evidence of strategic behavior in medicare claims reporting · author: strategic behavior in...

Electronic copy available at: http://ssrn.com/abstract=2630454

Submitted to Management Sciencemanuscript

Evidence of Strategic Behavior inMedicare Claims Reporting

Hamsa BastaniStanford University Department of Electrical Engineering, [email protected]

Joel GohHarvard Business School, [email protected]

Mohsen BayatiStanford University Graduate School of Business, [email protected]

Recent Medicare legislation has been directed at improving patient care quality by stopping reimbursement of

hospital-acquired conditions (HACs). However, this policy may be undermined if some providers respond by

upcoding, a practice where HACs are reported as present-on-admission (POA) to continue receiving full reim-

bursement. Identifying upcoding behavior from claims data is challenging due to unobservable confounders.

Our approach leverages state-level variations in adverse event reporting regulations and instrumental variable

techniques to discover contradictions between HAC and POA reporting rates that are strongly suggestive of

upcoding. We find over 11,000 upcoded infections a year, resulting in an added cost burden of $200 million.

Our findings suggest that, contrary to widely-held beliefs, increasing financial penalties alone may not reduce

HAC incidence and may even exacerbate the problem. We make several policy recommendations based on

our results, including a new measure for targeted HAC auditing and suggestions for effective adverse event

reporting systems.

Key words : Medicare, pay-for-performance, upcoding, hospital-acquired conditions, strategic behavior

History : This paper was first submitted on July 14, 2015.

1. Introduction

Hospital-acquired conditions (HACs), defined as conditions, infections, or complications developed

by patients as a consequence of medical treatment in a hospital, place a huge burden on society. In

2002 alone, there were an estimated 1.7 million HACs in the U.S., which contributed an estimated

98,987 deaths, placing HACs among the leading causes of deaths for that year (Klevens et al. 2007).

In addition, official estimates by the Centers for Disease Prevention and Control (CDC) estimate

the direct economic cost of HACs to be between $28 to $34 billion annually (Scott 2009).

Evidence has shown that most of these HACs are preventable through the use of better clini-

cal practices (see, e.g., Berenholtz et al. 2004, Berriel-Cass et al. 2006). However, until recently,

Medicare’s fee-for-service model reimbursed healthcare providers for these conditions regardless of

1

Electronic copy available at: http://ssrn.com/abstract=2630454

Author: Strategic Behavior in Medicare Claims Reporting2 Article submitted to Management Science; manuscript no.

whether or not they were due to an avoidable lapse in the provider’s quality of care. Furthermore,

Hsu et al. (2014) found that providers could increase their margins over eight-fold for a given ICU

patient if he or she incurred a HAC, since the patient would require an extended stay and more

services. This creates perverse incentives for providers to increase HAC rates.

This issue was addressed by the Center of Medicare & Medicaid Services (CMS) through the HAC

nonpayment policy (starting on October 1, 2008), which incentivized providers to invest in reducing

HAC incidence by placing the financial burden of treating HACs on the provider rather than on

Medicare. The policy targeted eight conditions, which were either high cost or high volume and

were considered to be reasonably preventable through better health care practices. When providers

submitted reimbursement claims diagnosing patients with one or more of these conditions, they

could indicate alongside their diagnosis whether the condition was present-on-admission (POA)

or not. If the condition was not POA, it was deemed preventable and would not be reimbursed,

causing a large financial loss to the provider for treatment of the HAC (Center for Medicare &

Medicaid Services 2014). The resulting financial incentives to providers were substantial, both

because Medicare is a major player in the U.S. healthcare system (for example, in 2011, Medicare

incurred 47.2% of all inpatient provider costs in the U.S. (Torio and Andrews 2013)), and because

private insurance companies have historically tended to adopt Medicare payment policies (Clemens

and Gottlieb 2013).

Unfortunately, multiple sources of evidence suggest that the HAC nonpayment policy has had

little impact on the rate of HACs (Lee et al. 2012, Schuller et al. 2014). It has been hypothesized

that this may be because the financial impact of the policy was too small to influence significant

change in practice (McNair et al. 2009). Consequently, public organizations that promote patient

safety have called for stronger financial penalties (see, e.g., Health Watch USA 2011). In response,

further Medicare legislation was issued in the form of the HAC Reduction Program, which created

harsher penalties (starting in October 1, 2014) for providers with high HAC rates.

In this paper, we investigate an alternate explanation for the lack of improvement in HAC

incidence: providers may have responded to the nonpayment policy by engaging in upcoding, the

practice of biasing claims reports towards higher-paying diagnoses, rather than taking steps to

reduce the true rate of HACs. In particular, providers can claim that HACs are actually present-

on-admission (POA) in order to continue receiving full reimbursement. It is important to identify

the extent of upcoding behavior (if it is present) because not only can it erode the effectiveness

of the current nonpayment policy, but it also raises questions about the veracity of self-reported

HAC rates; this is especially concerning when financial penalties are determined on the basis of

this data, as is the case in the upcoming HAC Reduction Program. Such policy measures might be

ineffective in the presence of significant upcoding. We note that we make no presumptions about

Author: Strategic Behavior in Medicare Claims ReportingArticle submitted to Management Science; manuscript no. 3

the intent (i.e. intentional or accidental) underlying upcoding behavior; rather, we will only focus

on finding evidence for upcoding, as well as its economic and policy implications1.

However, upcoding is difficult to detect since a patient’s true diagnosis is unobservable. Stan-

dard econometric techniques such as diff-in-diff estimates of HAC reporting rates before and after

the nonpayment policy do not apply because the distinction between HACs and POAs in claims

reporting did not exist prior to the nonpayment policy. Previous work has taken the approach of

manual auditing of claims data, but this is a time-consuming and expensive process that produces

high-variance results due to small sample sizes and the rarity of HACs. In particular, two previ-

ous studies on HAC upcoding that relied on auditing yielded conflicting results and could not be

generalized to a national scale (Meddings et al. 2010, Snow et al. 2012).

We approach this problem by analyzing national claims reporting statistics as a function of exist-

ing state-level adverse event reporting regulations. Our main empirical finding is that state-level

regulation is simultaneously associated with increased reporting of HACs and decreased reporting

of POAs. As we will argue in §2, the differential impact of state-level regulation on HAC and

POA rates is strongly suggestive of upcoding. We address endogeneity concerns through the use of

instrumental variables. In order to determine the consequences of upcoding, we make conservative

estimates of the rate of upcoding in Medicare inpatient claims for two important conditions (central

line-associated bloodstream infections (CLABSIs) and catheter-associated urinary tract infections

(CAUTIs)) that have been the focus of recent federal legislation (including the CMS nonpayment

policy and the HAC Reduction Program). We find that there are over 11,000 upcoded infections a

year, resulting in an added annual cost burden of $200 million to Medicare for reimbursing these

HACs. While this cost is small compared to other Medicare expenditures, it is important to note

that this money was intended as a penalty to providers to incentivize them to reduce HAC inci-

dence. The practice of upcoding has therefore eroded this financial incentive, thereby reducing the

effectiveness of the policy. Medicare’s current plan to increase penalties through the HAC Reduction

Program does not address these concerns, and may in fact exacerbate the problem since providers

with high HAC rates will face even greater financial pressure to engage in upcoding. Moreover,

providers who are trying to report more accurately than others will be unfairly penalized.

