In the context of hospital records, I have two dichotomous variables: sex (female/male) and admission circumstance (planned/unplanned) for different years, so subjects are not the same from one year to another. Which analysis can I use to compare if the proportion of planned and unplanned admissions changes depending on sex from 2015 to 2016? Can I use McNemar's test?
I would try the Fisher's exact test: https://mathworld.wolfram.com/FishersExactTest.html
Maybe if you want to see changes, you can make the difference of contingency tables and run the test on it.
I'd go for a logistic regression model with sex and time as well as sex:time interaction as predictors.
If the subjects aren't the same, you can't use McNemar's test. There's just no way to format the data so that it would work.
I did not recommended the regression models because you have binary values. 2 data points will result overfitting.
Thank you very much. How can I make the difference between contingency tables? I mean, I can make two contingency tables: one with unplanned admissions and the other with planned admission (I attach the example), but I can't see how to merge both tables.
The problem with the proposed regression is that those inputs are continuous, while here are binary. So binary input with binary output are given. To get difference of contingencies, I suggest division of tables to avoid negative values. In your example list we cannot conclude a solid result because 95% confidence (p < 0.05) needs minimum 20 data.
I would try this R code on your whole data:
library(readxl)
d
With this tiny dataset your statistical power is incredibly low. But there's another problem too : December of year 1 is actually just before January of year 2. The division of the data into years masks the continuous nature of the real variable – time – and reduces any study power by a considerable amount.
So applying the Mark I eyeball test to the data, there's nothing to see. And thinking about your time variable, dichotomising a continuous variable is never a good idea.
Laszlo Baranyai , can you provide more information about why you think regressions are not appropriate with binary data? Also, you suggest Fisher's test. Do you think the margins are fixed like in his tea example?
Binary input and output will draw a rectangle of 4 points what is proposed to fit with a logistic curve. If you summarize df, that will result overfitting. I think processing contingency is more robust. I tried. Did you?
This was hospital example, instead of tea. If you refer to the link, there is a Titanic example with many input variables containing category, binary, integer... Where is tea data?