the learning problem above. This is the true latent
distribution of values around which noise ηije is centered. The integral in (10) is over this latent
distribution.
To simulate choice probabilities, R1 draws are used for each employers’ valuation µi , transformed
2
. R2 draws specific to each
to come from a normal distribution with mean µ̄ + de and variance σµe
2
alternative and each draw r1 are then transformed to have mean µri 1 + de , and variance ση̂e
. The
simulated choice probability for this set of parameters is:
2
pSj|i,η̂,µ = Σr1 Σr2 exp η̂ r2 µrie1 , σµe
+ Xjt β − α log wij + CFjt ψ /
2
1 + ΣJj i exp η̂ r2 µrie1 , σµe
+ Xjt β − α log wij + CFjt ψ .
(11)
Letting dij = 1 if worker j is chosen and zero otherwise, the simulated GMM objective function has
empirical moment conditions:
1 Ji
Σi Σj=0 dij − pSj|i,η̂,µ Instij
N
where Instij = [Xjt , log wij , CFjt ]0 for the inside option and Instij = 1 for the outside option. The
model is just-identified and the objective function with identity weighting matrix is minimized using
gradient-based techniques. The model is estimated separately for employers who have not made
any prior hires on the site (N employers) and for employers that have hired on two prior jobs in
any job category on their next job posting in Data Entry (E employers).17 This sample split allows
us to compare the conditional distribution of f (ui + de ) for the subsamples of N and E employers.
2
2
With estimates of demand parameters θ̂ = β, α, µ̄ + de , σµe
, ση̂e
, ψ in hand for both N and E
employers, we can return to the worker’s optimal bid function given in equation (3). In this mixed
We are also in the process of estimating models that allow for truncation of f ui |ū, σµ2 , E , as employers selectively use the platform with experience. The equation for pij is modified in this case such that the integrand is
Πe pj|i . We don’t report results for this case, as it is more restrictive, but truncation is found to be important.
17
19
logit model, where the marginal choice probability pij is given by equation (10), the term
∂pij
∂ log wij
in
equation (3) is:
∂pij
=
∂ log wj
Z Z
2
2
−αpj|i,η̂,µ 1 − pj|i,η,µ f η̂ij |µi , ση̂e
f µi |µ̄e , σµ(e)
dη̂ij dµi .
Given that τ is a fixed 10 percent, workers’ outside wages (opportunity costs) plus the hassle costs
of working on job i, cij , can be estimated using equation (2). Equation (3) shows that, with an
estimate of cij , each wage bid that employer i receives can be decomposed into a term that reflects
the worker’s marginal cost and a term that reflects the worker’s optimal mark up over marginal
cost.
A final parameter that has not been recovered is dE , drift in employers’ mean market valuation
with experience. Normalizing d0 = 0, dE is estimated as the difference in µ̄ for the experienced
sample (here, just X̄β, which includes a constant for the inside choices, time effects, and worker
characteristics) and the mean of the distribution of µi given that the employer hired a worker in
the inexperienced sample. Then,
dE = X̄βE − E (µiN |HireN , XjtN , log (wij ) , CFjt ) ,
(12)
where
R
f (µi |HireN , XjtN , log wij , CFjt ) = R R
2
i
Pr (dij |µi , η, Xjt , log (wij ) , CFjt ) dF (η) f µi |µ̄N , σµN
ΣJj=1
i
i
pSj|i,η̂,µ
Pr (dij |µi , η, Xjt , log (wij ) , CFjt ) dF (η) dF (µi ) ≈ ΣJj=1
ΣJj=1
(13)
and the summation is over only the inside-alternatives.
4.4
Demand Results
Table 4 presents the results for the demand estimation for two different samples of employers. The
first column present estimates for employers who have not hired before on the site; the model is
estimated using data from 2215 postings. Column 2 presents results for the subset of employers
posting their next job on the site in Data Entry after their second hire. This subset comprises 2042
employers. These estimates include the inverse mills ratio in the first stage. The important results
to note deal with differences in estimated coefficients for the two groups. As is apparent from Panel
A, α, differs substantially between groups. The standard deviation of the conditional expectation of
log mijt is also smaller for experienced employers, consistent with learning about the environment.
20
The standard deviation of the conditional expectation falls from 1.57 for inexperienced employers
to 1.33 for experienced employers. Possibly reflecting selective participation on the platform, the
standard deviation of µ also falls with experience.
Importantly, experienced employers also seem to experience drift in valuations, suggestion that
there is an important element of increasing platform value that comes with prior use. In particular,
for those who hire in the inexperienced sample, valuations increase with drift by 1.79 units. To
put this in context, the drift is equivalent to about 0.95 standard deviations of the original latent
distribution of µ, reflecting a significant increase in values for the platform with experience. In
addition, calculation of the drift implies that the posterior mean of µi0 for those that hire is 1.86,
compared to an overall mean of −0.17. Those who hire as inexperienced employers are a selected
sample from the overall distribution. This reflects that inexperienced employers include those who
learn the market is of limited value to them and exit after the first job.
These parameters also allow for a discussion of markups. The mean own-bid elasticity for
inexperienced employers is -6.91. The mean own-bid elasticity is greater for experienced employers,
at -11.01. However, with limited sample sizes, the 95% confidence intervals for the elasticities
overlap in models without requiring simulation. Standard errors for the simulated GMM estimates
are in progress because the control function is a generated regressor. The greater wage elasticity for
experienced employers leads workers to offer them lower markups. Despite inexperienced employers’
lower probability of hiring, the estimated average mark up over cost for inexperienced employers
is 16 percent, compared to 10 percent for experienced employers. This reflects differing elasticities
across these segments of employers.
The estimated marginal costs of applying for job openings also differs by employer experience.
In the data entry sample, the mean cost of working for an inexperienced employer is 3.01 USD
per hours (before the oDesk fee) compared to 2.60 USD per hour for in the experienced sample.
This difference suggests that the additional expected hassle cost of applying to an inexperienced
employer is about 0.41 USD, and, at the mean, this represents a difference in estimated costs of
about 15%. The final few rows of the table show that mean costs differ in plausible ways across
countries. DISCUSS MORE!
In the sample of data entry jobs considered here, the mean bid gap between experienced and
inexperienced employers is 23.3%. Thus, using a mean difference in costs of 15%, about 66% of the
gap is due to cost differences, while about 34% of the gap is explained by differences in demand
elasticities and markups. EXPLAIN MORE!
21
5
Empirical Support for Employer Learning
5.1
Employers’ Number of Interviews Falls with Experience
Turning to investigate demand differences between inexperienced and experienced employers, several
findings support the notion that employers learn about market value through experience. One of the
immediate observations from Table 1 is that the number of interviews falls, on average, on successive
job postings. This is consistent with employers forming more precise beliefs about the distribution
of worker matches from their experience interacting with workers on the site. Table 5 presents
the results of a regression of the number of interviews conducted on the job posting number and
employer fixed effects. The first column does not include employer fixed effects, while subsequent
columns include combinations of employer fixed effects, controls for qualitative opening features and
fixed effects for expected job duration, and controls for the mean log bid on the opening. Even with
different levels of controls, in all specifications with employer fixed effects, the number of interviews
decreases, at a decreasing rate, on successive jobs. Figure 6 expands on this regression by plotting
the coefficients in Column 2. The figure shows that the predicted number of interviews falls by
more than 50 percent by the fourth job posting.
