Wednesday, May 20th, 2026
Here is a pattern that will be familiar to anyone who has worked with marine data. Sea surface temperatures rise in a coastal region. A few weeks later, a phytoplankton bloom appears. The two variables are correlated — beautifully, in fact, with a tight temporal lag and a high r-squared value. So warming causes the bloom, right?
Maybe. But consider the alternative. Dense phytoplankton blooms absorb solar radiation and can raise local surface temperatures by 0.5–1.5°C. The bloom might be amplifying the very warming signal you think is causing it. Or both variables might be responding independently to a third factor, say, a shift in wind-driven upwelling that simultaneously changes the thermal structure of the water column and delivers nutrients to the surface.
This is not a contrived example. It is an ordinary Tuesday in marine data analysis. And it illustrates a problem that sits at the heart of environmental science: the difference between correlation and causation.
Most of the statistical tools that marine researchers use daily, regression, principal component analysis, time-series decomposition, are designed to identify associations. They answer questions of the form: when X changes, does Y tend to change as well? These are useful questions. But they are not the questions that managers, policymakers, and conservation practitioners actually need answered.
The questions that matter for decision-making are causal. Will reducing nitrogen runoff decrease the frequency of harmful algal blooms? If we establish a marine protected area here, will fish stocks recover? If ocean temperatures rise by 2°C, what happens to this coral reef? These are questions about interventions, what happens if we do something, and correlation alone cannot answer them.
The reason is straightforward. A correlation between two variables can arise for at least three reasons: X causes Y, Y causes X, or some third variable Z causes both. Without additional information or assumptions, the data cannot tell you which of these is true. This is not a limitation of sample size or statistical power. It is a logical constraint. You can have a million data points and still not know whether the relationship is causal.
Marine datasets are especially vulnerable to confounding, the situation where an unobserved or unaccounted-for variable drives the apparent relationship between two others.
Consider the well-documented correlation between declining sea ice extent in the Arctic and increasing populations of certain sub-Arctic fish species moving northward. It is tempting to draw a direct causal arrow from ice loss to range expansion. But the picture is more complicated. Warming itself changes prey availability, alters current patterns, and affects reproduction, all independently of ice cover. Ice loss is also correlated with changes in light availability, which restructures primary production. The fish may be responding to any combination of these factors, and the relative importance of each is difficult to disentangle from observational data alone.
The problem is compounded by the fact that ocean variables tend to covary. Temperature, salinity, nutrient concentrations, light levels, and current velocities are all physically connected. When everything moves together, isolating the effect of any single variable becomes extremely difficult. Atmospheric scientists face a similar challenge, but marine ecologists have it worse because they are tracking living organisms whose responses are mediated by behaviour, physiology, and evolutionary history, none of which can be derived from first principles.
Causal inference is a family of methods designed to move beyond association and make defensible statements about cause and effect. The field has deep roots in statistics, philosophy, and computer science, and it draws on several distinct traditions, most notably the potential outcomes framework developed by Donald Rubin and the structural causal models formalised by Judea Pearl.
The core idea, stripped of technical detail, is this: to claim that X causes Y, you need to reason about what would happen to Y if you were to intervene on X, that is, change X while holding everything else fixed. This is what a randomised controlled experiment does. You randomly assign treatments, which breaks the link between the treatment and any confounders, and then compare outcomes.
In marine science, true experiments are sometimes possible. You can manipulate nutrient levels in mesocosms, transplant corals between sites, or exclude predators from patches of reef. But most of the big questions, the ones about climate impacts, large-scale management interventions, and long-term ecosystem trajectories, cannot be experimentally tested. You cannot randomly assign ocean warming to half the world’s coastlines and withhold it from the other half.
This is where observational causal inference comes in. The methods vary, instrumental variables, difference-in-differences, regression discontinuity, directed acyclic graphs, but they share a common logic. Each method relies on explicit assumptions about the structure of causal relationships in the system, and uses those assumptions to extract causal conclusions from non-experimental data. The assumptions must be stated, justified, and scrutinised. They are not hidden in the model; they are the model.
One of the most practical tools in causal inference is the directed acyclic graph, or DAG. A DAG is a diagram showing the variables in your system and the causal relationships you believe exist between them. Arrows point from causes to effects. The absence of an arrow is itself a claim, it says you believe there is no direct causal link between those two variables.
Drawing a DAG before you run a regression forces you to be explicit about your assumptions. It tells you which variables you need to control for, which you should not control for (because doing so would introduce bias), and which confounders you cannot address with your available data. This last point is particularly valuable. A DAG can reveal that the causal question you want to answer is not answerable with the data you have, a disappointing but genuinely useful finding that can save months of misdirected analysis.
For marine scientists accustomed to fitting models and interpreting coefficients, this represents a shift in practice. It asks you to think about the data-generating process, the real-world mechanisms that produced your observations, before you think about the statistical model. The statistics come second. The science comes first.
None of this implies that correlational studies are worthless. Identifying strong associations is often the first step towards understanding a system, and descriptive work remains the foundation of ecology. But the gap between “these two things are associated” and “changing one of these things will change the other” is large, and policy decisions that treat correlations as if they were causal can go badly wrong.
Fisheries management offers sobering examples. Stock-recruitment relationships, the correlation between spawning biomass and subsequent recruitment, have been used to set harvest levels for decades. Yet the causal link between the two is often weak, mediated by environmental conditions that vary from year to year. Managing a fishery on the assumption that more spawners always means more recruits is a correlational inference dressed up as a causal one, and it has contributed to stock collapses.
Causal inference methods will not solve all of these problems. They require strong assumptions, domain expertise to specify those assumptions correctly, and careful sensitivity analysis to understand what happens when the assumptions are wrong. But they make the reasoning transparent. They force researchers and decision-makers to articulate what they believe about how the system works and to confront the implications of those beliefs.
For those interested in the technical machinery, Bayesian networks, the do-calculus, graph neural networks applied to causal discovery, the EcoTwin project blog has covered these topics in earlier posts. This piece has deliberately stayed upstream of the mathematics because the conceptual foundation matters more than the formalism. If you understand why causation is different from correlation, and why that difference matters for decisions, you have the hardest part already.
The ocean is not going to become a simpler system. The data will remain noisy, confounded, and incomplete. But the questions we ask of that data can become sharper, and that starts with being honest about what our analyses can and cannot tell us.