Subject: Infections

Posted by Sarah Boyar on 10/15/2009
In Reply To:Infections Posted by Jim Duggan on 10/15/2009

Message:

Attached File: http://clexchange.org/ftp/K12Listserve/993_Sarah_SD-ABM_SIR.nlogo

Hello,
I think it's worthwhile taking a look at an Agent Based model (ABM) of infection. Netlogo is free and comes equipped with educational materials. I'm attaching a SIR (susceptible, infectious, recovered) model created in Netlogo which utilizes the SIR context to compare SD and ABM.
My notes on the exercise were as follows:
I think I learned more from interacting with the ABM vs. the SD model, although perhaps this is because I am more accustomed to the lessons taught by SD models.
While the SD model showed time path of behavior, it was nice to see in the ABM the characteristics of agents visually represented. When I played with the duration-of-disease and infectiousness, I developed a different sense of the dynamics and the specifics of how a time-path of behavior might unfold. Of course, a SD model could produce the same numbers, however, it was quite striking for me to see the ABM.
For example, it was possible in the ABM to not transition into a population that was 100% recovered, say, if the duration-of-disease was set to progress rapidly enough.
When I first noticed this, it was because of an ABM simulation wherein one agent simply did not become infected. So, 99.9% of agents had become infectious and then recovered, but one agent remained never-infected. In an SD model, I don’t think this would have been observed: the behavior-trend of the population would have been the focus.
Why is this important? Here’s one example of how this could have practical importance. If this happened in reality, the one person who never was infected might have been viewed as being immune or having some miracle gene which generated immunity. The ABM shows that spatiality of interactions combined with the timing and duration of the infection could produce this result randomly. Without this understanding of disease, a great deal of money, time, and effort might be put into studying that one person who was never infected to try and discover what about them granted them a miracle immunity.
So, ABM could temper the allure of examining outliers for special characteristics- that is, ABM can help in the comprehension of randomness in a way that SD cannot.

I have not fully formed an opinion about the best way to represent the SIR model. I think this is partly because, in this exercise, there was not an explicit reason that the model was built. (It was built purely for comparing SD with ABM.)
If the model was build to figure out the best way to invest resources into finding a cure, it’s possible that the ABM would be superior because of the miracle-immunity example.
On the other hand, a SD model might show that the policy of investing resources into finding a cure could be a fruitless policy, a ‘fix-that-fails’, because the virus happens to mutate rapidly. Such a SD-policy intervention model might show that a few easy dollars thrown at prevention could effectively wipe out the virus (say, if the virus was transmitted only through…playing tennis), whereas the system might show policy resilience to a policy of investing to find a cure.
I think that using both modeling technologies could probably present the best picture, as we’re essentially using these models as a way of groping in the dark in the hope of stumbling onto something either approximately true or else useful.

So, I think using SD and ABM in tandem (as in this exercise) would give a more robust picture than using one or the other in isolation.
Could a hybrid model be useful? I think that a hybrid model would probably be useful, although perhaps in a different way than using the two types of models in tandem would be. I cannot predict exactly how a hybrid model would be useful. This inability to predict the usefulness of a hybrid model is reasonable, I think, because the way that the SD model and ABM showed themselves as differently useful seemed to be idiosyncratic to the SIR model itself. (I did not do a cross-sectional study to see if the usefulness of the SD and ABM discovered herein is a generalizable usefulness of these types of models.)
I definitely think that NetLogo could be used to create hybrid ABM-SD models. I believe that it would be doable within the existing framework; however, it certainly could be done because of the extensibility that the authors of the software deliberately built into the program.
Doing this within the existing framework would likely require some tricky wiring. I would hypothesize this wiring as follows: use the reported variables from, say, the SD side of the model to update the patch characteristics in the ABM every tick. In this example, the patches could exhibit non-linear behavior generated by the SD model while the agents could continue to interact on a predominantly spatial basis.
The most difficult thing about this exercise was the synchronization of time across the two models for the comparison, and I don’t think I accomplished this very well. The various speeds of the models were difficult to control. That said, I think that this would be a challenge ifor creating a hybrid model- time calibration for the synchronization of time.
I’ve never calibrated a model for synchronous time before!! Only calibrated a mimic of history to actual history, or, calibrated a simulated possible future to a forecasted/desired future to see what the character of the present would need to be for the forecast to become true.
But calibrating for synchronous time between ABM and SD is exciting! It's like calibrating the universal time to the local time.
This is also very interesting to me:
I don’t think that ABM has a very good sense of time but an excellent sense of space. And SD has an excellent sense of time, delays, and the importance of time-paths, but is not sensitive to spatial considerations.
They seem to be potentially an excellent match, whether this takes the form of hybrid models or perhaps more elegantly, simple comparative models. It might in fact be more illuminating to just keep them comparative and learn something from the differences rather than make hybrid-model soup.