5.2
Employer Outcomes
Futher evidence consistent with employers using early experiences to learn about the market comes
from the data on hiring outcomes. To fix ideas, consider a sequential search model where an
employer is trying to find a good employee match. Imagine, first, that the employer has fixed
beliefs about the distribution of match quality, denoted by F (m|θ). The optimal policy rule in
this case is a reservation threshold rule, where the employer hires the first worker interviewed
whose quality exceeds a threshold level. With a mean-preserving spread of F (m|θ), the reservation
threshold increases. The problem is slightly more complicated when F (m|θt ) evolves based on the
prior θt ˜G̃t (·). As shown in Kohn and Shavell (1974), the reservation match quality is greater with
uncertainty in Gt (·) compared to F (m|E(θt )). This implies that the expected reservation quality
falls with Bayesian updates about θ due to employer learning, with implications for the relationship
between the number of interviews conducted before hiring (search intensity) and job outcomes.
The data reveal that inexperienced employers in data entry or web research who interview more
candidates on their first job posting tend to have less successful outcomes for the posting. The
sample consists of inexperienced employers who do at least one interview on their first job posting.
22
All specifications include job category and expected job duration fixed effects, along with yearmonth indicators. Table 6 shows that employers who interview fewer candidates are more likely to
hire, more likely to report having had a successful hire, and more likely to give good feedback to
the employed worker.18
Column 1 shows that inexperienced employers who interview between 5 and 8 workers are 9
percentage points less likely to hire than employers who interview only one worker. Columns 2 and
3 show that they are 3 percentage points less likely to hire and report success and 5 percentage
points less likely to hire and give good feedback to the worker. Columns 4 to 7 look at inexperienced
employers who conduct at least one interview on their first job posting and who also hire on that
opening. Employers who do two to four, or five to seven, interviews are significantly less likely to
report success on the job than employers who do one interview, excluding and also including the
hourly wage paid on the job as a control. This result holds for the sample of all such employers and
for the subset of inexperienced employers whose first and second job posting can be segmented in
the data because the second job posting (if any) occurs after the last application for the first job.
Table 7 examines a similar result for employers who go on to post a second job. Using the variation in interviews on the first job, these results show that employers who conduct more interviews
on the first job are less likely to report success or give workers good feedback on the second job
posting. Overall, Tables 6 and 7 show that the number of early interviews is negatively related
to job success. This is consistent with the hypothesis that, on average, employers are using early
interviews to learn about the distribution of unobserved match quality. In order to forgo the opportunity to learn from additional early interviews, and hire after only one or a few interviews,
employers must have drawn a very high-quality employee on an early interview, leading to a good
job outcome on their employment. A model where employers are not using interviews to learn about
the distribution would not predict the negative relationship between search effort and outcomes.
6
Counterfactual Estimates: Platform Fees
We now turn to the issue of differential platform fees. We are interested in understanding whether
a different oDesk fee structure would improve platform profits. The current fees are ad-valorem,
which, in this setting, makes analyzing counterfactuals computationally difficult. Instead we focus
18
There is support for this hypothesis using the subcategories featuring the simplest tasks, Data Entry and Web
Research. There appears to be no relationship for other categories, and this seems to be driven by employers that
are searching for a personal assistant. We suspect that an interview for a personal assistant may convey additional
important information about horizontal match rather than the distribution of workers available on the platform.
23
on specific fees and illustrate the trade off that the platform faces in the simple situation where
only employers who hire on the first job can hire on the future jobs.
The platform’s objective is to maximize total profits, and it can do so with different fees on
inexperienced and experienced employers, respectively denoted tI and tE . The platform’s problem
is:
max Pr (HI |wI [tI ]) × [tI + tE × Pr (HE |HI (wI [tI ]) , wE [tE ])] .
where Pr (HI |wI [tI ]) is the probability that an inexperienced employer hires given wages wI [tI ]
and Pr (HE |HI (wI [tI ]) , wE [tE ]) is the probability an experienced employer hires as a function of
wages wE [tE ] conditional on the first hire, HI (wI [tI ]).19 Notice that the platform does not set
wages, only fees, but wages that employers face vary with platform fees unless demand is perfectly
elastic.20
When users are uncertain about platform valuation, Pr (HE |HI (wI [tI ]) , wE [tE ]) is not simply
the truncated part of the demand curve above wages wI but instead may change depending on
how employers’ beliefs about the platform evolve with experience. That Pr (HE |HI (wI [tI ]) , wE [tE ])
specifically conditions on HI (wI [tI ]) captures the possibility that experienced hiring may be affected
by the identity of the marginal inexperienced employer. Variation in wages, induced by different
platform fees, induces variation in the identity of the marginal employer.
This formulation so far says nothing about how beliefs evolve with employer experience. This
leaves the learning process free, allowing models with myopic or anticipated learning. In the actual
calculations, foresight is assumed about drift dE .21
Using HI as shorthand for Pr (HI |wI [tI ]) and HE as shorthand for Pr (HE |HI (wI [tI ]) , wE [tE ]) ,
the first order conditions are:
HI + tI
19
∂HI ∂wI
∂HI ∂wI
∂HE ∂HI ∂wI
+ tE HE
+ tE HI
= 0
dwI ∂tI
dwI ∂tI
∂HI dwI ∂tI
∂HE ∂wE
HE × HI + tE
HI = 0
∂wE ∂tE
With multiple future jobs, the objective function becomes
max [tI + tE × F (N HE |HI (wI [tI ]) , wE [tE ])] × Pr (HI |wI [tI ]) .
20
For this analysis, wages that inexperienced employers face (after the tax) have an I subscript and wages that
experienced employers face (after the tax) have an E subscript. Wages bid by the worker have a j subscript.
21
If employers anticipate that they will learn about their individual valuations for the platform, initial hiring may
reflect employers’ recognition of an option to no longer use the platform if the initial experience is unsuccessful. On
the other hand, myopic employers who have low expectations of platform valuation may require inducement if the
option value of gaining information is not recognized.
24
The solution to the system of equations sets the fee for experienced employers equal to the
monetary value of the optimal markup for a monopolist with zero marginal cost:
HE
t∗E = − ∂HE ∂wE .
(14)
∂wE ∂tE
The fee for the inexperienced is:
HI
∂HE
t∗I = − ∂HI ∂wI − t∗E HE − t∗E HI
.
∂H
I
dw ∂t
I
(15)
I
The first term in t∗I is the standard markup for the segment of inexperienced employers in the
absence of dynamic considerations. This markup is reduced by the latter two terms, which adjust
tI to account for the spillover to future demand. The final term is of particular interest, which
accounts for composition effects.
It is possible that estimates of most terms of interest in equations (14) and (15) come directly
from the logit models, but some additional statistics are required to compute optimal taxes: the
average of
∂HE
∂HI
and
∂wI
∂tI
and
∂wE
.
∂tE
In particular, estimating
∂HE
∂HI
is difficult without fully simulat-
ing counterfactuals under different fee structures. We conduct these simulations assuming that
employers’ anticipatory behavior doesn’t change between the existing and simulated equilibrium.
