Generalizability is not only common to research, but to everyday life as well. In this section, we establish a practical working definition of generalizability as it is applied within and outside of academic research. We also define and consider three different types of generalizability and some of their probable applications. Finally, we discuss some of the possible shortcomings and limitations of generalizability that researchers must be aware of when constructing a study they hope will yield potentially generalizable results.
In many ways, generalizability amounts to nothing more than making predictions based on a recurring experience. If something occurs frequently, we expect that it will continue to do so in the future. Researchers use the same type of reasoning when generalizing about the findings of their studies. Once researchers have collected sufficient data to support a hypothesis, a premise regarding the behavior of that data can be formulated, making it generalizable to similar circumstances. Because of its foundation in probability, however, such a generalization cannot be regarded as conclusive or exhaustive.
While generalizability can occur in informal, nonacademic settings, it is usually applied only to certain research methods in academic studies. Quantitative methods allow some generalizability. Experimental research, for example, often produces generalizable results. However, such experimentation must be rigorous in order for generalizable results to be found.
An example of generalizability in everyday life involves driving. Operating an automobile in traffic requires that drivers make assumptions about the likely outcome of certain actions. When approaching an intersection where one driver is preparing to turn left, the driver going straight through the intersection assumes that the left-turning driver will yield the right of way before turning. The driver passing through the intersection applies this assumption cautiously, recognizing the possibility that the other driver might turn prematurely.
American drivers also generalize that everyone will drive on the right hand side of the road. Yet if we try to generalize this assumption to other settings, such as England, we will be making a potentially disastrous mistake. Thus, it is obvious that generalizing is necessary for forming coherent interpretations in many different situations, but we do not expect our generalizations to operate the same way in every circumstance. With enough evidence we can make predictions about human behavior, yet we must simultaneously recognize that our assumptions are based on statistical probability.
Consider this example of generalizable research in the field of English studies. A study on undergraduate instructor evaluations of composition instructors might reveal that there is a strong correlation between the grade students are expecting to earn in a course and whether they give their instructor high marks. The study might discover that 95% of students who expect to receive a "C" or lower in their class give their instructor a rating of "average" or below. Therefore, there would be a high probability that future students expecting a "C" or lower would not give their instructor high marks. However, the results would not necessarily be conclusive. Some students might defy the trend. In addition, a number of different variables could also influence students' evaluations of an instructor, including instructor experience, class size, and relative interest in a particular subject. These variables -- and others -- would have to be addressed in order for the study to yield potentially valid results. However, even if virtually all variables were isolated, results of the study would not be 100% conclusive. At best, researchers can make educated predictions of future events or behaviors, not guarantee the prediction in every case. Thus, before generalizing, findings must be tested through rigorous experimentation, which enables researchers to confirm or reject the premises governing their data set.
There are three types of generalizability that interact to produce probabilistic models. All of them involve generalizing a treatment or measurement to a population outside of the original study. Researchers who wish to generalize their claims should try to apply all three forms to their research, or the strength of their claims will be weakened (Runkel & McGrath, 1972).
In one type of generalizability, researchers determine whether a specific treatment will produce the same results in different circumstances. To do this, they must decide if an aspect within the original environment, a factor beyond the treatment, generated the particular result. This will establish how flexibly the treatment adapts to new situations. Higher adaptability means that the treatment is generalizable to a greater variety of situations. For example, imagine that a new set of heuristic prewriting questions designed to encourage freshman college students to consider audience more fully works so well that the students write thoroughly developed rhetorical analyses of their target audiences. To responsibly generalize that this heuristic is effective, a researcher would need to test the same prewriting exercise in a variety of educational settings at the college level, using different teachers, students, and environments. If the same positive results are produced, the treatment is generalizable.
A second form of generalizability focuses on measurements rather than treatments. For a result to be considered generalizable outside of the test group, it must produce the same results with different forms of measurement. In terms of the heuristic example above, the findings will be more generalizable if the same results are obtained when assessed "with questions having a slightly different wording, or when we use a six-point scale instead of a nine-point scale" (Runkel & McGrath, 1972, p.46).
A third type of generalizability concerns the subjects of the test situation. Although the results of an experiment may be internally valid, that is, applicable to the group tested, in many situations the results cannot be generalized beyond that particular group. Researchers who hope to generalize their results to a larger population should ensure that their test group is relatively large and randomly chosen. However, researchers should consider the fact that test populations of over 10,000 subjects do not significantly increase generalizability (Firestone,1993).
No matter how carefully these three forms of generalizability are applied, there is no absolute guarantee that the results obtained in a study will occur in every situation outside the study. In order to determine causal relationships in a test environment, precision is of utmost importance. Yet if researchers wish to generalize their findings, scope and variance must be emphasized over precision. Therefore, it becomes difficult to test for precision and generalizability simultaneously, since a focus on one reduces the reliability of the other. One solution to this problem is to perform a greater number of observations, which has a dual effect: first, it increases the sample population, which heightens generalizability; second, precision can be reasonably maintained because the random errors between observations will average out (Runkel and McGrath, 1972).