Ontology, Epistemology and Why They Matter in Practice

These two words are the reason many people give up on methods chapters. They sound like the gate-keeping vocabulary of a subject that has decided not to let you in.

They are actually the two most practical questions in research, and every study answers them whether or not it mentions them. The answers determine what gets measured, what counts as a finding, and — most importantly for a reader — what the study is structurally unable to see.

So we will do them the useful way round: the questions first, the words second, and then a single research brief run four times to show exactly where the answers show up.

One brief, one budget, four studies that share almost nothing.

An engineering firm with 4,000 employees wants to know why so few of its senior engineers are women. It has money for one study. Four teams pitch.

Team A proposes to assemble every personnel record for fifteen years: hiring, starting salary, performance ratings, promotion decisions, exits. They will build a model of promotion probability, control for qualifications, experience, department, tenure and rated performance, and report how much of the gap remains unexplained — the part that qualifications do not account for. They will also test whether the gap appears at hiring, at first promotion, or at the senior threshold.

Team B proposes to spend eight months inside two divisions, sitting in design reviews, promotion panels and the informal conversations that precede them. They want to record how competence gets attributed in the moment — who gets interrupted, whose idea is credited when it is repeated by someone else, how "not ready yet" is arrived at and phrased, and what makes an engineer legible as a future lead.

Team C proposes both, but organised differently. They will use the records to establish where in the pipeline the loss happens, then use interviews and observation at exactly that point to identify what is producing it — and then test whether the same thing is visible in the two divisions where the gap is smallest. Their output is a mechanism plus the conditions under which it operates.

Team D proposes to start from the women engineers themselves — not as subjects to be measured but as the vantage point. They will trace, from the everyday experience outward, how the firm's ordinary procedures — the project allocation system, the "high-visibility assignment", the promotion criteria document, the on-call rota — organise that experience from elsewhere.

All four teams are competent. All four would produce something true. The four reports would barely overlap.

And the disagreement is not really about method. It is about what they think "gender inequality" is — a measurable difference in outcomes between two categories of person, a thing accomplished in interaction, a structure with causal powers, or a set of institutional relations organising daily life from outside it.

Two people can agree on every fact and still design incompatible studies. What are they disagreeing about?

They are disagreeing about two things, and it helps enormously to keep them apart.

The first is about the world: what kind of thing is this? Is "gender" a property of individuals? A pattern in interaction? A structure? Is "the middle class" a real group with boundaries, or a name we give to a fuzzy bundle of people? Is "a recession" a thing that happens, or a way of describing a lot of separate things happening? These are questions about what exists — about the furniture of the social world.

The second is about knowledge: what could we possibly find out about it, and how would we know? Is direct observation the only reliable source? Can we know about things we cannot observe, like structures and mechanisms? Does knowing require occupying a particular position? Can we be certain of anything, and if not, what is good enough?

The first is ontology. The second is epistemology. And they constrain each other: if you think the thing is unobservable, you cannot hold that only the observable can be known and still study it.

The four-level ladder

Ontology, epistemology, methodology, methods — and each one narrows the next.

Ontology — what is there? What kinds of thing exist, and what is the nature of the thing you propose to study. Does it exist independently of our beliefs about it? Is it an individual property, a relation, a structure, a process?

Epistemology — what can be known, and what counts as knowing it? What sources of knowledge are legitimate, what makes a claim justified, how much certainty is available.

Methodology — what is the logic of a study that could answer this? Comparison, causal identification, immersion, reconstruction, triangulation. This is the strategy, not the technique.

Methods — what do I actually do on Tuesday? Survey, interview, experiment, archive, observation, coding, regression.

The ladder runs downwards and it is not decorative. If you hold that class is a real structure with causal powers that is not directly observable, then an epistemology permitting only observables is closed to you, a methodology of pure description will not reach it, and methods that record only self-reported identity will miss the thing you said you were studying.

Most confused research is a broken rung. A study announces that gender is socially constructed and accomplished in interaction, and then measures it with a tick-box on a form and treats the tick-box as the thing. The philosophy in the introduction and the operations in the method section are describing different objects.

The main ontological positions, in plain words

Six positions, each of which changes what you would go and look for.

Realism — the thing exists independently of our beliefs about it. There is a rate of unemployment whether or not anyone measures it, and our measurements can be better or worse at capturing it.

