1 Introduction
Constructing credible estimates of the current prevalence of SARSCoV2 in the United States is challenging. Despite growing since the start of the epidemic, testing rates remain low in most of the country: only a small fraction of the population is tested for SARSCoV2 on any given day. Moreover, tests are often allocated to people exhibiting COVID19 symptoms or who are thought to have come into contact with the virus (wsjTest). For example, New York State and Texas both use a selfdiagnostic tool to screen people for testing.^{1}^{1}1See https://covid19screening.health.ny.gov/covid19screening/ and https://txctt.force.com/ct/s/assessment?language=en_US. The low rate of testing in the general population means that the number of confirmed SARSCoV2 cases almost certainly understates the true number of infections in the population. At the same time, statistics like the fraction of tests that are positive likely overstate population prevalence because the tested population is more likely to be infected than the population as a whole.
In this paper, we propose a new approach to measuring the pointintime prevalence of active SARSCoV2 infections in the overall population using data on patients who are hospitalized for nonCOVID reasons. There is less uncertainty about SARSCoV2 prevalence for the nonCOVID hospital patient population because people in the hospital are tested at much higher rates than the general population, even if they are hospitalized for reasons unrelated to COVID19 (SuttonEtAl2020). We combine detailed testing data and hospital data from Indiana with a family of weak monotonicity assumptions that seem to have high credibility. The combination of these assumptions with linked testinghospital data leads to relatively tight upper and lower bounds on the prevalence of active SARSCoV2 infections in the overall population in Indiana in each week from midMarch to midJune. The detailed Indiana data allows us to conduct robustness checks that partially validate some of our assumptions. Our basic method (without the validation checks) could be implemented using data that many states are already collecting and partially reporting. Thus, our approach could help states extract timely prevalence information using existing surveillance data. Importantly, our paper is focused on estimates of the fraction of the population that would test positive in each week. These estimates are distinct from recent efforts to estimate the share of the population ever infected with SARSCoV2 (ManskiMolinari2020). The distinction between active prevalence and cumulative prevalence is important because the prevalence of active infections is a key determinant of the spread of the epidemic, given that the level of immunity in the population is thought to be quite low.
Estimates and forecasts of the prevalence of active SARSCoV2 infections are crucial for public and private responses to the disease. They have shaped decisions about disruptive nonpharmaceutical interventions such as school closures, nonessential business closures, gathering restrictions, stayathome mandates, and temporary increases in the generosity of the unemployment insurance system (gupta2020). Reported estimates of prevalence likely also motivate individual precautionary behaviors ranging from wearing a mask to reducing demand for goods and services that require physical interaction (gupta_brookings_2020; AllcottEtAl2020; philipson2000economic; philipson1996private; kremer1996integrating). Finally, estimates of prevalence are a necessary input into efforts to measure other quantities of interest, like the infection fatality rate and the infection hospitalization rate.
Given its importance, researchers have developed several approaches to measuring SARSCoV2 prevalence in light of the challenge of nonrepresentative testing. One of the most credible is to conduct a biometric survey in which tests are offered to a representative sample of the population (NirWave1; NirWave2; GudbjartssonEtAl2020). Other studies have tested a census of smaller populations such as cruise ships (e.g. DiamondPrincessFeb2020). However, it is difficult and costly to regularly implement a survey with accurate coverage and high response rates, especially a new survey that has not been in the field for long. A second approach involves backcalculation methods, which use data on observed hospitalizations or deaths to infer disease prevalence at earlier dates using assumptions about the unobserved parameters that determine the progression of the disease, hospitalization rates, and case fatality rates (brookmeyer1988method; egan2015review; FlaxmanEtAl2020; SaljeEtAl2020). Backcalculation may work well if hospitalizations or deaths are well measured, and if previous research has already reached consensus about key parameters related to the disease. However, backcalculation may be less credible for a novel virus because the scientific knowledge base is smaller. And even when backcalculation is based on credible assumptions, it may be of limited value for public health decision making because both hospitalizations and deaths lag current infections (VerityEtAl2020).
An alternative to biometric surveys and backcalculation is to combine nonrandom clinical testing data with weak distributional assumptions to construct bounds on population prevalence (e.g. manski1999identification; wing2010three). In the COVID19 epidemic, stock2020identification provide bounds on population prevalence using a testing encouragement design, which is not always available. ManskiMolinari2020 bound the share of the population that has ever been infected under a “test monotonicity assumption” that the infection rate is weakly higher among the tested than among the untested. This assumption is appealingly credible, and the bounds can be calculated from widely available test data. However, even when the focus is cumulative prevalence under test monotonicity, the prevalence bounds are often wide because testing is so rare.
We pursue a related strategy in this paper. Our analysis is based on the insight that test rates are especially high among hospitalized patients, even patients hospitalized for reasons that are apparently unrelated to COVID19, such as labor and delivery or vehicle accidents. The upper and lower bounds on prevalence in these populations will tend to be much tighter than similar bounds in the general population. In addition, it is plausible that some types of hospitalizations occur for reasons that are independent of infection risk. For example, people who are hospitalized for injuries sustained in a traffic accident might be expected to have about the same risk of SARSCoV2 infection as the general population. We build on this insight to estimate more informative upper and lower bounds on weekly SARSCoV2 prevalence in the population.
Our paper makes some methodological contributions that may be relevant to the development of more informative public health surveillance systems. We describe the conditions under which upper and lower bounds on active prevalence among nonCOVID hospitalizations are valid estimates of the upper and lower bounds on prevalence in the general population. We maintain the test monotonicity assumption throughout, and we derive upper and lower bounds on prevalence in the population under two alternative assumptions about the representativeness of nonCOVID hospitalizations for the broader population. The first assumption is a relatively weak “hospital monotonicity” assumption that prevalence is at least as high among nonCOVID hospital patients as it is in the general population. This assumption would be satisfied even if hospitalized patients are at greater risk of COVID because, for example, people who get into car accidents have more social interactions. The second assumption is a stronger ”hospital independence” assumption that prevalence is the same among nonCOVID hospital patients as it is in the general population. The resulting bounds are informative for prevalence at a point in time, not just cumulative prevalence. This is important because the information required to estimate the bounds is closely related to the simple statistics that many states already report. It appears possible for many states to use our method to report upper and lower bounds on prevalence in near real time using data that they already collect and report. In particular, states already report the COVIDhospitalization rate, as well as overall test rates and test positivity rates. To report a version of the upper and lower bounds we describe in this paper, states would simply have to compute testing and positivity rates among nonCOVID hospitalizations.
Our first empirical contribution uses this framework to estimate upper and lower bounds on weekly prevalence of SARSCoV2 in Indiana. To operationalize the basic idea, we work with two definitions of nonCOVID hospitalizations. The first, which we call nonInfluenza or COVIDlikeillness (nonICLI) hospitalizations, simply excludes all hospitalizations with diagnosis for ICLI (cdcCovidCodes; afhsc2015). This definition would be easy to implement in many different hospital data sets and yields a large population of hospital patients. However, we also work with a definition using a narrower set of patients who are hospitalized for six groups of clear nonCOVID causes: (i) cancer; (ii) appendicitis and vehicle accidents; (iii) labor and delivery; (iv) AMI and stroke; (v) fractures, crushes, and open wounds; and (vi) other accidents. The clearcause analysis is more intuitive and transparent, but it is based on smaller samples and might be harder to implement as part of a public health surveillance system.
We show that, in Indiana, test rates are much higher among nonCOVID hospital patients than in the general population. For example, in June, about 0.4 percent of the general population was tested in a given week, compared with 24 percent of nonCOVID hospital patients; test positivity rates are lower among nonCOVID hospital patients than in the general population. In the general population, 4.1 percent of tests were positive. In contrast, among nonCOVID hospital patients who were tested, only 2.6 percent of tests were positive. These testing and positivity rates can be combined to estimate upper and lower bounds on prevalence. Under the test monotonicity assumption, between 0.01 and 4.1 percent of the general population was infected with SARSCoV2 on June 15. Under the same test monotonicity assumption, the bounds are half as wide for nonCOVID hospital patients: prevalence was between 0.6 and 2.6 percent. Under the hospital monotonicity assumption, the upper bound on prevalence in the nonCOVID hospital population is a valid upper bound on population prevalence. And under the stronger hospital independence assumption, the upper and lower bounds on prevalence in the nonCOVID hospital population are valid upper and lower bounds on population prevalence. The bounds on prevalence based on the nonICLI definition are typically very similar to the bounds based on the clearcause definition. We present bounds under alternative hospital representativeness assumptions (none, monotonicity, independence) to allow readers to make their own judgements about which assumptions are credible, and to better understand how much information about prevalence is derived from the data versus assumptions.
We also calculate bounds among ICLI hospitalizations. We find, of course, much higher prevalence among this group, but our bounds still rule out very high prevalence. Specifically, we find that SARSCoV2 prevalence among ICLI hospitalizations is no higher than 50 percent at its peak, and no higher than about 30 percent by midJune. This shows that there is value to testing even highly symptomatic patients, as their SARSCoV2 rate is far from 100 percent, and testing outcomes would be informative for treatment and quarantine decisions.
Our second empirical contribution is to assess the credibility of key assumptions that would be difficult to study in other data sets. ManskiMolinari2020 point out that the accuracy of SARSCoV2 virological tests is not well understood. Incorporating information about testing errors alters the bounds on prevalence. We use data on people who were tested twice in a two day period to shed some light on the fraction of people who test negative but are actually infected. We tentatively conclude that test errors are have a negligible effect on the upper and lower bounds on prevalence reported in our paper.
We also assess the credibility of the hospital representativeness assumptions at the core of the paper. The most restrictive hospital independence condition assumes that SARSCoV2 prevalence is the same in the nonCOVID and general populations, and the weaker hospital monotonicity condition assumes that prevalence is at least as high among the hospitalized. Although we are not able to directly validate these assumptions, we probe their credibility in two ways. First, we compare the hospitalization bounds to estimates of population prevalence obtained from tests of a random sample of Indiana residents in April and June (NirWave1; NirWave2). Second, we examine the prehospitalization test rates of nonCOVID hospitalized patients. These validity checks are roughly consistent with the the hospital independence assumption, and highly consistent with hospital monotonicity. This suggests that these assumptions might be considered reasonable in other states where it would be easy to estimate the upper and lower bounds but harder to perform elaborate validation exercises.
Overall, our results indicate that combining testing data and information on nonCOVID hospitalizations may be a feasible and informative way of measuring SARSCoV2 prevalence. In the most recent weeks of our data (and under test monotonicity but without hospitalization data), we can only conclude that at most 4 percent of the overall Indiana population was actively infected. With hospitalization data and the hospital monotonicity assumption, we can conclude that at most 2.6 percent of the Indiana population was infected. Despite this substantial improvement, the bounds remain wide enough that we cannot conclude whether prevalence has fallen since early April. Similar bounds could be constructed in other states using aggregate data on nonCOVID hospitalizations and their testing and test positivity rates, potentially improving COVID surveillance systems across the country.
2 Inferring COVID prevalence from incomplete testing
The empirical goal of our study is to bound the fraction of the Indiana population that is infected with SARSCoV2 in each week. To fix ideas, we use to index the population of Indiana. Let indicate that person is currently infected with SARSCoV2 on date . The population prevalence of active SARSCoV2 infections in Indiana at date is , where we leave conditioning on date
implicit to reduce clutter. We are also interested in prevalence among hospital inpatients with various COVID and nonCOVIDrelated diagnoses. This is simply the probability that a person is infected, conditional on being hospitalized for a specified condition or set of conditions. Let
be a binary indicator set to 1 if the person was hospitalized with a nonCOVIDrelated diagnosis. Then is the prevalence of active SARSCoV2 infections in the subpopulation of people who were admitted with a nonCOVIDrelated diagnosis on date .A key inferential challenge in estimating prevalence is that values of are unknown for most people on most days, because testing is rare. Let if person was tested on and if the person was not tested. Let represent the proportion of the population tested on date , where conditioning on is implicit. Continuing with the notation laid out above, and represent prevalence among people who are tested and not tested, respectively. The value of is observed for people where , but unknown for people where , which means that is not identified by the data on testing and test outcomes.^{2}^{2}2Our analysis here assumes that there is no measurement error involved in COVID19 testing procedures. In practice, false negatives and false positives could create additional uncertainty. We abstract from such concerns throughout the paper. We define prevalence in the tested and untested hospital populations similarly, but with conditioning on both testing status and hospitalization status.
2.1 Worst Case Bounds
In the absence of any distributional assumptions, the observed clinical tests partially identify prevalence overall, and in any subpopulations that can be defined by observable covariates. Use the law of total probability to decompose population prevalence:
The only unknown quantity on the righthand side of the expression is , which is prevalence among people who were not tested. Without any additional assumptions or data, all that is known is that this value lies between and . Substituting 0 and 1 for the unknown prevalence yields worstcase lower and upper bounds and on population prevalence:
These bounds define the set of values for unknown population prevalence that are compatible with the observed data and the logical definition of prevalence. The lower bound is the confirmed positive rate, and the upper bound is that rate plus the untested rate. The width of the worstcase bounds on a given day is decreasing in that day’s testing rate. Testing more people can only increase the confirmed positive rate and decrease the untested rate. However if few people are tested, the bounds can be very wide.
2.2 Test monotonicity
To narrow the bounds, ManskiMolinari2020 propose the “test monotonicity” condition. This condition requires that the prevalence of SARSCoV2 is at least as high in the tested population as it is in the untested population. Formally, test monotonicity implies that . This is an appealing condition because virological tests are typically allocated to symptomatic individuals, who have a higher than average likelihood of infection.
Under test monotonicity, prevalence in the tested subpopulation represents an upper bound on the unknown prevalence among the untested subpopulation. The lower bound remains the worstcase lower bound. The lower and upper bounds under monotonicity, and , are:
The new upper bound will be lower than the worstcase upper bound as long as prevalence in the tested subpopulation is less than 1. In our data, test rates are less than 1 percent and positivity rates in the population are roughly 10 percent, so this assumption brings the upper bound down from 99 percent to 10 percent or less. A similar assumption could be made for the nonCOVID hospitalization population, yielding bounds and on prevalence among the nonCOVID hospitalized population.
2.3 Inferring population prevalence from nonCOVID hospital patients
Test monotonicity can be used to narrow the the bounds on prevalence in the population and among nonCOVID hospitalized patients. Because testing rates are much higher in hospitals than in the general population, the bounds on prevalence in hospitalized subpopulations are much narrower. Thus, assumptions that link hospital and population prevalence may be a powerful way to reduce uncertainty about population prevalence. We pursue two types of assumptions that enable extrapolation from nonCOVID hospital populations to the general population: (i) monotone selection into hospitalization and (ii) riskindependent hospitalization. These are both forms of hospital instrumental variable assumptions, and we refer to them collectively as hospital IV assumptions.
Hospital monotonicity lowers the population upper bound
The hospital monotonicity assumption requires that the prevalence of active SARSCoV2 infections among nonCOVID hospital patients is not lower than the prevalence of active infections in the general (nonhospitalized) population. Formally, . When prevalence is bounded in the hospitalized and general populations, the hospital monotonicity assumption may further reduce the width of both sets of bounds by ruling out values that would violate it.
In particular, under hospital monotonicity, the upper bound on population prevalence cannot be larger than the upper bound on hospital prevalence. When both the hospital monotonicity assumption and the test monotonicity assumption are imposed, the upper bound on SARSCoV2 prevalence in the population is
where is the upper bound on prevalence in the population under hospital monotonicity and is the upper bound on prevalence among nonCOVID hospital patients. In our data, the hospital upper bound is typically lower than the population upper bound, so the hospital monotonicity condition in practice implies that the positivity rate among nonCOVID hospitalizations is an upper bound population prevalence.^{3}^{3}3In principle, the same assumption could be used to derive a potentially higher lower bound on prevalence in the hospital population. But this lower bound condition is essentially nonbinding in practice, so we ignore it here.
Hospital independence tightens both population bounds
A stronger assumption that also facilitates extrapolation from a hospitalized subpopulation to the general population is a “hospital independence” assumption, which means that hospitalization for nonCOVIDrelated health conditions is mean independent of infection with SARSCoV2. Independence implies that people infected with SARSCoV2 have the same probability of being hospitalized for a nonCOVID condition as people who are not infected with SARSCoV2 so that . Equivalently, the independence assumption implies that SARSCoV2 prevalence is the same among people who are hospitalized for nonCOVID conditions and the general population. That is, under the hospital independence assumption: . The independence assumption implies that nonCOVID hospitalizations are an instrumental variable for testing. This assumption would be satisfied if nonCOVID hospitalizations arose randomly in the population, for example because of health conditions (such as pregnancy or heart disease) determined prior to the epidemic. The hospital independence assumption would fail, however, if hospitalization risk was systematically related to COVID19 risk, for example because essential workers have more social interactions and greater likelihood of hospitalizations.
Under the hospital independence assumption, the bounds on the common prevalence parameter are defined by the intersection of the hospital and population bounds. The lower and upper bounds under test monotonicity and hospital independence, and , are:
Under hospital independence (as well as test monotonicity), the upper bound on prevalence is the same as under hospital monotonicity. What the hospital independence assumption buys us is a tighter lower bound, which is now the greater of the lower bounds on population and hospital prevalence under test monotonicity. In practice we find that the lower bound is always higher in the nonCOVID hospitalization subpopulation than in the general population, so in practice this assumption implies that the lower bound on population prevalence is the confirmed positive rate among nonCOVID hospitalizations.
2.4 Measurement Error in Testing
Virological tests for the presence of SARSCoV2 may not be perfectly accurate, and so far there are no detailed studies of the performance of the PCR tests that Indiana is using to test people for SARSCoV2. To clarify how errorridden tests complicate our prevalence estimates, we augment our notation to distinguish between test results and virological status. We continue to use to represent a person’s true infection status, and we still use to indicate whether a person was tested at date or not. But now we introduce , which is a binary measure set to 1 if the person tests positive and 0 if the person tests negative. Using this notation, is called the Positive Predictive Value (PPV) of the test among people who are tested and who test positive. is called the Negative Predictive Value (NPV) among people who are tested and who test negative. is the fraction of people who test negative who are actually infected with SARSCoV2.
Our initial worst case bounds assumed no test errors. Relaxing that assumption yields a different set of upper and lower bounds on prevalence. Following ManskiMolinari2020, we assume that (i) so that none of the positive tests are false, but (ii) . The second condition imposes a bound on , which is the fraction of people who test negative who are actually infected. Under these two restrictions, the new worst case bounds work out to:
Allowing for test errors in this way increases the worst case lower bound by the bestcase fraction of missing positives, and increases the worst case upper bound by the worstcase fraction of missing positives. Similar expressions hold for prevalence bounds under test monotonicity and other independence assumptions.
The upshot of this analysis is that knowledge of test accuracy is important for efforts to learn about prevalence. In their study of the cumulative prevalence of SARSCoV2 infections, ManskiMolinari2020 computed upper and lower bounds on prevalence under the assumption that and , citing peci2014performance. ManskiMolinari2020 view this choice of as an expression of scientific uncertainty about test errors, and they refer to the resulting prevalence bounds as “illustrative”. However, the structure of the test error bounds makes it clear that assumptions about the numerical magnitude of test errors have inferential consequences. For example, setting implies that, regardless of the outcome of the test, at least 40 percent of the people who are tested for SARSCoV2 are infected.
Although there is little published evidence on the properties of the SARSCoV2 PCR test, previous research suggests that PCR test errors are uncommon in other settings. For example, peci2014performance study the performance of rapid influenza tests using PCRbased tests as a gold standard. PCR tests are used as a gold standard because they are expected to have very high PPV and NPV.
To shed more light on test errors, we constructed a sample of people who were (i) tested on day , (ii) not tested on day , and (iii) were tested again on day . A total of 16,401 test pairs met this criteria. Using and to represent the results of a person’s first and second test, we found that and among the people in the twicetested sample. The two tests were discordant for less than 1 percent of the twicetested sample. In Appendix C, we estimate prevalence and NPV in the twicetested sample under the strong assumptions that the tests have specificity equal to 1, sensitivity does not depend on initial test result, and retesting is random. Using this method, we find that prevalence was 12 percent and in the twicetested sample. With , the estimates imply only about 0.5 percent of people who test negative are actually infected.
The twicetested sample is likely not representative of the population, of course. People who are retested after a negative test are probably highly symptomatic.^{4}^{4}4Furthermore, it is possible that the first sample was drawn with less care than the second sample, suggesting that specificity may have differed across the repeated tests. This suggests that is probably higher in the twicetested sample than in the population. Accordingly, we think that a plausible value for is nearly zero, and a plausible value for is 0.005. Accounting for test errors in this range would have almost no effect on the upper and lower bounds reported in the paper.
2.5 Summary and data requirements
Our bounds turn out to be fairly simple objects. Under test monotonicity, the lower bound on prevalence is the confirmed positive rate, the share of the population that tests positive. The corresponding upper bound under monotonicity is the test positivity rate, the share of tests that are positive. Under hospital monotonicity, the upper bound becomes the test positivity rate among nonCOVID hospitalizations. And under hospital representativeness, the lower bound becomes the confirmed positive rate among nonCOVID hospitalizations.
An appealing feature of these bounds is that they can be calculated with little additional data beyond what public health organizations already report. Every state already reports the number of tests and the number of positive tests, and many states report the number of COVIDrelated hospitalizations.^{5}^{5}5See, e.g., covidTracking. States would only have to report test and positivity rates for nonCOVIDrelated hospitalizations. This appears possible because many states already report ”suspected” or ”under investigation” COVID hospitalizations, defined as hospitalized patients exhibiting COVIDlike illness. Some states actually report both the number of hospitalizations of patients with COVID or influenzalike illness and, separately, the number of hospitalizations of patients with a positive SARSCoV2 test (e.g. Arizona and Illinois (azDashboard; ilHosp; ilSyndromic)).^{6}^{6}6States reporting both confirmed SARSCoV2 hospitalizations and hospitalizations of suspected cases or cases under investigation include California (caDashboard), Colorado (coDashboard), Mississippi (msDashboard), Tennessee (tnDashboard), and Vermont (vtDashboard). Thus states have the capacity to identify ICLIrelated hospitalizations and link hospitalization and testing data.
3 Indiana Hospital and Testing Data
3.1 Data sets
Our analysis is based on two main data sources managed by the Regenstrief Institute. First we obtain data on the near universe of clinical virological tests for SARSCoV2 conducted in Indiana between January 1, 2020 and June 18, 2020. Second we obtained data on all inpatient hospital admissions from hospitals that belong to the Indiana Network for Patient Care (INPC), which is a health information exchange that centralizes and stores data from health providers across the state of Indiana, including all hospitals with emergency departments.^{7}^{7}7See grannisINPC2005 for more details. The hospital data are derived from the same database that the state uses for reporting hospitalizations on its dashboard (inCovidDashboard). We link the testing and hospital data using an encrypted common identifier. Of course, only a subset of hospital patients are tested and only a subset of tested people appear in the inpatient hospital data. For both data sets, we are also able to link the data with basic demographic data collected by the INPC; this information is available only for a subset of patients.
The test data contain individual records for nearly all of the SARSCoV2 tests conducted in Indiana during 2020. A small number of tests are excluded from our data because some institutions that conduct tests provide data to INPC but do not allow the data to be used for research purposes. The consequence of these exclusions is that we are missing some tests, which will result in a reduced lower bound in our framework. Despite these exclusions, our data set tracks the state’s official case counts fairly closely; see Appendix Figure A.1. As an aggregate summary, our data contain 39,472 total positive tests, and the state reports 41,541 positive tests as of June 18 (inCovidDashboard).^{8}^{8}8We compare positives rather than tests because some testing institutions appear not to report negative tests to the state. (Such institutions are not included in our data.) Each test record in our testing data includes information on the date the test was run, the outcome of the test (positive, negative, or inconclusive), and a patient identifier that we use to link the test data to demographic files and inpatient hospital files.
The hospital inpatient data contain separate observations for each admission. We always observe admission time and a patient identifier that we use to link to the test data and inpatient files. We observe discharge time and diagnosis information only for a subset of admissions. (Not all fields are available for all admissions because different institutions contribute different information to the INPC.) Because the INPC data come from health care providers and payers, the same hospitalization can appear in the data set multiple times. To deduplicate these records, we keep one observation per admission time (defined secondbysecond), keeping the observation with the most diagnosis codes.^{9}^{9}9This procedure could in principle result in us dropping relevant diagnostic information, but in practice this would happen only very rarely. There are 311,127 admissions that are unique in terms of patient identifier and time, and 270,576 nonunique admissions. In 579 of these nonunique cases are there multiple records with the same number of diagnostic codes; we keep one of these multiples at random. In 30 of the 270,576 cases there are multiple records with diagnosis codes that provide conflicting first diagnoses.
3.2 Key measures
Inhospital testing, positivity rate, and confirmed positives The fraction of people who are tested in the hospital is an important quantity of interest in our analysis because hospital patients are tested at a higher rate than the general population, and hospital testing may be less correlated with COVID symptoms. Our data do not distinguish whether a person was tested in the hospital or whether the test was initiated independently of the hospital visit. We say that a hospitalized patient was tested inhospital if she had at least one SARSCoV2 test dated between 2 days prior to admission to 4 days after admission. We chose to focus on this weeklong period, rather than strictly between admission and discharge, for three reasons. First, some patients will be tested prior to admission, as part of their preparation for admission. Thus it is valuable to look prior to admission. Second, we observe the date the test is run, not the the date the sample is collected. Backlogs in the testing system may mean that a patient is discharged before the test is run. Third, for some admissions, we lack discharge dates, but we can still define inhospital tests using this measure.
We say that a patient tests positive in the hospital if she has at least one positive COVID test between 2 days prior to admission and 4 days after admission. We define the positivity rate as the fraction of tests that are positive; the confirmed positive rate is the fraction of the population with a positive test.
Population Testing and Positivity Rates In some analyses we compare hospital testing and positivity to population testing and positivity. Hospital testing and positivity are defined over a weeklong span for a given hospitalization. To make the comparison with the general population clean, we examine test rates and positivity in a given weeklong period. We say that a person was tested if she was tested at least once in a given week, and we say she was positive if she was positive at least once in that period.
3.3 Sample construction
Throughout, a patient is in the “test sample” if they are tested at least once. We say a patient is in the “inpatient sample” if they are hospitalized at least once. We limit our analysis to admissions with nonmissing diagnostic information, and we say a patient is in the “inpatient diagnoses sample” if they meet this restriction. This limitation is important because diagnostic information is necessary for distinguishing COVIDrelated admissions from nonCOVIDrelated admissions.
ICLI and nonICLI Hospitalizations We construct three analytic samples from the inpatient data. We start by defining hospitalizations for influenza and COVIDlike illness (ICLI) using ICD10 codes. We identify admissions with any of a standard set of ICD10 codes for ICLI following afhsc2015. Then we identify admissions with any of the additional ICD10 codes that the CDC recommends using for coding COVID hospitalizations (cdcCovidCodes)
. Both the influenzalike and COVIDlike diagnoses include general symptoms such as cough or fever, as well as more specific diagnoses like acute pneumonia, viral influenza, or COVID19. We classify hospitalizations as ICLIrelated if they have any influenza or COVIDlike illness (ICLI) diagnoses, and we classify hospitalizations as nonICLI if they are not ICLIrelated.
Appendix B lists the ICD10 codes used to define the analytic samples.We view the nonICLI sample as a useful starting point for our analysis for two reasons. First, our hospital IV assumptions are most plausible for hospitalizations that are not obviously COVIDrelated, and this sample meets that criteria. Second, as we have noted, many states already classify hospitalizations as ICLIrelated; thus nonICLI hospitalizations are identifiable and measurable in nearreal time, so this sample can be studied more broadly.
We acknowledge, however, that the nonICLI sample may not satisfy the hospital IV assumptions for at least two reasons. First, it may condition on COVID itself, since a patient with a reported COVID diagnosis would be excluded from it. (In practice we observe many patients with positive COVID tests but no COVID diagnosis.) Second, COVID is a new disease with heterogeneous symptoms, so even if a patient is hospitalized because of COVID, she may not have one of our flagged diagnoses, and we may incorrectly call her hospitalization nonICLI (YangEtAl2020).
Clearcause sample To avoid these problems, we study a third sample, which we call the “clear cause” sample. These are hospitalizations with a clear cause that is not obviously COVIDrelated. We define clearcause hospitalizations as hospitalizations with a diagnosis code for labor and delivery, AMI, stroke, fractures, crushes, open wounds, appendicitis, vehicle accidents, other accidents, or cancer. For all of these conditions except cancer, we flag hospitalizations with a diagnosis at any priority. For cancer, we flag hospitalizations with a cancer diagnosis code as the admitting diagnosis, the primary final diagnosis, or any chemotherapy diagnosis. Appendix B lists the ICD10 codes used for these classifications.
We view the clearcause sample as important for two reasons. First, we believe the hospital IV assumptions are most plausible for this sample, so we believe the bounds on prevalence are most likely to be valid. Second, we view the clearcause sample as offering a test of the validity of the nonICLI sample. To the extent that the two samples generate similar bounds, we can be more confident that the nonICLI sample is informative of broader population COVID prevalence, despite the problems with the nonICLI classification. This would be valuable because classifying hospitalizations as ICLIrelated or not requires less information than ascertaining a clear cause of the hospitalization.
Summary statistics We show summary statistics for all of our samples in Table 1, as well as for the state as a whole (from Census Fact Finder and inCQF2019). The average tested and hospitalized patient is substantially older than the population as a whole, and also more likely to be female. Because the tested and hospitalized samples are not age representative of the general population, we reweight all samples to match the population age distribution. The tested and hospitalized samples are fairly similar to the general population in terms of racial composition. Limiting the inpatient sample to admissions with diagnoses reduces our sample size substantially, but it does not appear to change its demographic profile. Although our main analysis looks at test rates at the admission level (rather than the person level), the summary statistics show that hospitalized patients are vastly more likely to have ever been tested than the population as a whole – about 21 percent of everhospitalized patients, compared to about 280,000 out of 6.64 million (4.2 percent) for the general population. ICLIrelated hospitalizations are especially likely to have ever been tested.
4 Bounds on COVID19 prevalence
4.1 Bounds by broad sample
Figure 1 shows the SARSCoV2 testing rate for each of the subpopulations in our analysis. We report the exact values of each of the test rates and the weekly number of admissions in Appendix Table A.1. Because the tested and hospitalized populations have very different age distributions compared to the rest of the population, we reweight the hospital subpopulations to match the coarsened age distribution in the population. Specifically, we calculate test rates in weekbyagegroup cells, for age groups 017, 1830, 3050, 5064, 6574, and 75 and older. Then we average these agespecific testing rates across the age groups, weighting each group by its population share. We report unweighted test rates in Appendix Table A.2.
Figure 1 shows vastly higher test rates in the hospitalized samples than in the general population. The testing rate in the general population grew from 0.2 percent in April to between 0.4 and 0.6 percent in May and June. So despite tripling, the weekly test in the Indiana population rate remained below 1 percent. In contrast, people hospitalized for ICLI were tested at a very high rate, between 65 and 75 percent in most weeks. Testing rates among nonICLI hospital patients and among the clearcause nonCOVID hospital patients were lower than the ICLI sample but much higher than the population overall, about 20 percent in April and 2530 percent in May and June. In other words, testing rates among nonCOVID hospital inpatients are about 50 times higher than testing rates in the general population, but they are typically less than half as high as testing rates in the ICLI population.
These high test rates imply tighter bounds on population prevalence under test monotonicity, as we show in Figure 2 and report in Appendix Table A.3. We ageweight the bounds, analogously to our weighting of test rates, and we report unweighted bounds in Appendix Table A.4. Several patterns are clear in the figure. First, the ICLI hospitalized population has higher upper and lower bounds on prevalence than the other groups. For the ICLI patients, the prevalence bounds are 522 percent in the first week of our sample, then increase to 3040 percent in the last week of March, before declining steadily to 1118 percent in the final week of the sample. These bounds rule out the possibility that all or nearly all symptomatic patients are infected with SARSCoV2. In most weeks the ICLI bounds lie outside the other groups’ bounds, implying (unsurprisingly) unambiguously higher COVID prevalence. This separation shows that the data are sensible and that the bounds are informative enough to tell apart these highly distinct populations.
The second clear pattern in the figure is that the prevalence bounds are tighter for the nonICLI and clearcause hospitalization samples than for the alltest sample. In fact the bounds for both of these hospitalizations samples are always contained within the alltest bounds. Even at their tightest, the bounds for the alltest sample are as wide as 0.02 to 3.6 percent in the last week of our data. In that week the bounds for nonICLI hospitalizations are [0.6%, 2.4%] and for clearcause hospitalizations they are [0.5%, 1.9%]. This tight bound implies that the hospitalization data could be informative for population prevalence (under hospital monotonicity or mean independence).
The final pattern evident in Figure 2 is that the bounds for the nonICLI hospitalization sample and for the clearcause hospitalization sample are nearly indistinguishable. The only noticeable difference is that the upper bound for nonICLI hospitalization is perhaps slightly higher. This fact is important because nonICLI hospitalizations are potentially easier to measure, but they may be negatively selected in the sense that by construction they may exclude COVIDlikely cases. The similarity of the nonICLI bounds with the bounds for the clearcause sample (which is not selected based on COVIDlikelihood) provides some evidence in support of using nonICLI hospitalizations to measure general prevalence.
4.2 Bounds by cause of admission
Our clearcause hospitalization sample pools many distinct causes, including among others labor and delivery, vehicle accidents, and other accidents, including falls. In principle these hospitalizations may differ in their COVID likelihood. One might worry, for example, that pregnant women are especially cautious and careful not to become infected, whereas people getting into vehicle accidents may be less cautious (either because they are not careful drivers, or because they are out of the house at all).
We therefore report test rates and bounds for disaggregated causes as well as the overall clearcause sample. We focus on six sets of causes: cancers, appendicitis and vehicle accidents, injuries (fractures/crush/wounds), nonvehicle accidents, and AMI/stroke. These six groups have reasonable sample sizes throughout (they each have 1,0001,500 admissions per month), and the demographic profiles within each group are roughly similar; see Appendix Table A.5 for age profiles of admitted patients by cause of admission. All ages are represented in the cancer sample: appendicitis and vehicle accidents both afflict young people; AMI, stroke, and other accidents—primarily falls—afflict older people; and labor and delivery is limited, of course, to women of childbearing age.^{10}^{10}10These groups are not necessarily mutually exclusive, and in particular there is overlap between injury and accidents. Because not all age groups are represented in every category, we do not ageweight these results.
We report the monthly test rates for each of these groups in Figure 3. We also report sample sizes, exact tests rates, and bounds in Appendix Tables A.6 and A.7, by month and cause. All causes had relatively low test rates in March before rapid increases in April and May. By June there are some differences in the test rates, with higher rates for the injury, accident, and AMI/stroke admissions, and less testing for cancer and labor and delivery.
We report the bounds on prevalence by cause of admission in Figure 4. In March there is little testing; the bounds are wide and uninformative. The bounds tighten in April, May, and especially June. Importantly, we do not see obvious, systematic differences across the groups. Typically the bounds overlap, and there is no strong evidence that the upper or lower bounds differ by cause of admission. It is true that in June, cancer patients had zero positive tests, and the upper bound for appendicitis and vehicle accidents was below the lower bounds for injuries, other accidents, and AMI/stroke. However in April and May the cancer and appendicitis/vehicle patients had similar or even higher upper bounds than did the injury and other accident patients. This evidence shows that patients admitted to the hospital for different reasons and with different demographic profiles are all nonetheless tested at a high rate and with similar bounds on prevalence. This is perhaps reassuring for the view that pooling many distinct causes of admissions can nonetheless generate meaningful bounds on prevalence.
5 Assessing the hospital representativeness assumptions
The results so far show that the bounds on prevalence are much tighter for the nonICLI hospitalized population than for the population as a whole. This tighter bound is informative for general population prevalence only under assumptions about hospital representativeness, either a monotonicity assumption or an equal prevalence assumption. How valid are these assumptions? Assessing them directly is of course impossible because we lack data on prevalence in the population as a whole or in the hospital sample.
We have already provided one piece of indirect evidence in support of our hospital representativeness assumptions. The nonICLI and clearcause samples generate similar bounds, and, within the clearcause sample, there are not large differences in bounds across different causes of admission. This suggests that prevalence does not vary with the exact set of hospitalizations studied, although of course this does not prove monotonicity or representativeness.
In this section, we provide two additional pieces of evidence on the hospital IV assumptions. First we show that the hospital bounds are consistent with the estimates of population prevalence from the Indiana COVID19 Random Sample Study (NirWave1; NirWave2).^{11}^{11}11Our data do not contain the test results from the Random Sample Study, so we compare our bounds to the published results. Second, we compare the hospital sample to the general population in terms of their likelihood of prior testing (prior to the hospital data) and the test rate of their home counties. We take these to be proxies for their concern about COVID, although other interpretations are possible.
5.1 Comparison to random sample testing
A valuable benchmark for the hospitalbased prevalence bounds comes from a largescale study of SARSCoV2 prevalence in Indiana. The study invited a representative sample of Indiana residents (aged 12 and older) to obtain a SARSCoV2 test. The first wave of the study took place April 2529, and the second wave took place June 37. The preliminary results are reported in NirWave1 and NirWave2
. The response rate was roughly 25 percent, and no attempt was made to correct for nonrandom response. Nonetheless this survey appears to be the best benchmark available. We report the point estimates for prevalence (assuming random nonresponse) and their confidence intervals in the top panel of
Table 2. The first wave estimates 1.7 percent prevalence and the second 0.5 percent.^{12}^{12}12The estimates in Table 2 are slightly different from those reported by NirWave2. We report updated calculations, based on correspondence with the authors.We compare our prevalence bound during the same time periods in the bottom panel of the table. We limit our sample to tests of people aged 12 and older, for comparison with the population study. Using population testing we obtain very wide bounds that contain the random sample study estimates. This fact provides some support for the test monotonicity assumption. For both the nonCOVID hospitalization and clearcause hospitalization samples, the bounds are much tighter. Both sets of upper and lower bounds contain the April prevalence estimate, and both sets exclude the June estimate point estimates. However the confidence interval for the June 37 point estimates overlaps substantially with the nonICLI and clearcause bounds. Thus for both dates the prevalence point estimates are consistent with the bounds obtained from the nonCOVID hospitalizations. As a comparison we also report the bounds from the ICLIrelated hospitalizations, which always exclude the random sample estimates.
5.2 Comparison of prior testing and community testing
A standard way of measuring representativeness is to compare the distribution of covariates in a study population to their distribution in the target population. In our case, this approach is most convincing if we have wellmeasured covariates that proxy for having COVID19. Two candidate covariates are the community SARSCoV2 testing rate and the prior testing rate. The idea behind these proxies is that people who come from areas with high test rates, or who have been tested in the past, may themselves have a higher current likelihood of having COVID19.
To operationalize these measures, we define the community testing rate for person as the fraction of people in ’s county who have ever been tested, as of the end of our sample period. We define the prior test rate of person as of date as the probability that was tested at least once during the weeklong period . We focus on this window because it is the second week prior to our hospital testing window (which runs from to for a patient admitted at ). We allow for a week of time to elapse between the hospitalization and the “prior” testing because it is possible that some prehospital testing would occur in the window . When studying prior tests, we limit the sample to each person’s first hospitalization after March 1, 2020, to avoid picking up the higher testing that mechanically results from the fact that people hospitalized once are more likely than the general population to have been previously hospitalized. As with our bounds, here we weight the data to match the population age distribution.
Table 3 shows the community testing rate. The average county has a testing rate of 3.5%, with an interquartile range of 2.5% to 4.3%. The average person lives in a county with a test rate of 3.6%. The average nonICLI hospitalized patient comes from a county with a test rate of 4.1%: for clearcause hospitalizations it is 4.0%, and for ICLI hospitalizations it is 4.2%. These rates are all significantly different from the population average.
Figure 5 shows the prior testing rate as a function of admission date for the nonICLI hospitalization sample, the clearcause hospitalization sample, and the general population (for which the prior test rate on day is defined as the fraction tested between and ). The rates in the hospitalization samples are initially close to the population rate (when testing is low in general), but the lines diverge. By the last week of the sample, the prior testing rate is about 1.5 percentage points in the hospitalization samples, relative to approximately .75 percentage points in the population. Although the weekly differences are not individually statistically significant, overall the greater rate of prior testing is consistent with positive selection into hospitalization. However it is also consistent with the possibility that the hospitalization sample simply has more contact with the medical sector, resulting in greater testing at a given SARSCoV2 prevalence.
6 Conclusion
We have calculated weekly bounds on the prevalence of SARSCoV2 for the Indiana population as a whole and for three hospitalized populations: people hospitalized for influenza and COVIDlike illness, people hospitalized for other reasons, and people with clear (and clearly not COVID) causes of hospitalization. The bounds are valid under weak monotonicity assumptions. The bounds for the general population are wide but narrow over time. The bounds for the hospitalized population are much tighter, because the hospitalized populations are tested at a much higher rate than the general population.
The hospitalized populations are informative for the general population only under additional, stronger assumptions. In particular, if the hospitalized population is representative of the general population in terms of SARSCoV2 prevalence, then both the upper and lower bounds are valid. If the hospitalized population has a higher prevalence than the general population, then only the upper bound is valid. We assess these assumptions in multiple ways. We find that the nonCOVID hospitalized population bounds contained the point estimate of prevalence from a random sample in late April. By early June the point estimate from the random sample was below our lower bound, although we cannot reject that it was inside the bound. Hospitalized patients also appear to be tested for SARSCoV2 at somewhat higher rates than the general population, even outside the hospital. This last fact suggests that the representativeness assumption may be violated (although the magnitude may not be too severe), but both are consistent with a weaker monotonicity condition. Even under this monotonicity condition, the hospital data are still useful, bringing down the upper bound on population prevalence by a third or more.
We believe that the main promise of the nonCOVID hospitalization population is that it can provide nearreal time information about population prevalence. Only three numbers are necessary to calculate bounds from the nonCOVID hospitalizations: the count of nonCOVID admissions, the number of tests among this group, and the number of positive results. Although these numbers are not currently reported, many states already report related numbers, including both the number of COVID tests and the count of ICLIrelated hospitalizations. Thus the infrastructure largely exists already to calculate these bounds. The results here help validate this approach for realtime surveillance.
References
Appendix A Appendix figures and tables
Sample  Population  NonICLI  Clear Cause  ICLI  

