Predictive validity is raising and answering the question about how well the results of a test can predict a situation in the future or in the past.
Validity in general is seen as how (in)valid the research has been performed. If it is well performed it is very useful. There are several aspects to look at to evaluate the research. Content and construct validity have to do with the correct use of the terms. Predictive validity has to do with how well a test can predict a situation in the future or in the past. Internal validity is raising and answering the question how well the research has been performed. External validity is raising and answering the question how well the results of a research are useful for daily life. On this page I will explain how to view and score predictive validity.
Suppose you have made a test about abstract reasoning. Now it is easily suggested that the score on this test can predict school performance in the future. Students with a higher score are supposed to finish higher levels of education while students with a lower score will finish only lower levels. (I am sorry that this example is based on prejudice, but for the sake of simplicity it is helpful to explain predictive validity.)
To test this hypothesis a lot of student fill in the test and after a while – when they have finished school – the students fill in a questionnaire about the education level they finally reached. Now the predicted validity can be rated as the computed correlation coefficient between these two scores:
- If the correlation coefficient is 1, there is a perfect relationship.
- If the correlation coefficient is 0, there is no relation at all.
Most of the time a correlation is never exactly 0 or 1, but is somewhere in between. The computed correlation coefficient can be seen as the value for predictive validity.
Remarks about predictive validity
If the correlation coefficient has not been calculated, nothing can be said about predictive validity. So if you want to predict school performance, job performance, or anything alike, you need to know the correlation coefficient. If that is missing, no predictions can be made.
It is easier to make predictions about the near future. A prediction for school success next year is easier than a prediction for school success in ten years.
Predictions can also be post predictive. Now the prediction is about a situation that has happened in the past.
Instead of using a single test a whole battery of tests can also be used. Now the best indicator for predictive validity is the multiple correlation coefficient (R2).
You need to have variables that reflect data from the past and data from the future of the same objects. Therefore, a longitudinal research design is required that is usually quite expensive.