Early and accurate detection of community outbreaks is critical to address the threat of resurgent waves of COVID-19. Specifically, an epidemic outbreak is confirmed when its incident cases are estimated to grow exponentially. Furthermore, the potential impact of an outbreak is also measured by its exponential growth rate, as higher rates indicate more rapid disease spread. Last but not least, for a fixed epidemiology model (e.g. SIR, SEIR), there is an one-to-one correspondence between the exponential growth rate of an epidemic outbreak and its basic reproduction number , a common measure of intensity of epidemic outbreaks (lipsitch2003transmission). Therefore, the exponential growth rate of an epidemic outbreak’s incident cases is the most important ”model-free” parameter to estimate for detecting the outbreak (chowell2003sars).
It remains an epidemiological challenge to obtain accurate exponential growth rate estimates for disease outbreaks (ma2014estimating)
. Specifically, it is difficult to choose the fitting window size for the exponential growth rate estimation to balance the speed and accuracy of outbreak detection. On one hand, a longer fitting window is preferable as larger sample size would reduce variance of the exponential growth rate estimates of outbreaks. On the other hand, shorter fitting windows are better at detecting early-stage outbreaks, especially if these outbreaks were driven by recent policy changes such as school reopening. In the current practice, this fitting window size is treated as a hyperparameter that is either directly specified by the user (c.f.Mel) or determined by some cross validation methods (c.f. chowell2007comparative).
This paper develops a machine learning framework that balances the speed-accuracy tradeoff of outbreak detection via dedicated feature engineering and GRF (c.f. athey2019generalized), and apply it to detect county-level COVID-19 outbreaks. Specifically, the algorithm chooses an adaptive fitting window size for each county based on a rich set of features that affect the disease spread, such as face mask mandates, social distancing policies, the CDC’s Social Vulnerability Index, changes in tests performed and rate of positive tests. Furthermore, for counties with insufficient data to capture the recent policy changes, the algorithm pools together all relevant incident case growth trends across U.S. counties and throughout the COVID-19 pandemic history to adjust for these policy changes.
2.1 Exponential Growth Model and Exponential Growth Rate
During an epidemic outbreak, incident case number at county on day , i,e. , is governed by an exponential growth model,
(ma2014estimating).111See Appendix B for the epidemiological definition of incident case number and how we compute . Here is the incident cases exponential growth rate of outbreaks for all counties , while captures the initial incident case numbers at the beginning of the exponential case growth at county . To obtain the most recent exponential growth rate of county-level COVID-19 incident case number, we estimate the instantaneous counterpart of (1),
where the dependent variable is the log-linearized incident case number ; independent variable is day . The parameter of interest is the COVID-19 incident case exponential growth rate of county at day , . The intercept term, , is a parameter capturing the log-linearized initial incident case number of county at the beginning of the outbreak. is the error term.
Notably, the exponential growth rate in (2) varies in both day and county . In other words, we are interested in estimating the instantaneous county-level exponential growth rate of COVID-19 incident cases. Specifically, since COVID-19 related regulations are changing every day in the U.S. (c.f. CUSP), the county-level exponential growth rates, affected by these policies, also change from day to day. Therefore, in order to detect recent outbreaks, we need to estimate the most recent incident cases exponential growth rate. This instantaneous exponential growth rate , similar to the instantaneous reproduction number commonly used in the epidemiology literature (c.f. fraser2007estimating; cori2013new), captures the projected exponential growth rate of incident cases in county should the future COVID-19 regulations remain the same as those in day .
2.2 Relevant Features Affecting COVID-19 Disease Spread
The exponential growth rate of COVID-19 incident cases can be affected by many factors ranging from day-to-day changes in COVID-19 regulations to difference in population density and healthcare resources between counties. Hence, in order to estimate the instantaneous county-level exponential growth rate defined in (2), we need to control for these day-level and county-level heterogeneity. Provided that we have relevant features that captures these aforementioned factors affecting the instantaneous county-level exponential growth rate ,222Refer to Appendix C for the list of relevant features used in this study. we can identify the conditional average partial treatment effect as defined by wooldridge2010econometric. When data of these relevant features affecting COVID-19 disease spread is available, we can rewrite (2) as
following the common “redundant” assumption (c.f. wooldridge2010econometric), i.e.
3 Model Estimation
This section discusses how we estimate the exponential growth rate of COVID-19 incident cases defined in (3). Specifically, we first formulate the estimation problem in §3.1 and then present our estimation algorithm in §3.2.
3.1 Problem Formulation
First, we need the following “unconfoudness” assumption (c.f. rosenbaum1983central):
That is, day
is independent of all unobservable heterogeneity conditional on the feature vector. Under Assumption 3.1
, we can derive the following moment equations for (3): ,
through the instrumental variable , where is a Boolean function equal 1 if .
However, we note that the moment equations (4) cannot identify the exponential growth rate . Specifically, the moment equation system (4) is underidentified as the number of unknown parameters is exactly two times the number of moment equations. To address this problem, we assume that the exponential growth rate of day equal the local average causal response (c.f. abadie2003semiparametric; angrist1995two) from day to day , i.e.
