Reporting t-tests instead of simple linear regressions Many papers in health research based on observational data use the following format:
Does A cause B? 


*

*t-test A vs B (or chi-square, or whatever other test of dependence between two variables)

*multiple regression of B on A, adjusted for suspected confounders 

*make heroic inferences


Why is the apparent "standard" to first run the t-test (or what have you)  instead of just running the analogous unadjusted linear model, which can then be more clearly compared to the multivariable linear model of interest? Is there some other advantage to doing it the "standard" way that I am missing?
 A: Consider two situations: A is randomised and A is not randomised.
If A is randomised then the conventional wisdom is randomisation ensures lack of bias and subsequent regression adjustment will not hurt and may help estimate precision.  Against that Freedman's 'On regression adjustments to experimental data' paper argues that such adjustments can introduce bias in small samples and increase or misspecify estimate precision.  Lin 2013 (paywalled) provides a useful discussion of this dispute, summarised informally here, detailing the circumstances where regression adjustments will work well.  And in the comments to this answer @frank-harrell notes that regression adjustment may sometimes even be required to get an unbiased effect estimate.  (So much the worse for unadjusted estimates and initial t-tests...)
In any case, those on Freedman's side of this debate tend to require and appreciate having a t-test for differences in B as a function of A and will generally be willing to ignore subsequent modeling.
If A isn't randomised then you should in normal circumstances expect confounding -- why else do the multiple regression?  In that case I can't see any point to building a model e.g. the simple linear model for which the t-test is a test, that has an estimate that in advance you do not expect to identify the causal effect of A on B.  And then there are the precision considerations noted above.  It seems more sensible to cut straight to the task of identifying that effect by trying to statistically control for its confounders. 
It would seem to be a better idea to use the space the t-test took up to vary the list of confounders in a way that encompasses the set of theoretically reasonable stories about what confounds and how it does so.