% Tested  N  % Tested  N  % Tested  N  % Tested  
Week  (1)  (2)  (3)  (4)  (5)  (6)  (7) 
3/13/2020 
0.000  5,937  0.035  1,342  0.019  1,102  0.244 
3/20/2020  0.001  5,054  0.105  1,262  0.064  1,179  0.670 
3/27/2020  0.002  4,585  0.157  1,178  0.129  1,292  0.728 
4/03/2020  0.002  4,573  0.177  1,246  0.128  1,150  0.749 
4/10/2020  0.002  4,702  0.167  1,225  0.111  1,041  0.659 
4/17/2020  0.002  4,913  0.180  1,329  0.170  1,146  0.597 
4/24/2020  0.003  5,126  0.180  1,377  0.149  1,052  0.515 
5/01/2020  0.004  5,618  0.277  1,420  0.289  1,143  0.676 
5/08/2020  0.005  6,031  0.284  1,531  0.235  1,050  0.692 
5/15/2020  0.006  6,658  0.285  1,549  0.248  1,081  0.737 
5/22/2020  0.004  6,663  0.267  1,565  0.238  1,026  0.663 
5/29/2020  0.005  6,763  0.280  1,542  0.271  947  0.681 
6/05/2020  0.006  7,172  0.263  1,684  0.269  943  0.690 
6/12/2020  0.004  6,091  0.243  1,411  0.209  776  0.661 
Notes: Table reports the weekly test rate for the population, and the number of hospitalizations and test rate, by type of hospitalizations, weighted to match the population age distribution.. (The population size is 6.64 million in all weeks.) ICLI hospitalizations have at least one diagnosis for influenzalike or COVIDlike illness. Clear cause hospitalizations are hospitalizations for cancer, labor and delivery, AMI, stroke, fracture or crush, open wound, appendicitis, or accidents (vehicle or other). See Appendix B for definitions.
Sample  Population  NonICLI  Clear Cause  ICLI  

