Independence begins with information, not multiplication

For random variables X and Y with a joint density, ordinary independence is familiar as p(x,y) = p(x)p(y). Dawid starts there but treats the factorisation as a mathematical expression of a more basic idea: information about Y does not alter uncertainty about X. In conditional form, p(x|y) = p(x). He also allows looser factors, p(x,y) = a(x)b(y) or p(x|y) = a(x), because normalisation recovers the marginal distributions.

The asymmetric wording is useful. It speaks from the point of view of predicting X after seeing Y. Yet for variables with a proper joint distribution, the factorisation proves symmetry: if Y adds nothing about X, X adds nothing about Y. Dawid makes this Theorem 2.1. The theorem matters because it lets the analyst use whichever direction is easier to test or observe. In a retrospective study sampled by disease status, an exposure distribution conditional on disease may be accessible even though disease incidence by exposure is not directly sampled.

Symmetry should not be over-exported. Dawid also wants to speak about a fixed parameter Θ that indexes a family of sampling distributions. Without a prior, Θ is not a random variable and has no joint probability distribution with the data. He therefore uses conditional-distribution language such as S independent of Θ to mean that the sampling distribution of statistic S is the same at every parameter value. This expresses ancillarity, but the ordinary symmetry theorem no longer follows from a joint factorisation because no such joint law was assumed.

Conditioning adds a third place to the relation

Dawid writes X independent of Y given Z when X and Y are independent in their conditional joint distribution at every relevant value z. For densities, the joint conditional factorises:

Equivalently, p(x|y,z) = p(x|z): after Z is known, Y is superfluous for predicting X. The common conditional law may be written a(x,z), emphasising that it varies with Z but not Y. Under proper distributions these forms are equivalent and symmetry follows: X independent of Y given Z implies Y independent of X given Z.

The words “given Z” do real work. Marginal independence may disappear after conditioning, and marginal dependence may disappear after conditioning. Neither direction is paradoxical once the population being compared is kept visible. Conditioning can separate a common-cause path; it can also select a common effect and open a dependence. The symbol is therefore not a monotonic claim that adding information always removes association.

Dawid warns about improper distributions. An equality that resembles p(x|y,z) = p(x|z) may not safely imply the reversed conditional equality if the supposed joint distribution is not proper. The famous marginalisation paradoxes live near this boundary. His main treatment assumes proper distributions so the symmetric calculus is available, but his parameter examples deliberately use a broader, heuristic reading.

A small rule set replaces repeated density algebra

After establishing symmetry, Dawid states three elementary lemmas. In modern set-based language their consequences include decomposition, weak union and contraction. Decomposition permits a function or component to be discarded from an independent block. Weak union permits part of that block to move into the conditioning record. Contraction combines two compatible irrelevance claims into one larger claim.

The logic deserves slow reading. If X is independent of the pair (Y,W) given Z, then X is independent of Y given Z, and X remains independent of Y after W is also given. Conversely, if X is independent of Y given Z and independently adds no relation to W once Y and Z are given, contraction yields independence of X from the pair (Y,W) given Z.

Dawid’s own lemmas are phrased with random variables and measurable functions, not with the later vocabulary of semi-graphoids. That historical distinction matters. The 1979 paper supplies the probabilistic pattern; later work abstracts an axiomatic structure and analyses which additional laws hold in which models.

One tempting extra law is intersection: if X is independent of A given B and C, and independent of B given A and C, infer independence of X from (A,B) given C. Intersection is not valid for every probability distribution. Strict positivity is a familiar sufficient condition, and later work gives weaker support-connectivity conditions and counterexamples. A classroom mnemonic must not silently upgrade the calculus.

A Markov state is a screening boundary

Dawid uses the elementary rules to recover a nearest-neighbour property of a discrete-time Markov chain. Write the past before t, the present Xₜ and the future after t. The Markov property says the future is conditionally independent of the past given the present. By splitting past and future blocks and recombining statements, the conditional distribution of Xₜ given every other state is determined by its immediate neighbours Xₜ₋₁ and Xₜ₊₁.

This is a striking change of view. A Markov chain is usually introduced through transition probabilities. Conditional independence instead exposes the information boundary. The present state screens the future from the earlier past in the forward direction; for the full path distribution, the nearest neighbours screen an interior state from all remaining times.

The result does not claim that time itself has only local causes, nor that a low-dimensional state has been chosen correctly. If the recorded state omits memory, the conditional-independence claim fails. The calculus tells us what follows from a Markov representation, not whether nature supplied the representation.