Before detailing the simulations to compute optimal fees, model-based estimates of the passthrough rate of fees are required because oDesk has not, to date, varied it’s fee structure. Differentiating the workers’ first order conditions to derive the equilibrium pass-through rate of the fee
yields
−2
∂wj
1
= −cj 1 +
∂t
(1 − pij ) α
!
∂pij
∂t
[(1 − pij ) α]2
α,
(16)
where α is the coefficient on log price from the conditional logit model, pij is unit demand by
employer i for worker j, and
∂pij
∂t
∂p
ij
i
= ΣJj=1
∂wj
∂wj
.
∂t
After imposing symmetry and solving for the fixed
point of this system, estimates of the average change in the probability of hiring with respect to
a tax on all workers suggests that the amount workers receive is little changed. Simulating this
change,
∂pij
∂t
≈ −.008; plugging in other parameters and estimates of pij suggest that workers’ bids
are reduced by about about 0.4 percent for a unit tax. Because
E
pass through rate to employers, ∂w
or
∂tE
∂wI
,
∂tI
∂wj
∂t
is the wage to the worker, the
is approximately 1.
Our estimates of the composition effects to obtain
25
∂HE
∂HI
come from simulating the posterior
density of valuations in equation (13) at different values of the tax. We then use this posterior
density as the density of experienced employers’ µie under tax tI .22 This gives us demand in the
experienced segment and revenue for the experienced segment under optimal tax t̂E |tI .
Table 8 presents some preliminary results. The first results show counterfactual profits in the
estimation sample in Table 4 if oDesk were to charge optimal fees in each segment assuming that
the number of job openings stays constant. This analysis does not take into account how these fees
would distort future participation and assume that the sample composition is given. The second set
of results suggests that a 10% tax on inexperienced employers is approximately optimal for platform
profits if oDesk charges an optimal tax on experienced employers. These simulations suggest that
oDesk should charge higher fees to experienced employers. The suggested specific fee is around 0.51
per hour, or about 18% of the mean wage bid. While this fee structure would reduce the number
of transactions and result in monopoly deadweight loss, under the assumption that experienced
employers who hire do so three times in the future, on average, platform profits would increase by
about 44%.
7
7.1
Robustness Analysis
Different Application Rates
It is possible that the extent of competition on a job posting changes with employer experience
because the arrival rate of applicants for a job varies with the employers’ hiring experience. Table
1 showed that inexperienced employers received a larger number of applicants, suggesting that, on
average, competition might be greater for employers’ early jobs. This might lead workers to bid
lower wages for these jobs, which runs counter to what is observed in the data. Nonetheless, it
remains possible that staggered application times cause the perceived nature of competition to be
greater when applications are bunched close to the time of job posting. Table 9 examines whether
application arrival rates differ between inexperienced and experienced employers. The omitted
group is the first job opening. Columns 1 and 2 regress the number of applications received in the
first two hours that a job is open on indicators for the job posting number; Columns 5 and 6 extend
the window for the first day. In the cross section, employers with at least one prior job posting
receive more applicants per unit of time than employers on the first job posting. These parameters
22
We also employ an adjustment factor such that the variance of the posterior density for µi0 |Hiring under the
2
existing tax equals σµe
.
26
are not statistically different from zero once employer fixed effects are included. In specifications
using the 2-hour window, arrival rates increase by less than 5 percent at the mean for employers
with at least five job openings relative to employers on the first job post. It hence seems unlikely
that differences in the intensity of competition affect worker bids in the data.
7.2
Omitted Employer Characteristics
Because the order of interview requests is observed in the data, it is possible to examine whether
inexperienced employers’ early actions on a job posting are correlated with later choices and outcomes. If this were the case, some of the results presented so far may be driven by unobserved,
predetermined, employer characteristics that shape employers’ search processes rather than their
experience and learning in the market. Figure 7a examines the distribution of the hourly wage bids
of the worker selected for the first interview for employers who go on to do fewer than five interviews
on the first job and also for employers that go on to do five or more interviews in total on the first
job. The figure plots the residual of a regression of the log of the hourly wage on job category and
year-month fixed effects. A comparison of the two distributions shows very little difference in the
choice of first interviewee for employers who go on to interview few or many applicants. Figure
7b repeats this exercise by whether the employer hires on the first job. These comparisons cast
doubt on the hypothesis that inexperienced employers eventual differences in interviewing or hiring
behavior are driven by unobservable differences in information or preferences at the time of the
initial job posting or at the time of selecting the first interview candidate. Even more important,
the overlap in wages on the initial interview request sent by employers suggests that workers are
not able to segment inexperienced employers based on the eventual number of interviews that they
conduct or by the probability that they will hire.
8
Conclusion
This paper documents that potential employers on oDesk.com who have no prior experience hiring
in this labor market receive higher wage bids from workers compared to similar employers with
observable hiring experience. Employers who have made at least five prior hires receive bids that
are, on average, around 10% lower than employers who have not previously hired. This finding is
robust to controlling for employer, worker and for employer and worker fixed effects. The wages
of the workers hired are also lower for experienced employers. The analysis presented in the paper
27
shows that workers’ optimal bids to inexperienced employers are higher because of two channels:
one on the demand side and one on the supply side.
On the demand side, inexperienced employers’ hiring probabilities are less elastic to the wage
bids offered because, at greater levels of uncertainty about match quality, the variance of the highest
expected order statistic of the quality distribution is larger. A marginal bid increase has a smaller
effect on the probability of being hired, resulting in higher optimal mark ups. Estimating market
demand for both inexperienced and experienced employers yields estimated wage elasticities of -7
and -11.5, respectively.
On the supply side, inexperienced employers are likely to create higher costs for workers when
applying for the job or when hired. For example, inexperienced employers conduct more interviews
and are less likely to hire, creating hassle costs for the workers with less expected benefit. Inexperienced employers are also more likely to need help understanding the mechanics of the site. These
greater costs appear to be passed through to inexperienced employers in higher wage bids.
Making the assumption that observed wage bids are the equilibrium outcome of a differentiated
products Bertrand game allows us to decompose the wage bid difference into these demand and
supply side effects. Differences in equilibrium markups due to differences in the wage elasticity of
demand account for an estimated 34% of the wage bid gap between inexperienced and experienced
employers. The residual—or 66% of the gap—can be attributed to differences in workers’ costs.
Further empirical findings suggest that experience on the site enables employer learning, which
corroborates the hypothesis that inexperienced employers’ relative uncertainty about match quality
explains a share of wage bid differences. Inexperienced employers are shown to interview more
candidates but have job outcomes that are negatively associated with the number of interviews
done prior to hiring. A possible explanation for this somewhat counter-intuitive finding is that an
interview is valuable for both the information it provides about the quality of the interviewed worker,
and also about the distribution of match quality among all workers. When employers have greater
uncertainty about this distribution, an interview is more valuable. Opting to stop interviewing and
hire an interviewed worker, rather than conduct more interviews, suggests that the revealed match
quality of that worker is very high. That is, those that hire early have found good quality workers,
which explains their good job outcomes.