% Tested  N  % Tested  N  % Tested  N  % Tested  
Week  (1)  (2)  (3)  (4)  (5)  (6)  (7) 
3/13/2020 
0.000  5,937  0.035  1,342  0.019  1,102  0.244 
3/20/2020  0.001  5,054  0.105  1,262  0.064  1,179  0.670 
3/27/2020  0.002  4,585  0.157  1,178  0.129  1,292  0.728 
4/03/2020  0.002  4,573  0.177  1,246  0.128  1,150  0.749 
4/10/2020  0.002  4,702  0.167  1,225  0.111  1,041  0.659 
4/17/2020  0.002  4,913  0.180  1,329  0.170  1,146  0.597 
4/24/2020  0.003  5,126  0.180  1,377  0.149  1,052  0.515 
5/01/2020  0.004  5,618  0.277  1,420  0.289  1,143  0.676 
5/08/2020  0.005  6,031  0.284  1,531  0.235  1,050  0.692 
5/15/2020  0.006  6,658  0.285  1,549  0.248  1,081  0.737 
5/22/2020  0.004  6,663  0.267  1,565  0.238  1,026  0.663 
5/29/2020  0.005  6,763  0.280  1,542  0.271  947  0.681 
6/05/2020  0.006  7,172  0.263  1,684  0.269  943  0.690 
6/12/2020  0.004  6,091  0.243  1,411  0.209  776  0.661 
Notes: Table reports the weekly test rate for the population, and the number of hospitalizations and test rate, by type of hospitalizations. (The population size is 6.64 million in all weeks.) ICLI hospitalizations have at least one diagnosis for influenzalike or COVIDlike illness. Clear cause hospitalizations are hospitalizations for cancer, labor and delivery, AMI, stroke, fracture or crush, open wound, appendicitis, or accidents (vehicle or other). See Appendix B for definitions.
Sample  Population  NonICLI  Clear Cause  ICLI  