Simpson’s paradox locates a judgement the calculus cannot make

Dawid’s two contingency tables produce opposite aggregation patterns. In Table 1, treatment is associated with higher recovery within both sex strata, yet treatment and recovery are independent after sex is ignored. In Table 2, treatment and recovery are independent within each sex, yet they become associated after pooling. The arithmetic is exact. The scientific interpretation is not determined by the arithmetic.

To formulate a treatment effect at the level of an individual I, Dawid imagines the response Y under each possible treatment X and writes the null of no individual treatment effect as Y independent of X given I. This is conceptual and untestable because one individual cannot simultaneously receive both treatments. A covariate set Z is called sufficient when, for a given treatment and Z, further identity information I is irrelevant to Y: Y independent of I given (X,Z).

If Z is sufficient, the causal null is equivalent to the testable-looking statement Y independent of X given Z. But sufficiency is the bridge, not a consequence of the test. One may reject a candidate set by finding residual relevance in additional covariates, yet failure to reject does not prove sufficiency. Dawid explicitly says that at some point a subjective judgement must declare the covariate set sufficient. His discussants focus on this weak link, and Dawid’s reply accepts that causal inference remains subtle and incompletely resolved.

Randomisation offers a clearer route because treatment assignment can be independent of individual identity given design strata. Still, Dawid resists treating randomisation as a magic inscription. The conditional-independence statement identifies what the design is meant to guarantee and which inferential step uses it.

Sufficiency is a data-compression claim written as irrelevance

Let Θ index the distribution of full data X, and let T be a function of X. T is sufficient when the conditional distribution of X given T is the same at every Θ. Dawid writes this as X independent of Θ given T. Once T is known, the rest of X contains no further information about Θ. In a Bayesian model with a proper prior, the same statement means the posterior based on X is the posterior based on T.

Ancillarity is the dual-looking statement S independent of Θ: the sampling distribution of S does not change with the parameter. Identification is expressed on the parameter side. If X depends on a pair (Θ,Φ) but X is independent of Φ given Θ, then Φ is redundant once Θ is known. The language exposes a duality between reducing data and reducing parameter descriptions.

Prediction sufficiency introduces a future outcome Y. A statistic T(X) is adequate for predicting Y when it is sufficient for Θ and the distribution of Y given X depends on X only through T. Equivalently, X is independent of the pair (Y,Θ) given T. Under a prior, T retains all of X that matters for predicting Y. This is a stronger target than estimating one parameter well.

Transitivity and total sufficiency extend the idea through sequences. A running statistic that remains sufficient and adequate for the future can form a Markov chain of summaries. Dawid shows how the calculus transports such properties through functions and updates, turning lengthy special-case proofs into transformations of irrelevance statements.

Selection is harmless only for the target it leaves invariant

Let Q indicate whether generated data X are selected for observation, and let Y be the feature that governs selection. If Q is independent of X given Y, then within a fixed Y the conditional distribution of X is the same among selected and unselected cases. This can justify direct estimation of X given Y from selected data. It does not say the marginal distribution of X is preserved.

The target matters. In medical diagnosis, referral to a specialist may depend on presenting symptoms S₀ and background information G. If disease D is independent of the selection mechanism once the full symptoms S are known, the disease distribution given S may be transported from the clinic sample. The distribution of symptoms given disease can still be distorted. Conditional independence prevents the analyst from treating “selection is ignorable” as an unqualified property of a dataset.

Dawid applies the same logic to sufficient statistics after selection, prediction adequacy, randomised stopping times and missing data. Under an appropriate stopping condition, likelihood ratios or posteriors based on stopped data can match those from the corresponding fixed-length experiment. But the condition is a property of the stopping mechanism, not a universal blessing of optional stopping.

Invariance turns group structure into irrelevance

Suppose transformations act on both the sample and parameter spaces while preserving the form of the sampling model. A maximal invariant Z of the data removes changes generated by the group; a maximal invariant Ψ of the parameter records the parameter feature untouched by those transformations. Dawid writes Z independent of Θ given Ψ: the distribution of the data invariant depends on the parameter only through its invariant.

If S is sufficient and the group acts consistently on S, a maximal invariant U of S can itself be sufficient for the invariant parameter based on Z. This combines ordinary sufficiency with symmetry reduction. Examples include location models, scale models and directional data. The method clarifies exactly which orientations, offsets or scales have been factored out.

The paper’s final application replaces a single objective model with a family of subjective distributions held by individuals. An intersubjective parameter Θ is sufficient for the family when everyone agrees on the conditional distribution of observables Y given Θ, even if they disagree about the marginal distribution of Θ. Exchangeability yields a familiar example: observers may disagree about the mixing distribution of a Bernoulli chance while agreeing on the conditional Bernoulli model.

