Levels of analysis are the different scales or vantage points from which you can study the same phenomenon, and the central insight behind the idea is that an explanation that works perfectly at one scale can be incomplete or even misleading at another. A biologist studying aggression in animals might explain it through hormones, through evolutionary advantage, or through the social structure of a group. Each answer is correct, but none alone captures the full picture. The concept shows up across nearly every scientific discipline, from neuroscience to economics to ecology, and getting it wrong leads to some of the most persistent errors in research and everyday reasoning alike.
Two Frameworks That Launched the Idea
The two most influential versions of the concept come from very different fields but share a common structure. In cognitive science, the vision scientist David Marr proposed in the early 1980s that any information-processing system should be understood at three distinct levels: what problem the system is solving (the computational level), what step-by-step procedure it uses to solve it (the algorithmic level), and what physical machinery carries out that procedure (the implementation level). Marr’s framework was designed for understanding biological vision, but it has proven durable enough that researchers now apply it to artificial intelligence systems like large language models, arguing that the same confusion between “what it does” and “how it does it” plagues AI interpretability.1arXiv. Levels of Analysis for Large Language Models
In biology, the ethologist Niko Tinbergen published his own framework in 1963, identifying four questions you can ask about any animal behavior: how does it develop across the lifespan, what mechanism produces it, how and why did it evolve, and what adaptive function does it serve? Tinbergen emphasized that these levels are complementary, not competing, and that treating an answer at one level as if it refutes an answer at another level leads to pointless arguments.2PubMed. Leveling with Tinbergen: Four levels simplified to causes and consequences A territorial bird sings because of a hormonal trigger (mechanism), because it learned the song as a juvenile (development), because song helped ancestors attract mates (evolution), and because it currently increases reproductive success (function). Asking “which one is the real reason?” misunderstands what explanations are for.
Why One Level Cannot Simply Replace Another
A natural instinct when confronted with multiple levels is to assume one of them is fundamental and the others are just approximations. Physicists sometimes argue that everything is, in principle, reducible to particle interactions. Economists sometimes insist that macroeconomic patterns are just individual choices added together. This impulse toward reduction is powerful but runs into real limits.
The core problem is emergence: macroscopic properties that cannot be fully explained by the behavior of individual components. When billions of water molecules interact, you get fluid dynamics, surface tension, and weather systems. No amount of detail about a single water molecule predicts a hurricane. In complex systems research, emergence refers to this phenomenon where macro-level properties are not solely attributable to individual-level causes, and where new causal patterns arise at higher levels of abstraction.3PubMed Central. Emergence and Causality in Complex Systems: A Survey of Causal Emergence and Related Quantitative Studies
This is not just a philosophical curiosity. Philosophers of science have argued that explanations from different levels can sometimes have incompatible implications. Two true explanations of the same fact, drawn from different research approaches, may fail to combine into a single unified explanation because their counterfactual consequences clash. A neurochemical explanation of depression and a social-environmental explanation of depression may both be accurate, yet merging them into one account distorts both.4PubMed Central. Explanatory Pluralism and the (Dis)Unity of Science: The Argument from Incompatible Counterfactual Consequences The practical upshot is that while interdisciplinary work has clear value, pushing too hard toward a single integrated framework can actually destroy the insights that each separate approach provides.
Downward causation adds another wrinkle. The higher level does not just passively emerge from the lower level; it feeds back and shapes it. The structure of a whole system constrains and transforms what its parts do. This is not mysterious: the rules of a basketball game constrain how individual players move, even though the game is made up of those individual players. In more technical domains, some researchers argue that downward causation, understood as the effect of second-order structural relations on lower-level components, implies that complete micro-level reduction is impossible in principle, even though partial explanations across levels remain useful.5History, Philosophy and Theory of the Life Sciences. Emergence, Downward Causation, and Interlevel Integrative Explanations
The Fallacies of Jumping Between Levels
Mixing up levels of analysis does not just cause theoretical confusion; it produces concrete errors with real consequences. The most famous of these is the ecological fallacy: assuming that what is true of a group must be true of the individuals within it. If a region with higher average income also has higher rates of a certain disease, you cannot conclude that wealthy individuals in that region are the ones getting sick. The pattern at the group level may completely reverse at the individual level.
Less well known but equally dangerous is the reverse mistake, the atomistic fallacy. This occurs when researchers use individual-level data to draw conclusions about national or group-level outcomes. Political scientists have flagged that while the field has become skilled at avoiding the ecological fallacy thanks to better individual-level data, limited attention has gone to the risk that individual-level analyses may themselves produce misleading inferences about higher-level outcomes like election results or policy effects.6European Journal of Political Research. The (forgotten) atomistic fallacy in political science and its implications for how we interpret elections You can know exactly how each voter feels and still mispredict the election if you ignore how those voters are distributed geographically, how district boundaries interact with turnout patterns, or how media environments differ across regions.