Lower  Upper  Lower  Upper  Lower  Upper  Lower  Upper  
Week  (1)  (2)  (3)  (4)  (5)  (6)  (7)  (8) 
3/13/2020  0.0000  0.121  0.0022  0.055  0.0012  0.037  0.0468  0.171 
3/20/2020  0.0002  0.154  0.0130  0.116  0.0104  0.106  0.2128  0.312 
3/27/2020  0.0004  0.175  0.0264  0.147  0.0136  0.090  0.3050  0.402 
4/03/2020  0.0004  0.157  0.0217  0.119  0.0120  0.071  0.2480  0.325 
4/10/2020  0.0004  0.158  0.0193  0.112  0.0099  0.085  0.2127  0.312 
4/17/2020  0.0005  0.187  0.0089  0.044  0.0047  0.027  0.1417  0.230 
4/24/2020  0.0006  0.141  0.0129  0.068  0.0071  0.038  0.1873  0.348 
5/01/2020  0.0006  0.125  0.0148  0.052  0.0120  0.042  0.2074  0.312 
5/08/2020  0.0005  0.094  0.0111  0.037  0.0155  0.069  0.1414  0.210 
5/15/2020  0.0005  0.088  0.0099  0.033  0.0100  0.039  0.1590  0.221 
5/22/2020  0.0003  0.076  0.0141  0.053  0.0087  0.033  0.1411  0.202 
5/29/2020  0.0003  0.063  0.0079  0.027  0.0059  0.020  0.1227  0.179 
6/05/2020  0.0003  0.047  0.0085  0.032  0.0061  0.021  0.1348  0.195 
6/12/2020  0.0001  0.041  0.0062  0.026  0.0037  0.017  0.0970  0.153 
Notes: Table reports weekly bounds on COVID prevalence in the indicated sample under test monotonicity, weighted to match the population age distribution. The lower bound is the confirmed positive rate and the upper bound is the positivity rate. ICLI hospitalizations have at least one diagnosis for influenzalike or COVIDlike illness. Clear cause hospitalizations are hospitalizations for cancer, labor and delivery, AMI, stroke, fracture or crush, open wound, appendicitis, or accidents (vehicle or other). See Appendix B for definitions.
Sample  Population  NonICLI  Clear Cause  ICLI  