Showing that both demand and supply differ in the market faced by inexperienced and experienced employers allows us to find the platform profit-maximising fee that the platform should
charge each segment of employers. On the one hand, inexperienced employers’ more inelastic de-
28
mand would lead oDesk to impose a higher fee on these employers but, on the other hand, a lower
tax on inexperienced employers will increase the total number of transactions on the site in the
future, as these employers are more likely to remain in the market. The net effect is that oDesk
could increase profits by setting a fee of 1.05 dollars per hour to experienced employers and a much
lower fee to inexperienced employers.
Whether or not potential trading partners have experience transacting in a market is observable
in many kinds of markets, including many online product markets. This paper finds evidence that
employer (buyer) experience is a salient characteristic for workers (suppliers) in the oDesk labor
market because experience affects both the nature of employer demand and the cost of worker
supply.
Appendix
Appendix 1: Sorting into participation with variation in the IV.
To see how endogenous worker oDesk participation decisions affect the ability to estimate demand,
suppose the instrument using local exchange rates is denoted Zjt and the first stage regression is
log wij = a1 + a2 eit + Zjt γ1 + Xjt γ2 + ujt .
(17)
The residual in this regression contains worker and match quality, both of which enter the optimal
bid with unknown coefficients λ1 and λ2 , resulting in
ujt = υj λ1 + (ηijt + ξij ) λ2 + νjt .
For exposition purposes, we ignore correlation between Zjt and (ηijt + ξij ) , but illustrate the consequences of correlation between Zjt and common unobserved quality υjt . If Cov (Zjt , υjt ) 6= 0, as
would be the case if an exchange rate shift changed the participation decision of a non-random subset of workers from the distribution of unobserved quality, then the exclusion restriction is violated.
Consistent estimation in the control function setting requires that ujt contains any part of unobserved quality that is correlated with price, as fitted values of γ̂ are uncorrelated with unobserved
υjt by construction. When the exclusion restriction fails,
E (γ̂1 ) = γ1 + E
n
0
Zjt
Zjt
29
−1
o
0
Zjt
υj λ1 ,
which results in residuals for a control function that contain a part of unobserved quality
o
n
−1 0
0
Zjt (ξj ) λ1
CF = log (wijt ) − Zjt γ1 + Zjt Zjt
= log (wijt ) − Zjt γ1 − PZ υjt λ1
= νjt + (I − PZ ) υj λ1 + (ηijt + ξij ) λ2 .
In this case, due to inducing sorting into participation, the instrument does not eliminate endogeneity concerns. With sorting, only the part of unobserved quality orthogonal to the projection
matrix of υjt is absorbed in the residual.
30
References
[1] Ackerberg, Daniel. 2003. ”Advertising, learning, and consumer choice in experience good markets: an empirical examination.” International Economic Review, 44(3), 1007-1040.
[2] Agrawal, Ajay, Christian Catalini and Avi Goldfarb. 2011. ”Friends, Family, and the Flat
World: The Geography of Crowdfunding.” NBER Working Paper No.16820.
[3] Agrawal, Ajay and Avi Goldfarb. 2008. ”Restructuring Research: Communication Costs and
the Democratization of University Innovation.” American Economic Review, 98(4), 1578-1590.
[4] Agrawal, Ajay, Nicola Lacetera, and Elizabeth Lyons. 2012. ”How Do Online Platforms
Flatten Markets for Contract Labor?” Working Paper.
[5] Agrawal, Ajay, John Horton, Nicola Lacetera, and Elizabeth Lyons. 2013. ”Digitization and
the Contract Labor Market: A Research Agenda.” NBER Working Paper w19525.
[6] Akerlof, George. A. 1970. The market for” lemons”: Quality uncertainty and the market mechanism. The Quarterly Journal of Economics, 488-500.
[7] Altonji, Joseph, and Charles Pierret. 2001. ”Employer Learning and Statistical Discrimination.” The Quarterly Journal of Economics, 116(1), 313-350.
[8] Arcidiacono, Peter, Esteban Aucejo, Arnaud Maurel, and Tyler Ransom. 2013. ”College Attrition and the dynamics of information revelation.” Working Paper
[9] Autor, David. 2001. ”Wiring the Labor Market.” Journal of Economic Perspectives, 15(1),
25-40.
[10] Bajari, Patrick and Ali Hortaçsu. 2004. ”Economic Insights from Internet Auctions.” Journal
of Economic Literature, 42(2), 457-486.
[11] Bergemann, Dirk and Juuso Valimaki. 2006. ”Dynamic Pricing of New Experience Goods.”
Journal of Political Economy, 114(4), 713-743.
[12] Cabral, Luis, and Ali Hortacsu. 2010. ”The dynamics of seller reputation: Evidence from
ebay*.” The Journal of Industrial Economics, 58.1, 54-78.
31
[13] Erdem, Tulin, and Michael Keane. 1996. ”Decision-making Under Uncertainty: Capturing
Dynamic Brand Choice Processes in Turbulent Consumer Goods Markets.” Marketing Science,
15(1), 1-20.
[14] Farber, Henry and Robert Gibbons. 1996. ”Learning and Wage Dynamics.” The Quarterly
Journal of Economics, 111(4), 1007-1047.
[15] Ghani, Ejaz, William Kerr, and Christopher Stanton. 2014. ”Diasporas and outsourcing:
Evidence from oDesk and India.” Management Science, 60(7),1677-1697.
[16] Greif, Avner. 1993. ”Contract Enforceability and Economic Institutions in Early Trade: The
Maghribi Traders’ Coalition,” American Economic Review, 83(3), 525-548.
[17] Horton, John. 2010. ”Online Labor Markets.” Working Paper.
[18] Horton, John. 2013. ”The Effects of Subsidizing Employer Search.” Working Paper.
[19] Jin, Ginger and Andrew Kato. 2006. ”Price, Quality and Reputation: Evidence from An Online
Field Experiment.” The RAND Journal of Economics, 37(4), 983-1005.
[20] Kahn, Lisa. 2013. ”Asymmetric Information between Employers.” American Economics
Journal: Applied.
[21] Kahn, Lisa and Fabian Lange. 2014. ”Employer Learning, Productivity, and the Earnings Distribution: Evidence from Performance Measures.” Review of Economic Studies (Forthcoming).
[22] Lange, Fabian. 2007. ”The Speed of Employer Learning.” Journal of Labor Economics, 25(1),
1-35.
[23] Lewis, Gregory. 2011. ”Asymmetric Information, Adverse Selection and Online Disclosure: The
Case of eBay Motors.” American Economic Review, 101(4), 1535-1546.
[24] Luca, Michael. 2010. ”Reviews, Reputation, and Revenues: The Case of Yelp.com.” Harvard
Business School Working Paper No.13-042.
[25] Lyons, Elizabeth. 2014. ”Team Production in International Labor Markets: Experimental Evidence from the Field.” Working paper.
[26] Miller, Robert. 1984. ”Job matching and occupational choice.” Journal of Political Economy,
92(6), 1086-1120.
32
[27] Pallais, Amanda. 2013. ”Inefficient Hiring in Entry-Level Labor Markets.” Working paper.
[28] Stanton, Christopher and Catherine Thomas. 2012. ”Landing the First Job: The Value of
Intermediaries in Online Hiring.” Working Paper.
33
Figure 1a: Number of job postings per employer, truncated at 30.
