Your initial question was clear. The lack of answers may be because this area is rapidly evolving. One reason descriptive phylogeography is now more difficult to publish is that it has become clear the guesses made from simple descriptions (such as haplotype networks) are often misleading (see Lacy Knowles, Evolution 2008, pointing out the now famous 75% false positive rate). Such over-confident 'inference' is a very serious issue as it fills our literature with incorrect assertions. A second reason is that to follow a (scientific) hypothesis-driven approach in phylogeography we need spatially explicit models of the evolutionary process - and these are technically difficult to develop (see Feldsenstein 1975, Am. Nat, through to Barton, Kelleher & Etheridge Evolution 2010, Baird & Santos 2010). The situation is improving - I comment on a very encouraging study here:

http://www.cell.com/current-biology/retrieve/pii/S0960982212001297

and that (Next Generation Inference) pdf is available through my RG account.

You restated your question with much reduced scope. It might be summarised for inference in general as: Use all the information for inference, or use one part of it first, and then 'verify' that against the rest? The clean answer is: where possible use all the information. Why? Because when you decide to use one part and then verify, you are deciding (before any analysis) which part of your data is the 'important' part.... to check whether this decision influences your answers you should then also: first 'decide' the other part is the important part, and verify against the first... double the work, and difficult to integrate the answers.

To make this more concrete in a phylogeographic context: suppose a (non-pelagic life stage) intertidal mollusc spreads from one point on the Australian coast in both directions round the continent to 'meet' at the other side just at the time you sample the system. [You can change the geographic scale and organism as you wish, just keep the characteristics of the spreading history the same].

We now know the true context, but what might happen if (as in real life) you did not magically know the context and "you first investigate genetic structure and then go to refining hypotheses":

Analysis of the structure may fail because you have isolation by distance through most of the range, and most model based clustering algorithms have models which are very far from IBD. If you have geographically well separated sampling localities, however, clustering will suggest each sample comes from a distinct HWL unit - the samples on either side of the 'meeting point' being much more different from each other than other pairs with the same distance between them. How could this be interpreted?

A simple mistake, which often appears in the literature, is to interpret this as meaning there is some sort of barrier to geneflow between the sample localities at either side of the meeting point. We know, however, (magically) that the strong difference between these is due to phylogeographic history rather than current process. How could we distinguish these hypotheses without magically knowing the context?

The clean answer is we set up a spatially explicit model which allows simultaneously for these two possible causes of the big difference between the meeting point pair of locality samples, and infers the relative likelihood of the causes given all the data (dispersal rates, gene genealogies, GIS habitat suitability and barrier models).

Unfortunately such models are often difficult to develop (see above), and often require more data than can be collected in one research cycle. This should not stop us realising, however, what the best solution would be, and trying to get as close as possible.

One important lesson is: the genetic structure (as inferred using a model-based clustering algorithm) is not a phylogeographic hypothesis test: it is a description of (one part of) the data: It tells us what type of HWL units would my locality samples come from iff the system is made up of HWL units? The strong point of this description is: We can reason about the implications because we understand the explicit model underlying the inference (HWL ie genetic equilibrium within clusters). Often the clustering description allows us to construct a set of phylogeographic hypotheses (in the example, the constructed set has two alternatives) which can then be compared as best we can.

Several of the papers in my RG account touch on these issues - see particularly the 2 papers with Michael Fontaine on porpoises.

The short answer to your second phrasing of the question is then: Yes, in a preliminary analysis you can do the genetic structure first, but equally you could do a genealogical approach first, or model an explicit spatial spread first, the next step is not to refine hypotheses, but rather (using the hints from the preliminary analyses) to define (a list of) explicit hypotheses which are going to be the focus of your main analysis - and which should ideally all compete within one inference framework if you are to make sound relative judgements regarding them.

Hope some of this is helpful,

Stuart

Ok, no answer after few days, may be my question is not enough clear. Please consider it as a surveys and give your way to build hypothesis : do you use independant data (geology, climatic modelisation ...) or do you first investigate genetic structure and then go to refining of hypothesis ?

Your initial question was clear. The lack of answers may be because this area is rapidly evolving. One reason descriptive phylogeography is now more difficult to publish is that it has become clear the guesses made from simple descriptions (such as haplotype networks) are often misleading (see Lacy Knowles, Evolution 2008, pointing out the now famous 75% false positive rate). Such over-confident 'inference' is a very serious issue as it fills our literature with incorrect assertions. A second reason is that to follow a (scientific) hypothesis-driven approach in phylogeography we need spatially explicit models of the evolutionary process - and these are technically difficult to develop (see Feldsenstein 1975, Am. Nat, through to Barton, Kelleher & Etheridge Evolution 2010, Baird & Santos 2010). The situation is improving - I comment on a very encouraging study here:

http://www.cell.com/current-biology/retrieve/pii/S0960982212001297

and that (Next Generation Inference) pdf is available through my RG account.

