'Digital Medicine, the FDA, and the First Amendment' (MSU Legal Studies Research Paper No. 12-08) by Adam Candeub
comments
Digital medicine will transform healthcare more fundamentally than the introduction of anesthesia or the discovery of the germ basis of infectious disease. Inexpensive computerized DNA sequencers will allow practitioners to individualize drugs and treatments. Digitalization will "democratize medicine," enabling individuals to create and use their own medical data, even diagnose or treat themselves. Already, tens of thousands of "medical apps" are available for smartphones that can do everything from take echocardiograms, blood pressure, pulse, lung function, oxygenation level, sugar level, breathing rate and body temperature to diagnose skin cancer and analyze urine. Medical apps, aimed at practitioners but available all, such as Isabel, diagnose diseases.
In fall 2013, the Federal Drug Administration (FDA) has asserted regulatory authority over mobile medical applications and other digital medical services, threatening, to chill, if not, destroy this innovation — and guaranteeing lengthy, high profile litigation in the near future. This article argues that the FDA stands on firm legal ground regulating medical devices that invasively measure bodily functions or take physical specimens. On the other hand, the FDA’s exercise of jurisdiction over applications that simply process information, such as Isabel, or use approved medical devices to provide medical information, like 23andMe, a genome analysis firm which the FDA recently shut down in a high profile action, raise legal concerns. Because these medical applications simply process information, they stand beyond the FDA’s regulatory reach under the Food, Drug and Cosmetics Act and the Administrative Procedure Act.
This paper adds to the large debate on the First Amendment, information and computer code. Building on recent Supreme Court decisions, this paper shows how code and applications which create healthcare information are protected speech. Given digital applications’ capacity to produce pools of data which researchers can mine for clinical and epidemiological insights and given government funding of medical services, healthcare data is both scientific and political speech, deserving of full First Amendment protection.
'Cleaning House: The Impact of Information Technology Monitoring on Employee Theft and Productivity' [
PDF] by Lamar Pierce, Daniel Snow and Andrew McAfee considers
how investments in technology-based employee monitoring
impact both misconduct and productivity. We use unique and detailed theft and sales data
from 392 restaurant locations from five different firms that adopt a theft monitoring
information technology (IT) product. Since the specific timing of individual locations’
technology adoption is plausibly exogenous, we can use difference-in-differences models to
estimate the treatment effect of IT monitoring on theft and productivity within each
location for all employees. We find significant treatment effects in reduced theft and
improved productivity that appear to be driven by changing the behavior of individual
workers rather than selection effects. Although workers with past patterns of theft appear
more likely to leave treated locations than others, individual behavioral changes by existing
workers drive restaurant-level improvements. These findings suggest multi-tasking by
employees under a pay-for-performance system, as they increase effort toward sales following
monitoring implementation in order to compensate for lost theft income. This suggests that
employee misconduct is primarily a result of managerial policies rather than individual
differences in ethics or morality.
The authors comment
Employee theft and fraud are widespread problems in firms, with workers stealing roughly $200
billion in revenue from U.S. firms to supplement their income (Murphy 1993). A growing empirical
literature on forensic economics has clarified when and how theft and other misconduct occur (e.g., Jacob and
Levitt 2003; Fisman and Wei 2009; Zitzewitz 2012a), but says little about the overall impact of firms’ use of
forensics to monitor and reduce theft. This is a critical shortfall in the literature, given the substantial
investments made by firms in monitoring employees (Dickens et al. 1989), as well as the growing forensic and
monitoring capabilities enabled by information technology (IT) systems. This raises two important yet
unanswered questions about the economic impact of monitoring employee crime. First, if monitoring is
indeed effective in reducing theft, as theory (Becker 1968; Dickens et al. 1989) and some evidence (Nagin et
al. 2002) suggests, do these gains primarily result from changing worker behavior or instead from replacing
less honest workers with more honest ones? Second, if increased monitoring reduces theft of existing workers,
how do they adjust effort on other tasks in response to this lost income, and what is the overall impact on
firm productivity? Recent research on corruption suggests that reducing one type of misconduct through
monitoring might invoke a multitasking response that increases other corrupt activities that substitute for lost
income (Olken 2007; Yang 2008).
In this paper we address these questions by examining the impact of improved theft monitoring from
information technology in the American casual dining sector, using a unique dataset that details
employee-level theft and sales transactions at 392 restaurants in 38 American states. We focus on this setting
for several reasons. First, detailed theft and sales data allow us to identify specific worker-level productivity,
theft, and sorting responses to changes in firm monitoring. Second, unlike previous research on monitoring
(e.g. Duflo et al. 2012; Zitzewitz 2012b), restaurants provide a firm-based setting where workers receive
commission-based pay-for-performance compensation that incentivizes substitution from the monitored task
(theft) to the unmonitored and productive one (sales). Third, the increased monitoring in our setting results
from the staggered implementation of improved IT monitoring systems across multiple locations. Although
the impact of IT on productivity increases in firms is well documented (David 1992; Brynjolfsson 1993;
Grilliches 1994; Nordhaus 2001; Bharadwaj 2000; Bresnahan et al 2002), no research examines potential
productivity gains through reduced theft or other misconduct. Recent work by Bloom, Sadun, and Van
Reenen (2012) shows that the productivity gains from IT have been most substantial in industries, such as
restaurants, that have “tougher” human resource practices with higher-powered incentives.
