Stress levels in work shifts

[luckyb]

Hello

I'm interested in the effects of shift type on the workers' stress. In my study, I'm following the stress during a three-week period during which the participants have varying patterns of morning, day, or evening shifts and days off. Estimating the effects of particular shift types on stress should be adjusted for individual variance in stress levels.

Stress here is a dichotomous variable (markedly high stress or not).
Shift type is morning, day, or evening. There are far less day shifts than other types.
I've 20 subjects. Not all of them work all kinds of shifts.

I'm thinking shift type is a fixed factor and subject is random.

What kind of test would best suit this purpose?

 

[j58]

Yes, shift is fixed and subject random. Generalized (logitic) linear mixed model.

 

[luckyb]

Yes, shift is fixed and subject random. Generalized (logitic) linear mixed model.

Thanks. Another person suggested Friedman's test, but I'm not sure what to go with. Can you offer clarification as to why GLMM would be better? Friedman certainly looks easier to carry out.

 

[j58]

Well, modeling was your idea. Why else would you have brought up the issue of which effects were random and which were fixed? But modeling is always superior to mere significance testing, because modeling provides estimates of the effect sizes, as well as assessment of their statistical sigificance. Whereas mere significance testing will only tell you whether the effect is signficant or not, but will provide no estimate of the size of the effect.

 

[Karabiner]

But with small to medium sample sizes (here we have just n=20),
usually no reasonable estimate of an effect size can be made in
a frequentist framework, i suppose.

Apart from that, "Not all of them work all kinds of shifts" , but
Friedman would require that subjects provide data for all
conditions. So modeling would still be preferable.

With kind regards

Karabiner

 

[j58]

But with small to medium sample sizes (here we have just n=20),
usually no reasonable estimate of an effect size can be made in
a frequentist framework, i suppose.

Practically all null hypotheses are false, so testing them is usually silly and amounts to testing whether you had a large enough sample size to declare the observed effect "significant." Since the null hypothesis is almost certainty false, the only thing that makes sense to do is to estimate the effect size. If the sample size is too small to accurately estimate the effect size, that is something that is important to know, and it will be revealed by calculating the confidence interval.