Figure 1b: Number of hires per employer in the truncated sample
Figure 2A: Bid Decile of Employed Worker, Conditional on Hiring and having 5 or more applicants
Figure 2B: Bid Decile of Employed Worker, Conditional on Hiring and 5 or more applicants
Bid residual after controlling for job category and observable worker characteristics (country, feedback, education, english)
15
Residual Wage Bid
5
10
0
0
.005
Probability of Hire
.01
0 Prior Hires
2+ Prior Hires
Filipino Workers in Data Entry.
Figure 3: Wage Bids and Hiring Probabilities by Employer Experience. Points
are taken from a polynomial smooth of an indicator that the worker was hired on
their residual bid. The residuals remove time …xed e¤ects and priced observable
characteristics.
.015
8
Mean Bid for Inexperienced Employers
2
4
6
0
2
4
6
M ean Bid for Experienced Em ployers
Binned ScatterPlot of Bids
8
45-degree Line
Within-worker comparison of mean bids in data entry
Figure 4: Wage bids by observed employer hiring experience, calculated within
worker-month for workers who apply to both experienced and inexperienced
employers.
1.02
1.2
.99
1
Inverse Exchange Rate
1.01
1.15
Bid Ratio
1.1
1.05
.98
1
2008m 1
2008m 7
2009m 1
Log India to Phil Bid Ratio
2009m 7
2010m 1
2010m 7
Log Phil over Indian Ex
Figure 5: Ratio of Country-Level Mean Log Wage Bids and Log Exchange Rates
Figure 6: Number of interviews on successive job openings (DV is the log of the number of interviews +1)
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0
1
2
3
4
5
Notes: The series is the predicted number of interviews per posting on successive jobs calculated from the estimated constant term and coefficient estimates from a regression of job posting fixed effects on the total number of interviews. The estimation contains employer fixed effects and year‐month fixed effects. All coefficients are significantly different from zero. The sample is job postings where the prior job is complete by the time of posting.
Figure 7: Residual Hourly Bid made by first applicant selected for interview, after JC2 and YM
7a: Log Bids for employers who conduct fewer than five and five or more interviews on the first posting.
7b: Log bids ror employers who don't hire and who hire on the first posting.
Table 1A: Summary Statistics
On Postings where a Hire was made
Number of Candidates
Share of Employer‐
Initiated Candidates (1)
(2)
0
16903
1
4852
2
3583
3
2730
4+
27139
34.51
(59.56)
27.20
(59.31)
27.11
(67.82)
26.65
(60.08)
23.55
(67.14)
Previous Number of Hires
Job Openings
Mean Wage Bid of Hired Median Wage Worker Bid
Share with Share with Good Worker Missing Worker Feedback
FB
Mean Wage Bid Number of Interviews
Probability a Hire is Made
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
0.17
(0.35)
0.25
(0.42)
0.26
(0.43)
0.24
(0.42)
0.26
(0.43)
5.01
(3.09)
4.41
(3.09)
4.14
(3.05)
3.94
(2.89)
3.56
(3.08)
3.77
(8.12)
2.54
(5.50)
2.32
(5.37)
2.10
(4.48)
1.94
(5.07)
0.24
(0.43)
0.53
(0.50)
0.54
(0.50)
0.57
(0.50)
0.57
(0.50)
4.14
(4.00)
3.72
(3.29)
3.55
(3.34)
3.35
(3.12)
3.04
(3.45)
3.33
43%
36%
2.78
46%
32%
2.78
46%
30%
2.22
45%
31%
2.22
50%
28%
Table 1B: Summary Statistics for Sequential Openings
On Postings where a Hire was made
Number of Candidates
Share of Employer‐
Initiated Candidates (1)
(2)
0
10664
1
3337
2
2576
3
1963
4+
20365
35.68
(58.64)
28.29
(61.56)
27.48
(67.18)
27.06
(60.11)
24.82
(68.94)
Previous Number of Hires
Job Openings
Mean Wage Bid of Hired Median Wage Worker Bid
Share with Share with Good Worker Missing Worker Feedback
FB
Mean Wage Bid Number of Interviews
Probability a Hire is Made
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
0.21
(0.38)
0.28
(0.44)
0.30
(0.44)
0.27
(0.43)
0.28
(0.44)
5.19
(3.18)
4.49
(3.15)
4.21
(3.09)
4.00
(2.89)
3.56
(3.07)
3.89
(8.02)
2.57
(5.46)
2.37
(5.29)
2.15
(4.53)
1.98
(5.12)
0.24
(0.43)
0.53
(0.50)
0.53
(0.50)
0.56
(0.50)
0.56
(0.50)
4.32
(4.16)
3.83
(3.44)
3.62
(3.49)
3.46
(3.21)
3.04
(3.52)
3.33
44%
35%
2.78
46%
32%
2.78
47%
29%
2.50
46%
30%
2.22
51%
28%
Notes: Sample period is from January 2006 to June 2010, with most of the openings posted after 2008. Wage bids are winsorized at the 99th percentile, and applications closed by oDesk as likely spam are excluded. Standard deviations are given in parentheses.
Table 1C: Summary Statistics for Data Entry Job Postings
On Postings where a Hire was made
Number of Job Openings
Number of Candidates
Share of Employer‐
Initiated Candidates (1)
(2)
0
7098
1
1478
2
1161
3
913
4+
10757
49.47
(80.28)
47.15
(93.31)
47.41
(104.67)
43.16
(86.29)
33.50
(94.31)
Previous Hires
Mean Wage Bid Median Wage of Hired Worker Bid
Share with Share with Good Worker Missing Worker Feedback
FB
Mean Wage Bid Number of Interviews
Probability a Hire is Made
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
0.15
(0.34)
0.25
(0.42)
0.27
(0.44)
0.26
(0.43)
0.28
(0.44)
3.96
(2.31)
3.32
(2.10)
3.13
(2.03)
2.97
(1.89)
2.49
(2.09)
4.08
(9.79)
2.49
(5.65)
2.24
(5.29)
2.02
(4.54)
1.72
(4.87)
0.18
(0.38)
0.57
(0.50)
0.56
(0.50)
0.60
(0.49)
0.63
(0.48)
3.07
(2.96)
2.76
(2.09)
2.70
(2.20)
2.37
(1.78)
2.12
(2.20)
2.22
48%
33%
2.22
51%
29%
2.22
50%
27%
2.00
49%
29%
1.67
55%
25%
Table 1D: Summary Statistics for Sequential Openings for Data Entry Job Postings
On Postings where a Hire was made
Number of Job Openings
Number of Candidates
Share of Employer‐
Initiated Candidates (1)
(2)
0
4246
1
1028
2
854
3
652
4+
8364
54.07
(82.09)
50.83
(97.08)
47.26
(102.21)
45.41
(87.07)
34.07
(94.41)
Previous Hires
Mean Wage Bid Median Wage of Hired Worker Bid
Share with Share with Good Worker Missing Worker Feedback
FB
Mean Wage Bid Number of Interviews
Probability a Hire is Made
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
0.19
(0.38)
0.29
(0.44)
0.31
(0.46)
0.30
(0.45)
0.31
(0.46)
3.95
(2.32)
3.36
(2.03)
3.18
(2.07)
3.10
(1.93)
2.48
(2.04)
4.55
(10.25)
2.54
(5.48)
2.14
(4.28)
2.11
(4.59)
1.70
(4.81)
0.19
(0.39)
0.55
(0.50)
0.56
(0.50)
0.59
(0.49)
0.62
(0.48)
3.20
(3.20)
2.84
(2.20)
2.79
(2.33)
2.48
(1.83)
2.12
(2.12)
2.22
49%
32%
2.22
52%
28%
2.22
51%
27%
2.22
49%
27%
1.67
55%
24%
Notes: Sample period is from January 2006 to June 2010, with most of the openings posted after 2008. Wage bids are winsorized at the 99th percentile, and applications closed by oDesk as likely spam are excluded. Standard deviations are given in parentheses.
