HiSSE Model

How do new traits, or key innovations, shape the evolutionary trajectory of a species? This fundamental question in evolutionary biology seeks to understand the link between a lineage's features and its success, measured by speciation and extinction rates. For years, scientists have relied on compelling narratives and simple statistical models to connect traits to diversification, but this approach has been fraught with challenges.

A significant problem emerged with early methods like the Binary State Speciation and Extinction (BiSSE) model, which were found to have a high rate of false positives. These models often attributed diversification shifts to an observed trait when the true cause was an unmeasured, confounding variable correlated with that trait—a "ghost in the phylogeny." This created a critical need for a more nuanced framework that could distinguish true causation from mere correlation.

This article explores the Hidden State Speciation and Extinction (HiSSE) model, a sophisticated solution to this problem. In the following chapters, you will learn how HiSSE works and how it provides a more rigorous test of macroevolutionary hypotheses. The "Principles and Mechanisms" section will delve into the statistical underpinnings of HiSSE, explaining how it incorporates hidden states to create a fair comparison against null hypotheses. Subsequently, the "Applications and Interdisciplinary Connections" chapter will showcase how this powerful tool is used in practice to investigate key innovations, reconstruct evolutionary history, and forge connections across scientific disciplines.

To understand the great tapestry of life, we are driven by a simple, yet profound, question: does having a particular feature—a key innovation—change a group’s evolutionary destiny? Does growing wings, developing venom, or inventing a new way to photosynthesize put a lineage on the fast track to evolutionary success? For a long time, telling a compelling story was the best we could do. But science demands more; it demands a way to test these stories. The models we will explore here, particularly the HiSSE model, represent a sophisticated attempt to bring statistical rigor to these grand evolutionary questions. They provide a lens to peer into the past, but as we shall see, it is a lens that can create illusions as easily as it reveals truths, and understanding its properties is the key to telling the difference.

Imagine you are the CEO of a vast, branching corporation called "Life on Earth." Your corporation is composed of many independent companies, which we call ​​lineages​​. Every so often, a company might spin off a new, independent subsidiary; this is ​​speciation​​. At the same time, companies can go bankrupt and disappear; this is ​​extinction​​. Your job, as an evolutionary biologist, is to figure out what business strategies lead to success (high speciation, low extinction).

A business strategy might be a particular trait—for instance, some companies are "terrestrial" (trait state 0) while others are "aquatic" (trait state 1). We can build a simple model to see if this trait affects their fate. This is the essence of a ​​State-dependent Speciation and Extinction (SSE)​​ model. The most basic version for a two-state trait is called the ​​Binary State Speciation and Extinction (BiSSE)​​ model.

The BiSSE model is wonderfully straightforward. It assumes that each lineage with a given trait state has its own rates of "spin-offs" and "bankruptcies."

• Terrestrial lineages speciate at a rate $\lambda_0$ and go extinct at a rate $\mu_0$.
• Aquatic lineages speciate at a rate $\lambda_1$ and go extinct at a rate $\mu_1$.

Furthermore, companies can change their strategy over time. A terrestrial company might move into the water, changing from state 0 to 1 at a rate $q_{01}$, and an aquatic one might move onto land at a rate $q_{10}$. With a phylogenetic tree (the corporate family tree) and the traits of the currently existing companies (the tips of the tree), the BiSSE model uses the power of probability to estimate these six key parameters: $\lambda_0, \lambda_1, \mu_0, \mu_1, q_{01}, q_{10}$. If the model confidently concludes that $\lambda_1 > \lambda_0$, for example, we might be tempted to declare that being aquatic is a ​​key innovation​​ that drives diversification.

This simple picture is incredibly appealing. Unfortunately, it is also dangerously naive. For years, scientists using BiSSE and similar models found exciting correlations all over the tree of life. Yet, a nagging suspicion grew, culminating in a series of crucial studies that revealed a ghost in the machine: BiSSE has a tendency to find false positives, sometimes at an alarming rate.

Why does this happen? The problem is a classic villain in all of science: the ​​confounding variable​​. BiSSE is a simple-minded detective. It only knows about the trait it was told to investigate. It is blind to any other factor that might be influencing success.

Let's return to our corporate analogy, but make it more biological. Imagine we are studying plants, and our observed trait is flower color: blue ($X=0$) versus red ($X=1$). We notice that the red-flowered lineages seem to be diversifying much more rapidly. BiSSE, analyzing this pattern, would strongly support the conclusion that red flowers are a key innovation ($\lambda_1 > \lambda_0$).