Lower  Upper  Lower  Upper  Lower  Upper  Lower  Upper  
Week  (1)  (2)  (3)  (4)  (5)  (6)  (7)  (8) 
3/13/2020  0.0000  0.150  0.0036  0.112  0.0022  0.079  0.0534  0.228 
3/20/2020  0.0002  0.189  0.0153  0.158  0.0142  0.176  0.2713  0.416 
3/27/2020  0.0004  0.217  0.0303  0.196  0.0195  0.128  0.3927  0.510 
4/03/2020  0.0004  0.191  0.0252  0.141  0.0223  0.133  0.3131  0.410 
4/10/2020  0.0004  0.189  0.0232  0.128  0.0155  0.093  0.2620  0.374 
4/17/2020  0.0005  0.214  0.0124  0.068  0.0090  0.046  0.1922  0.315 
4/24/2020  0.0006  0.167  0.0149  0.081  0.0109  0.062  0.1896  0.369 
5/01/2020  0.0006  0.131  0.0157  0.057  0.0169  0.052  0.2047  0.308 
5/08/2020  0.0005  0.097  0.0126  0.046  0.0137  0.047  0.1698  0.248 
5/15/2020  0.0005  0.087  0.0122  0.043  0.0148  0.047  0.1616  0.224 
5/22/2020  0.0003  0.074  0.0134  0.050  0.0159  0.055  0.1654  0.236 
5/29/2020  0.0003  0.060  0.0088  0.032  0.0084  0.027  0.1309  0.188 
6/05/2020  0.0003  0.045  0.0089  0.033  0.0089  0.029  0.1391  0.207 
6/12/2020  0.0002  0.036  0.0058  0.024  0.0049  0.019  0.1121  0.183 
Notes: Table reports weekly bounds on COVID prevalence in the indicated sample under test monotonicity. The lower bound is the confirmed positive rate and the upper bound is the positivity rate. ICLI hospitalizations have at least one diagnosis for influenzalike or COVIDlike illness. Clear cause hospitalizations are hospitalizations for cancer, labor and delivery, AMI, stroke, fracture or crush, open wound, appendicitis, or accidents (vehicle or other). See Appendix B for definitions.
Number of  Age  

