Read on Twitter
Upon multiple request, here’s my thread regarding the boundaries of epidemiological modeling in English. https://twitter.com/theotherphilipp/status/1355164530960199684?s=20
1/22: #Models are simplified images of our world. They are useful until reality disproves them.
@ViolaPriesemann, a German physicist and government advisor, believes she can derive measures from her #Corona model with claim to truth that lacks any humility.
A thread
.
@ViolaPriesemann, a German physicist and government advisor, believes she can derive measures from her #Corona model with claim to truth that lacks any humility.
A thread
.
2/22: The quality of a model is determined by two criteria:
1. How well does the model represent reality?
2. How good are the assumptions (parameters) with which the model is executed?
It is one of the core tasks of epidemiology to develop such models.
1. How well does the model represent reality?
2. How good are the assumptions (parameters) with which the model is executed?
It is one of the core tasks of epidemiology to develop such models.
3/22: The Priesemann Group continues to shock the public and government with predictions of catastrophic further progression of the #Corona epidemic, should tough measures, such as extremely strict lockdowns, fail to materialize, and even calls for a #ZeroCovid strategy.
4/22: But how does the group come to such shocking conclusions?
It uses a specially developed computer model for this purpose. Anyone can download this model, examine it and also run it. The current codebase can be obtained via GitHub: https://github.com/Priesemann-Group/covid19_inference_forecast
It uses a specially developed computer model for this purpose. Anyone can download this model, examine it and also run it. The current codebase can be obtained via GitHub: https://github.com/Priesemann-Group/covid19_inference_forecast
5/22: There are many approaches when developing epidemiological models. Among the simplest are SIR models, first described by Kermack and McKendrick in 1927. Modern variants of this approach allow a simplified representation of epidemics with the help of stochastic methods.
6/22: A big disadvantage of SIR models (among many more) is that they do not take into account the fact that infected persons typically only become infectious after an incubation period. This circumstance is taken into account by its predecessors, so-called SEIR models.
7/22: No matter whether SIR or SEIR. What all these models have in common is that, while they can provide a good general understanding of the spread of a disease, they are rather unsuitable for concrete predictions for a country or region. Why is that?
8/22: S(E)IR models are dynamic (time-dependent) systems, but essentially assume a homogeneous spread of a disease through a population. The spread is determined by basic assumptions about basic reproduction number R0, generation time and a few more.
9/22: Unlike heterogeneous, nonlinear systems, these models know nothing about the specific mechanisms of spread of a disease in a population. They do not take into account age structure, the structure of contact networks, seasonality and other important factors.
10/22: A modern epidemiological model should take into account a large number of factors in order to be able to make concrete predictions about the course of an epidemic, and even more so to be able to make statements about the effectiveness of individual countermeasures.
11/22: A model that claims to assess the impact of school closures, for example, would need to map the age structure of the population, account for differences in contact networks between students and, say, retirees, map their different contagiousness, etc.
12/22: All the more astonishing is the decision of the Priesemann Group to build their simulation on the basis of a stochastic SIR model. Yes, right! The one from 1927, not even SEIR! Yes, it’s not a plain vanilla SIR-modell, there is a little more. But at its core, that’s it.
13/22: Now, on the basis of such a simple model, is it possible to make statements about the effect of
- school and daycare center closures,
- company closures,
- contact bans or
- the wearing of masks?
NO. You can't!
- school and daycare center closures,
- company closures,
- contact bans or
- the wearing of masks?
NO. You can't!
14/22: Then how does the group know how well the measures are working, how does it know what the impact of "relaxations" or "tightenings" will be?
Well, they don’t. They makes assumptions. To calculate the three different trajectories, they run three scenarios.
Well, they don’t. They makes assumptions. To calculate the three different trajectories, they run three scenarios.
15/22: Each scenario is run with its own set of parameters in such a way that different plots result. The time series parameters are to represent a different effect of the measures on the reproduction number R. The group calls this changepoints. In the code it looks like this:
16/22: But where do these values come from? The group cites a paper it penned in summer 2020, which it claims demonstrates the effectiveness of the NPI during spring in Germany. Also a modeling study, created with a similar code base.
17/22: The study is controversial. An important argument against the conclusions of the study is that those very "effects" of NPIs were absent during the 2nd lockdowns. The study ignores seasonality and is solely based on cases, another problem. https://arxiv.org/abs/2004.01105
18/22: Thus, the parameters used to run the three different scenarios, that lead to the published "shock graph“, are based on assumptions of questionable robustness.
19/22: Let’s summarize:
1. The model does not represent the heterogeneity of the studied population.
2. The model does not know anything about age differences, basic immunity, seasonality etc.
3. The model does not know anything about contact networks.
1. The model does not represent the heterogeneity of the studied population.
2. The model does not know anything about age differences, basic immunity, seasonality etc.
3. The model does not know anything about contact networks.
20/22:
4. The parameters used to run the model are pure estimates.
It is therefore not surprising that the developers of the model also warn about the reliability of their results.
4. The parameters used to run the model are pure estimates.
It is therefore not surprising that the developers of the model also warn about the reliability of their results.
21/22: But why Ms. Priesemann and her group believe they are allowed to propagate measures on the basis of this model, that have the most severe consequences for our society, plunge the national economy into a deep crisis and destroy entire livelihoods, remains a mystery.
22/22:
How wrote the once brilliant N.N. Taleb so beautifully?
»Academia is to knowledge what prostitution is to love.«
How wrote the once brilliant N.N. Taleb so beautifully?
»Academia is to knowledge what prostitution is to love.«
