# Which statistic test should I use?

For a task on data cleaning and analysis using SPSS I have two research questions for which I should pick an appropriate statistical test. It's been over a few years though since my last statistics course and I feel like some sources are contradicting each other. The analysis is done on a set of 5000 women giving birth.

The first question is whether there's a relationship between the Apgar-scores at 1 minute and 5 minutes, I interpreted this as two paired ordinal variables.

My first intention was to use Wilcoxon signed-ranks test, due to the description of the variables. But my statistics notes of back in the days mention a few times that when searching a relationship it is preferred to use correlation, rather than Wilcoxon, which would be more preferred when searching a difference. If it is correct to search correlation, I assume this would be the Spearman's rank correlation coefficient, due to the ordinal character of the variables.

As for the second question, is to find whether there's a difference in the need of extra care (nominal) when the duration of pregnancy is accounted for. The duration of pregnancy is divided into two categories, either less than 37 weeks or between 37 and 42 weeks. Due to the rather obvious greater than/lesser than character of the latter, I interpreted this as an ordinal variable.

Because I'm working with two categorical variables here, I was rooting for a Chi-squared test, but, as I heard from my colleagues, a binary logistic regression would also be possible, I never learnt anything about the latter in the past, so I don't know if this would be a better fit.

For the first question, I agree with you. Rank correlation is what I'd do.

For the second question, first, if at all possible, I would try to the actual number of weeks or days of the pregnancy. Dividing it into two categories is a really bad idea. Similarly, I'd try to get a better measure of care needed that "extra vs. not".

But, whether or not you can get that, this is a regression type problem. If your DV has two categories (extra care, no extra care) then logistic regression is a good choice.

• Thank you Peter! It was part of the assignment to transform the length of pregnancy from a continuous variable to the one as described above, for which I thought ordinal was rather better than nominal. Would it be possible to explain the motivation for logistic regression rather than Chi-squared, or why Chi-squared wouldn't be the best fit here?
– LD94
Mar 2, 2017 at 13:06
• For example, there was another optional question, which was to find whether there's a difference between a spontaneous delivery or a cesarean section in transfer to the intensive care for babies that had a breech presentation. As the method of delivery and whether or not there was a transfer are both nominal variables, wouldn't it be appropriate to use Chi-squared or Fisher's exact in this case?
– LD94
Mar 2, 2017 at 13:21
• I have been reading up on this and I think I came to the conclusion that both the question in the main post and the optional question I posed just above should be approached through binary logistic regression because I'm trying find if one variable can predict the other, rather than just "looking" at a possible relationship they might have?
– LD94
Mar 2, 2017 at 20:25
• Chi-square can be appropriate with two nominal variables, but here it seems like regression (in this case, logistic regression) is more suited to what you want to find out because you have a dependent variable. Mar 4, 2017 at 13:49

Agree with you and Peter regarding the first question.

Will comment more on the biology regarding the second one: categorizing gestational age into term/pre-term is a bad idea both statistically and biologically. Firstly, you end up treating a very wide range of gestational length as equal (25 weeks of pregnancy is very different from 37); secondly, a lot of deliveries are in the range 37-38 weeks, and gestational age estimation is usually not precise, so you will have a lot of misclassification; thirdly, it has been shown that a lot of subsequent health outcomes in fact have a smoother monotonous relationship with gestational age, as opposed to a threshold effect at GA of 37 weeks.

Source: been working on pregnancy phenotypes for the past few years.