Considering Cause and Effect

Determining if variation in one condition cause and effect in a second are core issues in analysis. The strongest research approach is experimental -- a process in which one condition is intentionally changed while all others are held constant. Drawing a conclusion that the changed condition is a cause of a change in a second condition is well supported by the logic of the process.

However, such experimental control is usually not possible in "real-world" applications. Five issues provide an outline of positing that an observed relationship between conditions can be judged as probably a cause and effect relationship.

	Time order
	Statistical relationship
	Assumption that all other things are equal
	Non-spurious
	Effect size -- not just statistically significance but also substantial effect
	An example of why statistical control is important in determining cause and effect: Unemployment and Crime

Time order:: The issue of "time order" is both simple to understand and yet complex to apply. The basic criteria is that any condition posited to be a cause must occur before its assumed effect. Logical? Yes! But the challenge arises when information is collected using respondent recall and single time survey.
For example, if a question asks about an outcome such as success in wining a sports event, and effort in preparing for that event a self report of effort prior to the race will be different depending on the outcome. The same person if they win will report higher effort in preparation for the event than will that person report if they have lost the event. In both of these situations the self-report may be different than a report of effort in preparing collected prior to the race.
Thus, the strongest procedure is to collect measures of conditions assumed to be causal prior to the time at which an outcome is determined. The less desirable procedure is to collect measures of conditions independent of measures of outcomes. For example, measure psychological traits in a survey, and outcomes such as earnings from an employer.
Statistical relationship: The issue of determining a statistical relationship between conditions is the focus of many of these pages and listed procedure. In order to judge, or conclude, a cause-effect relationship there must be some statistical measure of that relationship at a significant level.

That relationship must further meet two of the following conditions. The relationship needs to be non-spurious and substantial. We will approach the issue of non-spuriousness through first examining the idea of "other things being equal."
Assumption that all other things are equal: In an experimental process the use of a control group and experimental group; random assignment of subjects, and manipulation of a single condition or multiple experimental groups allows one to conclude that the conditions being manipulated are causes of changes since that procedure serves to prevent systematic biases. Thus, we claim that other than the manipulated condition "all other things are equal."

In the absence of this procedure two approaches are used in non-experimental procedures -- statistical control or alternative hypothesis testing. When there are conditions that we cannot control, but that we know have (or might have) significant affect on outcomes, one might introduce statistical controls. That is, we use statistical procedure to control for some of the other conditions in a real world setting while testing for a cause and effect relationship.

If, for example, we want to examine the effect of a training program on sales. And we believe that both effort (drive) and experience in the market also effect sales. Then a design that compares sales outcomes between those who attended the training program and those who did not attend, will not be a good evaluation of the training program. A better examination would use three variables: 1) training, 2) effort, and 3) experience. In this case the effort and experience measures sever as statistical controls for the evaluations of the training program.

Frequently we do not know what conditions we need to control for, or there are no measures of those conditions available in the data. In such a situation an alternative is to use theory and brainstorming to test if other conditions might show that the cause-effect relationship has a alternative causal explanation. Of course this is a far less desirable approach since we cannot test all possible alternative hypotheses.
Non-spurious relationship: Following from the need to control for conditions that may not be equal among cases discussed above, when such efforts fail the false assumption of cause and effect is labeled as "spurious." Thus, avoiding such a mistake is the issue of a relationship being "non-spurious."
Effect size -- not just statistically significance but also substantial effect: The first judgment in application of statistics to real-world settings is derived from a judgment of statistical significance. When the number of cases being examined is small that judgment is of primary importance. However, as the number of cases increases past hundreds then many relationships and differences attain a standard of significance from a statistical point of view. However, such patterns may not be of an substantial impact. In these cases the effect size may help one determine how important a cause may be.

Effect size is the ratio between the observed relationship and the standard error of that measure.