Statistics: OR v. frequencies

[Function]

Hello everyone

 

For my paper this year, consider a 2x2-table with following information:

 

37 premature babies needed postnatal ventilation. 168 didn't.

145 babies born at term needed postnatal ventilation. 2 819 didn't.

 

A Chi²-test conform Pearson (SPSS) has indicated a significant difference in the distribution of the need for ventilation in both subsamples (at term v. premature).

 

Is it more interesting to report an odds ratio of 4.28, or to report a frequency of 18.05 % in premature babies v. a frequency of 4.89 % in babies born at term?

 

Thanks!

 

F

 

[Prometheus]

In a publication you could expect to find both: the frequencies maybe buried in a table somewhere and a ratio of the proportions explicitly (OR or RR).

 

Generally the difference in proportions is not discussed. Consider, the difference between 0.01 and 0.001 is more noteworthy than the difference between 0.41 and 0.401 even though the difference in proportions is 0.009 in both cases. Thus the OR (with confidence intervals!) is more informative and is more easily comparable across different models too (e.g. logistic regression).

[studiot]

What is your null hypothesis?

 

That the two measurements come from the same population

 

or

 

That they come from different populations

 

Prometheus also makes a good point about what confidence levels your are testing to.

Edited October 20, 2016 by studiot

[Function]

So I should include CI? Confidence intervals of the OR themselves? I'll look into it. What will it imply for my tests?

 

Null hypothesis is that the two measurements come from the same population; study hypothesis is based on a former Swedish study which states and concludes that babies born by means of vacuum extraction are at higher risk of falling victim to cerebral hemorrhages.

 

CI should be 95%, ergo, alpha should be 0.05.

[Prometheus]

Yes, report something like OR (95% CI) = 4.28 (2.36, 5.94). The OR is the ratio of the odds . If the OR = 1, then there is no difference between the odds as would be expected under the null. If you then had an OR (95% CI) = 4.28 (0.92, 12.22) you would say that the OR is not statistically significant (at alpha = 0.05), as the interval between 0.92 and 12.22 includes 1.

[Function]

Thank you already for your valuable feedback.

 

Hmm strange, when I calculate the OR myself, I get 4.28:

 

(37/145)/(168/2819) = 4.28

 

When I run a crosstabs in SPSS, I get the following results:

 

• OR for Postnatal ventilation? (Yes / No) = 0.234, 95% CI = [0.158, 0.346]
• For cohort at term = 0.844, 95% CI = [0.784, 0.909]
• For cohort premature = 3.615, 95% CI = [2.617, 4.992]

What's going on? Why doesn't "For cohort premature" give me 4.282 and what should I do now?

 

EDIT: I now see that 1/4.282 equals the SPSS OR of 0.234. This is due to another set-up of exposure (premature) vs. non-exposure (at term) and of outcome (ventilation needed) vs. no outcome (ventilation not needed).

How do I make it clear to SPSS that premature = exposure and that at term = non-exposure?

 

EDIT (bis): I managed to get the 4.282 in SPSS by changing the value coding for the variable "Postnatal ventilation required?" from 1 = yes, 2 = no to 0 = no, 1 = yes.

 

Ergo, everything is alright. Thanks for the help!

Edited October 21, 2016 by Function

[Prometheus]

If you have the time and inclination i'd recommend learning R for your stats. Its far more versatile and gives you a toe into computational medicine which is a fast developing field.

[Function]

Time's a factor we don't possess We're to make every statistical analysis using SPSS (version 23.0 or 24.0), and tbt, I got used to it so I'm not sure if I'd be so eager in switching to another

[Function]

Update: made a 'mistake' (rather a formality error): since basically the study conducted is a retrospective cohort study, a preference is given to risk ratio (RR) instead of odds ratio (OR), so I've replaced all OR with RR.