Our results suggest that in order for HAC reduction policies to be effective and fair, federal

regulation must be introduced to induce accurate reporting. To this end, we provide two policy

recommendations: (1) targeted audits based on a new measure we introduce for identifying poten-

tially upcoding providers, and (2) federal implementation of certain features of current state-level

1 Some literature, such as Silverman and Skinner (2004), suggests that upcoding may be intentional profit-maximizingbehavior, while other literature, such as Meddings et al. (2010), has suggested that upcoding may be a result ofmiscommunication between nurses and medical coders (specialized hospital staff who translate medical records toclaims reports).


regulations that we find to be effective at eliciting truthful reporting. More broadly, we emphasize

the importance of ensuring the veracity of self-reported data as Medicare moves towards additional

data-driven pay-for-performance policies in the future.

1.1. Related Literature

There has been much interest in the operations management literature regarding policy design in

the presence of strategic agents. For instance, prior work has studied pricing strategies with forward-

looking consumers (Li et al. 2014) and payment mechanisms that incentivize profit-maximizing

medical providers (Fuloria and Zenios 2001). We are interested specifically in the strategic behavior

of healthcare providers in response to Medicare payment mechanisms. KC and Terwiesch (2011) find

empirical evidence that specialized hospitals cherry-pick easy-to-treat patients. Similarly, Ata et al.

(2013) show how the current hospice reimbursement policy may cause providers to engage in adverse

selection by preferentially admitting short-lived patients. Our work focuses on providers altering

their claims reporting behavior rather than patient admissions. Powell et al. (2012) study a single

hospital’s reimbursement patterns and find that the proportion of patients who are assigned high-

severity reimbursement is reduced when physician workload is high; they attribute this change in

coding behavior to time-constrained providers being unable to perform complete claims paperwork.

In contrast, we are interested in upcoding by strategic providers to increase claims reimbursement.

Previous studies in the medical literature have looked at provider upcoding behavior in response

to Medicare’s traditional fee-for-service system (Silverman and Skinner 2004) as well as the HAC

nonpayment policy (Meddings et al. 2010, Snow et al. 2012). These studies identify upcoding

through manual reviews of medical records by costly medical experts, and are consequently lim-

ited by small sample sizes. In contrast, our approach studies claims reporting statistics and finds

evidence of upcoding occurring at a national level. To the best of our knowledge, this is the first

study to show and quantify upcoding behavior across hospitals at a national scale.

Two of the aforementioned studies were targeted towards detecting HAC upcoding, but yielded

conflicting results. Meddings et al. (2010) examined CAUTI infection reports and found that hospi-

tals often engaged in upcoding by reporting HACs as POAs when filing claims; thus, they concluded

that compliance with the CMS nonpayment policy was lacking. On the other hand, the Office of

the Inspector General conducted a second study examining 5 different HACs (including CAUTIs)

and found that HACs were indeed reliably reported and that there was very little evidence of

upcoding (Snow et al. 2012). Our work helps resolve this conflict by providing evidence for HAC

upcoding as well as conservative estimates of its magnitude.

The remainder of the paper is organized as follows. In §2, we present our argument for HAC

upcoding in Medicare claims data based on our empirical findings. We then describe our various


sources of data on patients, providers, and state regulations in §3. We establish our key empirical

results for upcoding and estimate the resulting monetary losses incurred by Medicare in §4. We

perform robustness checks in §5. We conclude by discussing the policy implications of this work

for CMS and Medicare providers in §6.

2. Empirical Strategy for Detecting Upcoding

Ideally, we would be able to detect HAC upcoding from national claims reporting statistics. A naive

approach would identify upcoding providers as those who have low HAC reporting rates and/or

high POA reporting rates. However, these effects may also be caused by provider improvements

in detection and prevention of HACs in response to the nonpayment policy, i.e., (1) improved

measures for HAC prevention would naturally lead to lower HAC rates, and (2) improved infection

detection would lead to higher POA reporting rates (because many POA infections may have gone

undetected before). In other words, this naive approach cannot distinguish the effects of upcoding

and provider quality. Furthermore, claims data before 2008 did not distinguish between HACs

and POAs because the present-on-admission indicator was introduced at the same time as the

nonpayment policy. Thus, we cannot calculate the changes induced by the nonpayment policy

on POA and HAC reporting rates separately. Therefore, we argue that studying claims reporting

statistics alone is not sufficient for detecting upcoding, and a more nuanced approach is necessary.

Our approach for assessing upcoding exploits an external data source: variations in existing state-

level adverse event regulation. Many states passed laws that mandated the reporting of various

HACs prior to the federal nonpayment policy in 2008. As documented by the Office of Inspector

General (OIG) of the Department of Health and Human Services, regulations on the contents of

these reports varied significantly from state to state (Levinson 2008), thereby creating a natural

quasi-experiment. While the primary aim of these reporting systems was to track HAC incidence

across providers, a subset of states included measures to ensure accurate reporting (e.g. detailed

patient and event information monitoring and root cause analysis). We will refer to this subset of

states as strongly-regulated states and all other states as weakly-regulated states. Because strongly-

regulated states required provider accountability for accurate HAC reporting as well as follow-up

corrective strategies, these reporting regulations indirectly mandated that providers in such states

had to improve their capabilities to detect and try to prevent these targeted conditions.

Consequently, we expect that providers in strongly-regulated states would report, on average,

higher POA rates and lower HAC rates compared to providers subjected to fewer regulations (after

risk-adjusting for appropriate confounding factors). However, we actually find the opposite effect:

providers in strongly-regulated states have lower POA reporting rates and higher HAC reporting

rates (see Fig. 1 for the unadjusted reporting rates in a random sample of almost a million Medicare


inpatient stays from 2009-10). We will show in §4 that this effect persists with high significance

after adjusting for various confounders such as patient risk factors and provider quality metrics.

Figure 1 Average (unadjusted) POA and HAC reporting rates for strongly- and weakly-regulated states in a

random sample of almost a million Medicare inpatient stays from 2009-10. Providers in strongly-regulated states

have lower POA and higher HAC reporting rates compared to providers in weakly-regulated states.

There are several potential explanations for the finding that providers in strongly-regulated states

have lower POA reporting rates: (1) providers in weakly-regulated states may indeed have better

infection detection ability, therefore reporting higher POA infection rates, (2) our risk adjustment

may be biased due to unobserved variables, and the discrepancy may be because patients in weakly-

regulated states are more susceptible to infection, or (3) providers in weakly-regulated states may

be engaging in upcoding by untruthfully reporting non-POA infections as POA, thereby receiving

increased reimbursement. In the first two cases, we expect that the higher (reported) POA rates in

weakly-regulated states would be accompanied by higher (reported) HAC rates. In particular, (1) if

providers in weakly-regulated states are better at detecting infections, then they should detect more

HACs as well (since the detection mechanism for these infections is the same for HACs and POAs),

and (2) if patients in weakly-regulated states are more susceptible to infection, providers would

also observe higher HAC rates for these patients. However, the first two explanations contradict

the finding that providers in strongly-regulated states have higher HAC reporting rates. Thus,

the evidence supports the third explanation: providers in weakly-regulated states are engaging in

upcoding relative to providers in strongly-regulated states.

