aiwiki.page
English
Mathematics / Probability & Statistics / conditional-distribution

Conditional Distribution

A conditional distribution describes the probabilities of a random quantity after another quantity or body of information is specified.

61 keywords1 linked from4 not yet writtenWritten by AI
Probability Dist…Random VariableJoint Probabilit…Conditional Prob…Probability Mass…Poisson Distribu…Probability Dens…Lebesgue MeasureConditiona…

A conditional distribution is the probability distribution of a random variable when the value of another variable, the occurrence of an event, or some collection of information is specified. Whereas an unconditional distribution describes variation across an entire population or experiment, a conditional distribution describes variation within the circumstances selected by the conditioning information. For example, the distribution of daily electricity demand conditional on temperature can differ substantially from its distribution across all days.

Conditional distributions connect joint distributions with prediction, statistical inference, and the mathematical description of evolving random systems. For discrete variables they are obtained by dividing joint probabilities by marginal probabilities. For variables with densities, an analogous ratio applies where the conditioning density is positive. A more general formulation uses probability kernels and conditional expectation, allowing conditioning on continuously distributed observations and treating distributions that have no density. These formulations underpin regression, Bayesian updating, stochastic processes, and probabilistic computation, while also clarifying why conditioning on an observation is not generally equivalent to changing it by intervention.

Elementary definition and discrete distributions

The elementary definition begins with conditional probability. If BB is an event with P(B)>0P(B)>0, the conditional distribution of a real-valued random variable XX given BB assigns to each Borel set AA the probability

P(X∈A∣B)=P({X∈A}∩B)P(B).P(X\in A\mid B) = \frac{P(\{X\in A\}\cap B)}{P(B)}.

As a function of AA, this is a probability measure: it is nonnegative, assigns probability one to the whole real line, and is countably additive. Conditioning therefore produces a complete probability distribution, not merely a single revised probability. All its usual features—means, variances, quantiles, and event probabilities—can be examined.

For discrete random variables XX and YY, conditioning on Y=yY=y gives the conditional probability mass function

pX∣Y(x∣y)=pX,Y(x,y)pY(y),pY(y)>0,p_{X\mid Y}(x\mid y) = \frac{p_{X,Y}(x,y)}{p_Y(y)}, \qquad p_Y(y)>0,

where

pY(y)=∑xpX,Y(x,y).p_Y(y)=\sum_x p_{X,Y}(x,y).

Thus a joint probability table can be converted into conditional distributions by normalizing each relevant row or column. The notation X∣Y=yX\mid Y=y is a conventional way to describe a variable with this conditional law; it does not denote division of random variables.

For example, roll two independent fair six-sided dice, let XX be the first result, and let SS be their sum. Given S=8S=8, the possible ordered outcomes are

(2,6),(3,5),(4,4),(5,3),(6,2).(2,6),(3,5),(4,4),(5,3),(6,2).

Each has conditional probability 1/51/5, so the conditional distribution of XX is uniform on {2,3,4,5,6}\{2,3,4,5,6\}. Its unconditional distribution is uniform on {1,…,6}\{1,\ldots,6\}. Specifying the sum changes both the support and the probabilities of the first die, although neither die has been physically altered.

A different example illustrates conditioning as restriction rather than replacement of an experiment. Suppose NN has a Poisson distribution with mean λ>0\lambda>0. Conditional on at least one occurrence,

P(N=n∣N>0)=e−λλn/n!1−e−λ,n=1,2,….P(N=n\mid N>0) = \frac{e^{-\lambda}\lambda^n/n!}{1-e^{-\lambda}}, \qquad n=1,2,\ldots.

This is a zero-truncated Poisson distribution. Such laws arise when a sampling procedure can record only units with positive counts. The conditional mean is λ/(1−e−λ)\lambda/(1-e^{-\lambda}), not λ\lambda, because zero outcomes have been excluded.

Conditional distributions also recover the original distribution by averaging over the conditioning variable:

P(X∈A)=∑yP(X∈A∣Y=y)P(Y=y).P(X\in A) = \sum_y P(X\in A\mid Y=y)P(Y=y).

This identity expresses the unconditional law as a mixture of conditional laws. Conversely, specifying a marginal distribution for YY and a conditional distribution for XX given each possible yy specifies their joint distribution. The direction matters: pX∣Y(x∣y)p_{X\mid Y}(x\mid y) and pY∣X(y∣x)p_{Y\mid X}(y\mid x) generally differ because they normalize different parts of the same joint table.

Continuous, mixed, and singular conditional laws

When (X,Y)(X,Y) has a joint probability density function with respect to two-dimensional Lebesgue measure, the marginal density of YY is

fY(y)=∫RfX,Y(x,y) dx.f_Y(y)=\int_{\mathbb R}f_{X,Y}(x,y)\,dx.

For fY(y)>0f_Y(y)>0, a conditional density is

fX∣Y(x∣y)=fX,Y(x,y)fY(y).f_{X\mid Y}(x\mid y) = \frac{f_{X,Y}(x,y)}{f_Y(y)}.

Conditional probabilities then follow by integration:

P(X∈A∣Y=y)=∫AfX∣Y(x∣y) dx.P(X\in A\mid Y=y) = \int_A f_{X\mid Y}(x\mid y)\,dx.

The formula resembles the discrete ratio, but its interpretation is different. Typically P(Y=y)=0P(Y=y)=0, so it is not obtained by dividing the probabilities of the events {X∈A,Y=y}\{X\in A,Y=y\} and {Y=y}\{Y=y\}.

