Phylogeography has emerged over the last two decades as an exciting and informative field that spans the chasm between population genetics and systematics. Population genetics has traditionally been interested in the processes of evolution, including estimates of associated parameters such as migration rates, genetic diversity, and measures of selection; systematics has focused on historical patterns of divergence and associated morphological, temporal, and ecological changes correlated with such patterns. Phylogeography attempts to merge these perspectives and capitalizes on gene divergence patterns and their fit to predictions derived from different models in population genetics. The seminal work of Avise et al. (1987) ushered in this approach to merging concepts from these two fields to study speciation and population divergence (Hickerson et al. 2009). Over the past two decades, an impressive set of analytical tools have been developed with advances in both population genetics and phylogenetics that have had a significant impact in phylogeography (Pearse & Crandall 2004; Excoffier & Heckel 2006; Kuhner 2009). Yet phylogeographers still face a fundamental question of whether they are testing a priori hypotheses associated with the species in question, estimating important population genetic parameters of interest, developing post hoc hypotheses of divergence, or attempting to distinguish among an array of alternative models to explain the observed data from the population(s) of interest (or some combination of any or all of these). In this issue of Molecular Ecology, Carstens et al. (2009) advance the use of model selection and multimodel inference as a means of exploring alternative models of phylogeographic patterning and thereby add significantly to the phylogeographer’s toolbox.
Model selection is currently enjoying increased use in several biological disciplines (Johnson & Omland 2004); many readers of this journal will be aware of its application in molecular systematics (Huelsenbeck & Crandall 1997; Sullivan & Swofford 1997; Posada & Crandall 2001; Posada & Buckley 2004; Sullivan & Joyce 2005; Posada 2008). A hallmark of this statistical approach is that verbal hypotheses represented as mathematical models are evaluated for fit to observed data (Burnham & Anderson 2002). This allows researchers to simultaneously evaluate multiple hypotheses, shifting the focus away from accepting or rejecting a single null hypothesis, and towards comparing the relative strength of support for competing explanations. Successful application of model selection relies on carefully articulating an a priori set of hypotheses that can be represented as mathematical equations and that can be fit to observed data, often using a maximum likelihood approach in an information theoretic framework or using a Bayesian framework. The key challenge is defining the candidate set of hypotheses and ensuring that the parameters and model construction accurately reflect the verbal constructs. Multi-model inference is an extension of model selection wherein the relative importance of different parameters can be determined (e.g. base frequency differences vs. transition/transversion differences in models of molecular evolution). Clearly, this statistical approach could have important application in the field of phylogeography, where several historical scenarios might explain current patterns of spatial genetic variation within species with associated differences in historical and current parameter estimates (e.g. migration rates or genetic diversity).
The application of model selection by Carstens et al. (2009) illustrates the potential utility of this approach in phylogeography. In their study, these authors use multilocus data from a species of salamander from the Pacific Northwest (Fig. 1) to rank 17 different models to explain evolutionary processes impacting this species. Interestingly, their two best-fitting models showed comparable support, but differed dramatically from one another: one incorporates migration and the other does not (with implications for determining divergence times). Consequently, the authors used simulation to demonstrate that migration or incomplete lineage sorting probably best explained their results. Carstens and colleagues suggest that model selection is useful when historical data are not available to generate hypotheses. However, practitioners in other disciplines argue against deploying model selection without first justifying a biologically plausible set of competing models. Failure to do so amplifies the risk of favoring trivial hypotheses, especially when all possible (rather than plausible) models are considered—an activity some have referred to as a form of data dredging (Burnham & Anderson 2002).
The use of model selection in phylogeography is promising and adds significantly to the phylogeographer’s toolbox. Model selection provides a clean and efficient way to evaluate multiple competing models. It also provides a mechanism to estimate the values of parameters that compose these models, in the case of Carstens and colleagues these include divergence time, migration rates, and genetic diversity. Curiously, in this study, the authors did not find good discrimination between alternative models that differed greatly in their associated migration rate parameters. Models with poor fit to the data (those rejected without the Bonferroni correction) had equal levels of ancestral to current genetic diversity, whereas those models with good fit to the data showed an increased level of current genetic diversity compared to ancestral diversity. However, the models examined here treat the multi-locus data as if they have the same history (e.g. providing a single genetic diversity estimate), yet we know that mtDNA and nuclear genes can have very different histories (e.g. Shaw 2002) and averaging these histories might not be advisable under such circumstances. Thus, a limitation of the model selection approach as currently employed here is that the models remain narrowly defined—each model included in the candidate set assumes only two populations, no recombination, constant migration, and constant population sizes (albeit a different ancestral population size). Other phylogeographic methods allow for more flexibility, including divergence time estimation and fluctuating population sizes (e.g. BEAST Drummond et al. 2005) or partitioning of historical impacting events across time and space (e.g. NCPA Templeton 2008). Hence, this study exposes a need for the candidate set of models to reflect the full spectrum of evolutionary processes that the phylogeographer might want to consider.
The application of model selection in phylogeography will inevitably stir discussion about the appropriateness of different analytical approaches. The ongoing philosophical debate about the appropriateness of post hoc inferences vs. a priori null hypotheses might be supplanted by adopting a model selection framework. For example, a strength of post hoc inference is that these approaches generate testable hypotheses that can be used to generate a candidate set of competing models. A priori null hypotheses are valuable in experimental settings, but are often of more limited use in historical studies. For example, researchers often suppose that rejecting a null hypothesis (usually one without any biological meaning) lends credence to its alternative, when in fact rejecting the null tells us nothing about the relative support for the alternative. What we really want to know is the degree of confidence that we can place in a particular hypothesis, which is exactly the strength of the model selection framework.
Carstens and colleagues deserve praise for moving model selection further into mainstream phylogeography. The challenge now is for researchers to build upon this promising start. We still lack a cohesive framework for model construction in phylogeography. Converting verbal hypotheses of migration, changes in population size over time, local extinction, recolonization, and all other potentially important demographic processes into mathematical models remains a formidable task. Equally important is the implementation of these mathematical models into useful software programs. However, once these challenges are met we anticipate an important paradigm shift will occur in phylogeographic analysis, resulting in more complete understanding of the influence of past events on current patterns of genetic diversity. Hence, although the application of model selection by Carstens and colleagues is limited in scope, it offers a clear framework for moving forward. This is an exciting advance that deserves attention.