But what if there's a hidden factor? Suppose that the real driver of diversification is habitat. Some plants live in the lowlands (hidden state $A$) and others live in a newly formed, geographically complex mountain range (hidden state $B$). The mountains offer countless opportunities for isolation and adaptation, leading to a much higher speciation rate ($\lambda_B \gg \lambda_A$). Now, suppose that by historical accident, the ancestor of the plants that colonized the mountains happened to have red flowers. As this lineage diversified into a spectacular array of new species, they all carried their ancestral red flowers with them.

What does our BiSSE detective see? It sees a large, rapidly diversifying clade of red-flowered plants. Since it is blind to the mountain habitat, it makes the only inference it can: the red color must be the cause of the success. The model has confused correlation with causation. The observed trait ($X$) had no causal effect on diversification, but because it was correlated with the true, unmeasured cause ($H$), BiSSE was duped. This problem is especially severe when the trait has evolved only a few times. If the "red flower" trait only appeared once on the tree, we really only have a single data point—not a strong basis for a general rule about all red flowers.

To make a better inference, we need a smarter detective—one that is aware of the possibility of ghosts. This is the conceptual leap of the ​​Hidden State Speciation and Extinction (HiSSE)​​ model.

HiSSE works by explicitly including the possibility of an unobserved, or "hidden," factor that influences diversification. It expands the simple state space of BiSSE. A lineage is no longer just "red-flowered" or "blue-flowered." It is now described by a composite state that includes both the observed trait and the hidden state. For example, with two hidden states $A$ and $B$ (e.g., lowland vs. mountain), our model now has four possible states for a lineage:

• Blue-flowered, Lowland ($0A$)
• Blue-flowered, Mountain ($0B$)
• Red-flowered, Lowland ($1A$)
• Red-flowered, Mountain ($1B$)

Each of these four composite states is allowed to have its own speciation rate ($\lambda_{0A}, \lambda_{0B}, \lambda_{1A}, \lambda_{1B}$) and extinction rate ($\mu_{0A}, \mu_{0B}, \mu_{1A}, \mu_{1B}$). The model also includes rates for transitioning between these states (e.g., a blue-flowered plant evolving red flowers, or a lowland lineage moving into the mountains). Because the hidden state is, by definition, unobserved, the model uses a clever mathematical trick: it calculates the likelihood of the tree by summing, or ​​marginalizing​​, over all possible histories of the hidden states. This is like our new detective considering all possible scenarios involving the hidden factor to explain the observed evidence.

Simply adding more parameters isn't the solution. The true genius of the HiSSE framework is that it allows us to set up a fair contest between competing hypotheses. The central question is no longer "are the diversification rates different between red and blue flowers?" but rather, "does flower color explain diversification rates better than a model where some other, unobserved factor is at play?"

To formalize this, we construct a special type of HiSSE model called a ​​Character-Independent Diversification (CID)​​ model. In this model, we enforce the rule that the diversification rate depends only on the hidden state, not the observed one. For our example, we would impose the following constraints:

• The speciation rate for any plant in the lowlands is $\lambda_A$, regardless of its flower color ($\lambda_{0A} = \lambda_{1A} = \lambda_A$).
• The speciation rate for any plant in the mountains is $\lambda_B$, regardless of its flower color ($\lambda_{0B} = \lambda_{1B} = \lambda_B$).
• (And similarly for extinction rates $\mu$).

This CID model represents our "ghost in the machine" hypothesis. It allows the diversification rate to vary across the tree (since we can have $\lambda_A \neq \lambda_B$), but it stipulates that this variation is completely independent of flower color.

The test is now a direct comparison. We fit several models to our data:

1. The original ​​BiSSE​​ model (which has no hidden states).
2. The ​​CID​​ model (where only the hidden state matters).
3. The full ​​HiSSE​​ model (where both the observed and hidden states could matter).

Let's imagine we do this for a real dataset, as in the scenario from problem. We compare the models using a metric like the ​​Akaike Information Criterion (AIC)​​, which rewards models that fit the data well but penalizes them for having too many parameters. Suppose we get the following (corrected AIC, or AICc, scores):

• $AICc_{BiSSE} = 1053.3$ ($k=6$ parameters)
• $AICc_{HiSSE} = 1046.0$ ($k=10$ parameters)
• $AICc_{CID-2} = 1044.3$ ($k=8$ parameters)

The model with the lowest AICc score is the preferred one. Here, the CID-2 model wins. What does this tell us? It says that a model with two hidden rate regimes that are unrelated to our observed trait provides the most parsimonious explanation for the patterns in our phylogenetic tree. The BiSSE model, with its much higher AICc, is a very poor explanation. Even the full HiSSE model doesn't fit as well as the simpler CID model. Our conclusion: we have found no evidence that the trait is a key innovation. The diversification heterogeneity is real, but it is driven by a "ghost"—an unobserved factor that HiSSE helped us detect and account for.