Constructionism — the thing exists because of collective human activity: agreements, categories, practices, institutions. Money is the standard example. A banknote's value is not a physical property of the paper. It exists because enough people treat it as existing, and it would stop existing if they stopped.

Materialism and idealism — whether the fundamental furniture is material conditions and relations of production, or ideas, beliefs and meanings (see 1.3.6). Almost nobody holds either purely; the question is what does the explaining when they conflict.

Individualism — ultimately only individuals and their actions exist; talk of "society" or "the system" is shorthand that must be cashable in terms of what people do. Methodological individualism is the research rule that follows: explanations must be built up from individual action.

Holism / emergence — collectives have properties not present in their members, which are therefore not reducible to them (see 4.2.1). Durkheim's whole programme. A traffic jam has properties no car has; a norm persists as everyone who held it is replaced.

Relationism — the fundamental units are neither individuals nor wholes but relations. Bourdieu's slogan was that the real is relational: a "field" is a structure of positions, and a position has its properties only by its difference from the others (see 5.8.1). Elias said the same about figurations.

The single most useful distinction in this entire lesson: constructed does not mean unreal.

This confusion wrecks more discussion than any other, in and out of the academy. "That's a social construct" is used to mean "that's not real, so it doesn't matter" — and it does not follow, at all.

Money is constructed and it will get you evicted. Borders are constructed and people die at them. Caste is constructed and it determined who could draw water from which well (see 6.2.2). Race has no biological basis as a system of discrete natural kinds, and racial classification has organised property, policing, medicine and mortality for four centuries (see 8.4).

Constructed things are made by human activity, and being made by human activity is not a weakness — it is how the most powerful social facts get their power. They are real in the sense that matters most: they have effects, and you cannot escape them by disbelieving in them.

What construction claims do imply is something different and much more interesting: contingency. If a thing was made, it was made by a process, at a time, by particular people, and it could have been made differently or not at all. That is a historical claim, and it is checkable — which is what separates a serious constructionist argument from a slogan. Ian Hacking's advice in The Social Construction of What? (1999) is exactly right: whenever you meet the phrase, ask what, precisely, is claimed to be constructed — the thing, the category, or the experience of it — and what the argument is that it could have been otherwise.

Hacking's other contribution, which sociology should use far more than it does.

He distinguished indifferent kinds from interactive kinds.

An indifferent kind does not know what it is called. Quarks are unaffected by the theory of quarks. Plutonium behaves the same whether or not it has been classified.

An interactive kind is a classification of people, and people react to being classified. They accept the label, resist it, organise around it, perform it, avoid the behaviour that attracts it, or claim the benefits attached to it. So the category changes the people, and the changed people change the category. Hacking called this the looping effect.

Examples: "gifted child", "refugee", "at-risk youth", "ADHD", "middle class", "NEET", "OBC". Each began as a description, was attached to consequences — a service, a stigma, a quota, a treatment — and then reshaped the population it described.

Why this matters for research design: a study of an interactive kind is intervening in what it studies, in a way that a study of rocks is not. And the discipline's own categories are interactive kinds too — 6.1.1's colonial census is the largest example in this course, where categories devised for administration were adopted by the people classified and became the basis of political organisation.

The main epistemological positions, in plain words

How you think knowing works.

Empiricism — knowledge comes from sense experience. What cannot in principle be observed cannot be known. The purest form is the positivism of 7.1.1 and it has largely been abandoned, because science is full of unobservables that we have excellent reasons to believe in.

Rationalism — reason and theoretical argument are independent sources of knowledge; not everything worth knowing is arrived at by looking.

Foundationalism — knowledge rests on a base of certainties, and everything else is built on them. Coherentism — there is no such base; a belief is justified by fitting with everything else we hold, and the whole web is revisable. Most working researchers are coherentists whether they say so or not.

Fallibilism — Peirce's position, and the sane default: every claim we hold might be wrong, and holding it anyway is rational as long as we hold it in a way that lets it be corrected. Not scepticism — scepticism refuses to believe; fallibilism believes provisionally and keeps the door open.

Constructivism about knowledge — knowledge itself is produced socially, in communities, with conventions about what counts. (Distinct from constructionism about the world: you can hold that the object is mind-independent and still hold that our knowledge of it is a social product. That combination is exactly Bhaskar's, from 7.1.1.)

Standpoint epistemology — position shapes what can be seen; starting from subordinated lives yields less partial knowledge (see 7.1.1).

The payoff: one brief, four designs

Now watch the assumptions turn into decisions.

