First of all $\min (|E|) = 0$ since the graph can be disconnected. $\min(|E|) = |V| - 1$, is true only for connected graphs. Whether a vertex is needed to be connected to the graph depends on the problem being considered. In general, a vertex can be isolated. So, in general, $0 \leq |E| \leq {{|V|}\choose{2}}$, if the graph is not a multi-graph. If the graph is a multi-graph then there is no upper limit to $|E|$.
Secondly if we are comparing $O((\log|V|^2||E|))$ against something like $O(|V|^2+|E|)$, former will be always better than the latter , since $O(\log(|V|^2)|E|) = O(\log|E|+2\log|V|)$.
So the question is, whether to analyze algorithms in terms of $|E|$ (i.e $|E|$ and $|V|$) or only in terms of $|V|$.
Of course what you say about needing to relate $|V|$ and $|E|$ is correct. However, whenever a complexity is stated in terms of $|E|$, it is better than, say, substituting $|V|^2$ for $|E|$. Note that $|E|$ is always $O(|V|^2)$ even for the cases when it is smaller, for example, for trees. $O()$ is upper bound.
If the analysis of an algorithm is in terms of $|E|$ than we get the complexity bound for all the cases, whether the graph is dense or sparse. Thus $|E||V|$-time algorithm is considered better than a $|V|^3$-time algorithm.
In the cases, if the graph is a tree, or a graph with a max degree $d=O(1)$, then $|E|$ is only $O(|V|)$. So, as an example, an $O(|E||V|)$ algorithm will be considered better than $O(|V|^3)$ algorithm. Thus, if we are able to do a tighter analysis, we are sure that the algorithm will fare better in case of non-worst case input.
As pointed by Raphael, $\Theta(|E|)$ is not same as $O(|V|^2)$. $\Theta$ analysis is better wherever applicable. But usually we don't give $\Theta$ analysis, because if we say some algorithm is $\Theta(f(n))$, and for some easy input the algorithm runs in $o(f(n))$ time, our statement will be wrong.