Table 2: Log Wage Bids Decline with Observable Employer Experience
Employer Fixed Effects
Worker Fixed Effects
Worker Fixed Effects
Employer and Worker Fixed Effects
Employer and Worker Fixed Effects
OLS
OLS
Employer Fixed Effects
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
‐0.0816***
(0.00972)
‐0.0930***
(0.0112)
‐0.109***
(0.0124)
‐0.135***
(0.0179)
‐0.165***
(0.0122)
‐0.0875***
(0.00953)
‐0.0991***
(0.0111)
‐0.117***
(0.0123)
‐0.142***
(0.0175)
‐0.170***
(0.0120)
‐0.0456***
(0.0154)
‐0.0635***
(0.0158)
‐0.0695***
(0.0172)
‐0.0887***
(0.0227)
‐0.0997***
(0.0171)
‐0.0424***
(0.0150)
‐0.0618***
(0.0153)
‐0.0702***
(0.0169)
‐0.0868***
(0.0224)
‐0.0963***
(0.0169)
‐0.0357***
(0.00200)
‐0.0457***
(0.00221)
‐0.0538***
(0.00245)
‐0.0667***
(0.00284)
‐0.0909***
(0.00155)
‐0.0380***
(0.00199)
‐0.0482***
(0.00221)
‐0.0569***
(0.00245)
‐0.0694***
(0.00284)
‐0.0928***
(0.00155)
‐0.0151***
(0.00723)
‐0.0311***
(0.00786)
‐0.0417***
(0.00992)
‐0.0506***
(0.01856)
‐0.0605***
(0.00884)
‐0.0135*
(0.01196)
‐0.0303***
(0.00546)
‐0.0421***
(0.00983)
‐0.0497***
(0.01189)
‐0.0587***
(0.01066)
No
982,799
0.168
Yes
982,799
0.170
No
982,799
0.281
Yes
982,799
0.282
No
982,799
0.710
Yes
982,799
0.711
No
982,799
0.752
Yes
982,799
0.753
‐0.163***
(0.0116)
‐0.254***
(0.0283)
0.138***
(0.0286)
0.0870***
(0.0277)
‐0.125***
(0.0265)
0.0393*
(0.0236)
0.0198
(0.0198)
‐0.122***
(0.0263)
0.0406*
(0.0238)
0.0189
(0.0198)
‐0.146***
(0.0198)
0.0923***
(0.0200)
0.0585***
(0.0195)
‐0.147***
(0.0197)
0.0912***
(0.0199)
0.0570***
(0.0193)
‐0.0731***
(0.0188)
0.0255
(0.0166)
0.0107
(0.0129)
‐0.0713***
(0.0106)
0.0265**
(0.0142)
0.0104
(0.0111)
No
753,131
0.176
Yes
753,131
0.181
No
753,131
0.299
Yes
753,131
0.300
No
753,131
0.728
Yes
753,131
0.728
No
753,131
0.770
Yes
753,131
0.770
‐0.163***
(0.0116)
‐0.208***
(0.0224)
0.0418**
(0.0194)
0.0556**
(0.0283)
‐0.141***
(0.0256)
0.0453***
(0.0148)
0.0458*
(0.0266)
‐0.137***
(0.0254)
0.0449***
(0.0149)
0.0448*
(0.0269)
‐0.122***
(0.0150)
0.0354***
(0.0129)
0.0429**
(0.0175)
‐0.125***
(0.0149)
0.0363***
(0.0130)
0.0436**
(0.0178)
‐0.0934***
(0.0154)
0.0406***
(0.0070)
0.0333**
(0.0114)
‐0.0912***
(0.0177)
0.0404***
(0.0085)
0.0327
(0.0134)
No
753,131
0.176
Yes
753,131
0.179
No
753,131
0.299
Yes
753,131
0.300
No
753,131
0.727
Yes
753,131
0.727
No
753,131
0.770
Yes
753,131
0.770
Panel A
On posts after making 1 hire
2 hires
3 hires
4 hires
5+ hires
Third order polynomial of job description length
Observations
R‐Squared
Panel B
On posts after making 5+ hires
5+ hires and no observable employer feedback
5+ hires and good observable employer feedback
Third order polynomial of job description length
Observations
R‐Squared
Panel C
On posts after making 5+ hires
5+ hires and no worker feedback given
5+ hires and good worker feedback given
Third order polynomial of job description length
Observations
R‐Squared
Dependent variable is the log of the hourly wage bid. The sample is limited to worker‐initiated applications on sequential job openings. Robust standard errors are clustered by employer. Columns 7 and 8 contain bootstrapped standard errors with resampling over employers. All specifications contain the worker's feedback score, an indicator for zero feedback, number of jobs completed, an indicator for past work and zero feedback, and fixed effects for expected duration of job, expected duration of job interacted with hours per week, time, and detailed job category. The sample in Panels B and C is employers' first jobs and all jobs after their fourth hire. Table 3: First Stage Regression of Log Hourly Bids and Participation on Exchange Rate Instruments (Data Entry Openings)
Dependent Variable
Model
Sample
Log Local Currency to Dollar (monthly; standardized)
Log Hourly Wage Bid (Job Opening Level)
Linear Regression
Experienced Experienced First Job Post
First Job Post
Sample
Sample
(1)
(2)
(3)
(4)
(5)
(6)
‐0.0501***
(0.00536)
‐0.0503***
(0.00565)
‐0.0527***
(0.00525)
‐0.0458***
(0.00575)
0.0788***
(0.00509)
0.102***
(0.00864)
‐0.0649***
(0.00861)
0.149***
(0.00839)
0.108***
(0.00866)
‐0.0147
(0.0101)
0.431***
(0.0468)
0.0659***
(0.00419)
0.0662***
(0.00691)
‐0.0680***
(0.00726)
0.153***
(0.00785)
0.0777***
(0.00705)
0.00505
(0.00943)
0.452***
(0.0390)
0.0688***
(0.037)
‐3.021***
(0.0395)
0.5779***
(0.003)
0.062***
(0.007)
0.5975***
(0.009)
0.0939***
(0.042)
‐1.637***
(0.0447)
0.560***
(0.0027)
0.100***
(0.0072)
0.5837***
(0.0087)
54,122
0.212
606.2
54,122
0.213
586.8
73,056
0.169
101.1
73,056
0.171
33
1,837,515
0.3007
1,475,701
0.3068
Mean Bid from Workers in Other Countries (Aggregate)
Feedback Score (Out of 5)
Agency Affiliate Indicator
Prior Work Experience
Monthly Participation Inverse Mills Ratio
Number of Observations
R‐Squared
F Statistic on Excluded Instruments
1+ application during the month
Probit
Experienced First Job Post
Sample
Notes: The inexperienced sample is employers on their first job post. The experienced sample is employers who have hired 2 or more previous workers from any job category and have posted at least 2 previous jobs in adminstrative support. Robust standard errors in parentheses. All models contain a fifth order polynomial in calendar time, fixed effects for 6 country groups, controls for English skills, and an indicator for having zero feedback. The last country group includes many countries with very small application shares. Models in columns 1 ‐ 4 also include a piecewise‐linear spline with 4 knots for the application number, an indicator for an employer‐initiated application, and an indicator that the worker only applies to this job during the month. The Log Local Currency to Dollar exchange rate is calculated using monthly data and z‐scores from the aggregate macro data are used to make the measure comparable across countries. The inverse mills ratio in columns 2 and 4 is taken from columns 5 and 6. Other workers' average bids (aggregate) in the probit models in columns 5 and 6 are first calculated excluding own‐country and then a monthly time fixed effect is removed. A separate interaction for workers in the United States and the basket of other currencies (not reported) because these workers do not have any variation in own exchange rates. Table 4: Mixed Logit Parameter Estimates, Elasticities, Costs, and Markups (1)
