Question 766:

Asked on June 23, 2009

Tags: Correlation , Non-Linear Correlation

1

Answer:

No answer provided yet.

When conducting a correlation the underlying structure of the relationship should be roughly linear (that is, it looks like a line). Non-linear relationships exist. For example, consider alertness and test-scores. When someone is not at all alert they get poor scores, when they are extremely alert (hyper-alter) they also get poor scores. It is only at the optimum level of alertness (or non-extreme alertness) where good scores are achieved. This relationship if graphed would look like an upside down u and would not show up well using linear regression or the correlation coefficient (which we see are related). Decisions based on linear regression and non-linear relationships would be inaccurate and likely lead to incorrect decisions. This depends on how severely non-linear the data are.

 

The second area of concern is using individual items from a lager test and the using these to make inferences about another variable (e.g. frequency of absenteeism). It is often the case that test of ability are designed and tested with multiple questions (which allow for tests of internal reliability) and ideally these test-questionnaires do a good job of measuring whatever the construct of interest is. When you deal with individual items, it's not necessarily incorrect, you just increase the chances of measuring something you don’t think you’re measuring. If there are big consequences for decisions that come from the analysis, it isn’t wise to use individual items. In short there is some ethical concern here.