One concern in this analysis is that state adverse event regulation may be endogenous to HAC

reporting rates. Specifically, states may have introduced adverse event reporting regulation directly

in response to high HAC rates, in which case it is conceivable that providers in strongly-regulated

states report relatively higher HAC rates. We address this issue by using an instrumental variable


approach: our instruments are various measures of state taxation levels (known as the Economic

Freedom Index (Ashby et al. 2010)) which are correlated with the “strength” of a state’s regulatory

environment but bear no direct relationship with patient infection rates. We find our results remain

consistent despite accounting for this endogeneity. A second concern is that strongly-regulated

states may simply have lower quality of care, which would lead to higher HAC rates. However,

we find evidence through provider-specific risk-adjusted mortality rates (a widely-used proxy for

provider quality, see for e.g. Selim et al. (2002), Forthman et al. (2010)) that provider quality is

higher in strongly-regulated states, suggesting that HAC rates in weakly-regulated states should

in fact be higher. Therefore, we conclude that this concern is unlikely to impact our findings.

3. Datasets

In this section, we describe our various sources of data and define our key variables for the treatment

effect estimation. We also discuss potential confounders and and our approach to control for these

effects. We reproduce state reporting system features in Table 1, and we list all of our controls and

instruments in Table 2. We report summary statistics of all our variables in Table 3.

3.1. Data Sources

Our main sources of data were Medicare inpatient claims data and state adverse event system

classification by the OIG. We also use zipcode-level census data and Medicare provider data from

Hospital Compare2 for controls, and data on state-level economic freedom indices for instruments.

3.1.1. Medicare Patient Data We used the MedPAR Research Identifiable Files (RIF) made

available by the Centers of Medicare and Medicaid Services (CMS). This dataset contains infor-

mation on every inpatient stay between 2007 and 2010 of a randomly selected 5% sample of all

Medicare beneficiaries in the United States. Of the approximately 2.5 million Medicare beneficiaries

in our 5% sample, 492,218 had at least one inpatient stay between 2007 and 2010.

Our dataset contains records on 3,865,734 inpatient stays over the four years. Each inpatient stay

record includes anonymized beneficiary and provider IDs, diagnoses (ICD-9 codes) and procedures

associated with the stay, patient demographic information, and claims and billing information.

Patients (i.e., beneficiaries) are assigned unique IDs, allowing them to be tracked across multiple

inpatient stays over the four-year period. This allows us to compute health risk measures for

individual patients based on claims histories.

The unit of observation is an individual Medicare inpatient stay. We perform our treatment effect

estimation on inpatient stays in 2009-10. However, we use a rolling two-year window of claims

histories to compute various measures of patient risk for each inpatient stay, and so data from

2 http://www.medicare.gov/hospitalcompare


2007-08 are used indirectly. We also limit our sample to short stays under the prospective payment

system served by providers in the United States (which is the healthcare setting that was targeted

by the nonpayment policy) as well as patients with at least one prior Medicare inpatient stay in

the past 24 months (the length of our rolling window) so that we can better assess patient risk. We

note that these filters affect all states uniformly, and therefore do not create bias in our analysis.

3.1.2. State Reporting System Classification As of January 2008, 26 states had imple-

mented adverse event reporting systems in the absence of federal guidelines. The OIG performed a

detailed comparison of these systems based on telephone interviews with the staff responsible for

each state’s reporting system (Levinson 2008). The OIG report describes key features of the state

reporting systems, including the type of information that must be reported by each state regarding

(1) the affected patient, (2) the adverse event, and (3) the root cause of the adverse event. All

26 states with reporting systems enforced at least reporting the identity of the hospital and the

adverse event that had occurred. We reproduce the information reported in each category and the

number of states that had implemented each requirement in Table 1.

Category Information # States

Any Reporting Event and Hospital 26

Patient-Specific

Impact of Event on Patient 12Patient Age or Date of Birth 19Patient Diagnosis 16Patient Medical Record Number 5Patient Billing Number 2

Event-Specific

Type of Event 26Location within Hospital 20Date of Event 24Date of Discovery 10Summary Description 18Detailed Description 11

Root Cause AnalysisRoot Cause Analysis Team Name 7Identified Cause 12Contributing Factors 16

Table 1 Different types of information reporting requirements used in state adverse event reporting systems

and the number of states that had implemented each requirement. Reproduced from Levinson (2008).

3.1.3. Other Sources of Data We used data from the American Community Survey (2008-

12) by the US Census Bureau to obtain household characteristics aggregated at the zipcode-level.

These measures included population statistics for age, birthplace, education, income, and insurance

status. We also obtained data on individual life expectancies aggregated at the county-level from


the Institute of Health Metrics and Evaluation. We used Hospital Compare data on provider-level

risk-adjusted mortality rates for Medicare patients. Finally, for our instrumental variable analysis,

we used the three areas of the 2010 federal North American Economic Freedom Index by state: (1)

Size of Government, (2) Takings & Discriminatory Tax, and (3) Labor Market Freedom (Ashby

et al. 2010).

3.2. Treatment Variable

One possible definition of the treatment variable is simply having an adverse event reporting

system. Interestingly, our results show that merely having an adverse event reporting system did

not have a significant effect on POA and HAC claims reporting rates for CLABSIs and CAUTIs

(see Table 7). This is because the quality of the reporting systems varied widely from state to state.

Instead, our approach is to look for states that impose meaningful requirements on the quality

of the reporting. We construct a treatment variable that is an indicator for whether the provider

is located in a state that had strong regulations on adverse event reporting prior to the federal

nonpayment policy in 2008.

As previously noted, we use data from an OIG report which lists each state’s information report-

ing requirements (see Section 3.1.2). We are particularly interested in regulation that enforced

truthful reporting. The OIG report claims that states identified cases of underreporting by

“analyzing reported data, comparing hospital reports against complaints, referrals, and admin-

istrative databases, and conducting onsite audits” (Levinson 2008, pg. 4).

These methods are greatly aided by the availability of more detailed data. In particular, we argue

that the more data a state has regarding the circumstances of an adverse event, the harder it is

for a provider to be untruthful about the event. Thus, we used the amount of required informa-

tion reported to states in each category as a proxy for increased regulatory pressure for truthful

reporting. For simplicity, we chose the most informative reporting requirement from each of the

three information categories (see Table 1), namely,

1. Patient-specific: patient medical record number or billing number

2. Event-specific: detailed description of the adverse event

3. Root cause analysis: identified cause of adverse event

We define our treatment variable based only on these three reporting requirements, which helps

us better interpret our results in order to make concrete policy suggestions. Since there are many

other ways to define the treatment variable, we perform a robustness check where we consider

several alternate definitions of the treatment variable that yield consistent results (see Section

4.4.1). This alleviates the concern that a particular definition of the treatment variable gave rise

to our results by chance.


In order to construct the treatment variable, we compute a binary3 “strength” for each state’s

regulation of its adverse event reporting system based on the number of these three features

adopted. The median state with a reporting system adopted one of these features so we considered

the set of strongly-regulated states to be those with two or more of these features. According to

this definition, the strongly-regulated states were CT, FL, MA, MN, NJ, NY, RI, and SD.

Thus, we defined a binary treatment variable S for providers:

• S = 0: Provider was located in a weakly-regulated state, i.e. either had no adverse event

reporting system, or had an adverse event reporting system that had zero or one of the reporting

requirements described above.

• S = 1: Provider was located in a strongly-regulated state, i.e. had an adverse event reporting

system with two or more of the reporting requirements described above.

