Summary of Initial Modeling Efforts – COVID-19 in Iowa
Whitepaper Prepared for the Iowa Department of Public Health
by the University of Iowa College of Public Health COVID-19 Response Group

Submitted by
Grant Brown, Ph.D., Assistant Professor of Biostatistics
Joseph Cavanaugh, Ph.D., Professor and Head of Biostatistics
Aaron Miller, Ph.D., Assistant Professor of Epidemiology
Jacob Oleson, Ph.D., Professor of Biostatistics
and Director of the Center for Public Health Statistics
Michael Pentella, Ph.D., Clinical Professor of Epidemiology
and Director of the State Hygienic Lab
Eli Perencevich, M.D., M.S., Professor of Internal Medicine and Epidemiology,
and Director of the Center for Access and Delivery Research and Evaluation
Daniel Sewell, Ph.D., Assistant Professor of Biostatistics

The COVID-19 modeling team at the University of Iowa has taken a multi-disciplinary, multimodel approach to estimating the progression of the COVID-19 outbreak in the State of Iowa.
Rather than be limited by a single modeling approach, utilizing different models can allow us to
ask different nuanced questions, which will give us a more comprehensive understanding of the
outbreak and allow for more effective forecasting of the impact of interventions on the outbreak.
We are concurrently developing and evaluating three modeling paradigms which can each be
used to accomplish separate but complementary goals. The three different approaches and
their goals are as follows.
(M1) A short-term prediction model with few modeling assumptions designed to make
accurate one-week ahead forecasts. This model additionally serves to evaluate
the plausibility of longer-term predictions based on the current short-term
trajectory of the COVID-19 outbreak.
(M2) A medium-term prediction model providing the potential range of the outbreak,
yielding estimates of disease characteristics and important prediction bounds for
the ongoing outbreak in Iowa. The disease characteristics included in the model
are latent and infectious periods of COVID-19 as well as mortality rate.
(M3) A model designed to evaluate various policy changes related to nonpharmaceutical
interventions such as increasing or relaxing social distancing measures or
implementing universal face shields. This model will be able to provide
projections of how we expect the disease to spread if social distancing measures
are implemented at various points in time.

 Taken together, these approaches should allow the team to understand the current and
immediate state of the COVID-19 outbreak in Iowa, quantify the size and scope of the outbreak
in the coming months, and characterize how various actions may affect the trajectory of the
outbreak. All results presented here are based on publicly available state-level data using
models developed by the team. The team has not yet had sufficient time to develop extensions
using the supplied IDPH data.
Conclusions: We have found evidence of a slowdown in infection and mortality rates due to
social distancing policies, but not that a peak has been reached. The current data does not
support a conclusion that the reproductive number has been brought below 1. There is
considerable uncertainty still in how many cases and deaths Iowa could eventually have, with
possible projections between 150 and >10,000 total deaths. Therefore, prevention measures
should remain in place. Without such measures being continued, a second wave of infections is
likely.

Model 1: Short term forecasts and trajectory assessments
Overview. This model projects the longer-term outbreak epidemiology (e.g., out to 6 months) by
estimating the most-likely outbreak trajectory in the short-term (e.g., 3-14 days) and comparing
trajectories over the near-term period. This approach uses only location-specific data (e.g., from
Iowa) to estimate the short-term trajectory. Thus, it aims to avoid possible misspecification or
conforming to the trajectories observed in other states or locations that have characteristics or
social distancing policies different from Iowa. Note that the validity of short-term forecasts is
continuously evaluated by comparing forecasting performance against a similar model applied
to different locations or time periods; this allows the model to avoid “over-fitting” the limited
amount of information that is currently available in Iowa. Model fitting is also performed
dynamically so that changes in an outbreak may be detected and more quickly integrated into
the projections.
The methodological approach can be described in three parts.
•

