Determining Significance - in desperate need!

[kretz]

Dear all,

First post and I'm afraid I'm so out of my depth I don't even know if my title is correct. I am very new to statistics and have been struggling through but am now completely stuck. Any help would be massively appreciated. My basic question is what test can I use to determine significance between two cohorts and a binary variable?

Some background: I am trying to publish a research paper on what catheters (tube to wee through) are placed in patients following a TURP - resection of the prostate gland. Common practice is to put a three way catheter (3WC), which is done everywhere except my hospital, where we put a two-way catheter (2WC). We have collected data for a decade and our outcomes look equivalent to other centres and in some aspects better - and it's cheaper to do it our way too.

So, over the decade we have placed both two- and three-way catheters so we have two cohorts in our own data set to compare - that I think I've done fine. The next step is comparing data from our set to an externally published source. The current variable I am looking at is:

What proportion of patients required blood transfusion intra-/post-operatively. So looking at the 2WC cohort of my patients: 3/601 (0.17%) required transfusion, in the externally published cohort of 3WC patients, 0/247 required transfusion. I tried a Chi-square calculator but that came back saying you can't use zero as a value? (I don't understand why?).

I understand that these numbers are tiny and therefore will have limited significance but I would like to show that with a massive p number somehow.

Any help would be massively appreciated, many thanks in advance for anyone that takes the time.

L

 

[Romsek]

Suppose we are trying to estimate the proportion of a population that meets some criteria. Such as requiring a transfusion.
We collect a sample population of size [MATH]N[/MATH] and we note the number meeting the criteria as [MATH]X[/MATH]With a large enough population our estimate of the proportion

[MATH]p = \dfrac{X}{N}[/MATH] is approximately a normal random variable with parameters

[MATH]\mu = p\\ \sigma = \sqrt{np(1-p)}[/MATH]
This is why the chi square test with [MATH]p=0[/MATH] doesn't work. The standard deviation becomes 0 which is meaningless.

So how to fix this? Well, ideally you would be able to increase the sample size of the 3WC patients until at least 1 required a transfusion.
I'm assuming that this isn't possible at this point. So I would jiggle things a bit. Let your estimate for \(\displaystyle p\) for 3WC be [/tex]0.01[/tex]
for example, instead of 0, and see what that gets you.

I'm not sure I understand why you think 2WCs are better than 3WCs given that the 3WCs had 0 required transfusions.

 

[kretz]

Romsek thank you so much I'll put this to the data tomorrow and see what we get.

The reason 2WCs are better than 3WCs is due to other variables where they perform better, this is one small variable and its exceedingly rare that someone needs a tranfusion and if they do its likely to do with other pre-existing health conditions/medications and therefore unlikely to be related to the catheter. the reason I wanted to get a p value for this was to demonstrate that this likely wasn't statistically significant.

Thanks again, I'll post tomorrow and let you know how it goes!

L