Correlational Design vs. Correlational Analysis

It is commonly assumed that if you use a non-experimental (i.e., correlational design, where no independent variable is manipulated) then you will use a correlational analysis. However this is not necessarily the case: correlational design and analysis are not the same thing.

Correlational designs involve simply measuring, but not manipulating, the variables. In some situations they are more practical or ethical, or they may allow us to explore new relationships. When we have a correlational design, we may use a correlational analysis (for example, if both variables are continuous) or we may end up using something like a t-test (if one variable is nominal, with two categories, and the other variable is continuous).

Power in Correlational Designs

Just as we should consider power in an experimental design (how to maximize power and minimize the effect of extraneous variables), we should do the same in a correlational design. Therefore, we should try to minimize extraneous variables as much as possible (remember that extraneous variables are any variables that influence the dependent variable but that are not measured or manipulated in our study). Let’s say we are interested in the connection between hours of sleep (sleep quantity) and grumpiness upon waking. Extraneous variables that could also influence grumpiness upon waking up in the morning might be things like age, hours worked the day before, employment status, and so on. Therefore, we might aim to keep our sample as homogeneous as possible, by only selecting people of a certain age range and with similar employment to minimize variability in happiness that is due to those variables versus sleep quantity.

Another way to maximize power in a correlational design is to use reliable and valid measures. For example, if we measure sleep quantity by asking people at the end of the day how many minutes of sleep they had the night before, they might be fairly inaccurate in their response, resulting in an unreliable measure (compared to if we use EEG to measure time spent sleeping based on brain activity). This causes more error variance in our design and, as a result, we shall have a less consistent relationship between the two variables of interest.

Finally, we should ensure to avoid restriction of the range. Restriction of the range occurs when the difference between the lowest and the highest scores on one or both of our variables of interest is small. For example, in our study of sleep quantity and grumpiness, if we select people for our study who are almost always short on sleep (e.g., junior doctors), we will have restricted the range of sleep quantity scores and will likely also have little variability in morning grumpiness, making it difficult to detect if there is a relationship between sleep quantity and grumpiness.