•

•

First, a relatively simple short-term model is built to forecast the outbreak trajectory over
the next 3-14 days, depending on the optimal state-specific lag between input and
outcome measures. In the example below, the short-term model leverages the
relationship between the lagged number of positive cases and the cumulative mortality
rate. (Note: this approach can be extended to other outcomes.)
Second, “feasibility bounds” around these short-term projections are constructed by
evaluating “out-of-sample” performance, which is derived by comparing many different
prior short-term forecasts to the observed values in different locations and time periods.
Third, long-term trajectories are estimated by evaluating those that best conform to outof-sample performance. For example, a long-term trajectory may be deemed “too
extreme” if it represents a deviation from the short-term forecast that exceeds 95% of all
out-of-sample estimates. This third step can be done using external long-term models
(e.g., IHME model) or by fitting various models through the plausible points contained in
the short-term feasibility bounds described in step 2.

Preliminary Findings. We have evaluated this model in two ways. First, we compared this
approach to updated projections of the IHME model since it was unveiled on March 25th. This
short-term fitting approach does a good job of anticipating revisions in the IHME projections, by

 identifying state forecasts that appear to be over or underestimating the short-term trajectory in
the coming days. Second, we applied this model to the most recent counts in the state of Iowa
and constructed various long-term trajectories through different feasibility bounds, using the
basic “curve-fitting” approach in the IHME model. Figure 1 depicts short-term forecast feasibility
bounds for cumulative mortality in the coming week based on out-of-sample performance using
the previous two weeks of forecast estimates in the states of IA, MN, WI, NE, SD, MO and IL.
The feasibility bounds correspond to 50%, 80% and 90% of the out-of-sample prediction errors
(e.g., 50% of the out-of-sample forecasts errors would suggest a true value within the 50%
percent feasibility bound.)

Figure 2, below, depicts five possible trajectories using the short-term forecast bounds depicted
above for the next 7 days. These trajectories represent plausible trajectories within the 95%
prediction interval (based on out-of-sample performance) that are also consistent with the most
recently observed data. This figure also depicts the corresponding peak dates (for peak daily
mortality) and total mortality by July 1st that corresponds to these points.

These figures demonstrate that a wide range of plausible trajectories still exist for the state of
Iowa, ranging from 151 deaths (peaking on April 18th) to 965 deaths (peaking on May 8th).
However, given the continued increased in positive cases, hospitalizations, ICU admissions and
patients on ventilators, it is our opinion that Scenarios 1 and 2 are far less likely. Note that

 Scenarios 1 and 2 represent peak mortality occurring this past Saturday or this coming
Wednesday; these represent peaks occurring earlier than most existing forecast estimates for
the state of Iowa. However, it is important to note that the long-term trajectories plotted in
Figure 2 assume the parametric shape and underlying social distancing assumptions of
the IHME model, and this short-term model fitting approach should only be used to
assess the short-term plausibility of a given trajectory and not to generate long-term
forecasts.

Model 2: Long term forecasts and disease characteristics
Overview. M2 takes a population-level approach to the spread of COVID-19 in Iowa, focusing
on the reported mortality as a reliable, though lagged indicator of epidemic spread. This model
is designed to synthesize available information about the latent and infectious periods of the
COVID-19 infection, as well as existing public health interventions in the United States and
abroad. With these tools, we evaluate the preliminary projections available by considering Iowaonly mortality, and also describe a range of plausible future projections assuming that the
patterns observed elsewhere are applicable to the situation in Iowa.
Uncertainty remains about the mortality rate of COVID-19, and the best information indicates
that it is heterogeneous by age, among other factors. Nevertheless, reported COVID-19
mortality may provide a useful perspective on the epidemic more broadly. This model posits that
the true mortality rate is around 2%, but allows the estimate to be updated based on available
data. This prior mean is consistent with Lai et. al. (2020), but may be conservative (Baud et. Al.,
2020). M2 is a stochastic SEIR model, tracking population counts in the Susceptible, Exposed,
Infectious, and Removed disease states. The latent period, during which individuals in the
Exposed (E) category are infected but not yet actively transmitting the pathogen, was assumed
to last between 2 and 14 days (95% probability), with typical values around 5 days (Lauer et. al.,
2020). Much less is known about the infectious (I) period, which was assumed to last between
10 and 32 days (95% probability), with typical values around 14 days, based reports about
recovery times.
Preliminary Findings. Regions affected by COVID-19 vary widely with respect to both the
baseline transmission intensity, as well as the degree/type of community and policy response to
the epidemic. We therefore consider two approaches to understanding the current situation and
expected trajectory in the state of Iowa. First, we consider models of the trajectory of mortality in
Iowa under varying assumptions and compare these trajectory results to comparable fits in
other states and countries further along in the outbreak. This comparison gives rise to the
results presented in Figure 3, in which we see substantial variability in the Basic Reproductive
Number (R0) by location. We estimate R0 for each location using the Empirically Adjusted R0
(EA-RN) calculated prior to the start of public health interventions. The EA-RN captures the
expected number of secondary infections caused by each infectious individual in a specific
context, sharing many of the same features as the commonly cited “Effective Reproductive
Number”.