3.3. Outcome Variables

We focus on CLABSIs and CAUTIs, the only two conditions directly targeted by both the HAC

nonpayment policy and the recent HAC Reduction Program. We define two outcome variables:

• POAi is an indicator variable for whether either a CLABSI or a CAUTI was diagnosed along

with the present-on-admission indicator in the claims record for inpatient stay i

• HACi is an indicator variable for whether either a CLABSI or a CAUTI was diagnosed without

the present-on-admission indicator in the claims record for inpatient stay i

3.4. Controls

We define a variety of controls to account for potential confounders.

3.4.1. Patient Risk. States that implement strong regulation for HACs are likely to have

also implemented other measures towards improving population health; this may, in turn, affect

downstream patient infection rates. To account for this effect, we control for an extensive list

of patient-specific factors that are computed from their claims histories. Age, sex, and race are

obtained from MedPAR’s summarized beneficiary demographic information. We use a rolling win-

dow of 6 months of each patient’s claims history to identify risk-associated quantities such as the

number of days since the patient’s last admission, the number of prior admissions, the number

of prior procedures performed on the patient during those admissions, the number of previous

CLABSI and/or CAUTI infections sustained during that time, and the total length of hospital

stay days. These quantities can be directly assessed from the MedPAR data.

We also use patient history to compute the Charlson comorbidity index, which predicts the

likelihood of patient mortality within 6 months (Deyo et al. 1992); the Charlson score is a widely-

accepted measure of patient risk in the medical community.

3 We define the treatment variable to be binary to improve the interpretability of our results. In §5.2, we perform arobustness check to show that our results are consistent if the treatment is a continuous variable.


3.4.2. Demographic Factors. States that did not implement strong HAC regulation may

generally be poorer or more resource-constrained. This may, in turn, affect the completeness of

patient claims data; in particular, poor patients may not have access to frequent healthcare due

to lack of health insurance or other resource constraints, and thus their health risks may not be

completely captured from claims histories. We address this by using health-related controls from

census data based on the patient’s listed zipcode. These controls included the average household

income as well as fractions of individuals in the population who were above 65, uninsured, unem-

ployed, near the poverty line, foreign-born and/or had not completed high school aggregated at

the zipcode level.

3.4.3. Provider’s Billing Aggressiveness. Providers that code their claims more aggres-

sively to achieve the highest possible reimbursement rates may generally have different reporting

trends. Note that this is different from upcoding, since the codes are still accurate but possibly

optimized for reimbursement. We control for variations in coding practices by defining a provider-

specific measure of billing aggressiveness:

payratioj =

∑i∈Tj

charges to Medicare for inpatient stay i∑i∈Tj

Medicare payment for inpatient stay i

where j is the index of the provider and Tj is the set of all inpatient stays under the care of provider

j. This heuristic is intended to capture the provider’s aggressiveness in obtaining reimbursement.

3.4.4. Provider Quality. Low-quality providers may be associated with higher HAC rates

due to lapses in quality and lower POA rates due to patient choice (a sicker patient with a present-

on-admission infection may choose to admit herself with a higher quality provider). We account

for this effect by including provider-level risk-adjusted pneumonia mortality rates in our control

set. We chose pneumonia because it is associated with significant catheter use (and therefore

catheter-associated infections as well) and is one of 3 conditions for which Medicare published

provider-specific mortality rates in that time period.

3.5. Instrumental Variables

Our treatment variable may potentially be endogenous if states passed regulation on adverse

event reporting in response to high HAC rates and poor provider quality. We address this issue

through the use of instrumental variables for our HAC analysis. Note that high POA rates cannot

have impacted a state’s decision to regulate since these adverse event reporting systems targeted

hospital-acquired conditions and, to the best of our knowledge, there were no state agencies that

even collected information on present-on-admission infection rates. Thus, we only use an instru-

mental variable approach for our analysis of HAC rates.


Following the example of Mukamel et al. (2012), we use all three areas of the Economic Freedom

Index as our instruments. States with strong adverse event regulation tend to have a smaller size of

government (Area 1) and are more stringent with respect to takings and discriminatory taxation

(Area 2) and labor market freedom (Area 3). Summary statistics are reported in Table 3.

We believe these instruments meet the necessary conditions that they are (i) correlated with the

treatment variable (states with a stronger government presence, i.e. less economic freedom, tend

to have more stringent regulations on patient adverse event reporting) and (ii) uncorrelated with

HAC infection rates except through the treatment variable and controls.

We check the first condition through a weak identification test, but the second condition cannot

be verified empirically. However, there is no evidence of a direct relationship between a state’s

level of economic freedom and hospital-acquired infection rates. Moreover, since we are using more

instruments than endogenous variables, we perform an overidentification test that helps support

the validity of this condition.

Type Variable Definition Variable Name

Patient-Level Age ageSex: male, female sex xRace: white, black, asian, hispanic, native american, race xother, unknownCharlson 6-month comorbidity score charlsonDays since last admission (up to 6 months) days sinceNumber of prior admissions in the last 6 months num admitNumber of prior procedures in the last 6 months num prcdrNumber of prior CLABSIs in the last 6 months num clbiNumber of prior CAUTIs in the last 6 months num cautiNumber of days of hospital stay in the last 6 months tot los

Provider-Level Billing aggressiveness payratioRisk-adjusted pneumonia mortality rate mort pneu

Zipcode-level Fraction of population above 65 above65Fraction of local population uninsured uninsuredFraction of local population below 1.38× poverty index poorFraction of local population foreign born foreignbornFraction of local population unemployed unemployedFraction of local population did not complete high school nohighschlAverage household income of local population incomeLife expectancy of local population: male, female lifeexp x

State-level Economic Freedom Index: Area 1, 2, and 3 efi x

Table 2 Definitions of our control and instrumental variables. Note that ‘x’ at the end of a variable name is a

placeholder for the type of a categorical variable.


Variable Strong Weak

All POA HAC All POA HAC

# Observations 229,425 1529 197 677,497 5737 527

sex female 59% 48% 65% 60% 50% 62%race white 81.1% 75.9% 74.6% 81.0% 74.7% 78.9%race black 12.6% 17.3% 19.3% 14.2% 20.1% 16.9%race other 1.4% 1.6% 1.0% 1.0% 1.0% 0.6%race asian 0.9% 1.1% 0.5% 1.0% 0.9% 1.5%

race hispanic 3.5% 3.7% 4.1% 1.9% 2.5% 1.5%race native 0.2% 0.2% 0.0% 0.7% 0.5% .6%

race unknown 0.3% 0.2% 0.5% 0.2% 0.3% 0%

age 75 71 73 74 70 70(14) (16) (16) (14) (16) (15)

charlson 1.8 2.6 2.1 1.8 2.6 2.2(2.3) (2.5) (2.4) (2.3) (2.6) (2.5)

days since 89 56 65 91 61 65(75) (65) (71) (75) (67) (70)

num admit 1.7 2.8 2.1 1.6 2.6 2.1(2.0) (2.5) (2.3) (1.9) (2.5) (2.0)

num prcdr 2.2 4.4 2.9 1.9 3.9 3.0(3.6) (5.1) (3.8) (3.4) (4.8) (3.9)

num clbi .0065 .12 .046 .0072 .093 .023(.095) (.44) (.23) (.098) (.39) (.16)

num cauti .0049 .12 .025 .0068 .17 .011(.078) (.40) (.19) (.099) (.52) (.12)

tot los 16 40 28 15 34 22(34) (53) (48) (32) (47) (33)

payratio 4.9 4.9 4.9 4.5 4.6 4.5(2.4) (2.3) (2.4) (2.1) (2.1) (1.8)