(2)
Data Entry Job Post 1
Experienced Employers
Log Hourly Bid
‐6.95
‐11.12
Control Function
6.29
10.18
Std Dev of Random Effect (Std Dev of E(η))
1.57
1.33
Average Mean of μ
‐0.17
3.65
Panel A: Parameter Estimates
Sample:
Drift 1.79
Std Dev of Permanent Heterogeneity ( σμ )
Number of Job Openings
Probability of Inside Hire (Data)
Probability of Inside Hire (Model)
Mean Wage Bid (Pre‐oDesk‐Fee)
1.84
1.75
2215
0.1968
0.1974
3.55
2042
0.3433
0.3434
2.88
Panel B: Estimated Elasticities, Markups and Costs
Mean Own‐Price Elasticity
‐6.91
‐11.01
Mean Lerner Markup, Pre‐oDesk‐Fee
Mean Cost (USD, Pre‐Fee)
Median Cost (USD, Pre‐Fee)
0.16
3.01
2.11
0.10
2.61
1.83
Mean Cost in India (USD, Pre‐Fee)
Mean Cost in the Philippines (USD, Pre‐Fee)
Mean Cost in the United States (USD, Pre‐Fee)
2.87
2.43
5.74
2.69
2.20
5.05
Notes: The sample is employers on their first job posting (column 1) or on their job postings after having hired 2 previous workers (column 2). Employers must have also had 5 or more worker‐initiated applications to be included in the sample. Employers conducting more than 110 interviews were dropped. Mixed logit model estimates, costs and markups are estimated via GMM as described in the text. Drift is estimated as the difference in the mean of the density of mu for experienced employers and the mean of the conditional density of mu for those who hire on job 1. Table 5: Log interviews per job opening fall with hiring experience
DV: Log Number of Interviews +1
One previous hire
Two previous hires
Three previous hires
Four previous hires
Five or more previous hires
OLS
Employer Effects
Employer Effects
Employer Effects
Employer Effects
(1)
(2)
(3)
(4)
(5)
‐0.0988***
(0.0198)
‐0.142***
(0.0208)
‐0.162***
(0.0221)
‐0.169***
(0.0234)
‐0.244***
(0.0194)
‐0.503***
(0.0524)
‐0.601***
(0.0610)
‐0.674***
(0.0673)
‐0.735***
(0.0711)
‐0.883***
(0.0806)
‐0.501***
(0.0525)
‐0.601***
(0.0611)
‐0.674***
(0.0673)
‐0.733***
(0.0711)
‐0.879***
(0.0808)
0.826***
(0.0400)
1.489***
(0.0894)
1.446***
(0.0892)
‐0.491***
(0.0531)
‐0.587***
(0.0618)
‐0.656***
(0.0682)
‐0.718***
(0.0719)
‐0.864***
(0.0818)
0.117***
(0.0122)
1.388***
(0.0919)
‐0.490***
(0.0531)
‐0.587***
(0.0618)
‐0.657***
(0.0683)
‐0.717***
(0.0720)
‐0.860***
(0.0819)
0.113***
(0.0120)
1.350***
(0.0917)
No
38,868
0.041
No
38,868
0.557
Yes
38,868
0.559
No
38,864
0.560
Yes
38,864
0.562
Mean Log Bid
Constant
Includes job duration fixed effects and third order polynomial of job description length
Observations
R‐Squared
Notes: Robust standard errors are clustered by employer. All specifications contain year‐month fixed effects as well as job category and job duration fixed effect.
Table 6: Productivity and Search Effort on the First Job, for employers who interview on first job
Data Entry and Web Research Sample
Hires on Openings with 1+ Interview
Success, also Success, also add Controls add Wage for Worker Control
Characteristics
Hires with Good Feedback
Success
Hires a Worker
Hires and Reports Success
Success, Sequential Openings
(1)
(2)
(3)
(4)
(5)
(6)
(7)
‐0.032*
(0.017)
‐0.051**
(0.023)
‐0.090***
(0.032)
‐0.185***
(0.027)
‐0.021
(0.016)
‐0.053***
(0.019)
‐0.075***
(0.024)
‐0.109***
(0.020)
‐0.025
(0.015)
‐0.044**
(0.021)
‐0.065**
(0.025)
‐0.102***
(0.025)
‐0.047
(0.030)
‐0.094**
(0.037)
‐0.129**
(0.058)
‐0.015
(0.053)
‐0.028
(0.031)
‐0.107**
(0.042)
‐0.154**
(0.066)
0.006
(0.063)
‐0.025
(0.031)
‐0.100**
(0.041)
‐0.151**
(0.066)
0.013
(0.062)
‐0.025
(0.031)
‐0.102**
(0.041)
‐0.151**
(0.066)
0.010
(0.063)
No
0.333
5,144
0.085
No
0.208
4,829
0.098
No
0.188
4,634
0.076
No
0.708
1,420
0.108
No
0.745
1,198
0.111
No
0.745
1,198
0.114
No
0.745
1,198
0.115
‐0.070***
(0.023)
‐0.107***
(0.030)
‐0.158***
(0.041)
‐0.248***
(0.035)
‐0.041**
(0.020)
‐0.084***
(0.026)
‐0.123***
(0.028)
‐0.153***
(0.026)
‐0.060***
(0.020)
‐0.085***
(0.027)
‐0.124***
(0.029)
‐0.162***
(0.028)
‐0.035
(0.036)
‐0.074
(0.046)
‐0.147**
(0.067)
0.009
(0.074)
‐0.001
(0.043)
‐0.038
(0.059)
‐0.117
(0.077)
0.069
(0.079)
0.004
(0.043)
‐0.025
(0.060)
‐0.116
(0.077)
0.084
(0.075)
0.003
(0.042)
‐0.029
(0.059)
‐0.114
(0.076)
0.079
(0.077)
Yes
0.347
4,212
0.145
Yes
0.218
3,950
0.161
Yes
0.198
3,778
0.136
Yes
0.709
1,217
0.222
Yes
0.747
1,033
0.231
Yes
0.747
1,033
0.236
Yes
0.747
1,033
0.239
Panel A
2‐4 Interviews
5‐8 Interviews
9‐11 Interviews
12+ Interviews
Includes Buyer Group Fixed Effects
Mean of DV
Observations
R‐Squared
Panel B
2‐4 Interviews
5‐8 Interviews
9‐11 Interviews
12+ Interviews
Includes Buyer Group Fixed Effects
Mean of DV
Observations
R‐Squared
Notes: Panel B includes fixed effects for groups of employers who interview workers with similar characteristics on the first job. Standard errors clustered by detailed job category x time. All specifications contain fixed effects for expected duration of job, time, and detailed job category. Columns 1‐3 include all first openings. Columns 4 to 7 include all first openings where a hire is made. Table 7: Productivity on the Second Job and Search Effort on the First Job, for employers who interview on the first job.