 Figure 3. Estimated Empirically Adjusted Reproductive Numbers in selected regions, comparing the unmitigated
epidemic intensity. Iowa is included here twice, once with interventions starting on 3/17/2020, and again with
interventions assumed to begin on 4/4/2020. As expected, choices of intervention start time did not affect the
estimated baseline (pre-intervention) epidemic intensity.

Building models purely on reported Iowa mortality data, we find that predictions are highly
sensitive to assumptions about when contact behavior (social distancing) became widespread in
the state, especially for longer term forecasts. In Figure 4, we illustrate the mortality curve of a
model assuming that contact behavior began to shift on March 17th, when a Public Health
Disaster Emergency was declared in the state. Here, we present the quantiles of fitted/projected
cumulative mortality curves to show the range of likely/possible futures under this assumption.
By May 28th, this model predicts a median mortality of 747 with a 95% interval between 156 and
9,734, reflecting a huge degree of uncertainty in trajectory and magnitude. This uncertainty is
further shown by considering the model in Figure 5, where we modify the assumed date of
behavior change to April 4th, at which point formal measures to close educational institutions
were taken, and individuals and businesses had benefitted from more time to respond to
changing circumstances. Despite being fit to the same data, this shift in the date on which
contact behavior was allowed to change produced dramatic modifications to the six-week
mortality prediction. The model estimated a median cumulative mortality of 1,369 on May 28th,
with a 95% interval between 315 and 28,599. These projections will become more precise as
more fatalities are observed, and hypothetical epidemics with incompatible shapes are
discarded. In Figure 6, we present the Empirically Adjusted Reproductive Number curve for the
March 17th intervention model. This measure is designed to capture the number of secondary
infections per infectious individual observed in a particular population, and is related to the
common “effective reproductive number”, more flexibly capturing the change over time. As
illustrated in the figure, we do not see strong evidence that the reproductive number has been
brought below one. These initial results communicate several important points:
1. We see preliminary evidence for the impact of current imposed and adopted social
distancing and contact mitigation efforts in Iowa, with cumulative mortality increasing
more slowly than would be expected without such mitigations, and the reproductive
number dropping over time.

 2. We observe a huge range of possible outcomes, from relatively low fatalities to
catastrophic loss of life. These models assume that the current measures taken to limit
contact in Iowa remain in place, so while we see evidence that the epidemic intensity
has decreased from the initial period of uncontrolled spread, there is not sufficient
evidence in this data set alone to conclude that current measures will necessarily be
sufficient to prevent a return to higher rates of transmission and corresponding mortality.

Figure 4: Projected cumulative 2-week mortality in Iowa using only Iowa specific data, and assuming that a shift in contact
behavior occurred on March 17th.

Figure 5: Projected cumulative 2-week mortality in Iowa using only Iowa specific data, and assuming that a shift in contact
occurred on April 4th.