Group  Admissions  Newborn  017  1829  3049  5064  6574  74 
All  236,267  0.085  0.029  0.108  0.185  0.226  0.176  0.191 
Has diagnosis  106,354  0.108  0.025  0.117  0.182  0.209  0.167  0.192 
ICLI  14,992  0.004  0.024  0.039  0.153  0.274  0.231  0.276 
NonICLI  91,362  0.125  0.026  0.130  0.187  0.198  0.157  0.178 
Clear cause  19,721  0.003  0.032  0.173  0.183  0.194  0.178  0.237 
Cancer  3,159  0.000  0.052  0.027  0.119  0.330  0.288  0.184 
Labor/delivery  4,690  0.010  0.015  0.595  0.380  0.000  0.000  0.000 
AMI  2,567  0.000  0.000  0.005  0.110  0.324  0.273  0.288 
Stroke  2,652  0.000  0.005  0.011  0.088  0.286  0.263  0.346 
Fracture  4,147  0.002  0.038  0.061  0.122  0.175  0.193  0.409 
Open wound  1,125  0.000  0.050  0.101  0.186  0.208  0.172  0.283 
Appendicitis  633  0.000  0.212  0.196  0.275  0.185  0.085  0.047 
Vehicle accident  582  0.002  0.089  0.194  0.304  0.215  0.134  0.062 
Other accident  3,072  0.002  0.035  0.034  0.079  0.154  0.213  0.483 
Notes: Table reports the number and age distribution of admissions, for different categories of admissions, over the time period March 13, 2020 through June 18, 2020. See Appendix B for definitions of the different causes of admissions (Cancerother accident).
Month  # Admits  % Tested  Lower bound  Upper bound 

