In bbolker/McMasterPandemic: Pandemic Model

model structure

The "McMasterPandemic" model is a compartmental model built for the purposes of modeling the COVID-19 pandemic. Its main features are:

• moderate epidemiological complexity, incorporating asymptomatic and presymptomatic infectious classes as well as subdividing symptomatic infections into moderate and severe
• health-care utilization structure, including acute-care and ICU compartments
• structure for calibrating to data
• flexibility in incorporating time-varying parameters

library(McMasterPandemic)

vis_model(method="igraph",do_symbols=FALSE)

basic operation (simulation)

time-varying parameters

breakpoints

log-linear models

The

splines

calibration

model expansion

Testing and tracing

To incorporate testing and tracing mechanistically, we have expanded the basic model in order to be able to

• continue to make use of case reporting data (\ie positive test confirmations per day) without having the results confounded by variation in testing intensity (by incorporating both positive and negative tests into our calibration);
• facilitate analysis of scenarios involving changes in testing intensity and contact tracing.

Our basic approach is to add an additional set of compartments to our model, expanding each compartment into the following sub-compartments:

• untested (_u extension) People with negative test results move back into this category since they can be tested again if there are new reasons for the test.
• tested, awaiting report, sample is negative (_n extension) either because the tested individual was uninfected, or because of a lack of test sensitivity. We can assign compartment-specific probabilities of a positive test (low for $S$ [currently set as 0], high for $I$ compartments [currently set as 1], and possibly lower for non-symptomatic/early stages than for later stages).
• tested, awaiting report, sample is positive (_p extensions) either because the tested individual was infected, or because of a lack of test specificity.
• tested positive (_t extension) Individuals who have a positive test will not be tested again, and may change their behaviour (isolate), depending on degree of compliance to public health advice.

In addition to these compartments, we also add accumulator compartments for negative and positive test reports. The appropriate accumulator is incremented at an appropriate time.

• If test results are indexed to the time of sampling then the accumulators are incremented during the _u to _n and _u to _p transitions
• If test results are indexed to the time of reporting then the accumulators are incremented during the _n to _u and _p to _t transitions

We can then recover the daily numbers of negative and positive tests recorded by differencing the cumulative totals (and add observation error if desired).

source(system.file("pix","testing_flow_graph.R",package="McMasterPandemic"))

Individuals can also move between epidemiological compartments while awaiting test results, e.g. from "presymptomatic, awaiting positive test results" (Ip_p) to "mild infection, awaiting positive test results" (Im_p). This does not change their testing status, except that when untested, severely symptomatic individuals are hospitalized for COVID, we assume they are tested immediately (i.e. `

In order to reflect the range of possible testing strategies, we assign a testing weight to each compartment in the model that specifies what fraction of the current testing intensity is allocated to that compartment. We take the current testing intensity $T$ (overall per capita tests/day) as a model input. Then, given testing weights $w_i$, the *per capita} rate at which individuals in epidemiological compartment $i$ move from _u to _n or _p (awaiting test results) is \begin{equation} \frac{T w_i}{\sum_j w_j P_j} \,, \end{equation} where $P_j$ is the proportion of the population in compartment $i$. The weights depend on the testing strategy and population state in complicated ways; if only confirmatory testing is being done (no screening or contact tracing or surveillance testing), then the weights will be skewed toward symptomatic compartments -- although not entirely, as observed test positivity rarely goes above 20%, and sensitivity is thought to be much higher than this. More contact tracing will increase the weights of non-symptomatically infected people. More random-surveillance testing will make the weights of all the groups more similar.

Including the testing structure increases the number of compartments substantially, and consequently yields a much larger flow matrix

vis_model(testify=TRUE,aspect="fill")

Explicit testing structure is enabled if the parameter vector/list contains an element testing_intensity which is set >0. (As a side note, if you are using read_params("PHAC_testify.csv") to capture our most recent set of default parameters, and you don't want an explicit-testing model, you should use update(., testing_intensity=0).) The argument testing_time to the make_ratemat() function determines when testing is counted ("sample" or "report"); this can be passed to run_sim() (or from farther upstream) in the ratemat_args (list) argument; the default is to set counting time to "sample", with a warning.

age structure

• we can construct a larger rate matrix
• how does beta work? Suppose we have $n_a$ age categories and $n_I$ infectious categories (e.g. $I_a$, $I_p$, $I_m$ etc.) We want a beta matrix with dimensions $n_a \times (n_a n_I)$ (because there is only one $S$ compartment per age category, at least until we introduce testing)

goals

The ability to simulate/forecast

• time and height of peak health utilization (acute care, ICU beds, ventilators) regionally and locally
• testing
• ability to model local differences in epidemic curves based on locale-specific information (e.g. age-demographics, population density, etc.)
• estimation of $R(t)$!

why another simulator?