These fallacies are not limited to political science. In public health, concluding that a neighborhood is healthy because the average individual health metrics look fine ignores community-level factors like pollution exposure, access to parks, or food deserts. In business, assuming that a company will perform well because each employee is talented ignores organizational culture, communication failures, and leadership dynamics. The error is always the same: treating one level’s data as if it transparently translates to another level.
Ecology and the Scaling Problem
Ecology is a field where levels of analysis are not just a conceptual tool but a daily practical problem. An ecologist might study a single organism’s physiology, a population’s growth rate, a community’s species diversity, or an entire ecosystem’s nutrient cycling. Each of these is a legitimate level with its own patterns, and the relationships between levels are often loose rather than tight.
Hierarchy theory offers a way to make sense of this. It suggests that ecological systems are “nearly decomposable” systems, meaning the coupling between levels in both structure and function is loose enough that you can simplify by studying each level somewhat independently, using what researchers call time-space decomposition.7Canadian Journal of Remote Sensing. Hierarchy and scaling: Extrapolating information along a scaling ladder Processes at a given scale operate at characteristic speeds: cellular processes in fractions of a second, individual organism behavior over hours or days, population dynamics over years, and landscape changes over centuries. Because these timescales are well separated, you can study each level without tracking every detail of the levels above and below.
The danger comes when you try to extrapolate findings across those scales without accounting for the loose coupling. A physiological stress response measured in a lab animal does not straightforwardly predict population decline in the wild, because the population level introduces new variables like migration, competition, and habitat availability that simply do not exist at the individual level. Getting the scaling wrong leads to wildly inaccurate predictions, which matters when you are making conservation or land-management decisions.
Multilevel Selection in Evolution
One of the most contentious applications of levels of analysis has been the debate over whether natural selection acts only on individuals (or genes) or also on groups. Multilevel selection theory proposes that selection can operate at multiple levels simultaneously: genes, individuals, groups, and potentially even larger units.8PubMed Central. Multilevel Selection in the Margins: A Review of Its Representation in Undergraduate Biology Textbooks A behavior that is costly to an individual but benefits the group can spread if between-group selection is strong enough to overpower within-group selection against it.
This idea has been controversial for decades, partly because of genuine scientific disagreements and partly because of level-of-analysis confusion. Critics have argued that group-level traits are just aggregations of individual traits, making group selection unnecessary as a concept. Proponents counter that group-level properties can be genuinely emergent and exert their own selection pressures. A genetical approach to multilevel selection has shown that even when group traits can be decomposed into individual contributions, the decomposition does not make the group-level dynamics disappear.9PubMed Central. The genetical theory of multilevel selection
Empirical work has added weight to the multilevel picture. In studies of wild animal populations, researchers using contextual analysis have found that the social structure of a group can exert selection pressures on individual fitness that are just as strong as, or stronger than, the effects of individual social behavior alone.10Proceedings of the Royal Society B: Biological Sciences. Multilevel selection on individual and group social behaviour in the wild In other words, who you are surrounded by can shape your evolutionary trajectory as much as what you personally do. That finding only makes sense if you are willing to take the group level seriously as a unit of analysis rather than dissolving it into individual-level data.
The Brain as a Multilevel System
Neuroscience may be the field where levels of analysis create the most everyday confusion. You can study the brain at the level of individual ion channels, single neurons, local circuits, brain regions, large-scale networks, or whole-brain dynamics. Each level has its own methods: molecular biology at the bottom, functional MRI and EEG at the top, electrophysiology and optogenetics somewhere in between.
Network neuroscience has emerged as an approach that tries to work across these levels simultaneously, mapping and modeling the elements and interactions of neural systems from molecules up to social systems.11PubMed Central. Network neuroscience The appeal is that brains are genuinely multiscale networked systems, and understanding any single level in isolation gives a distorted picture. A neurotransmitter imbalance changes not just individual synapses but network-wide communication patterns; a brain lesion in one area can disrupt the function of distant, physically intact regions because it severs network connections.
This multiscale reality has direct consequences for psychiatry. Computational psychiatry has been developing tools that try to bridge cellular and synaptic mechanisms with observable behavior, creating models that connect what happens at the level of molecules to what a patient experiences as a symptom.12PubMed Central. Computational psychiatry The motivation is that decades of treating mental illness at a single level, whether purely pharmacological (targeting molecules) or purely behavioral (targeting symptoms), have produced limited progress. Conditions like schizophrenia or depression likely involve disruptions at multiple levels simultaneously, and effective treatment may require understanding how those disruptions interact across scales rather than fixing any single level in isolation.
The Micro-Macro Gap in Economics
Economics has its own version of the levels-of-analysis problem, and it has generated one of the discipline’s most stubborn puzzles. Microeconomics studies individual consumers and firms; macroeconomics studies national output, inflation, and unemployment. In principle, macro patterns should emerge from micro behavior. In practice, the two often disagree.
