What statistical tests and models are appropriate for your research questions?

The level of measurements should be investigated carefully for the data you are handling since it needs to be considered not only when you are describing the data, but also the proper tests for comparisons. For example, if we want to compare the how age distribution differ between male and female diabetes patients, usually we treat age as ratio data with continuous features. Thus, a two sample t-test would be proper to test the differences between two groups. But if we collect the sample with two variables as gender and diabetes diagnosed or not, a proper test to see if the gender is associated with the diagnosis of diabetes would be a chi-square test for independence.

From my experience, most of the clients do not like complicated or advanced analytics. 3 things that helped me was- 1. Keep it simple! Not everyone's a statistics guru. Break it down with analogies or snappy summaries. 2. Don't just serve the data; serve it with a side of "how to use it." ("So, if we increase A, sales go up"). It's not just about what you found; it's about how it helps achieve their business goal. 3. Talk with visuals. A picture's worth a thousand words, and in analytics, it's worth even more. Even highlighting the top 3 numbers in a 50-column table can reenforce your point and its your job to tell 'why' these 3 numbers matter.

Rule: measure as precise as possible and on as few measurable features as you need for building a model. Know what the error sources are and see if there are good ways to mitigate the noise.

Your research question is a good place to start here. What are you looking to uncover with data? What decisions to make? Which statistical model would give you the most confidence if shown significant? This will determine the data you need to gather or access. And the type of data you have, in turn, will determine the level of measurement.

Preferably less than five according to John von Neumann as recalled by Enrico Fermi to Fresman Dyson. I reality you can do solid work with 10. Using more variables makes you extremely sensitive to missing values.

In my past research experience, I have encountered different types of research questions. The number of variables vary in terms of the questions you are dealing with. For example, for a study that aimed model to claim sizes, only the claim sizes are available then we can only pursue univariate approaches such as distribution modeling; for a questionnaire based study that aimed to investigate how people differ psychologically when facing pandemic, there are 50-100 variables and possibly a factor analysis/structural equation modeling would be good approach; For a study that tries to investigate the genetic effect on a particular disease, there might be millions of SNPs as each of them can be treated as a variable in a regression model.

Look at the raw data. Think about the problem is poses. Try to understand which underlying process created the data. Argue against all your own ideas and be your own worst enemy.

The nature of the relationship is something uncovered by the statistical tests and models, not the other way around. The level of measurement of your variables determines the appropriate models you can create; these, in turn, will give you clues as to the nature of the relationship. The interpretation is key—it's where you, the researcher, really get to shine!

I have encountered a very interesting problem in my recent research in terms of nature of relationship. Me and the group were investigating the potential effect of some recommended behaviors that on a particular illness condition (binary: with or without the conditions). A very simple logistic regression model was built with proper covariates and we do see the recommended behaviors are strongly related to the illness condition. The shocking part is that, the ones who claimed that they did the behaviors based on the instructions were the ones with the higher risk of the illness condition. This seems counterintuitive but this also tell us when people sensed the conditions, they were more likely to practice the recommended behaviors.

I would like to know the Precision, Recall and F1 accuracy metrics. Your inference is model-based so how good was the model is good to know.