The HiSSE framework is a powerful tool for avoiding foolish conclusions, but it's not a magical crystal ball. It requires scientific humility and a clear understanding of its own limitations.

First, the hidden states are just labels. The model might tell us there are two rate regimes, 'A' and 'B', but it can't tell us what they are. It could be mountains vs. lowlands, wet vs. dry climates, or some internal physiological factor we haven't even thought of. Because the labels are arbitrary—we can swap 'A' and 'B' everywhere in the model and the fit will be identical—we cannot interpret the states themselves. We can only interpret the process: for example, "the data support a model with two distinct rate regimes, but these regimes are not correlated with the observed trait.".

Second, HiSSE does not solve all of the inherent difficulties of peering into the deep past. A fundamental mathematical property of these models is that, when you only have data from living species (an extant tree), it is extremely difficult to separately identify the birth rate ($\lambda$) and the death rate ($\mu$). An infinite number of different time-varying speciation and extinction histories can produce the exact same tree of survivors. What we can estimate more reliably are combinations like the net diversification rate ($r = \lambda - \mu$). So, while HiSSE allows for more robust comparisons between hypotheses, we should remain cautious about the absolute values of any single estimated rate, especially extinction.

Finally, how do scientists build even greater confidence? They can turn the model on itself. Using a technique called ​​posterior predictive simulation​​, a researcher can take their best-fitting CID (null) model and use it to simulate hundreds of new trait datasets on their real phylogeny. They then analyze each fake dataset with BiSSE. This shows how often, under a "true" world of trait-independent diversification, a spurious correlation would be detected. If BiSSE frequently finds a link when we know there isn't one, it gives a stark measure of how likely it is that our original, exciting result was just a statistical phantom.

This journey, from a simple idea to a complex and nuanced statistical toolkit, is the hallmark of scientific progress. It is a story of building tools to match our ambitions, and then learning the humility to understand what those tools can and cannot show us. HiSSE doesn't give us final answers, but it teaches us how to ask much better questions.

Now that we have explored the intricate machinery of the Hidden-State Speciation and Extinction (HiSSE) model, let us embark on a journey to see it in action. The principles and mechanisms we've discussed are not mere theoretical abstractions; they are powerful lenses through which we can view the grand tapestry of life's history with newfound clarity. The true beauty of a scientific tool lies in its application, in the new questions it allows us to ask and the old puzzles it helps us solve. From the venom of a snake to the intricate dance between a flower and its pollinator, HiSSE and its conceptual cousins allow us to move beyond simple observation and begin to untangle the complex causal web that has shaped the world's biodiversity.

Perhaps the most common and compelling use of the HiSSE framework is in playing the role of a great detective, investigating the long-standing mystery of "key innovations." A key innovation is a trait that is thought to have opened up new ecological opportunities, sparking a so-called adaptive radiation—a rapid burst of speciation. For decades, biologists have pointed to species-rich groups and nominated a "key" trait as the cause. Cichlid fishes in the African Great Lakes, with their specialized pharyngeal jaws, seem to have radiated into hundreds of species filling countless ecological niches. Many plant lineages that underwent whole-genome duplication (WGD) appear to be more diverse than their relatives, suggesting WGD might be a powerful evolutionary catalyst. The evolution of venom in reptiles or elaborate horns in scarab beetles used for male-male combat are also frequently associated with impressive radiations.

But as any good detective knows, correlation is not causation. Is the trait the hero of the story, or merely a bystander who happened to be at the right place at the right time? This is the crucial question. An unmeasured factor—a change in climate, the colonization of a new continent, a shift in body size—might be the true cause of the radiation, and the "key innovation" might just be along for the ride.

This is where HiSSE's central logic comes into play. The analysis becomes a form of statistical cross-examination. We fit a full HiSSE model, which proposes that diversification rates (speciation $\lambda$ and extinction $\mu$) depend on both the observed trait (e.g., venom present/absent) and a hidden, unobserved factor. Then, we compare this to a special kind of null model, the Character-Independent Diversification (CID) model. The CID model is a clever construct: it has the exact same complexity and hidden states as the HiSSE model, but with one critical constraint—the diversification rates depend only on the hidden state, not the observed trait.

The comparison, typically using a metric like the Akaike Information Criterion (AIC), is like asking the data: "Can you explain the patterns of diversification on this tree just as well by invoking some hidden, unmeasured factors, without giving any special role to the trait I'm interested in?" If the CID model fits the data just as well as or better than the full HiSSE model, the case for the key innovation is severely weakened. It suggests the observed correlation was likely spurious. However, if the full HiSSE model provides a substantially better explanation, and the estimated rates consistently show that the trait is associated with higher diversification across all hidden background states, then we have strong evidence that the trait itself is a genuine driver of macroevolutionary dynamics.

