There is one strategy that is usually seen in scientific papers: to do a univariate analysis for every single variable and include in the multivariate analysis those with a certain level of significance (

If I understand correctly, you are trying to confirm whether or not the factors you believe are important in determining a certain outcome (e.g. hypertension/heart rate --> cardiovascular disease) do actually impact said outcome (in your dataset).

Now, if you would like to do this for each variable in and of itself (E.g. Hypertension alone or elevated heart rates alone), then you could simply calculate a measure of association for each of these variables and see how significant the association is.

The question of which measure of association to use depends on the type of data you have. For instance, if you have a binary risk factor (E.g. Hypertensive/Non-hypertensive) and a binary outcome (E.g. Cardiovascular disease/No CVD), then you can simply calculate the odds ratio (or the risk ratio, with the latter having the advantage of being more easily interpretable).

Other types of combinations (E.g. Binary risk factors and continuous outcomes) may be assessed using different measures of association, all depending on the type of data you have.

Now, if you'd like to input several risk factors into a single model, then the choice of which factors are the most important (and thus deserve inclusion in the model) can be determined by a number of approaches . As Ángel Herráiz-Adillo mentioned above, one approach would be to assess each variable in and of itself, calculate a p-value for each, and then include those variables having a significant p-value, but as mentioned, this has been criticized as being a somewhat inappropriate method.

A more holistic approach (E.g. Does it make sense, in terms of pathophysiology (for instance), to include this particular factor as a predictor of this particular disease? That decision can be based on your understanding & reading of previous literature) has been advocated.

However you choose to construct your model (there are a number of methods (e.g. forward or backward stepwise regression) that you can read about, and a number of pitfalls/caveats to avoid), you'll finally arrive at a model containing a number of variables, each of which has a certain measure of association. (Again, the measure of association depends on the type of data you have)

In that model, some variables may have a stronger association than others (For instance, the odds ratio for blood pressure might be 3.4, whereas for heart rate it might be 2.1, which may lead us to conclude that BP is a more important variable than heart rate). It's obviously not that simple as there are a number of pitfalls during model-building that can cause you to underestimate each variable's contribution (Collinearity being one example), so I would recommend that you take a bit of time to read a little bit about these pitfalls.

A fantastic course I've enjoyed on this topic is linked here: https://www.coursera.org/learn/linear-regression-r-public-health

My collkeague Dr. Tiago Paggi de Almeida is a specialin this. Please try to contact him at... [email protected].