mort pneu 10.8 10.8 10.8 11.3 11.2 11.2(1.7) (1.6) (1.8) (1.9) (1.8) (1.9)

above65 16.2% 16.0% 17.7% 14.1% 13.5% 13.2%uninsured 13.0% 13.0% 13.4% 15.2% 15.3% 15.4%

poor 68.3% 68.2% 68.6% 68.3% 68.6% 68.3%foreignborn 17.5% 15.6% 16.0% 8.7% 9.3% 9.4%unemployed 9.6% 9.8% 9.6% 9.7% 9.9% 9.8%nohighschl 13.4% 13.1% 13.6% 15.2% 15.2% 15.2%

income 29,446 28,619 27,655 25,053 26,827 25,377(19,685) (18,215) (18,267) (17,481) (17,750) (16,532)

lifeexp female 81.5 81.3 81.4 79.8 79.8 79.9(1.3) (1.4) (1.5) (1.7) (1.7) (1.8)

lifeexp male 76.6 76.5 76.5 74.8 74.8 74.9(1.7) (1.8) (1.9) (2.4) (2.3) (2.5)

efi 1 7.4 7.5 7.5 7.1 7.1 7.1(0.3) (0.3) (0.3) (0.8) (0.9) (0.8)

efi 2 5.7 5.7 5.6 6.2 6.2 6.2(0.4) (0.4) (0.4) (0.5) (0.5) (0.6)

efi 3 7.0 7.0 7.0 7.2 7.2 7.2(0.6) (0.5) (0.6) (0.6) (0.6) (0.6)

Table 3 Summary statistics for all variables. Standard deviations are shown in parentheses.


4. Estimation & Results

We take a treatment effect estimation approach to determine the causal effects of strong state

regulation through adverse event reporting systems on Medicare POA and HAC claims reporting

rates for CLABSIs and CAUTIs.

We apply regression techniques under a linear model4 to determine the effects of strong state

regulation in adverse event reporting on POA and HAC reporting rates. We find that the presence

of strong state regulation of adverse event reporting was associated with decreased POA rates and

increased HAC rates. As argued in Section 2, this suggests that states with weak or no regulations

are engaging in upcoding behavior by reporting HAC infections as POAs in Medicare claims.

4.1. POA Regression

Let Ci denote the vector of controls (including an intercept term) for inpatient stay i. We use a

linear model with the econometric specification:

POAi = βPOAS Si +βTCi + εi

where εi is the error term. The coefficient of interest is βPOAS , which represents the effect of strong

state regulation on POA reporting rates. Specifically, if βPOAS is negative, this would indicate that

after controlling for potential confounders, providers in states with strong regulations have a lower

probability of reporting POAs than providers in states with little or no regulation.

The standard OLS estimator makes the assumption that all errors in the POA model are

homoskedastic and independent. However, this is unlikely to be the case as hospital stays served by

the same provider may have correlated heteroskedastic errors due to unobserved provider-specific

variables. To account for this, we cluster our data at the provider level, and use cluster-robust

standard errors that relax our assumptions to allow both arbitrary heteroskedasticity and arbi-

trary within-provider correlation. (In §5.2, we perform a robustness check with coarser state-level

clustering and confirm that our results remain significant.)

The regression coefficients, and robust standard errors clustered by provider are shown in Table 4.

Our results show that, after controlling for patient risk, provider-specific, and demographic fac-

tors, strong state regulation on adverse event reporting is associated with significantly lower POA

reporting rates (p= 7.2× 10−7).

4 Although our outcomes are binary, we use a linear probability model (see Section 15.2 of Wooldridge (2010) for ajustification) rather than a logit or probit model so that we can perform instrumental variable validity tests in thepresence of clustered errors (e.g., see Cameron and Miller (2015)). We perform a robustness check in §5.2 to showthat our results remain consistent under a probit model.


4.2. HAC Regression

Let Ii denote the vector of instrumental variables for inpatient stay i. We use two-stage least squares

(2-SLS) to estimate a linear model with instruments. In the first stage, we fit our endogenous

variable

Si = βT1 Ci +βT

I Ii + εi,1

In the second stage, we fit our outcome variable using the predicted Si from the first stage

HACi = βHACS Si +βT

2 Ci + εi,2

where εi,1, εi,2 denote error terms. In this case, if βHACS is positive, it would indicate that after

controlling for potential confounding variables and the endogeneity of regulation, providers in states

with strong regulations have a higher probability of reporting HACs than providers in states with

little or no regulations.

Once again, we use cluster-robust standard errors clustered at the provider level. We also perform

weak- and over-identification tests to support the validity of our chosen instruments.

The regression coefficients, and robust standard errors clustered by provider are shown in Table 4.

Our results show that, after controlling for patient risk, provider-specific, and demographic factors,

strong state regulation on adverse event reporting is associated with significantly higher HAC

reporting rates (p= 1.4× 10−2).

Tests of Instrument Validity Our first-stage regression for HACs (R2 = 0.40) produced a

Kleinberg-Paap Wald F -statistic of 140, which is well above the Stock-Yogo weak ID test critical

values for the maximal IV relative bias (13.91 at the 5% level) and for the maximal IV size (22.30

at the 10% level), indicating that our instruments are not weak (Baum 2007). Moreover, our

overidentification test produced a Hansen J statistic of 0.374 with a χ2 p-value 0.83; thus we failed

to reject the null hypothesis that our model is correctly specified, lending credence to the validity

of our instruments. Finally, we performed an endogeneity test on the treatment variable and found

evidence (p = 0.02) rejecting the null hypothesis that the treatment variable is exogenous with

respect to HAC outcomes; this result justifies our instrumental variable approach.

4.3. Loss Estimates

We estimate lower bounds on the number of annually upcoded CLABSIs and CAUTIs among

Medicare inpatient stays in the United States, as well as the associated costs to Medicare. We

take the number of upcoded infections to be the relative excess of POA reports by providers in

weakly-regulated states. This implicitly makes two conservative assumptions:

1. Providers in strongly-regulated states engage in no upcoding

2. All providers have similar capabilities for infection detection


Variable (1) POA Reports (2) HAC Reports

Estimate SE Estimate SE(intercept) 6.42× 10−3 8.53× 10−3 5.03× 10−3* 2.95× 10−3

sex female −2.40× 10−3*** 2.34× 10−4 1.68× 10−4*** 5.99× 10−5

age −1.00× 10−4*** 9.52× 10−6 −1.14× 10−5*** 2.65× 10−6

race unknown 7.21× 10−4 2.25× 10−3 −1.28× 10−4 5.72× 10−4

race white 1.07× 10−3 1.05× 10−3 3.04× 10−4 3.45× 10−4

race black 2.87× 10−3** 1.12× 10−3 4.58× 10−4 3.55× 10−4

race other 1.09× 10−3 1.40× 10−3 2.94× 10−5 4.10× 10−4

race asian 4.28× 10−4 1.41× 10−3 6.60× 10−4 4.87× 10−4

race hispanic 2.67× 10−3** 1.32× 10−3 2.60× 10−4 3.97× 10−4

charlson 1.87× 10−4*** 6.44× 10−5 −1.76× 10−5 1.86× 10−5

days since −1.12× 10−5*** 1.68× 10−6 −3.40× 10−6*** 5.76× 10−7

num admit −8.33× 10−4*** 1.13× 10−4 −7.58× 10−5** 2.95× 10−5

num prcdr 5.65× 10−4*** 5.34× 10−5 3.96× 10−5*** 1.36× 10−5

num clbi 7.03× 10−2*** 4.60× 10−3 1.40× 10−3** 5.58× 10−4

num cauti 1.37× 10−1*** 5.52× 10−3 6.46× 10−4 4.82× 10−4

tot los 9.45× 10−5*** 6.68× 10−6 4.17× 10−6*** 1.35× 10−6

payratio 1.11× 10−4* 6.62× 10−5 2.16× 10−8 1.50× 10−5

mort pneu −7.88× 10−5 6.71× 10−5 −2.22× 10−5 1.98× 10−5

uninsured 9.70× 10−4 2.37× 10−3 1.36× 10−3** 6.81× 10−4

above65 −2.91× 10−3* 1.59× 10−3 −1.40× 10−4 5.52× 10−4

foreignborn −2.94× 10−3** 1.49× 10−3 −5.06× 10−4 4.24× 10−4

unemployed −5.46× 10−4 2.95× 10−3 −6.01× 10−4 8.59× 10−4

poor 2.41× 10−3 1.78× 10−3 −2.79× 10−5 5.15× 10−4

nohighschl −5.13× 10−3*** 1.86× 10−3 −3.75× 10−4 5.11× 10−4

income 2.01× 10−8*** 7.08× 10−9 −1.16× 10−9 2.06× 10−9

lifeexp female 2.36× 10−4 2.39× 10−4 −1.54× 10−4* 7.88× 10−5

lifeexp male −1.23× 10−4 1.76× 10−4 1.17× 10−4** 5.48× 10−5

S −1.66×10−3*** 3.34×10−4 5.29×10−4** 2.15×10−4

*p < 0.10, **p < 0.05, ***p < 0.01

Table 4 Results of regressions. Point estimates and cluster-robust standard errors (SE) of coefficients for (1)

OLS regression of POA reports and (2) 2-SLS regression of HAC reports against strength of state reporting

system and controls.

We believe our estimates are conservative since providers in strongly-regulated states likely have

better infection detection due to the increased investigative and reporting requirements as discussed

earlier. In this case, the number of excess POA reports by weakly-regulated providers is larger than

what we estimate. Secondly, it is unlikely that providers in strongly-regulated states do not engage

in upcoding at all; in this case, the overall amount of upcoding is again larger than our estimate.

We perform two linear regressions on CLABSI-POA and CAUTI-POA outcomes respectively.

We find the absolute value of the treatment effects, i.e. excess POA reporting rates, of:

• CLABSI-POA: 2.28× 10−4 with standard error 1.08× 10−4

• CAUTI-POA: 1.45× 10−3 with standard error 2.54× 10−4


Our data comprises 677,497 inpatient stays in weakly-regulated states over 2 years. Since we have

a random 5% sample of all Medicare inpatient stays, we estimate that there are 6,774,970 Medicare

inpatient stays per year in weakly-regulated states that meet our criteria. Thus, we compute a

lower bound

[# patients in weakly-regulated states per year]× [excess POA rate]

on the number of upcoded POA reports claimed annually by weakly-regulated states for each

condition (see Table 5).

We also obtain estimates of Medicare’s added reimbursement cost burden from these infections

from Umscheid et al. (2011). They find that CLABSIs result in an estimated incremental cost of

$110,800 (95% CI: $22,700 - $327,000) on average, and CAUTIs result in an estimated incremental

cost of $2950 (95% CI: $1200 – $4700). Note that these estimates only account for the immediate

hospital service costs, and do not incorporate long-term effects on patient health or lost patient

time. We use these estimates to compute lower bounds

[# patients in weakly-regulated states per year]× [excess POA rate]× [cost of infection]

on the cost burden to Medicare due to upcoding. Results are shown in Table 5.

Infection Minimum # Upcoded Cases Minimum Added Cost to Medicare

Estimate 95% CI Estimate 95% CI

CLABSI 1545 [111, 2979] $ 171 million $2.5 million – $974 millionCAUTI 9824 [6451, 13197] $ 29 million $7.7 million – $62 million

Table 5 Conservative estimates are shown for the number of upcoded cases per year and the associated cost

burden to Medicare for both CLABSIs and CAUTIs.

Thus, we estimate a total of 11,369 upcoded infections with an associated cost burden of $200

million in annual Medicare reimbursements.

4.4. Policy Comparison

We defined our original treatment variable based on three reporting requirements that we con-

sidered informative. We now alter the definition of the treatment variable based on reporting

requirements along three dimensions: patient, event, and cause (see Table 1). This serves two

purposes:

• We show that our results are robust to the choice of treatment variable as long as it captures

the stringency of regulations on truthful reporting.

• We draw inferences about which types of reporting requirements may be most effective at

reducing upcoding behavior.


4.4.1. Alternative Definitions of Treatment Variable We construct alternative defi-

nitions of the treatment variable through the following procedure. For every combination of

patient/event/cause, we consider the relevant set of reporting requirements and compute the

median number implemented by the states with adverse event reporting systems (see Table 6). We

define all states with more than the median number of requirements as “strongly regulated.”

Treatment Definition Median Requirements # States

Patient 2 out of 5 11Event 4 out of 6 11Cause 1 out of 3 11

Patient & Event 6 out of 11 12Patient & Cause 3 out of 8 12Event & Cause 6 out of 9 9

Patient, Event, & Cause 8 out of 14 10

Table 6 Different definitions of the treatment variable based on the number of reporting requirements along

three dimensions (patient, event, and cause), as well as the number of states that satisfied this condition.

We also investigate an alternative definition where a strongly-regulated state is one that simply

has an adverse event reporting system. These states include CA, CO, CT, DC, FL, GA, IN, KS,

ME, MD, MA, MN, NJ, NV, NY, OH, OR, PA, RI, SC, SD, TN, UT, VT, WA, and WY.

For each of these definitions of the treatment variable, we ran a linear regression and a 2-SLS

regression for POA and HAC outcomes respectively, as described in Sections 4.1–4.2. We list the

estimated treatment effect along with cluster-robust standard errors and p-values in Table 7. The

“Original” definition refers to the measure that was defined and used earlier in the paper.

Treatment (1) POA Reports (2) HAC ReportsDefinition Estimate SE p-value Estimate SE p-value

Patient −1.31 · 10−3 3.06 · 10−4 0.000 6.02 · 10−4 2.36 · 10−4 0.011Event −1.27 · 10−3 3.56 · 10−4 0.000 2.82 · 10−4 2.41 · 10−4 0.242Cause −1.01 · 10−3 3.27 · 10−4 0.002 7.12 · 10−4 2.95 · 10−4 0.016

Patient & Event −1.29 · 10−3 3.04 · 10−4 0.000 6.06 · 10−4 2.36 · 10−4 0.010Patient & Cause −1.28 · 10−3 3.19 · 10−4 0.000 8.85 · 10−4 3.53 · 10−4 0.012Event & Cause −1.66 · 10−3 3.64 · 10−4 0.000 1.84 · 10−4 2.63 · 10−4 0.484

Patient, Event & Cause −1.85 · 10−3 3.24 · 10−4 0.000 7.05 · 10−4 3.03 · 10−4 0.020

Has Reporting System? −5.57 · 10−4 2.93 · 10−4 0.058 5.20 · 10−4 2.90 · 10−4 0.073

Original −1.66 · 10−3 3.34 · 10−4 0.000 5.29 · 10−4 2.15 · 10−4 0.014

Table 7 Point estimates and cluster-robust standard errors for the coefficient of the treatment variable are

shown for alternative definitions of strong state regulation.


