Post by Caleb Aldridge.— Recently I have been curious to know if researchers consider decision analysis when designing experiments. Obviously, not all research is geared towards problem solving or may not have an immediate application; but what about experiments that are and that do?
Let us think about a hypothetical, but probable, ecological problem: invasion of a non-native plant species. The problem we are facing is that an invasive grass (e.g., cheatgrass, cogongrass) is out competing native herbs on the wildlife area we manage. We have observed malnutrition and decreased suitable habitat for several vertebrates, including game species. The invasive grass is absent at a nearby wildlife area, and there are no signs of malnutrition or decreased suitable habitat from previous surveys. We infer that the invasive grass is having a negative impact at our wildlife area—literature supports our suspicion.
Image source: USDA National Invasive Species Information Center (https://www.invasivespeciesinfo.gov/profile/cogongrass)
We decide to conduct an experiment to see if one of the two herbicides is more effective at killing and suppressing the invasive grass. A simple experimental design might contrast the herbicides at their respective concentration recommendations. We could compare the expected outcomes for each herbicide (e.g., frequency of percentage of dead stems; less than 60%, between 60 and 80%, and above 80%). We could then use a chi-squared test to help clarify if the proportions we observe between herbicides differ-offering support that one performs better than the other. But what if the most effective herbicide is also the most expensive and our budget is limited to $500 (USD). If we purchase the most effective herbicide, we will only be able to treat a small area. Should we buy the less effective herbicide and apply it to a larger area?
Perhaps could we modify our experimental design to help solve this problem? Let us first reconsider our objectives (equally important): maximize percentage of dead stems/m2 and maximize area treated (ha). We presume, from our inference and the literature we reviewed, that meeting these objectives should help bring us closer to our more fundamental objectives: increase suitable habitat and decrease malnutrition. With a small change in the experimental design to include concentration as an additional factor (e.g., below suggested and suggested) we can better approach our decision problem. Now we can experiment and analyze results again using expected outcomes and a chi-squared test. More importantly we can provide some probabilities of the outcomes for each herbicide by concentration condition. We can then calculate the area we will be able to treat from application rate and volume of each herbicide by concentration condition, then include these as attribute utilities. We can then place values on the different outcomes and use a decision tree to clarify the best decision from our parameters.
I’ll be demonstrating how that is done in a future post (i.e., Design and Decisions, Part II), so you’ll have to keep an eye out.