This post should be required reading for all modelers. And no, I’m not going to reproach sloppy modeling practices. This is much more interesting than that.
Sloppy models are an idea that formalizes a statement Jay Forrester made long ago, in Industrial Dynamics (13.5):
The third and least important aspect of a model to be considered in judging its validity concerns the values for its parameters (constant coefficients). The system dynamics will be found to be relatively insensitive to many of them. They may be chosen anywhere within a plausible range. The few sensitive parameters will be identified by model tests, and it is not so important to know their past values as it is to control their future values in a system redesign.
This remains true when you’re interested in estimation of parameters from data. At Ventana, we rely on the fact that structure and parameters for which you have no measurements will typically reveal themselves in the dynamics, if they’re dynamically important. (There are always pathological cases, where a nonlinearity makes something irrelevant in the past important in the future, but that’s why we don’t base models solely on formal data.)
Now, the required part. Continue reading “Sloppy System Dynamics”
Vi Hart’s interesting comments on the dynamics of political polarization, following the release of an innocuous video:
I wonder what made those commenters think we have opposite views; surely it couldn’t just be that I suggest people consider the consequences of their words and actions. My working theory is that other markers have placed me on the opposite side of a cultural divide that they feel exists, and they are in the habit of demonizing the people they’ve put on this side of their imaginary divide with whatever moral outrage sounds irreproachable to them. It’s a rather common tool in the rhetorical toolset, because it’s easy to make the perceived good outweigh the perceived harm if you add fear to the equation.
Many groups have grown their numbers through this feedback loop: have a charismatic leader convince people there’s a big risk that group x will do y, therefore it seems worth the cost of being divisive with those who think that risk is not worth acting on, and that divisiveness cuts out those who think that risk is lower, which then increases the perceived risk, which lowers the cost of being increasingly divisive, and so on.
The above feedback loop works great when the divide cuts off a trust of the institutions of science, or glorifies a distrust of data. It breaks the feedback loop if you act on science’s best knowledge of the risk, which trends towards staying constant, rather than perceived risk, which can easily grow exponentially, especially when someone is stoking your fear and distrust.
If a group believes that there’s too much risk in trusting outsiders about where the real risk and harm are, then, well, of course I’ll get distrustful people afraid that my mathematical views on risk/benefit are in danger of creating a fascist state. The risk/benefit calculation demands it be so.
I’ve been working with pharma brand tracking data, used to calibrate a part of an integrated model of prescriptions in a disease class. Understanding docs’ perceptions of drugs is pretty important, because it’s the major driver of rx. Drug companies spend a lot of money collecting this data; vendors work hard to collect it by conducting quarterly interviews with doctors in a variety of specialties.
Unfortunately, most of the data is poorly targeted for dynamic modeling. It seems to be collected to track and guide ad messaging, but that leads to turbulence that prevents drawing any long term conclusions from the data. That’s likely to lead to reactive decision making. Here’s how to minimize strategic information content:
- Ask a zillion questions. Be sure that interviewees have thorough decision fatigue by the time you get to anything important.
- Ask numerical questions that require recall of facts no one can remember (how many patients did you treat with X in the last 3 months?).
- Change the questions as often as possible, to ensure that you never revisit the same topic twice. (Consistency is so 2015.)
- Don’t document those changes.
- Avoid cardinal scales. Use vague nominal categories wherever possible. Don’t waste time documenting those categories.
- Keep the sample small, but report results in lots of segments.
- Confidence bounds? Bah! Never show weakness.
- Archive the data in PowerPoint.
On the other hand, please don’t! A few consistent, well-quantified questions are pure gold if you want to untangle causality that plays out over more than a quarter.