Data Entry and Web Research Sample
Gives Good Gives Good Reports Success, Feedback, Reports Success, Feedback, Sequential Openings Sequential Openings Sequential Openings Sequential Openings and Wage Control
and Wage Control
(3)
(4)
(5)
(6)
Reports Success
(1)
Gives Good Feedback
(2)
‐0.066**
(0.029)
‐0.097**
(0.047)
‐0.076
(0.047)
‐0.142***
(0.050)
‐0.092**
(0.035)
‐0.191***
(0.039)
‐0.052
(0.055)
‐0.106**
(0.046)
‐0.114**
(0.047)
‐0.186***
(0.066)
‐0.038
(0.055)
‐0.251***
(0.071)
‐0.147***
(0.049)
‐0.232***
(0.067)
0.003
(0.070)
‐0.220**
(0.085)
‐0.111**
(0.047)
‐0.180***
(0.067)
‐0.034
(0.058)
‐0.241***
(0.068)
‐0.145***
(0.049)
‐0.224***
(0.068)
0.009
(0.068)
‐0.206**
(0.081)
No
0.614
1,462
0.095
No
0.687
1,221
0.060
No
0.656
610
0.149
No
0.708
518
0.138
No
0.656
610
0.151
No
0.708
518
0.142
‐0.055
(0.043)
‐0.081
(0.063)
‐0.072
(0.063)
‐0.110*
(0.066)
‐0.068
(0.055)
‐0.170***
(0.060)
‐0.072
(0.078)
‐0.106*
(0.063)
‐0.084
(0.082)
‐0.129
(0.107)
‐0.109
(0.107)
‐0.186*
(0.110)
‐0.101
(0.100)
‐0.178*
(0.102)
0.040
(0.126)
‐0.149
(0.133)
‐0.082
(0.083)
‐0.128
(0.109)
‐0.107
(0.109)
‐0.185
(0.112)
‐0.098
(0.102)
‐0.176*
(0.103)
0.041
(0.126)
‐0.146
(0.132)
Yes
0.609
1,246
0.211
Yes
0.688
1,040
0.178
Yes
0.648
517
0.319
Yes
0.708
439
0.303
Yes
0.648
517
0.319
Yes
0.708
439
0.304
Panel A
2‐4 Interviews
5‐8 Interviews
9‐11 Interviews
12+ Interviews
Includes Buyer Group Fixed Effects
Mean of DV
Observations
R‐Squared
Panel B
2‐4 Interviews
5‐8 Interviews
9‐11 Interviews
12+ Interviews
Includes Buyer Group Fixed Effects
Mean of DV
Observations
R‐Squared
Notes: Sample is employers who do at least one interview on the first job. Panel B includes fixed effects for groups of employers who interview workers with similar characteristics on the first job. Standard errors clustered by detailed job category x time. All specifications contain fixed effects for expected duration of job, time, and detailed job category. Table 8: Counterfactual Profits
Fee
Hires (Relative to Current Baseline)
Expected Platform Expected Wage Bill Expected Profits Profits for 1 Hour Plus Tax per Hour per Job Opening of Work on Each for 1 Hour of Work
of Work
Job
Current Fee Structure, Assuming 3 Expected Jobs if Hiring as Experienced
Current Ad‐Valorem Fee on Experienced
Current Ad‐Valorem Fee on Job Post 1
Total
10% Ad Valorem
10% Ad Valorem
100.0%
100.0%
1980
1427
0.07
0.06
143.6 x 3 = 430.8
129.7
416.9
0.09
0.08
173.6 x 3 = 520.8
187.92
N/A
0.08
0.05
159.1 x 3 = 477.3
120.96
598.26
Naïve (Infeasible) Static Fees without Composition Accounting, Assuming 3 Expected Jobs if Hiring as Experienced
Optimal Specific Fee on Experienced (Static Estimate)
Specific Fee on Job Post 1 (Static Estimate)
Total
0.52
1.1
47.6%
39.6%
1360
1017
Optimal Dynamic Fees Accounting for Composition, Assuming 3 Expected Jobs if Hiring as Experienced
Optimal Specific Fee on Experienced
Optimal Specific Fee on Job Post 1
Total
0.51
0.30
44.6%
94%
1367
All figures in this table are based on calculations using estimates from Table 4 and the techniques for recovering the posterior distribution as described in the text. Calculations done per
hour of work weight by the probability of hiring and do not gross up by expected job durations. The naïve estimates assume that the platform knows demand elasticities by segment but
does not account for dynamic spillovers across segments. The number of job openings is naively and incorrectly taken as given, so total profits are not reported because of the lack of accounting for spillovers.
Table 9: Arrivals of Applicants by Job Opening Number
Posting Number
Two
Three
Four
Five
Six
Seven
Eight
Nine
Ten or greater
Constant
Errors clustered by
Employer fixed effects
Observations
R‐squared
Number of Number of Number of Number of Applications in First Applications in First Applications in First Applications in First Two Hours
Two Hours
Day
Day
1.334***
(0.245)
1.319***
(0.281)
1.193***
(0.335)
1.989***
(0.436)
1.333***
(0.395)
2.802***
(0.578)
1.738***
(0.442)
2.322***
(0.559)
1.985***
(0.198)
13.87***
(0.631)
0.494
(0.627)
0.503
(0.784)
0.935
(0.970)
0.633
(0.998)
0.839
(0.941)
0.994
(1.117)
1.147
(1.055)
0.670
(1.170)
0.341
(0.907)
16.46***
(1.762)
2.066***
(0.621)
2.260***
(0.788)
1.920**
(0.903)
3.082***
(1.061)
1.543
(0.995)
5.121***
(1.406)
3.256***
(1.222)
3.383**
(1.326)
4.035***
(0.551)
33.64***
(1.513)
‐0.972
(1.591)
0.330
(2.215)
1.799
(2.639)
‐0.0779
(2.540)
‐0.0308
(2.413)
1.164
(3.125)
1.501
(2.889)
‐1.540
(3.083)
‐0.0824
(2.299)
40.51***
(3.977)
Employer
No
19,469
0.152
Employer
Yes
19,469
0.646
Employer
No
19,469
0.129
Employer
Yes
19,469
0.632
Notes: Robust standard errors in parentheses. The sample is sequential jobs where the vacancy duraction exceeds one hour. All specifications contain fixed effects for expected duration of job, time, and detailed job category.