Team A — realist about categories, empiricist about knowledge, individualist in unit.

Gender is an attribute of individuals. Inequality is a difference in outcome distributions between two groups. Knowing means measuring and estimating with stated uncertainty.

So they must define the outcome precisely, find comparable individuals, control for everything that could confound, and quantify what is left. Their finding takes the form: a gap of X, of which Y is attributable to measured characteristics and Z is not.

What they cannot see: what produces Z. The residual is where the interesting thing is, and their design converts it into a number rather than an explanation. They also cannot see anything the records do not contain — and promotion panels do not minute the reasons that matter.

Team B — constructionist ontology, interpretive epistemology, interaction as unit.

Gender is not a property people have but something done — West and Zimmerman's "Doing Gender" (1987) — accomplished continually in interaction, and assessed by others against normative expectations. Inequality is the cumulative product of countless small attributions of competence.

So they must be present where the attributions happen and record them in fine detail. Their finding takes the form: here is the mechanism, in operation, on the following occasions, and here is how participants themselves account for it.

What they cannot see: prevalence. They cannot tell the firm how much of the gap this produces, or whether the divisions they did not observe work the same way.

Team C — critical realist.

Gender is a structure of social relations with causal powers, operating as a tendency in an open system, only some of whose effects surface as measurable events (see 7.1.1). Inequality is what those powers produce when nothing counteracts them.

So they must use the measurable pattern to locate the site, then go beneath it for the generative mechanism, then look at the cases where the outcome differs to find what counteracted it. Their finding takes the form: this mechanism, operating under these conditions, produces this outcome; where condition Q holds, it does not.

What they cannot see — or rather, what they risk: with no requirement to produce a clean regularity, an unfalsifiable story about hidden mechanisms is always available. The discipline of their design is entirely in the contrast cases.

Team D — feminist standpoint, institutional ethnography.

Dorothy Smith's programme. The starting point is the everyday world as actually lived, treated as problematic — that is, as something to be explained rather than described. What organises it are the ruling relations: the texts, forms, categories and procedures through which activity in one place is coordinated from another.

So they must begin with the women's own accounts, then follow outwards to the documents and procedures that shape them — the criteria sheet, the allocation system, the rota — and show how these coordinate what happens without anyone intending it locally. Their finding takes the form: here is how the everyday is put together from elsewhere, and here is the textual machinery that does it.

What they cannot see: they will find it hard to say whether the firm is better or worse than others, and the design does not aim to.

None of the four is the correct one, and none can substitute for another.

Team A's number is what a regulator or a board needs, and no amount of ethnography produces it. Team B's mechanism is what an intervention needs, and no amount of regression produces it. Team C's contrast is what tells you whether the intervention will transfer. Team D's account is what tells you what the criteria document is doing that nobody wrote it to do.

The mistake is not choosing wrongly. The mistake is not knowing you chose — and then claiming what the design cannot deliver, which is what almost every over-reaching study in the world is doing.

And notice how early the choice was made. By the time anyone picked a method, the decisive commitment — what kind of thing is gender inequality — had already been made, usually without discussion, usually by whoever wrote the brief.

Four ways this goes wrong.

One — philosophy as a badge. Researchers announce a position in the introduction as identity, then do whatever their training equipped them to do. The tell is a paragraph on constructionism followed by a design that treats every category as a fixed attribute. A philosophical commitment that changes no decision is decoration.

Two — the paradigm wars. In the 1980s and 1990s a great deal of energy went into arguing that quantitative and qualitative traditions were incommensurable — that they rest on incompatible ontologies and cannot be combined. The overclaim did real damage, discouraging exactly the designs that answer the most questions (see 7.8.3). Kuhn's incommensurability was never meant to license this, and he spent years narrowing it.

Three — ontology smuggled in through measurement. The moment you decide how to measure something you have made a claim about what it is. Measure poverty by household income and you have said poverty is a property of households at a moment. Measure it over ten years and you have said it is a trajectory. Measure it by what people cannot do and you have said it is a capability shortfall (see 7.2.2). Nobody debates this at the operationalisation stage, and it is where the ontology is actually decided.

Four — reification. Treating the measure as the thing. IQ is a score on a test; "intelligence" is not that score. GDP is an accounting convention; "the economy" is not it. Crime statistics are records of institutional processing (see 7.4.4). The measure is a trace of the thing, produced under conditions, and the conditions are part of the finding.