A classic example involves labor supply. When economists study how individuals respond to tax changes using detailed personal data, they tend to find small effects: people do not dramatically change how much they work when tax rates shift. But when macroeconomists compare countries with very different tax regimes, they sometimes find much larger differences in total hours worked. This micro-macro gap has been a longstanding puzzle. One multi-country study using cross-sectional data from several decades found that the evidence does not strongly support the view that macro-level elasticities are systematically higher than micro-level ones, suggesting the gap may partly reflect methodological differences in how the two levels are measured rather than a genuine mismatch in economic behavior.13Journal of Public Economics. Estimating labour supply elasticities based on cross-country micro data: A bridge between micro and macro estimates?
Still, the deeper issue persists: aggregating individual behavior does not always reproduce the macro pattern, because macro-level constraints feed back on individual choices in ways that individual-level models do not capture. The price level, the money supply, and institutional rules create a context that shapes individual decisions from the top down, much like the ecological constraint of habitat shapes what individual organisms can do.
AI Interpretability and Borrowed Frameworks
One of the more recent domains to grapple with levels of analysis is artificial intelligence, specifically the challenge of understanding what large language models are actually doing when they produce human-like text. These systems are powerful but opaque: they have billions of parameters, and no one fully understands how those parameters combine to produce coherent answers to questions, generate creative writing, or occasionally fabricate facts.
Researchers have argued that this problem is structurally similar to the historical challenge of understanding the human mind, and that cognitive science’s toolkit, organized around Marr’s three levels, can help. At the computational level, you ask what problem the model is solving. At the algorithmic level, you ask what representations and procedures it uses internally. At the implementation level, you look at the actual weights and activations in the neural network.14arXiv. Levels of Analysis for Large Language Models The same analytical tools developed for biological neural systems, organized according to Marr’s levels, have been adapted to study artificial ones.15arXiv. Multilevel Interpretability Of Artificial Neural Networks: Leveraging Framework And Methods From Neuroscience
Much of the confusion in the public conversation about AI stems from level-mixing. When people debate whether a chatbot “understands” language, they are often conflating the computational level (it produces appropriate responses) with the algorithmic level (does it represent meaning the way humans do?) or the implementation level (its architecture is nothing like a brain). Keeping the levels separate does not resolve the debate, but it clarifies what exactly is being argued about, which turns out to be half the battle.
Health Disparities and the Need for Multi-Scale Models
Public health is a field where the practical cost of ignoring levels of analysis is measured in lives. Social determinants of health, including factors like poverty, racism, neighborhood safety, and access to care, operate at scales ranging from the societal and community levels down to the cellular and molecular levels.16PLoS Computational Biology. How multi-scale modeling can help examine social determinants of health and resulting disparities Chronic stress from living in a high-crime neighborhood does not just affect your mood; it alters cortisol regulation, inflammatory markers, and gene expression. A full account of why certain populations get sicker and die younger requires connecting those biological changes to the community and societal forces that produce them.
Traditional approaches have tended to pick a level and stay there. Biomedical research focuses on molecular and cellular mechanisms. Social epidemiology focuses on population-level patterns. The emerging push for multi-scale computational models aims to connect these, simulating how a policy change (macro level) ripples through community structures (meso level) to alter individual stress responses (micro level) and ultimately cellular processes (molecular level). The ambition is enormous and the technical challenges are steep, but the recognition that no single level tells the whole story is itself a significant shift in how health research is conducted.
When Levels of Analysis Collide in Everyday Thinking
You do not need to be a scientist to stumble over level-of-analysis errors. They show up constantly in how people think about causes and blame. When a corporation pollutes a river, you can analyze the problem at the level of individual executives making bad decisions, at the level of corporate incentive structures that reward cost-cutting over safety, at the level of regulatory frameworks that fail to enforce standards, or at the level of an economic system that externalizes environmental costs. Each of these is a legitimate level of analysis, and each points to a different intervention. Blaming only individuals misses systemic causes; blaming only the system ignores that specific people made specific choices.
The same pattern appears in education. A student’s poor performance can be analyzed at the individual level (study habits, motivation), the classroom level (teaching quality, peer effects), the school level (resources, curriculum), or the societal level (funding inequities, cultural attitudes toward education). Interventions at one level may accomplish nothing if the binding constraint is at another level. Tutoring a student who is hungry because the school has no lunch program is working at the wrong level of analysis.
Dynamic systems theory offers a different critique of rigid level-separation. Rather than treating levels as static layers, some cognitive scientists argue that cognition should be understood as a continuous, time-dependent interaction between brain, body, and environment, where the boundaries between levels are not fixed but emerge from the dynamics of the system itself.17PubMed Central. Grounding cognitive-level processes in behavior: the view from dynamic systems theory On this view, Marr’s clean separation into three levels may be too tidy for systems where the implementation constantly reshapes the algorithm, and the algorithm constantly reshapes the computational goal. The levels do not sit neatly on top of each other; they bleed into one another in real time. This does not mean the concept of levels is useless, but it does mean that any particular carving of a system into levels is a modeling choice, not a fact discovered about the world. And the best modeling choice depends on what question you are trying to answer.