We conceptualize the employee theft issue as a stylized multitasking problem (e.g., Holmstrom and
Milgrom 1991), where workers under a pay-for-performance scheme (such as tips) can derive earnings from
two tasks: sales productivity and theft. Earnings from each task are increasing and concave in effort. The cost
of effort from each task is convex and increasing, but theft bears two additional costs. First, the employee will
be detected and punished by management (the principal) with some probability p that is increasing in theft.
Second, the employee may suffer moral or ethical costs based on identity or preferences that make theft costly
even when it is effortless and unmonitored (e.g. Akerlof and Dickens 1982; Mazar et al. 2009; Bénabou and
Tirole 2011; Dal Bó and Terviö 2013).
Such a setup has three immediate implications for the impact of increased IT monitoring on
employee effort allocation. First, any employee with existing non-zero theft levels will reduce effort allocated
to theft in response to increased monitoring by management. Second, the resulting decrease in earnings will
thus motivate them to increase effort allocated toward productivity. Third, employees with existing non-zero
theft levels will be more likely to leave the firm as outside employment options become relatively more
attractive than before.
We use approximately two years of detailed theft and sales data from 392 restaurant locations from
five restaurant firms (hereafter referred to as “chains”) that adopt an IT monitoring product, NCR
Corporation’s Restaurant Guard, that reveals theft by specific employees. Restaurant servers (also called
waiters) use multiple techniques to steal from their employers and customers, including voiding and
“comping” sales after pocketing cash payment from customers, and transferring food items from customers’
bills after they have paid. Restaurant Guard alerts managers to egregious examples of these actions in a
weekly report. These alerts represent the “tip of the theft iceberg”, since the product is designed to identify
instances of theft that are so obvious as to be indefensible by servers. Consequently, while the weekly alerts in
our data average only $108 per location, interviews with managers indicate the losses to be considerably
larger.
Our data provide the identity of each server, as well as the revenue, theft alerts, tips, shifts, and food
items sold for each day. The data also provide the date on which Restaurant Guard was implemented at each
location. The Restaurant Guard product was rolled out to individual store locations in a plausibly piecemeal
way not related to individual store needs or theft levels. Rather, the rollout pattern was driven by the schedule
and week-to-week geographic location of the vendor’s Restaurant Guard implementation team. This rollout
strategy allows us to treat adoption dates as plausibly exogenous to the individual restaurant location and not
correlated with revenue or theft levels. Our quasi-experimental setting thus enables us to estimate behavioral
and productivity changes within each location across all employees. We use difference-in-differences models
to estimate the treatment effect of the monitoring technology on theft, sales productivity, employee turnover,
and other performance metrics at both the individual and restaurant level. The different implementation dates
for each location allow us to control for time trends and time-invariant location-specific and worker-specific
fixed effects.
Our empirical models identify a 22% (or $24/week) decrease in identifiable theft after the
implementation of IT monitoring. This treatment effect is persistent, with the magnitude growing from $7 in
the first month to $48 in the third month. The treatment effect on total revenue, however, is much larger.
Total revenue increases by $2,975/week (about 7% for the average location) following implementation of
Restaurant Guard, suggesting either a considerable increase in employee productivity or a much larger latent
theft being eliminated by the IT product. Furthermore, the implementation of Restaurant Guard increases
drink sales (the primary source of theft) by $927/week (about 10.5%). This result is particularly important
because the profit margins on drinks in casual dining are between 60 and 90 percent, representing
approximately half of all restaurant profits. Furthermore, we observe an increase in average tip levels of 0.3%,
which represents one sixth of a standard deviation improvement from a base rate of 14.8%. This result
suggests improvement in customer service from IT monitoring.
While these results show considerable impact on theft, revenue, and profitability for the restaurants,
they do not explain the mechanisms through which these improvements are gained. To disentangle these
mechanisms, we examine the impact of the IT product on individual employee outcomes. We employ a
similar difference-in-differences approach, alternatively including worker and restaurant fixed effects to
examine whether our results are due to behavioral changes in existing workers or selection effects as the worst
workers leave the restaurants (e.g. Lazear 2000; Hamilton et al. 2003).
