Chapter 16 Selecting the Right Test

Although both the research design and the level of measurement of the variables determine which statistical technique can be used, level of measurement is more likely to limit what technique you can use.

Regression and correlation are limited to continuous variables, dichotomous variables, or ordinal level variables. Discrete categories, like drug A versus drug B versus drug C, cannot be included in a regression model or a correlation. These types of comparisons require an ANOVA (or a t-test, depending on the number of groups). Conversely, a continuous variable cannot be used as an IV in an ANOVA model. You could get around this by dichotomizing the variable (convert the continuous variable into a two-value variable).

There is some overlap between t-tests and ANOVA. You should use a t-test whenever possible, as it is a simpler technique. Mathematically, the results will be the same, however. T-tests can also be one-tailed or two-tailed, while ANOVA (anything using an F statistic) is always two-tailed.

Finally, all the techniques we have covered in this course have been parametric techniques. Parametric techniques make assumptions about the underlying population distribution. Often, but not always, parametric techniques assume a normally distributed population from which your sample is drawn. Because of the central limit theorem, statistics with normality assumptions tend to be robust with random samples that are sufficiently large (30 or more). We can avoid this altogether at a cost of some of our statistical power. Nonparametric techniques have relaxed rules about the required levels of measurement and assumptions. The cost is reduced power. We will look at two nonparametric techniques next.

Important: See the summary table of statistical techniques posted to Canvas for a handy reference and summary of the differences between the techniques.