There is no rule of thumb for your problem (despite some colleagues might disagree), since each marker have its own type of information. I will recommend a multi-marker approach where you use a couple of moderately variable markers (one mitochondrial and one nuclear), as well as two highly variable (one mitochondrial, as the control region; and one nuclear as microsatellites). Then in your analyses will have to explain the different patterns they will show. Beware of the F-statistics, since they are very sensitive to these differences among markers.

I will also use a "total evidence" approach, where you concatenate all the info you have.

I totally agree with Jose, the combined use of variable markers having different evolutionary history and with each one having its own unique genealogy may indicate concordance or divergence from the species history. Indeed, mitochondrial DNA only reveals a species maternal demographic history. However, nuclear-derived data are influenced by sex-biased dispersal because of the character of biparental inheritance. Thus, the combination of these different markers may actually provide complementary information to test the population demographic history and extant population structure.

The patterns of migration connectivity among different compartments of the lake-river hydrographic network might be explored via the software MIGRATE on both markers when you will examine both historical and contemporary gene flow. On the other hand if you want to test the hypothesis of individuals coming from each source being genetically homogeneous, but being different across sources, I would advise you to make BAPS (Bayesian analysis of population structure) analysis on both markers. STRUCTURE and GeneClass softwares would enable you to strengthen your analysis based on nuclear data. I would also advise to make BSP (Bayesian skyplot) analysis, implemented in BEAST, on mitochondrial data to get a clear picture on population demographic history, assessing the evolution of effective population size through time. 

Good luck!

Perhaps you could take a look at some publications of Sasa Maric, University of Belgrade, Institute of Zoology. He combines mitochondrial and nuclear markers for some freshwater fish species. Maybe it can be useful. cheers

http://www.biomedcentral.com/content/pdf/1297-9686-43-2.pdf

The key is to have a good mutational model for whatever regions you type. Then you can apply that model to coalescent simulation, of some other means of simulating data, to generate different expectations of your data under different models you want to explore / test / compare.

Single loci systems (e.g. mtDNA) would be a poor choice as  its genealogy is usually only weakly constrained by the demographic history, so logically it only contains very limited information on that demographic history. BEAST  is very good at squeezing information out of mtDNA data, but please remember that it is so elegant precisely because it assumes a single population model.

This frustration has lead many into the murky and pseudo-scientific world of interpretative phylogeography, and a collapse in their scientific credibility.

If you have many thousands of genomic loci (SNPs, STRs) there are plenty of clustering-based approaches that will sharpen your intuitions about underlying population structure. FineSTRUCTURE is pretty good, and has a time component, and for pretty plots you could do PCA / ADMIXTURE / others mentioned Temim Deli above. But do remember that most of the genome-wide analysis methods don't allow you to estimate parameters for, or compare different bespoke models.

It is from your observation that you should decide the genetic markers. It will not necessarily be the same for all researchers working in this area.

Thanks for all of your replies. My general impression is that if I had to pick just one marker type, msats have the best value. I like the idea of using genome-wide SNPs, but the costs are likely more that I can ask funding for.  And I feel that mtDNA markers are a poor bet, as they are not likely to be variable enough anyway.

Now comes my follow up question: I have found papers using msats for each of the three species I would like to study, but they are all from populations in other continents. I have no experience working with msats, and relatively limited resources to devote to an initial pilot for this project. In your experience, how likely is a given msat region to be useful across geographically distant populations of the same species?

Fully agree. I am just looking to maximize the chance the pilot will give me enough results to support a grant proposal. Make it a really well informed guess :)