Estimate sample size needed for valid study of medical test

#1
Hi all,
We are performing a follow-up study of a medical test, investigating if it's useful for predicting mortality. I've spent quite a few hours trying to estimate the sample size needed to perform a valid study but I'm still very unsure how to go about this, and I believe I'm at the top of confusion right now... I therefore seek your help!

The problem: Supposedly a medical test can be used to split patients into two groups; low risk and high risk of dying. We wish to test this hypothesis through a prospective study where the test is performed and patients followed up after a given time interval. For this study we wish to estimate the sample size needed to obtain a valid result.

Several values are known from a previous, smaller study, but currently I'm so confused that I'm unsure which ones are actually needed for this calculation. I've tried several different approaches but have a hard time figuring out what actually makes sense in this setting.

Can anyone shed a bit of light on this and hopefully send us in the right direction? :)


Best wishes,
Rune Jensen
 
Last edited:

hlsmith

Omega Contributor
#2
You are likely looking at a two-sided Fisher's exact test. After getting your sample size, don't forget to control for loss-to-follow-up if applicable, by likely adding an additional percentage of the sample to the calculated total. Also, if these aren't inpatients how do you know if they died versus a no-contact (loss-to-follow-up)?
 
#3
You are likely looking at a two-sided Fisher's exact test. After getting your sample size, don't forget to control for loss-to-follow-up if applicable, by likely adding an additional percentage of the sample to the calculated total. Also, if these aren't inpatients how do you know if they died versus a no-contact (loss-to-follow-up)?
Thank you, I'll have a look into this!

Regarding loss-to-follow-up: In Denmark we access to a national personal ID registry which is updated when people die - pretty ideal for studies like this.
 
#5
Rune!

What do you mean by this? Do you have a single variable e.g. a blood sample and for high/low values you would get high risk or low risk? Then I come to think of roc-curves, receiver operating characteristic.

If you want to combine many variables to classify to high/low risk other methods might be needed.
Yes exactly, the test corresponds to a blood test with an established cut-off value based on a pilot study - so in that sense very standard :)