You restated your question with much reduced scope. It might be summarised for inference in general as: Use all the information for inference, or use one part of it first, and then 'verify' that against the rest? The clean answer is: where possible use all the information. Why? Because when you decide to use one part and then verify, you are deciding (before any analysis) which part of your data is the 'important' part.... to check whether this decision influences your answers you should then also: first 'decide' the other part is the important part, and verify against the first... double the work, and difficult to integrate the answers.

To make this more concrete in a phylogeographic context: suppose a (non-pelagic life stage) intertidal mollusc spreads from one point on the Australian coast in both directions round the continent to 'meet' at the other side just at the time you sample the system. [You can change the geographic scale and organism as you wish, just keep the characteristics of the spreading history the same].

We now know the true context, but what might happen if (as in real life) you did not magically know the context and "you first investigate genetic structure and then go to refining hypotheses":

Analysis of the structure may fail because you have isolation by distance through most of the range, and most model based clustering algorithms have models which are very far from IBD. If you have geographically well separated sampling localities, however, clustering will suggest each sample comes from a distinct HWL unit - the samples on either side of the 'meeting point' being much more different from each other than other pairs with the same distance between them. How could this be interpreted?

A simple mistake, which often appears in the literature, is to interpret this as meaning there is some sort of barrier to geneflow between the sample localities at either side of the meeting point. We know, however, (magically) that the strong difference between these is due to phylogeographic history rather than current process. How could we distinguish these hypotheses without magically knowing the context?

The clean answer is we set up a spatially explicit model which allows simultaneously for these two possible causes of the big difference between the meeting point pair of locality samples, and infers the relative likelihood of the causes given all the data (dispersal rates, gene genealogies, GIS habitat suitability and barrier models).

Unfortunately such models are often difficult to develop (see above), and often require more data than can be collected in one research cycle. This should not stop us realising, however, what the best solution would be, and trying to get as close as possible.

One important lesson is: the genetic structure (as inferred using a model-based clustering algorithm) is not a phylogeographic hypothesis test: it is a description of (one part of) the data: It tells us what type of HWL units would my locality samples come from iff the system is made up of HWL units? The strong point of this description is: We can reason about the implications because we understand the explicit model underlying the inference (HWL ie genetic equilibrium within clusters). Often the clustering description allows us to construct a set of phylogeographic hypotheses (in the example, the constructed set has two alternatives) which can then be compared as best we can.

Several of the papers in my RG account touch on these issues - see particularly the 2 papers with Michael Fontaine on porpoises.

The short answer to your second phrasing of the question is then: Yes, in a preliminary analysis you can do the genetic structure first, but equally you could do a genealogical approach first, or model an explicit spatial spread first, the next step is not to refine hypotheses, but rather (using the hints from the preliminary analyses) to define (a list of) explicit hypotheses which are going to be the focus of your main analysis - and which should ideally all compete within one inference framework if you are to make sound relative judgements regarding them.

Hope some of this is helpful,

Stuart

Nice, big, answer, I am going to read it carrefuly and I will come back

Stuart I read your long but very interesting answer. What I feel is that phylogeography is now a step within a more an approach focusing on the relation between the genome and the environnement, searching for local adapation or causes for phenotypic polymorphism. Former goals of phylogeography were more "naturalist", as reviewed by Avise phylogeography was the link between macroecology (biogeo, paleo) and microevolution study but may be the aims are now getting closer to micro-evol side.

I try to modelize the actual approach depicted in your answer and also in recent litteratutre :

(1) First use occurrence data (present or past) and environemental data to set hypothesis on the process leading to the current distribution of populations.

(2) Use phylogeography to choose the best hypothesis. This step help to define what is the result of migration or of persistence or what have been structured by secondary contact.

(3) in next steps results of phylogeography can be used to formulate precise question about stability in time and space (refugia) or to set independant samples for research of gene under selection, or physical barrier to migration or gene flow.

Dear Baumel,

I got some interesting results which contributed a bit to resolve the complex puzzle of the biogeographic distribution patterns of rhizobia (DOI: 10.1111/j.1365-294X.2011.05368.x). In our case, we got some phylogeography which raised question about the geographic distribution of rhizobia.

Tks for the ref ! What is the consequence of this knowledge on your researches ? Does the geographical structure of rhizobia diversity help you to refine or formulate new hypothesis ?

yes! we hypothesized that a physical barrier (high mountain i.e. Kunzum pass) separates two valleys (Lahaul and Spiti) is the place where rhizobia of East tries to meet West but fail to meet. Resulting isolates of Rlv from Lahaul valley are genetically similar to strains found widely in the western world, whereas those in Spiti valley are more like types that have only been reported from China. We need more information to prove thi.