 Figure 6: Estimated Empirically Adjusted Reproductive Number curve, based on the March 17th Iowa only intervention model.
Only when this value drops below one do we expect the epidemic growth to be sub-exponential. While it’s clear that the
reproductive number has dropped over time, we do not have evidence that it has fallen below one, or that it is expected to
unambiguously do so in the immediate future.

In order to improve the precision of our forecasts, we can introduce assumptions about how the
intervention in Iowa relates to that implemented in other locations. This approach allows us to
borrow information about how the spread of COVID-19 has been fought in other places, without
requiring that we assume the entire disease course in Iowa will be similar. Nevertheless, every
state and nation can reasonably be expected to show different transmission dynamics, and the
formal and informal responses to COVID-19 have certainly been highly varied in type, timing,
implementation, and adoption. Thus, this approach is better suited for describing a range (from
optimistic to pessimistic) of possible futures than for adopting a specific best prediction, unless
strong evidence is presented to justify the substantial similarity between the responses of the
locations under study. Table 1 presents this range of scenarios, varying by which state is used
to inform prior knowledge of intervention efficacy, and which date is deemed most comparable
to the intervention in that state. Figures range from extremely optimistic to moderate, and are
compared to the results from the Iowa-only models given above.

Table 1: Expected Mortality Scenarios by 5/28 (NOT Total Mortality)
Intervention Magnitude
Extremely Optimistic
Extremely Optimistic
Optimistic
Optimistic
Moderate
Moderate
Iowa Only
Iowa Only

Intervention
Estimated From:
New York
New York
Washington
Washington
Minnesota
Minnesota
Iowa
Iowa

Iowa Intervention
Beginning On:
3/17/2020
4/4/2020
3/17/2020
4/4/2020
3/17/2020
4/4/2020
3/17/2020
4/4/2020

Median Mortality by 5-28
(95% Cr-I)
107 (67-296)
334 (169-632)
341 (133-1955)
492 (240-1997)
560 (115-6219)
871 (200-10008)
747 (156-9734)
1369 (315-28599)

 Overall, the results of M2 support evidence of “curve flattening”, but do not definitively indicate
that this will be quickly followed by the containment of virus spread. While improved and
ongoing mitigation is indicated as a plausible outcome, M2 also identifies risks of
returning to a mode of increasingly rapid spread. These results indicate that great caution is
needed at this early stage before loosening of potentially insufficient containment measures is
considered.
Next Steps. M2 is a population-level model, and as such it misses important heterogeneity in
state populations. Extensions of this work should apply more granular data to better understand
the distribution of COVID-19 cases in Iowa. In addition, in the next few weeks as more
information about the mortality curve in Iowa becomes available, estimates for the peak time
and final size of the epidemic based on M2 are expected to become more accurate, and the
length at which forecasts will be meaningfully precise will increase. M2 is also able to
accommodate a hypothetical relaxing of interventions, if expressed in relative terms (e.g.,
statements like “reducing intervention intensity to 85% of current levels”), or to similarly infer the
likely effect of such policies if adopted first elsewhere.

Model 3: Evaluating polices
Overview. M3 is designed to describe disease transmission dynamics over a large contact
network corresponding to the 3.155M residents of Iowa. By explicitly modeling how individuals
contact one another, contact based interventions or the relaxing thereof can be tested virtually
to understand how we might expect such interventions to affect the disease trajectory. While
this model may be used to construct long-term predictions, its primary function is to understand
how the disease curve can be manipulated by public policy decisions.
This model leverages information collected in prior studies (Kwok et al., 2018; Zhaoyang et al.,
2018) to accurately reflect the human-to-human contact patterns. It computes the expected
number of infectious individuals at any given date due to disease transmission through these
contacts. The effects of increasing or decreasing social distancing can be implemented directly
in a time-varying and individual-specific manner. Additionally, time-varying infective-susceptible
transmission probabilities can be modeled explicitly, thereby allowing the estimation of the effect
of, e.g., universal face shields on the disease curve.
Preliminary findings. An early utilization of M3 can shed light on several critical areas. First,
we show that barring additional interventions such as contact tracing, we expect that reopening
society to its pre-COVID-19 status anytime in the next four months will lead to a second wave of
infections that will necessitate reimplementing our current social distancing policies. These
estimates were based on an adaptive strategy of reopening society if the number of daily cases
(both reported and unreported) fell below a threshold of 150, 100, 50, or 10, and reinstating our
current level of social distancing if we observed 7 contiguous days of increasing number of daily
cases. Figure 7 shows these results by plotting the expected number of daily cases according to
the various thresholds. In all cases, a second peak emerges necessitating the reinstatement of
social distancing policies.