4.4.2. Results & Observations First, we find that our results are largely consistent for

different definitions of strong regulation that capture the magnitude of the providers’ reporting

burden in that state. In particular, stringent regulation on adverse event reporting is associated with

reduced upcoding levels. On the other hand, merely having regulations for adverse event reporting

is not associated with significant changes in reporting rates or upcoding behavior. These findings

support the hypothesis that laws cannot create proper incentives without sufficient accountability.

Second, we infer that reporting patient information is most valuable, while only reporting infor-

mation on the event has limited value. This may be because reporting patient information (such

as the medical record number) may allow state entities to more easily audit hospital records. Our

findings also suggest that reporting along all three dimensions is best; in particular, reporting

patient, event, and cause information was associated with the highest reduction in upcoding.

5. Robustness Checks

We perform two robustness checks to provide further evidence justifying our assumptions and to

show that that our empirical results are consistent under alternate specifications.

5.1. Provider Quality Variation

As noted earlier in Section 2, one source of endogeneity is provider quality. In particular, strongly-

regulated states may have higher HAC rates due to low provider quality. To address this issue, we

checked how risk-adjusted mortality rates (which are the most direct measure of provider quality)

varied between strongly- and weakly- regulated states. During 2009-10, Medicare publicly reported

these rates for three conditions: heart attack, heart failure, and pneumonia. We performed a t-test of

provider-specific risk-adjusted mortality rates to compare provider quality in strongly- vs. weakly-

regulated states. Results are shown in Table 8. We find that strongly-regulated states have lower

risk-adjusted mortality rates across all 3 conditions with high statistical significance. Therefore, it

is unlikely that they report relatively higher HAC rates due to relatively poor provider quality.

Condition Mean Mortality Mean Mortality 95% CI of p-value(Strong States) (Weak States) Difference

Heart Attack 15.80% 16.32% [-0.69%, -0.35%] 1.7× 10−9

Heart Failure 10.63% 11.17% [-0.69%, -0.39%] 2.4× 10−12

Pneumonia 11.10% 11.55% [-0.62%, -0.28%] 4.1× 10−7

Table 8 T-test results are shown comparing Medicare risk-adjusted mortality rates in strongly vs. weakly

regulated states for heart attack, heart failure, and pneumonia patients.


5.2. Regression Specification

In addition to checking the robustness of our results to different definitions of the treatment variable

(§4.4), we also ensure that our results are consistent under alternative regression specifications:

1. Since the outcomes are binary, we use a probit model specification rather than a linear model.

2. We use a continuous (rather than binary) definition of the treatment variable. We define the

treatment variable to simply be the number of total reporting requirements (out of 14) adopted

by each state (see Table 1).

3. We employ coarser state-level (rather than provider-level) clustering of standard errors.

Coarser clustering is believed to yield more conservative estimates (Cameron and Miller 2015).

We redo our POA and HAC analyses (as decribed in §4.1–4.2) under each of these alternative

specifications. Again, we find that our results are consistent (see Table 9).

Change in (1) POA Reports (2) HAC ReportsSpecification Estimate SE p-value Estimate SE p-value

Probit Model -0.089 0.018 0.000 0.19 0.075 0.012Continuous Treatment −1.10 · 10−4 3.34 · 10−5 0.001 5.14 · 10−5 2.24 · 10−5 0.022State-Level Clustering −1.66 · 10−3 5.65 · 10−4 0.005 5.29 · 10−4 2.35 · 10−4 0.024

Table 9 Point estimates and cluster-robust standard errors for the coefficient of the treatment variable are

shown for alternative specifications of the POA and HAC regressions.

6. Discussion & Concluding Remarks

Our results show that providers in states with stronger regulations on adverse event reporting

have (1) lower risk-adjusted POA rates and (2) higher risk-adjusted HAC rates for CLABSIs and

CAUTIs. This effect is statistically significant even after controlling for a wide range of patient-level,

provider-level, and demographic characteristics as well as arbitrary intra-provider correlations and

endogeneity of regulation for HAC outcomes. While the POA results can potentially be explained by

weakly-regulated states having better infection detection capabilities or more susceptible patients,

these explanations are not consistent with the HAC results. In particular, both better infection

detection and increased patient susceptibility should translate to higher HAC rates in weakly-

regulated states as well. It could be argued that providers in weakly-regulated states have lower

HAC rates due to better quality of care. However, we find that risk-adjusted mortality rates (a

widely-used proxy for provider quality) for all conditions reported by Medicare are significantly

lower for providers in strongly-regulated states, making it unlikely that they have lower quality

facilities. Thus, the empirical evidence suggests that providers in weakly-regulated states were

upcoding during the sample period by reporting HACs as POAs. In particular, we conservatively


estimate that over 11,000 infections were upcoded per year, resulting in an added annual cost

burden of $200 million to Medicare reimbursements.

Our work suggests that financial incentives alone are not sufficient to reduce HAC incidence;

these policies must be accompanied by regulation to enforce truthful reporting. This hypothesis is

supported by recent evidence that the nonpayment policy has not reduced HAC rates (Lee et al.

2012). In fact, increasing financial incentives (e.g. HAC Reduction Program) or reputation incen-

tives (e.g. published infection rates on Hospital Compare) may worsen the problem as providers

may simply increase their rate of upcoding. Increased upcoding would have a number of negative

consequences:

1. Truthful providers are unfairly penalized and face greater financial pressure to upcode as well

2. Upcoding biases medical records resulting in a loss of accurate information. This interferes

with tracking harmful conditions and evaluating the effectiveness of policies aimed at improving

quality (Saint et al. 2009)

3. Publishing biased quality metrics may harm patients by routing them to providers who are

engaging in upcoding rather than providing better quality of care

Thus, we recommend that CMS implement measures to enforce truthful reporting by providers.

To this end, our results suggest two policy recommendations to help mitigate upcoding. First, we

suggest that CMS perform targeted audits of providers with a high POA-to-HAC reporting ratio.

As discussed in §2, providers with higher risk-adjusted POA reporting rates and lower risk-adjusted

HAC reporting rates are more likely to be engaging in HAC upcoding. Second, we recommend that

the federal government implement certain features of current state-level regulations that we find

to be effective at eliciting truthful reporting. Our analysis establishes the causal effect of stronger

regulation on decreased upcoding, and helps isolate some of the state adverse event reporting sys-

tem features that were successful in reducing upcoding. These include reporting patient-identifying

information (medical record number or billing number), a detailed description of the adverse event,

as well as the identified root cause of the adverse event. On the other hand, we note that simply

having a reporting system without stringent requirements produced no significant effect on report-

ing rates; we find that it is crucial that the regulation creates sufficient provider accountability.

We hypothesize that simply requiring providers to report detailed information on how and why

an adverse event occurred forces providers to implement the necessary infrastructure for detect-

ing and preventing HACs. Moreover, reporting more detailed information increases the threat of

setting off red flags when upcoding, and thus possibly diminishes the rate of upcoding. CMS may

benefit by implementing such detailed information reporting requirements in addition to existing

financial incentives to help improve hospital infrastructure and truthful reporting nationally. These


measures are especially important as Medicare moves towards more pay-for-performance policies

that rely on self-reported quality metrics.

Acknowledgments

The authors gratefully acknowledge Centers for Medicare & Medicaid Services for providing us with our

dataset.