A. Cancer  
March  595  0.069  0.002  0.024 
April  873  0.166  0.008  0.048 
May  1,029  0.269  0.005  0.018 
June  662  0.183  0.000  0.000 
B. Labor and delivery  
March  904  0.009  0.000  0.000 
April  1,456  0.052  0.003  0.053 
May  1,489  0.163  0.007  0.041 
June  841  0.188  0.004  0.019 
C. AMI  
March  437  0.188  0.023  0.122 
April  706  0.344  0.041  0.119 
May  839  0.451  0.041  0.090 
June  585  0.373  0.012  0.032 
D. Stroke  
March  469  0.102  0.021  0.208 
April  744  0.198  0.013  0.068 
May  932  0.303  0.015  0.050 
June  507  0.249  0.010  0.040 
D. Fractures and crushes  
March  616  0.055  0.002  0.029 
April  1,109  0.210  0.011  0.052 
May  1,483  0.373  0.014  0.038 
June  939  0.408  0.011  0.026 
Notes: Table reports monthly number of admissions for cancer, along with the percent of admissions tested in hospital, and the bound on prevalence under test monotonicity. The lower bound is the confirmed positive rate and the upper bound is the positivity rate. See Appendix B for definitions.
Month  # Admits  % Tested  Lower bound  Upper bound 

