When nothing is something

When complex adapative systems are involved, very often nothing is something and something is nothing. This makes it hard to get the correct model engaged to use to interpret the data.

A nice, clean example of the first principle, where "nothing" is "something" is a very common regulatory feedback loop often known as a "clamp." Probably the most common examples of these are thermostats and automobile speed "cruise control." The system is engineered on purpose to observe the world, to see a difference between some observable property of the world and some goal, and to take the appropriate action to bring the world's property back into line with the goal that is currently set. No big deal.

However, when the regulatory feedback clamp is the hidden model, it may be much harder to detect and realize what you are looking at. A good example might be, say, a tobacco control ad campaign. Say you measure the historical prevalence of smoking in some area, then intervene with a big educational ad campaign, then measure the final prevalence of smoking, and there is no change. The question is, "Did the ad campaign work?"

One possible underlying model is that no one watched the ads, and so it flopped for that reason. Another possible underlying model is that many people watched the ads and started to change their behavior, which was instantly picked up by the tobacco companies, which responded by seriously ramping up their ads until smoking returned to the level the industry wanted for that area. And, for every dollar spent in the tobacco control ad, the industry had to spend $1,000 to counter the effect. So, now, again the question is, "Did the ad campaign work?"

This is a case where "nothing is something." Because of the intervening active external systems effect, the impact of the intervention netted out to zero, as measured by one particular outcome variable. So, if you assume an underlying linear regression model where
prevalence = some constant plus some factor times your ad campaign dollars, you will measure zero effect for the coefficient factor. If you're sloppy, you would say that the ad had no effect on prevalence. If you're careful, you'd say that you cannot demonstrate any statistically valid impact on prevalence. These conclusions are correct, but misleading, as your ad campaign was very effective, but wasn't the only thing going on.

( Image is from "deadgirl1121" at flickr )

technorati tags:smoking, tobacco_control, causality, reasoning, GLM, general_linear_model, systems_thinking, feedback, feedback_loops, regulatory_feedback_loops, regulation, reasoning, modeling