An executable Simpson-table laboratory

The laboratory reconstructs Dawid’s two published 2×2×2 tables. Table 1 contains treatment association within men and women but exact marginal independence after pooling. Table 2 contains exact treatment-recovery independence within both sex strata but marginal dependence after pooling. Select a table, then transpose X and Y to test the symmetry of ordinary conditional independence.

For each 2×2 table, independence is checked by the exact integer cross-product equality n₁₁n₀₀ = n₁₀n₀₁. Where every cell is positive, the odds ratio gives the same answer: one means independence. The display also reports recovery or treatment proportions so that equalities can be inspected rather than merely labelled.

How to inspect it. Start with Table 1 and compare the two treatment bars within each sex, then inspect the pooled pair. Switch to Table 2 and watch the logical pattern reverse. Finally transpose the reading direction. Counts move across the 2×2 cells, but the cross-product equality and independence verdict do not change.

The laboratory proves no causal effect. It verifies arithmetic relations in two finite tables and demonstrates why a conditioning decision cannot be chosen from the pooled association alone. Sampling uncertainty is also absent: these are the paper’s constructed counts, not estimates with confidence intervals.

The paper’s reach is real because its boundary is visible

The published discussion tests Dawid’s thesis rather than merely celebrating it. David Lindley doubts whether independence is always the most satisfactory primitive and warns that claims involving fixed parameters can become dangerous without clear semantics. James Kadane accepts the unification but stresses that declaring a covariate set sufficient is a sensitivity-laden judgement. Other discussants distinguish association, dependence and causality, question notation, and propose weaker relations.

Dawid’s reply does not claim victory over these objections. He says conditional independence has not resolved all philosophical problems of association and causation, agrees that Simpson’s paradox still has much scope for improvement, and defends the language as a suitable matrix for further work. This response sets the right evidential boundary: a useful formal vocabulary can expose assumptions without automatically validating them.

The immediate sequel supplied more rigorous foundations for conditional independence in statistical operations. Later graphoid and separoid work abstracted symmetry, decomposition, weak union and contraction beyond probability. Graphical models gave conditional-independence claims a spatial calculus through separation. Yet stronger deductions retain conditions: intersection can fail when the support is disconnected or has structural zeros, and composition is not universal.

The non-random-parameter issue also survived. Modern extended and transitional notions distinguish stochastic variables from parameters, treatments, regimes and environments without inventing a prior over those inputs. That work vindicates Dawid’s diagnosis while refining his heuristic notation: a relation that compares all parameter values may need an asymmetric kernel factorisation rather than ordinary symmetric conditional independence.

Exact conditional independence is also statistically severe. In finite samples, a failed test may reflect weak power; a passed test cannot prove equality of full distributions. In broad nonparametric settings with continuous conditioning variables, a uniformly valid conditional-independence test can have no nontrivial power against alternatives. Practical systems therefore need an explicit model class, error criterion and sensitivity analysis in addition to the algebra.

Memory studio

Next three papers to read

1. Canon 0020, Rosenbaum and Rubin (1983). See how a scalar propensity score becomes a balancing score under strong ignorability.
2. Canon 0021, Geman and Geman (1984). Use local conditional distributions to build a stochastic image-restoration process.
3. Dawid (1980). Replace the 1979 heuristics with a more formal theory of statistical operations.

Conditional independence is a language for disciplined forgetting

The paper’s deepest contribution is representational. It asks statisticians to name which information becomes redundant, relative to which retained information, and for which target distribution. Once named, the statement can be transported through a small calculus and recognised in fields that had developed separate vocabularies.

That same representation prevents overclaiming when used carefully. A sufficient statistic is sufficient for a parameter in a model, not universally informative. A selection mechanism is ignorable for a specified conditional distribution, not for every question. A covariate set screens individual differences only under an assumption that must be defended. A Markov state is a memory boundary only if the state description is adequate.

For intelligence, agency and consciousness research, these distinctions are practical. A behavioural state may appear to screen future action from history, an internal representation may look sufficient for prediction, or a reported mental state may seem independent of architecture after task variables are conditioned on. Each is a model-relative irrelevance statement. None by itself establishes agency, phenomenal consciousness or the reality of the proposed latent categories.

Dawid’s enduring lesson is therefore neither “condition on everything” nor “graphs reveal causes”. It is to turn vague claims of relevance into explicit three-place relations, deduce only what the justified rules allow, and keep the scientific choice of conditioning information visible.