Granger in logged fiirst differences or in VECM

[DaanN]

I have one variable of which I'd like to understand the relationship to three other variables better. I want to test their relationships on bivariate basis each time, which gives me three pairs. All these variables have a unit root and are I(1). I want to to Granger causality tests on them

For one pair, the cointegrating relationship is very strong.
For two other pairs, the Johansen cointegration test is just barely significant of rank 1, but almost rank 0 (no cointegrating relationship).

1. To do Granger causality tests, how can I best proceed?
a. do Granger causality directly in logged first differences
b. do all of them in VECM View/Lag Structure/Granger causality
c. do only the strong cointegrating relationship via VECM and the others in logged first differences

2. I believe I read somewhere that Granger Causality should be done in a VECM framework if there actually is a cointegrating relationship between the two variables. Could somebody explain me why?