Because it converts a vague unease about a study into a specific question you can actually answer.

You read that a study "found no evidence of discrimination". Instead of believing it or dismissing it, you can now ask a chain of questions with definite answers.

What did they take discrimination to be? An unexplained residual after controls? A difference in treatment of identical applications? A pattern of attributed competence? Each is a different object, and the study measured one.

What follows from that? If it was a residual after controls, then the better their controls, the smaller the discrimination they can find — because any channel through which discrimination operates that they controlled for has been subtracted from the estimate. A study can control discrimination out of existence and report the result as an absence.

What could this design never have shown? If the answer is "the thing being claimed", you have your reading, and you did not need to know anything about the topic to get it.

This is the reader's version of the whole lesson: identify what the study thinks the thing is, and everything else follows.

Ontology is about the thing; epistemology is about the knowing. Everything else in a design descends from them, through methodology (the logic of the study) to methods (what you do on Tuesday). Broken rungs are the commonest defect in research: a constructionist introduction with a tick-box measurement.

Ontological positions: realism (independent of belief), constructionism (made by collective activity), materialism and idealism, individualism versus holism and emergence, and relationism (the real is relations, not substances).

Constructed does not mean unreal. Money, borders, caste and racial classification are constructed and consequential. What construction claims imply is contingency — that it was made, by a process, and could have been otherwise, which is a historical claim and therefore checkable. Hacking's rule: ask what exactly is claimed to be constructed. His second contribution is the looping effect — people react to being classified, so social categories change the populations they name, and the discipline's own categories are among them.

Epistemological positions: empiricism, rationalism, foundationalism versus coherentism, fallibilism (hold everything provisionally and correctably), constructivism about knowledge, and standpoint.

One brief, four designs: gender inequality as an attribute of individuals (measure the gap), as accomplished in interaction (observe the attribution of competence), as a structure with causal powers (locate the mechanism and the counteracting conditions), or as everyday life organised by ruling relations (follow the texts outward). All four are legitimate; none substitutes for another; and the decisive choice was made before anyone mentioned a method.

And the reader's payoff: ask what the study took the thing to be, and what a study built on that assumption could never have shown.

Ontology — the account of what kind of thing exists and is being studied.

Epistemology — the account of what can be known about it and what justifies a claim.

Methodology / methods — the logic of the design, and the techniques used within it.

Realism — the object exists independently of beliefs about it.

Constructionism — the object exists through collective human activity; implies contingency, not unreality.

Interactive kind / looping effect — Hacking: classifications of people change the people classified, who then change the classification.

Indifferent kind — a classification whose objects are unaffected by being classified.

Methodological individualism — the rule that explanations must be built from individual action.

Emergence — the position that collectives have properties not reducible to their members.

Relationism — Bourdieu's and Elias's position that the fundamental units are relations, not substances.

Fallibilism — every claim may be wrong; rational belief is provisional and correctable.

Coherentism — justification comes from fitting with the rest of what we hold, not from a certain foundation.

Doing gender — West and Zimmerman: gender as accomplished continually in interaction and assessed against normative expectations.

Institutional ethnography / ruling relations — Dorothy Smith: starting from everyday experience and tracing outward to the texts and procedures that organise it from elsewhere.

Reification — treating a measure, or an abstraction, as though it were the concrete thing itself.

One — name the object. Take any research finding you have seen reported and write one sentence answering: what did this study think the thing was? An attribute, an interaction, a structure, a process, an institutional relation. Then write what that choice made invisible.

Two — do the four-team exercise yourself. Pick a question you care about — why a neighbourhood changed, why a rule is ignored, why people leave a profession. Sketch four one-paragraph designs on the four ontologies above. You will find one comes easily and the others are hard, and that tells you your own default.

Three — hunt a looping effect. Find a category applied to people that carries consequences — a diagnosis, an eligibility class, a school label. Ask how people respond to being placed in it, and whether those responses have changed what the category picks out.

Four — catch an ontology in a measurement. Take any indicator in the news — a poverty line, an unemployment rate, a literacy figure, a crime statistic. Write down what it implies about what the thing is, then write one alternative measure that implies something different.

Five — apply the controls test. Find a study reporting that a gap "disappears once we control for X". Ask whether X is a confounder or a channel — that is, whether X is a separate cause or one of the ways the thing being studied operates. If it is a channel, the control removed the finding rather than the bias, and the reported absence is an artefact of the design.