 Figure 7. Expected number of daily COVID-19 cases over time with an adaptive strategy for reopening society.
Social distancing will be relaxed if the number of daily cases drops below 150 (pink), 100 (purple), 50 (red), or 10
(orange) cases, and reinstated if 7 consecutive days of increasing daily case counts occurs. The black curve
provides a baseline which maintains the current level of social distancing indefinitely.

We have also used M3 to evaluate the efficacy of universal PPE, such as face shields. To
illustrate this approach, we evaluated universal PPE in terms of reduction in the probability that
an infective-susceptible contact would lead to a transmission event. Figure 8 shows these
results using the adaptive scheme described above with a threshold of 100 daily cases to
reopen, where the universal PPE has been implemented starting on May 1, 2020. From this
figure we see that even with low efficacy, the current peak drops considerably more rapidly than
our current trajectory. However, reopening society is still highly likely to lead to a second peak
unless the PPE efficacy is strong (at least 80% reduction in transmission probability).

 Figure 8. Expected number of daily cases due to varying levels of efficacy for universal PPE implemented on
May 1st as an illustration. Efficacy is measured in terms of reduction in the probability of a susceptible-infective
contact leading to a new transmission. The adaptive strategy for reopening and reimplementing shows that only
with high efficacy rates does a second peak (and subsequent reinstatement of social distancing) fail to emerge.

Finally, we used M3 to evaluate the effect of opening society for all but those individuals with the
highest contact rates. Figure 9 shows how reopening society based on contact rates works
conjointly with PPE effectiveness. This figure shows that if we can maintain the current
quarantining levels for those with the 20% highest contact rates through, e.g., limiting large
social events, we can greatly diminish the impact of the second wave, and a lower level of PPE
efficacy is required to eliminate the second wave entirely (50% reduction in transmission
probability).

 Figure 9. Expected number of daily cases due to varying levels of efficacy for universal PPE implemented on
May 1st as an illustration. Efficacy is measured in terms of reduction in the probability of a susceptible-infective
contact leading to a new transmission. The adaptive strategy used here maintains quarantining at the current level
for those with the highest number of contacts while reopening society for the rest. Pairing this with medium
efficacy universal PPE, a second peak (and subsequent reinstatement of social distancing) may not emerge.

Future directions. The current model is limited in that the model is being trained on the true
number of cases which can only be estimated. The preliminary results presented here are
based on publicly available data alongside current estimates of unreported cases in the United
States (Jagodnik et al., 2020). The data received by IDPH this past week will allow us to better
estimate the true number of cases, thereby increasing the accuracy of our assessments and
forecasts.
In addition, we are developing a web-based tool for IDPH and other interested parties to use in
order to evaluate specific changes in social distancing practices and universal PPE use. We will
notify IDPH as this tool becomes available for use.

 References
Baud, David, et al. "Real estimates of mortality following COVID-19 infection." The Lancet
Infectious Diseases (2020).
Jagodnik et al., “Correcting under-reported COVID-19 case numbers: estimating the true scale
of the pandemic.” medRxiv (2020). https://doi.org/10.1101/2020.03.14.20036178
Lai, Chih-Cheng, et al. "Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) and
corona virus disease-2019 (COVID-19): the epidemic and the challenges." International Journal
of Antimicrobial Agents (2020): 105924.
Lauer, Stephen A., et al. "The incubation period of coronavirus disease 2019 (COVID-19) from
publicly reported confirmed cases: Estimation and application." Annals of Internal Medicine
(2020).