# How does reliability and validity affect the results (descriptive statistics)?

Okay, I think this might be a basic question but I just can't quite seem to get there. I understand what reliability and validity is, but I don't understand just what I should do with it when looking at the results.

Imagine a super-simple test examining how much people spend on groceries every week.

For example, say that that test gives a significant result that is reliable but with low validity. So I get that that means that the test wasn't good for looking at the thing we wanted to look at, but that we would get roughly the same results if we repeated it. What can I do with that result then? Does it make the result completely useless? Anything else than draw the conclusion that we can't say anything about the subject that we wanted to look at?

The opposite as well: a valid test with low reliability. Anything else than drawing the conclusion that we can't say anything?

I'd be very grateful for any help. Thank you.

A reliable result that has low validity can't really be redeemed. In your example, it would be like recording something silly like people's height to determine how much money they spent instead of just recording how much money they spent.

An unreliable result that has high validity can be redeemed because it's a problem of measurement error. In your example, suppose you made a numerical mistake in converting currency. In less silly examples, you can integrate measurement error in your model.

• Thank you. So a follow up question: how does validity and reliability affect whether we get a significant result or not? Say that we're doing a study and looking for a correlation between two things, and we get a significant result: if we find out that our method had high reliability but low validity, how did that affect the result? Or the other way around: same test, significant result but with high validity but low reliability. What effect did that have on the test? I'm suspecting that it should decrease the chance that we get a significant result? Thank you for your help. Commented Dec 30, 2013 at 12:13
• Sorry for the late response. I am going to edit my answer above within 24 hours to include an example that will illustrate your concerns. For now, I don't think there is a universal answer yes/no that applies. As far as I know, in real life problems, models have measurement errors integrated into the before doing statistical tests. This is because you know you're making a measurement error since you're not able to measure the thing you want to measure. E.g. in animal studies, you may not be able to measure because it would either kill the animal or require you to remove it from its habitat Commented Dec 30, 2013 at 18:52