Cilicia Nascimento, I think the simplest, and probably most sensible, thing you could do is run two independent-samples t-tests. In both of them, pregnant women would comprise one group, and nonpregnant women the other group.

In the first t-test, you would compare the two groups on % colonization; in the second t-test, you would compare the two groups on % serotype.

I would be interested to see whether others disagree with me.

Cilicia Nascimento, I think the simplest, and probably most sensible, thing you could do is run two independent-samples t-tests. In both of them, pregnant women would comprise one group, and nonpregnant women the other group.

In the first t-test, you would compare the two groups on % colonization; in the second t-test, you would compare the two groups on % serotype.

I would be interested to see whether others disagree with me.

I agree with Robert Trevethan. Two separate independent-samples t-tests were also my first idea for implementation of your analysis.

The only thing I wanted to add was that your specification "it's a sneaky little bug" really made me laugh so thank you Cilicia Nascimento for that and good luck with your research :)

Two models come to mind, if you operationally define the target Dependent Variable as "Infected = Yes(1) else = No (0)". This is a classic dichotomous variable suitable for predictive models such as Binary Logistic Regression (BLR) or Chi Square Automatic Interaction Detection (CHAID) aka "classification/regression decision trees" . Both models are non-parametric and can accommodate long lists of Independent Variables (predictors of the target DV). With Infected Yes/No as the dichotomous DV, all other Independent Variables in the model could be dummy coded as in "Pregnant Yes (1) or No (0)". Serotype A Y/N, Serotype B Y/N, Serotype C Y/N, Serotype XYZ...). Percent colonization can be entered as a scale proportion 0.10 (10%), 0.25 (25%) for each Serotype. The heuristic for BLR is that scale IV should not be "linear in the logit" (not in the form of a log function). CHAID has no such limitations. The results of both models are a c-statistic of model probabilities that can be evaluated as an AUROC (area under the receiver operating characteristic) curve, representing the plot of sensitivity versus specificity, where anything above 0.75 AUROC is considered to be a good model. These type of predictive models answer the question "What factors like preganacy (or not) interactively predict the Odds of Infection?" BLR tells you how much (what are the Odds at 95% confidence) that being pregnant (or not) resulted in Infection (or not), along with all of the other demographics that should be controlled for (age, gender, co-morbid conditions, procedures, diagnosis codes, smoking, drinking, drug abuse, and anything else that could confound Infection). CHAID tells you what BLR never shows: what is the hierarchy of interaction? What factor accounts for the largest ranking overall effect upon the "Odds of the prevalence"? That is, the overall population infection rate?

One further note. If you are going to use General Linear Model methods (ANOVA, t-tests), you need to demonstrate that all groups and subgroups in your data do not violate the key assumption of the GLM tests, namely that the data are distributed normally. Use the Kolmogorov-Smirnov test for normality. If the data fail this test, then use non-parametric methods, otherwise any GLM test results will be invalid and contested during peer review for publication. Healthcare data from non-stratified convenience samples are infamous for being non-normally distributed. BLR requires around n1,200 observations, but can work with as few as n200 if there are not too many IVs. CHAID will work with quite small samples (n50).

I hope this helps answer your question.

Cilicia Nascimento, you have just been offered some quite sophisticated advice about how to analyse your data. I don't want to get into a slinging match, but I suspect it is beyond what you need.

For one thing, the t-test has been found to be quite robust to departures from normality. For another, the Kolmogorov-Smirnov test isn't the bee's knees under all circumstances. You might like to check out the document that I've attached to this post.

I also suspect that, given the nature of your question, you are not familiar with ins and outs of statistics, so the simpler - but valid - means of analysis might be the most appropriate and effective.

That said, it would be helpful if you provided information about how many women are in each group and what your data look like (e.g., what the means and standard deviations are, and whether the data distributions are noticeably skewed or not).

Do either of the percentage variables have a lot of data near 0 or 100%? What is the correlation between the two variables? What I think would be useful is showing us a scatterplot of the two percentage variables (jittering if needed) separately for the two group. Do you believe that there is a causal relationship between colonization and infection (I am assuming so). THe answers to these questions and those plots will be useful for anyone to answer this question.

The question needs more detail on the dependent variables of % colonization and % infection and also how you are integrating different serotypes with these independent variables.

I agree with Viswa Chandu , we need more details about your study. What is the hypothesis and whether you are studying different serotypes, if so how many? This will determine your dependent and independent variables and the number of the categories your analyzing which is important to decide what statistical test you should use. Good luck!

First, thank you ALL for the amazing answers. I did not expect such great insights on my problem.

Robert Trevethan you are right, a do not have a great background in statistics and many of these answers got me more confused. But thank you for the PDF I did a Q-Q plot on my data. It's in this file I attached.

Also, thank you Richard E Gilder for the response. But I really don't think i need all of that, my data is much simpler.

I just want to clarify: I'm counting all samples as positive, but depending on the site where the bacteria was isolated, I'm considering as an infection or normal colonization.

There are ten serotypes but I found only 7 and some of them are not-typeable so there are eight variables. Three serotypes are more common so the percentage is concentrated in there.

As more details were requested =) Viswa Chandu Ebtihal Isleem

My hypothesis now is that: non-pregnant women carry different infection strains than pregnant women. And pregnant women carry different colonizing strains than non-pregnant (let me know if this is too confusing, i often have that problem).

This attached file has some descriptive data on my numbers (and a statter plot as mentioned by Daniel Wright ), hope it clarifies what I'm dealing with.

Thanks again you all for the HUGE help =)

Cilicia Nascimento, I would like to be able to keep helping you, but I have to confess that the output you've provided mystifies me. Furthermore, I now have to confess that I don't know enough about the nature of the data you have collected, so I am "disadvantaged" at a fundamental level.

I hope someone else can help you.