These individual worker models show that Restaurant Guard reduces average hourly theft by between
$0.05 and $0.06 in both models. This suggests that all the decrease in theft found in our restaurant-level
models can be explained by employees changing their behavior, as opposed to a change in the group of
employees working at the restaurant. We also find that IT monitoring also increases hourly sales by $2.02 for
existing workers, with similar increases for drink sales and tip percentage. In each case, the worker fixed
models suggest behavioral changes by workers rather than a selection effect. Given the pay-for-performance
compensation policy of our restaurants, these results are consistent with multi-tasking and principal-agent
models of worker behavior (Alchian and Demsetz 1972; Holmstrom and Milgrom 1991). When a worker’s
ability to gain money from theft is reduced due to increased monitoring, he or she reallocates effort toward
increasing sales and customer service in order to regain some of that loss.
Finally, our models shed light both on how management responds to the new theft information and
on workers’ endogenous choices to leave the firm. To do so, we separate workers into “known thieves” and
“unknown” groups based on their observed (by the researchers, not by the managers) pre-treatment theft.
Known thieves are those with observable pre-treatment theft. Cox hazard models show employees with known
pre-treatment theft levels have higher attrition rates than do employees without observable theft. The
observation that these exits are unlikely to happen within two weeks of a theft report to management suggests
that this attrition is voluntary and not due to termination following theft revelation to management. This
voluntary attrition by thieves following increased monitoring is consistent with workers selecting out of jobs
after monitoring limits theft income. The apparent rarity of termination also echoes Dickens et al.’s (1989)
observation that firms infrequently employ the efficient low-monitoring, high-punishment crime deterrence
strategy described in Becker (1968). We also observe that while known thieves’ weekly hours remain
unchanged following the IT implementation, other workers’ weekly hours increase on average by 2.25 hours,
which is consistent managers reallocating the hours toward more honest workers.
This paper has implications for several important research streams. First, we contribute to the
literatures on forensic economics and corruption. Only a few studies focus on explicitly illegal behavior by
employees of private firms, and those that do almost exclusively rely on empirical evidence aggregated at the
firm level (Fisman and Wei 2004; 2009; Zitzewitz 2006; Heron and Lie 2007; DellaVigna and La Ferrara
2010; Chen and Sandino 2012; Pierce and Snyder 2012).8 Our worker-level data, like Nagin et al.’s (2002)
study of call center fraud, allow us to disentangle firm-level misconduct from individual-level decisions that
run counter to firm profitability. Unlike their work, however, our multi-firm longitudinal data allow us to
more comprehensively examine the impact of monitoring on selection and treatment across multiple tasks,
including productivity. Our results show that employee productivity and misconduct are linked through
organizational policies such as compensation or information technology monitoring. This unique finding is
particularly important because it has roots in foundational models of compensation that allow for both
productivity and sabotage (Lazear 1989).
Our results also contribute to work in personnel and organizational economics on employee response
to compensation systems. While theory modeling counterproductive employee behavior is extensive (Alchian
and Demsetz; Jensen and Meckling 1976; Holmstrom 1979; Lazear and Rosen 1981; Holmstrom and
Milgrom 1991), only recently has empirical work examined incentives impact explicitly illegal behavior in
firms. A growing literature on bonus gaming examines employees’ strategic responses to incentive systems
(e.g. Oyer 1998), but these behaviors are not clearly corrupt or illegal. The fundamental difference between
counter-productive and explicitly illegal behaviors goes beyond standard principal-agent and multi-tasking
models because effort allocated toward illegal behaviors not only indirectly hurts the firm through foregone
production, but also directly hurts the firm through such costs as stolen revenue and legal liability.
Furthermore, our study suggests that the effort that workers allocate toward corrupt or illegal behavior can be
redirected toward more productive behavior through incentives. Interventions can simultaneously reduce theft
and improve productivity, a result that to the best of our knowledge has not been observed in the field.
Finally, we contribute to the literature showing the impact of technology on productivity
(Brynjolfsson 1993; Brynjolfsson and Hitt 1996; David 1992; Griliches 1994; Athey and Stern 2000;
Nordhaus 2001). One of the key findings from this research has been the impact of IT on labor productivity
growth (Jorgenson and Stiroh 2000; Oliner and Sichel 2000). While other studies show that IT can also
improve productivity by reducing mild forms of misconduct such as shirking and absenteeism (Hubbard
2000; Baker and Hubbard 2003; Duflo et al. 2012), our paper is the first to show both the direct impact in
reducing explicitly illegal behavior such as theft as well as the secondary effect of incentivizing increased
productivity. Furthermore, our paper supports the view that the impact of IT systems is intimately tied to
other elements of firm policy such as asset ownership (Baker and Hubbard 2003; Rawley and Simcoe 2013),
human resource policy (Bloom et al. 2012), and other organizational practices such as products and services
(Bresnahan et al. 2002). The impact of IT monitoring on sales and customer service increases in our setting is
likely dependent on the tip-based compensation system that incentivizes wait staff to increase productivity
after theft is constrained by monitoring.