E. Open wounds  
March  182  0.110  0.005  0.050 
April  294  0.167  0.010  0.061 
May  401  0.314  0.025  0.079 
June  248  0.298  0.016  0.054 
F. Vehicle accidents  
March  61  0.049  0.000  0.000 
April  136  0.154  0.000  0.000 
May  222  0.212  0.009  0.043 
June  163  0.288  0.000  0.000 
G. Other accidents  
March  530  0.094  0.025  0.260 
April  815  0.221  0.028  0.128 
May  1,085  0.368  0.011  0.030 
June  642  0.336  0.011  0.032 
H. Appendicitis  
March  120  0.033  0.017  0.500 
April  153  0.203  0.000  0.000 
May  228  0.333  0.022  0.066 
June  132  0.417  0.008  0.018 
Notes: Table reports monthly number of admissions for cancer, along with the percent of admissions tested in hospital, and the bound on prevalence under test monotonicity. The lower bound is the confirmed positive rate and the upper bound is the positivity rate. See Appendix B for definitions.
Appendix B Defining causes of admissions
This section provides more details on our definition of ICLI, nonICIL, and “clear cause” hospitalization, listing the ICD10 codes used to define each.
Following afhsc2015, the codes for influenzalike illness are B97.89, H66.9, H66.90, H66.91 H66.92, H66.93, J00, J01.9, J01.90, J06.9, J09, J09.X, J09.X1, J09.2, J09.X3, J09.X9, J10, J10.0, J10.00, J10.01, J10.08, J10.1, J10.2, J10.8, J10.81, J10.82, J10.83, J10.89, J11, J11.0, J11.00, J11.08, J11.1, J11.2, J11.8, J11.81, J11.82, J11.83, J11.89, J12.89, J12.9, J18, J18.1, J18.8, J18.9, J20.9, J40, R05, and R50.9. We say a hospitalizaiton is for an influenzalike illness if it has any of these diagnosis codes in any position. We say a hospitalization is for a COVIDlike illness if it has any ICD10 code among those that is among the CDC’s lists of diagnosis codes for COVID19 cdcCovidCodes. These codes are J12.89, J20.8, J22, J40, J80, J98.8, O95.5, R05, R06.02, R50.9, U07.1, Z03.818, Z11.58, and Z20.828.
We define ICLIrelated hospitalizations as ones with at least one ILI or CLI diagnosis code. We define nonICLI related hospitalizations as hospitalized with diagnosis codes, but no ILI or CLI code.
We also define “clear cause” hospitalizations. These are hospitalizations for labor and delivery, AMI, stroke, fractures and crushes, wounds, vehicle accidents, other accidents, appendicitis, or cancer. With the exception of cancer, we define a hospitalization as belonging to one of these groups if it has any diagnosis codes for that group, listed below. Cancer is treated differently because it can be a comorbidity. We say a hospitalization is for cancer if a cancer diagnosis (listed below) is an admitting diagnosis, the primary final diagnosis, or if chemotherapy diagnosis is present. We use the following ICD10 codes.

AMI I21, I22.

Appendicitis K35K38.

Cancer C00C97 (in primary or admitting diagnosis), or Z51.0Z51.2 (in any position).

Fracture/Crush S02, S12, S22, S32, S42, S52, S62, S72, S82, S92, T02, S07, S17, S37, S47, S57,S67, S77, S87, S97, T07.

Labor and delivery O60O75, O80O84.

Other accidents W00W99, X00X59.

Stroke I61I64.

Vehicle accident V01V99.

Wound S01, S11, S21, S31, S41, S51, S61, S71, S81, S91, T01.
Appendix C Calculating negative predictive values with testretest data
Setup and identification Here we show how to use data on multiple tests to simultaneously identify prevalence, test error rates, and how to use this information to obtain the negative predictive value, NPV. Assume in particular that people have been tested exactly twice, with the outcome of the first test and the outcome of the second test for person , and person ’s true infection status, which we assume is fixed between the tests. Let be the prevalence of active SARSCoV2 infections in this twice tested population.
Test outcomes may differ from true infection status because of test errors. In general, therefore, there are four possible sequences of test outcomes: . We let for .
We make three strong assumptions to simplify the analysis.

The specificity of the test is 1. That is, .

The sensitivity of the test, , does not depend on the initial test result.

Retesting is random, i.e. independent of and .
Assumption 1 is the weakest of these assumptions. It implies that there are no false positives, which is consistent with typical practice (ucsfDiagnostic). The remaining assumptions are stronger. Assumption 2 says that the test errors are independent of the initial test result. It would be violated, for example, if false negatives are more common for patients with high levels of mucus, and mucus levels are correlated across test results. Assumption 3 says that retesting rates do not depend on possible testing errors. We would expect this condition to fail if highly symptomatic people with negative tests are especially likely to test negative. We view this assumption as the most suspect.
Under these assumptions, the probabilities simplify considerable. As the probabilities sum to one, and the assumptions imply that , the only nonredundant probabilities are
We can observe and . Solving for the unknowns and , we have
This shows how to get and from two tests and the assumption that specificity equals 1. Our goal is to find the negative predictive value, but we can calculate it given knowledge of and . In general the NPV of a single test is . Applying Bayes rule,
Results To implement this approach, we construct a sample of all people who are tested on a given day, not tested the previous day, and then tested again in the next day. There are 16,401 such test pairs. We find and . Nearly all the mass is on the diagonals; test results switch less than 1% of the time. This fact, together with the assumption that specificity is equal to 1, implies very low false negative rates. Plugging these values into our formula, we have and , which imply . Using instead, all people who are retested once within a three day period, we find similar results: .
We emphasize that these estimates are valid for the twicetested population and under assumptions 13, in particular, random retesting. The prevalence estimate is the prevalence among people tested twice, not the population prevalence. And it is only a valid estimate under assumptions 13. In reality, it is likely that retests are most common among suspected false negatives (i.e. when a highly symptomatic patient tests negative). We see some evidence for this: and , inconsistent with the random retesting assumption. We therefore do not view our estimates of prevalence and sensitivity as definitive; rather we think of the sensitivity estimate as a lower bound on sensitivity, because we have selected a retest sample which has a disproportionate number of false negatives. As is increasing in sensitivity, , our implied estimate of is likely an upper bound on .
Comments
There are no comments yet.