Behavioral Modelling

[mikkokotila]

Background

At the moment, the situation appears roughly as follows:

• behavior change interventions focus single dimensionally on epidemic effects
• a limited number of possible interventions have been considered
• at the moment interventions are considered as a single lump as opposed to stand-alone interventions
• while we know restrictions work, we don't know which restrictions work the best; there is not even a theoretical framework for robustly evaluating interventions
• intervention modeling assumes that countries are homogenous in terms of the effect

It seems fair to assume that such an approach leads to a suboptimal epidemic effect while causing severe economic and psychological damage. Simply put, the current approach, widely implemented throughout the globe, is an overly simplistic way to tackle the problem at hand.

A more meaningful situation would be:

• interventions focus on epidemic, economic, and psychological effects
• a large number of possible interventions have been considered
• interventions are considered individually, with further granularity per intervention
• there is a way to attribute the desired effect to a given intervention in comparison to other interventions (i.e. "pandemic mitigation attribution")
• a robust country-level weighting system for interventions is available

Examples

• Closing restaurants cause significant economic damage and some psychological damage. Instead of closing restaurants, a similar epidemic effect could be achieved by restricting capacity and imposing strict guidelines on restaurant owners, such as prohibiting buffets and any other cases where patrons are required to handle food with common utensils. This would entirely remove the psychological damage, limit economic damage, and would likely not be notably worse in terms of epidemic effect.
• At the moment grocery shops are open, and kindergartens are closed. The latter harms economy importantly; remote worker productivity suffers, and non-remote workers can't go to work. A small portion of households have kindergarten-age children, and kindergartens tend to be relatively small. In stark contrast, grocery shopping is heavily centralized (typically an entire town/locale goes to 2-3 shops) and every household frequently goes to the shop. If brick-and-mortar grocery shopping was replaced temporarily with delivery/pick-up/etc and special care was taken to make sure that nobody is left without a meaningful way to purchase groceries, there would be almost no economic effect, and the psychological effect would be notable. The epidemic effect would be far greater than what we get from closing kindergarten.

Proposal

Can we add two different models upstream from REINA:

• A country-specific weighting model overview
• A bespoke and highly granular behavioral model overview

For the latter, we have manually labeled each behavior for its perceived economic, psychological, and epidemic behavior. We then take probability distribution based on the individual labels, and using normally distributed random function, pick one to decide how much effect there will be.

Our vision is as follows:

• a person has a list of behaviors and can choose how much each will be restricted (0 to 100%)
• based on the inputs, the reproductive number is produced as an input to REINA
• at the end of the REINA simulation, economic damage, psychological damage, and a list of the effect that each behavior had on the outcomes

[mikkokotila]

Instead of the complexity in the "bespoke and highly granular behavioral model" illustrated above, we can start with following a simpler approach:

• Home
• Work
• School
• Transport
• Leisure
• Other

Each of these will have a restriction factor, where 0 means there is no epidemic, economic, or psychological effect. Each will also have a sensitivity factor, which will introduce randomness to the value (allows introducing doubt into the model).

Each of these will be described in human language in terms of the kinds of behaviors that each consist (which can later be used to allow more granular control), and each will be classified on scale 1 through 5 (or 7) where 1 is "strongly disagree" and max value is "strongly agree". Classification takes place based on statements like "restricting workplace contacts will cause significant economic damage".

This would plug into REINA through two interfaces:

• the interface where it is decided if contact will happen
• the interface where it is decided if a contact leads to infection

Slightly related #22.

We had discussed, that Autonomio covid-19 group can do all the work up until providing this as a Python API, which could then be cythonized by @juyrjola (if needed for performance reasons).

[mikkokotila]

Update

The proposed:

• Home
• Work
• School
• Transport
• Leisure
• Other

...is not going to work as it is. None of this is actionable. In fact, there are mere category headings for actionable items. For example, under education we have:

• kindergarten
• pre-school
• elementary school
• middle school
• high school
• secondary school
• university

Each has quite a different epidemic effect, psychological effect, and economic effect. Closing high schools will have close to zero economic effects, but kindergarten is one of the worst things overall that can be done in terms of economics. Also, many secondary school is not possible to do remotely (think car mechanics and plumbers for example) but most of university is possible remotely. And so forth.

I also want to challenge the position that "everything must be based on literature". REINA itself is out of literature, and very good that it is.

The short summary is that behavioral granularity is as useful as the output is actionable, and the inputs relevant to the line items policymakers are already considering. White House's Opening America Again Guidance is a good example of an overly simple, yet highly granular approach to behavioral intervention planning.

To shed a little bit light into how behavior change intervention planning actually works, it's useful to understand what goes into asking "is this behavior change reasonable". For the past decade or so, we've used Stanford Persuasion Lab's behavior change framework for similar cases.

The first consideration is about feasibility:

Whereas typically behavior change focuses on triggers (tell people what to do), sustainable interventions consider motivations (why would people do it), and most importantly ability (are people equipped to do it).

The second part is about identifying familiarity, intensity, duration, and so on:

Based on this kind of evaluation, behavior change interventions become actionable and measurable.

While we will not be able to, at least initially, to bring this level of depth into the Pandemik behavioral model itself, it's valuable to know that this is the kind of thinking that is driving the effort. The key point is actionability.

[juyrjola]

If I understood correctly, one major problem that you identify is that the category school is not granular enough to be able to used in your refined behavioral model. You may well use the age of the person in determining what type of "school" it is. Here is the plot of the age distribution of all the contacts that are categorized as school:

In Finland, people < 7 years old could be considered to have their school contacts in a kindergarten or a pre-school.

[mikkokotila]

Ah sorry, I did not realize you had already replied. I have been thinking about how to make the behavioral part entirely based on POLYMOD as we had initially discussed. Even though "school" does not break down to all the different schools, it's much better than current "restrictions" :)

I will spend today sometime working on the POLYMOD data to test some ideas.

[mikkokotila]

I think the data becomes immediately more meaningful if we don't use the provided 10 age groups, but we just break it down to maybe four (pre-school, school, adult, elderly).