References

Ashby, Nathan J, Fred McMahon, Avilia Bueno. 2010. Economic Freedom of North America 2010 . Fraser

Institute Vancouver, BC.

Ata, Baris, Bradley L Killaly, Tava Lennon Olsen, Rodney P Parker. 2013. On hospice operations under

medicare reimbursement policies. Management Science 59(5) 1027–1044.

Baum, C. F. 2007. Enhanced routines for instrumental variables/generalized method of moments estimation

and testing. Stata Journal 7(4) 465–506(42). URL http://www.stata-journal.com/article.html?

article=st0030_3.

Berenholtz, Sean M, Peter J Pronovost, Pamela A Lipsett, Deborah Hobson, Karen Earsing, Jason E Far-

ley, Shelley Milanovich, Elizabeth Garrett-Mayer, Bradford D Winters, Haya R Rubin, et al. 2004.

Eliminating catheter-related bloodstream infections in the intensive care unit*. Critical care medicine

32(10) 2014–2020.

Berriel-Cass, Dorine, Frank W Adkins, Polly Jones, Mohamad G Fakih. 2006. Eliminating nosocomial

infections at ascension health. Joint Commission Journal on Quality and Patient Safety 32(11) 612–

620.

Cameron, A Colin, Douglas L Miller. 2015. A practitioner’s guide to cluster-robust inference. Journal of

Human Resources 50(2) 317–372.

Center for Medicare & Medicaid Services. 2014. Statute regulations program instructions. URL http:

//www.cms.gov/Medicare/Medicare-Fee-for-Service-Payment/HospitalAcqCond/. Last Accessed

March 14, 2014.

Clemens, Jeffrey, Joshua D Gottlieb. 2013. Bargaining in the shadow of a giant: Medicare’s influence on

private payment systems. NBER Working Paper 19503 .

Deyo, Richard, Daniel Cherkin, Marcia Ciol. 1992. Adapting a clinical comorbidity index for use with

icd-9-cm administrative databases. Journal of Clinical Epidemiology 45 613–619.

Forthman, M Thane, Robert S Gold, Henry G Dove, Richard D Henderson. 2010. Risk-adjusted indices for

measuring the quality of inpatient care. Quality Management in Healthcare 19(3) 265–277.

Fuloria, Prashant C, Stefanos A Zenios. 2001. Outcomes-adjusted reimbursement in a health-care delivery

system. Management Science 47(6) 735–751.


Health Watch USA. 2011. Medicare’s payment policy for hospital acquired condi-

tions. URL http://www.healthwatchusa.org/HWUSA-Initiatives/PDF-Downloads/

20110705-HAC-Medicare-Non-Payment-comments.pdf. Last Accessed April 6, 2015.

Hsu, Eugene, Della Lin, Samuel Evans, Kamran Hamid, Kevin Frick, Ting Yang, Peter Pronovost,

Julius Cuong Pham. 2014. Doing well by doing good assessing the cost savings of an intervention to

reduce central line associated bloodstream infections in a hawaii hospital. American Journal of Medical

Quality 29 13–19.

KC, Diwas Singh, Christian Terwiesch. 2011. The effects of focus on performance: Evidence from california

hospitals. Management Science 57(11) 1897–1912.

Klevens, R Monina, Jonathan R Edwards, Chesley L Richards, Teresa C Horan, Robert P Gaynes, Daniel A

Pollock, Denise M Cardo. 2007. Estimating health care-associated infections and deaths in US hospitals,

2002. Public Health Reports 122(2) 160–166.

Lee, Grace M, Ken Kleinman, Stephen B Soumerai, Alison Tse, David Cole, Scott K Fridkin, Teresa Horan,

Richard Platt, Charlene Gay, William Kassler, et al. 2012. Effect of nonpayment for preventable

infections in us hospitals. New England Journal of Medicine 367(15) 1428–1437.

Levinson, Daniel R. 2008. Adverse events in hospitals: State reporting systems. URL https://oig.hhs.

gov/oei/reports/oei-06-07-00471.pdf.

Li, Jun, Nelson Granados, Serguei Netessine. 2014. Are consumers strategic? structural estimation from the

air-travel industry. Management Science 60(9) 2114–2137.

McNair, Peter, Harold Luft, Andrew Bindman. 2009. Medicare’s policy not to pay for treating hospital-

acquired conditions: The impact. Health Affairs 28 1458–1493.

Meddings, Jennifer, Sanjay Saint, Laurence McMahon. 2010. Hospital-acquired catheter-associated urinary

tract infection: documentation and coding issues may reduce financial impact of medicare’s new pay-

ment policy. Infection Control and Hospital Epidemiology 31 627–633.

Mukamel, Dana B, David L Weimer, Charlene Harrington, William D Spector, Heather Ladd, Yue Li. 2012.

The effect of state regulatory stringency on nursing home quality. Health services research 47(5)

1791–1813.

Powell, Adam, Sergei Savin, Nicos Savva. 2012. Physician workload and hospital reimbursement: Overworked

physicians generate less revenue per patient. Manufacturing & Service Operations Management 14(4)

512–528.

Saint, Sanjay, Jennifer A. Meddings, David Calfee, Christine P. Kowalski, Sarah L. Krein. 2009. Catheter-

associated urinary tract infection and the medicare rule changes. Annals of Internal Medicine 150(12)

877–884.

Schuller, Kristin, Janice Probst, James Hardin, Kevin Bennett, Amy Martin. 2014. Initial impact of medi-

care’s nonpayment policy on catheter-associated urinary tract infections by hospital characteristics.

Health Policy 115 165–171.


Scott, R Douglas II. 2009. The direct medical costs of healthcare-associated infections in US hospitals and

the benefits of prevention. Division of Healthcare Quality Promotion National Center for Preparedness,

Detection, and Control of Infectious Diseases, Centers for Disease Control and Prevention.

Selim, Alfredo J, Dan R Berlowitz, Graeme Fincke, Amy K Rosen, Xinhua S Ren, Cindy L Christiansen,

Zhongxhiao Cong, Austin Lee, Lewis Kazis. 2002. Risk-adjusted mortality rates as a potential outcome

indicator for outpatient quality assessments. Medical care 40(3) 237–245.

Silverman, Elaine, Jonathan Skinner. 2004. Medicare upcoding and hospital ownership. Journal of Health

Economics 23 369–389.

Snow, Catherine, Linda Hotlzman, Hugh Waters, Nancy McCall, Michael Halpern, Lisa Newman, John

Langer, Terry Eng, Carolyn Reyes Guzmann. 2012. Accuracy of coding in the hospital-acquired con-

ditions - present on admission program. Centers of Medicare and Medicaid Services .

Torio, Celeste M., Roxanne M. Andrews. 2013. National Inpatient Hospital Costs: The Most Expensive

Conditions by Payer, 2011. HCUP Statistical Brief #160 . Agency for Healthcare Research and Quality,

Rockville, MD.

Umscheid, Craig, Matthew Mitchell, Jalpa Doshi, Rajender Agarwal, Kendal Williams, Patrick Brennan.

2011. Estimating the proportion of healthcare-associated infections that are reasonably preventable

and the related mortality and costs. Infection Control and Hospital Epidemiology 32 101–114.

Wooldridge, Jeffrey M. 2010. Econometric analysis of cross section and panel data. MIT Press.

evidence of strategic behavior in medicare claims reporting · author: strategic behavior in...

Documents