Given the large number of available simulators, why did we write another one?

• maybe laziness/stubbornness
• desire for a flexible and clear model structure that might nevertheless go beyond existing models (age and geographic structure were higher priority goals when we started)

Platform

Familiarity/convenience/speed and ease of implementation, debugging, maintenance are most important. For us, this rules out Julia and Python and strongly suggests R. Speed may eventually be important, for (1) running stochastic replicates and varying-parameter ensembles and (2) incorporating the model in simulation-based estimation methods (ABC, iterated filtering/pomp). If we are focused on estimation, we have a choice among Gibbs-based platforms (NIMBLE, JAGS) and platforms that only allow continuous latent variables (TMB, Stan).

• Estimation: TMB, Stan, NIMBLE, pomp + ?, ABC + ?
• Forward simulation: we could write in base R or try to couple with something faster (e.g. odin, Rcpp). Lots of possibilities for vectorization and matrix operations to speed up base R implementations, although some operations like convolution might be hard to vectorize. pomp provides a Csnippets interface (used by the "COVID interventions" project). I am tempted by odin, a platform that allows translation/compilation in C++ via an R-like syntax, e.g. see the discrete-simulation vignette

Model

Compartmental structure

• we need at least SEIR + hospitalization, ICU
• probably important to include asymptomatic and presymptomatic compartments as well
• may want to include compartments that model the testing as well as the treatment pipeline
• deaths

Interventions

Some of these are simple. Want to set up general structure for importing a dated list of changes in parameters (date, focal parameter, new value or proportional change). Testing rates can be included as part of this "non-autonomous"/forcing/external part of the parameterization.

Parameterization

• get parameters from existing model implementations
• list from MIDAS web page
• age-specific death rates (Hubei) -- from Riou et al. https://www.medrxiv.org/content/10.1101/2020.03.04.20031104v1.full.pdf
• Hospitalization parameters: https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(20)30566-3/fulltext

Asymptomatic/undetected cases

• could try to integrate new testing-distribution framework?
• specify underreporting probability (ad hoc), beta-binomial process
• time-varying reporting probability?
• related to compartment structure, testing/intervention

Stochasticity

• demographic stochasticity ([beta]binomial, Poisson/NB)
• autoregressive variation in parameters? (important for estimation)
• iteration over parameter uncertainty (hypercube/Sobol/etc.)

Geographic structure

• including spatial structure at the local level doesn't seem too hard. Has implications for size of state vector/speed (mitigated by a vectorized/matrix-based set of machinery?), and raises questions about parameterization (spatial variation in parameters, travel/contact matrix)

Age structure

• important for predictions of health care utilization/severity
• also for effects of interventions (school, work closures, etc.): use info from Prem et al. 2013 WAIFW matrices divided by activity {school, home, work, other} (DOI: 10.1371/journal.pcbi.1005697; machine-readable info at https://doi.org/10.1371/journal.pcbi.1005697.s002)
• interactions with space? can use "other" category WAIFW matrix only (Kronecker product)

Delay distributions

(not sure how important this is?)

• most existing methods are using exponential periods
• Erlang/linear chain?
• generation intervals/convolution

Disadvantages of linear-chain and convolution models are speed and potentially complexity/transparency (modeling convolution with GI handles arbitrary distribution of infectiousness over time, but making it interact with testing and treatment pipelines seems complicated?)

Relevant existing models (see model list)

• Covid19tauleapmodel [compartments, interventions, age, parameters]
• COVID_interventions [compartments, interventions, parameters] Uses "Csnippet" (pomp-style C code) for speed.
• CSH-covid19 [interventions]
• Riou et al. [compartments, parameters, age, estimation] (but slow: uses ODEs in Stan, ~ 1 day run time)
• Song lab [estimation?] (but some weird statistical choices?)
• Georgia CEID

bbolker/McMasterPandemic documentation built on Aug. 25, 2024, 6:35 p.m.