3.2 Estimation Algorithms
Given the above assumptions, the exponential growth function of COVID-19 incident cases (3) can be estimated using the GRF algorithm by athey2019generalized. Specifically, Assumption 3.1 implies that the exponential growth rate of a county on day is identified by a targeted data “block” . Hence, when estimating , we can partition the panel dataset into data blocks, and pool those data blocks “similar” to the targeted block to construct an adaptive window size for this estimation. Particularly, the similarity measure is provided by the GRF algorithm. In Appendix A, Algorithm LABEL:alg:block provides the pseudocode to construct these data blocks. Algorithm LABEL:alg:grf explains how we feed these data blocks into the GRF algorithm to obtain the estimates of interest.
4 Performance Evaluation
To benchmark our method against non-adaptive window size choices, we compare the Mean Absolute Percentage Error (MAPE) of these methods’ 7-day ahead predictions. Specifically, we use the NYTimes COVID-19 Dataset (c.f. NYTimes2020) as the source of daily reported cases per county. We note that there are many data quality issues with these reported case counts, especially during the first few months of this pandemic (c.f. NYTimesIssue). As an example, bergman2020oscillations found that parts of the oscillation in U.S. COVID-19 case counts can be explained by increased testing during some week days and test backlogs during the weekend. To alleviate these issues, this section preprocesses the raw case data using 4-day moving average and only compares the median MAPEs. Additional results for performance evaluation are available at Appendix D.
As shown in Figure LABEL:fig:MAPE4 and Table LABEL:tab:4, our method outperforms methods with fixed 2-, 4-, 8- or 16-day fitting window sizes. Specifically, Figure LABEL:fig:MAPE4 shows that our method provides a uniformly better performance roughly 100 days after the first recorded COVID-19 case in the NYTimes COVID-19 Dataset, when there were enough historical data for the GRF algorithm to conduct meaningful partition. Furthermore, Table LABEL:tab:4 demonstrates that even if the early-day MAPEs are included in the comparison, our method still has the best median MAPE among the 4 methods. Last but not least, we note that when only comparing the performance of non-adaptive window size choices, there are no obvious best choice. While choosing shorter fitting window sizes could in general lead to lower median MAPEs (c.f. Table LABEL:tab:4), it was still frequently outperformed by longer fitting window size choices (c.f. Figure LABEL:fig:MAPE4). This inherent difficulty in non-adaptive window size choices suggests that treating fitting window size as a hyperparameter and choosing it through cross validation methods (c.f. chowell2007comparative) cannot balance the the speed-accuracy tradeoff in outbreak detection.
5 Conclusion and Future Work
In this work, we developed a novel framework in feature engineering that allows GRF to match observations across time and space, and adequately balance the the speed-accuracy tradeoff in COVID-19 outbreak detection. This estimation framework can be readily extended to other panel data estimation problems in epidemiology where fitting window size choices play an important role, e.g. estimating the county-level instantaneous reproduction number (c.f. fraser2007estimating; cori2013new).
Appendix A Pseudocode for Estimation Algorithms
Appendix B Epidemiological Definitions
For this work, where we wish to provide useful estimates of case trajectories for epidemiologists and decision makers in the government, we ideally wish to measure the growth rate of active case numbers
where is the cumulative number of cases so far, is the cumulative number of deaths, and is the cumulative number of recovered cases. This intuitively captures the number of still ”infectious” cases, as we can assume for COVID-19 that those who have recovered or died are no longer capable of infecting others. Unfortunately, as the number of recovered cases are no longer reported, we are unable to directly calculate the number of active cases. While other works (c.f. Mel) approximate the number of recovered cases with
Here the underlying assumption is that those who were infected but not dead after 22 days have recovered. However, upon testing this assumption at the county-level, we find that this assumption does not hold as there will be some days where this newly approximated active case number becomes negative.
We hence rely upon a commonly used proxy: the incident case count, , defined as
which is the number of new cases within a 22 day time period and is a useful proxy for infectious cases.
Appendix C Feature Data
All feature data we used, i.e. , are publicly available online. In this section, we briefly describe what these datasets are and how we incorporated them in our work.
c.1 2019 US Census Gazetteer Files
The 2019 United States Census Gazetteer Files (c.f. USCensus2019) were used to obtain the geographic locations (latitude-longitude centroids) of each county officially registered by the United States Census Bureau. This provides a spatial feature space for the GRF to further split upon.
c.2 Centers for Disease Control and Prevention Social Vulnerability Index 2018 Database
The Social Vulnerability Index (SVI) database (c.f. CDCSVI2018) is a compilation of socio-economic factors such as unemployment rate, poverty rate, education attainment level etc. at the county level. These features are also included in our methodology.
c.3 COVID-19 US state policy database (CUSP)
The CUSP database (c.f. CUSP) tracks when each state implemented and ended policies such as mask mandates, lockdowns, economic policies in response to the COVID-19 pandemic. As such, it is a vector of features capturing daily policy changes in each state. As these policies is state-wide, we naturally extend them to the respective county level.
c.4 The COVID Tracking Project
From the COVID Tracking Project (c.f. COVIDTracking), we obtained the he daily numbers of PCR, Antibody and Antigen tests performed and their positivity rates at each state. These are used as features in our framework as well.