10.4.2 Randomized target design  10.4 Some model experimental  10.4 Some model experimental

10.4.1 Target-control designs

One way to reduce the problem of high natural variability is to find a covariate, a quantity that is correlated with the quantity of interest so that it can be used to reduce the expected variance in that quantity. For example, if the rainfall in a particular area is correlated with some other independent quantity, like instability or low-level moisture, then knowledge of that other quantity can be used to predict the desired rainfall to some accuracy that is better than the climatological average. Such a covariate can make it possible to detect smaller effects of seeding on rainfall. (The same principle applies to other observed weather phenomena and to causes other than seeding.)

One covariate often used is the rainfall in a nearby area. If the nearby area is unaffected by seeding but historically shows a high correlation with the rainfall in the target area, then that rainfall can be used to predict the rainfall expected in the target area in the absence of seeding. For example, a linear regression model of the relationship between target and control precipitation might indicate a correlation coefficient of r between target and control precipitation, thereby reducing the variance about the regression prediction by a factor of 1-r².

There are serious dangers in use of target-control designs. Persistent weather patterns can produce short-term departures from historical relationships, so the linear regression model (applicable to randomly occurring events) can lead to serious errors when applied to such correlated sequences. Another flaw is that meteorological phenomena seldom occur with Gaussian distributions, but more commonly have more outlying events (e.g., with high precipitation) than a Gaussian distribution. Because of this, tests that rely on Gaussian distributions for their justification are usually not applicable. (Sometimes this problem can be alleviated by working with data transformed to new variables, such as square roots or cube roots of rainfall amounts, selected to give distributions closer to Gaussian in character, but this does not help the problems introduced by correlated sequences.)

   
 10.4.2 Randomized target design  10.4 Some model experimental  10.4 Some model experimental