This framework is not limited to a single confounding factor. Modern extensions of HiSSE can explicitly incorporate other measured variables. For instance, when investigating whether beetle horns drive diversification, we might suspect that body size and habitat are also important. Advanced models allow us to build these covariates directly into the rate equations, letting us ask if the horns have an effect after we've already accounted for the influence of size and ecology. Similarly, when testing the role of pharyngeal jaws in cichlids, we can simultaneously model the effect of living in a lake versus a river, disentangling the key innovation from the ecological opportunity. This integrated approach, which combines testing for the influence of the focal trait, measured confounders, and unmeasured "hidden" factors, represents a powerful and rigorous way to test a complete causal hypothesis about adaptive radiation.

While HiSSE is often used to test hypotheses about diversification rates, its utility doesn't end there. The model also provides a more nuanced way to reconstruct the evolutionary history of the traits themselves. Traditional methods for reconstructing ancestral states—what a long-extinct ancestor was like—often implicitly assume that the trait's evolution is independent of the process of speciation and extinction.

But what if this isn't true? Imagine a trait that dramatically increases the extinction rate. A lineage that gains this trait is less likely to have a long and successful history. Therefore, if we see this trait in two distantly related living species, it's more likely that it evolved independently in both lineages rather than being inherited from a very ancient common ancestor that somehow managed to survive for eons with a high-risk trait.

HiSSE and other state-dependent diversification models naturally account for this. By jointly estimating the parameters of trait evolution and diversification, the model "knows" that certain states are more perilous than others. When reconstructing the past, it weighs these probabilities accordingly. The final reconstruction is not just a simple average; it's a sophisticated inference that has been corrected for the biasing effects of state-dependent birth and death. The real power comes from embracing the full complexity of the HiSSE model. By fitting the model and then "marginalizing" over the hidden states—summing up the probabilities of all hidden possibilities—we can obtain a robust estimate for the probability of the observed trait at each node in the tree. This approach, often combined with model averaging to account for uncertainty in which model is "best," gives us our most reliable picture of the past.

The most exciting applications of a scientific tool are often those that build bridges between different fields. HiSSE is no exception. It serves as a vital component in larger analytical pipelines that connect macroevolution with ecology, genetics, and even large-scale data science.

Consider the intricate coevolutionary dance between plants and their pollinators. A classic hypothesis suggests that the evolution of a key innovation in a plant, like a floral nectar spur, might not only cause the plant lineage to radiate but might also trigger a corresponding radiation in the pollinator guild that learns to exploit this new resource. Testing this "diffuse coevolution" hypothesis requires a multi-pronged approach. First, we could use HiSSE on the plant phylogeny to rigorously test if the nectar spur was indeed a key innovation that elevated plant diversification rates. Then, using time-dependent diversification models on the pollinator phylogeny, we can search for a burst in pollinator diversification that occurs significantly after the nectar spur first appeared. By comparing this to null models where the timing is randomized, we can build a statistical case for a causal link, connecting evolutionary events across kingdoms.

Furthermore, as our ability to generate phylogenetic data grows, we are no longer limited to single case studies. We can now aspire to a grand synthesis. Imagine applying the HiSSE framework to dozens of independent plant and animal clades, each with its own candidate key innovation. How can we synthesize these results to ask if, for example, traits related to reproduction are more potent drivers of diversification in animals than in plants? The answer lies in connecting the world of phylogenetics to the statistical field of meta-analysis. By treating the diversification contrast (e.g., the difference in speciation rate, $\Delta \lambda = \lambda_1 - \lambda_0$) from each HiSSE analysis as a single data point, we can use hierarchical models to estimate the average effect size for plants and animals, while accounting for the fact that some estimates are more certain than others. This allows us to move from storytelling about a single clade to drawing robust, quantitative conclusions about the general rules of evolution.

Looking forward, the principles underlying HiSSE are helping to build even more sophisticated models that test entire mechanistic pathways. Instead of just asking if mycorrhizal associations in plants increase diversification, we can build a causal model that tests the specific pathway: Mycorrhizal type $\to$ Enhanced nutrient uptake rate $\to$ Higher diversification. This links macroevolutionary modeling with physiology and functional ecology in a single, unified statistical framework.

From a single tree to the entire tree of life, the HiSSE framework and the ideas it embodies are transforming our ability to interpret evolutionary history. It pushes us to think more critically about causation, to account for the unseen, and to build bridges between disciplines. It is a tool that helps us appreciate not just the patterns of biodiversity, but the beautiful and complex processes that generated them.