95% Confidence intervals differ between statistical programs

#1
Dear all,

I am currently writing a paper for which I have performed some logistic regression analyses using graphpad prism.
In these models we evaluated whether a biomarker was able to predict disease progression in neurological disease patients.
We found that the biomarker was able to predict the disease progression, both in a univariate model and in multivariable models including demographics.
(for example, we found that in the multivariable model biomarker levels were significantly associated with disease progression with an OR of 1.097 (95% CI, 1.045 - 1.169; p = .001))

Now, my PI always wants us to confirm eachother statistical results using different statistical software.
My colleague therefore also performed the logistic regression using Medcalc (with the same database) which resulted in the same OR of 1.097 and same p-value of .001, but in slightly different confidence intervals (1.0373 - 1.1595).
When I repeat the analyses in SPSS it gives the same OR, p-values and confidence intervals as Medcalc.

Is it possible that graphpad calculates the 95% CI for the odds ratios in a different manner than SPSS and Medcalc, thus explaining these results?
If so, which program would you recommend to follow?
 
Last edited:
#2
I would not trust graphpad prism that much as a statistical computing program. I would trust SPSS more.

I guess that one of them computes the simplified Wald interval, (and the most common) of dividing the parameter estimate with its standard error. It is possible to get a more precise estimate. (Pawitan "In All Likelihood") R is using a profile likelihood.
 

hlsmith

Less is more. Stay pure. Stay poor.
#3
I would second @GretaGarbo recommendation. I will also note these are trivial differences at the hundredths place, so it isn't like one is spitting out bizarre values.