Normality testing when comparing between groups using independent t test/ mann whitney U

[Catherine101]

Hi there, I'm hoping someone may be able to clarify:

I have two independent groups one patient group and one control group and I am wanting to compare their scores on the Beck Depression Inventory (BDI)

When doing normality testing to check whether to use a parametric t test or non parametric mann whitney U, should I be checking the normality of each groups scores separately or the whole sample (including patients and controls)?

1. As in do you check normality for the the patient groups BDI scores and then seperately the control group BDI scores?

2. If this is the case and one group comes back as their BDI scores being normally distributed and the other as abnormal do you do non parametric testing as one of the groups isnt normally distributed on that variable?

Thank you in advance for your help on this,

BW
Iona

 

[obh]

Hi Iona,

First, even if the distribution is not normal but reasonably symmetrical you may use the t-test.

1. As in do you check normality for the patient groups BDI scores and then separately the control group BDI scores?

For paired t-test, the differences need to distribute normally.
For independent two-sample t-test, I think both need to be normal. (or reasonably symmetrical)

2. If this is the case and one group comes back as their BDI scores being normally distributed and the other as abnormal do you do non-parametric testing as one of the groups isn't normally distributed on that variable?
Iona

If at least one of them is not reasonably symmetrical you should use the Mann-Whitney U test.

Even if both distribute normally you can use the Mann-Whitney U test. I think The Mann-Whitney U test has 95% efficiency in comparison to a Two-sample T-test so it better to use the t-test if possible, but it isn't so bad to use the MW test.

 

[Karabiner]

Clinical scales are not distributed normally or at least symmetrically in general populations.

The description of the original problem lacks characterization of the populations the samples
were drawn from, and of sample sizes.

With kind regards

Karabiner