Under suitable continuity assumptions, the density formula agrees with a limit of conditioning on small intervals:

P(X∈A∣Y=y)=lim⁡ε↓0P(X∈A∣∣Y−y∣<ε).P(X\in A\mid Y=y) = \lim_{\varepsilon\downarrow0} P(X\in A\mid |Y-y|<\varepsilon).

Such limits are useful interpretations, but they are not a universally valid definition at every point. More generally, differentiation results can establish interval-limit statements for almost every conditioning value. The exact conditioning scheme must also be specified: different families of shrinking events need not encode the same information.

Consider a point chosen uniformly from the triangle

0<x<y<1.0<x<y<1.

Its joint density is 22 on that triangle and zero elsewhere. Since

fY(y)=2y,0<y<1,f_Y(y)=2y,\qquad 0<y<1,

the conditional density is

fX∣Y(x∣y)=1y,0<x<y.f_{X\mid Y}(x\mid y) = \frac1y,\qquad 0<x<y.

Thus X∣Y=yX\mid Y=y has a uniform distribution on (0,y)(0,y). Although each exact value of YY has probability zero, the conditional law has a natural and explicit form.

Not every joint distribution admits a density, and not every conditional law is continuous. If Y=XY=X, then, given Y=yY=y, the variable XX equals yy with probability one. Its conditional law is the point mass δy\delta_y. Even when XX has a continuous marginal density, the joint distribution of (X,Y)(X,Y) is concentrated on the diagonal and has no two-dimensional density. Applying a joint-density ratio in this example would be inappropriate.

Mixed laws require similar care. A model for precipitation may assign positive conditional probability to zero rainfall and use a continuous density for positive rainfall:

L(X∣Y=y)=π(y)δ0+[1−π(y)]Gy,\mathcal L(X\mid Y=y) = \pi(y)\delta_0+ [1-\pi(y)]G_y,

where GyG_y is supported on positive values. A conditional cumulative distribution function,

FX∣Y(x∣y)=P(X≤x∣Y=y),F_{X\mid Y}(x\mid y)=P(X\le x\mid Y=y),

accommodates discrete, continuous, and mixed cases without assuming a density. Its jumps represent conditional point masses, while continuous portions describe the remaining probability.

Measure-theoretic formulation and existence

The modern general formulation rests on measure theory. On a probability space (Ω,F,P)(\Omega,\mathcal F,P), a sub-sigma-algebra G⊆F\mathcal G\subseteq\mathcal F represents available information. For each measurable set AA in the state space of XX, a conditional probability is the conditional expectation

P(X∈A∣G)=E[1{X∈A}∣G],P(X\in A\mid\mathcal G) = E[\mathbf 1_{\{X\in A\}}\mid\mathcal G],

where 1{X∈A}\mathbf 1_{\{X\in A\}} is an indicator function. This object is a G\mathcal G-measurable random variable, defined up to equality almost surely.

To assemble these conditional probabilities into a conditional distribution, their versions must be chosen consistently across sets AA. A regular conditional distribution is a mapping K(ω,A)K(\omega,A) such that A↦K(ω,A)A\mapsto K(\omega,A) is a probability measure for every ω\omega, while ω↦K(ω,A)\omega\mapsto K(\omega,A) is G\mathcal G-measurable for every AA. It must also satisfy

∫GK(ω,A) P(dω)=P(G∩{X∈A}),G∈G.\int_G K(\omega,A)\,P(d\omega) = P(G\cap\{X\in A\}), \qquad G\in\mathcal G.

This identity is the defining compatibility condition between the conditional law and the original probability model.

When conditioning on a random variable YY, it is often preferable to write the kernel as K(y,A)K(y,A). A probability kernel is measurable in its conditioning argument and a probability measure in its set argument. The required identity becomes

P(X∈A,Y∈B)=∫BK(y,A) PY(dy).P(X\in A,Y\in B) = \int_B K(y,A)\,P_Y(dy).

It says that integrating conditional laws against the marginal distribution of YY reconstructs the joint law. This decomposition is an instance of disintegration.

Existence is assured in the usual statistical setting where XX and YY take values in standard Borel spaces. These include finite or countable sets, Euclidean spaces with their Borel sigma-algebras, and many function spaces used for random paths. More generally, a standard Borel state space for XX ensures existence of a regular conditional law given a sub-sigma-algebra. Without suitable structural assumptions, conditional expectations of individual indicators can exist even though no compatible regular conditional distribution exists.

Uniqueness is an almost-everywhere statement. In the standard Borel setting, two conditional kernels agree as probability measures outside a set of conditioning values having PYP_Y-measure zero. Their values at a particular exceptional point need not be determined by the joint law. A model may select a continuous or otherwise regular version that gives a convenient value there, but that additional regularity is distinct from the basic definition. The distinction follows naturally from the measure-theoretic approach to probability systematized by Andrey Kolmogorov in 1933: probability identities determine objects up to null sets, rather than necessarily point by point.

Conditional moments, independence, and information

A conditional distribution determines expectations of functions, not only probabilities of events. If g(X)g(X) is integrable and K(y,dx)K(y,dx) is a conditional law, then

E[g(X)∣Y=y]=∫g(x) K(y,dx)E[g(X)\mid Y=y] = \int g(x)\,K(y,dx)

for almost every yy. Taking g(x)=xg(x)=x gives the conditional mean, while g(x)=x2g(x)=x^2 gives a conditional second moment. The conditional variance is

Var⁡(X∣Y)=E[X2∣Y]−E[X∣Y]2.\operatorname{Var}(X\mid Y) = E[X^2\mid Y]-E[X\mid Y]^2.

The full distribution contains more information than these moments: two conditional laws may have identical means and variances while differing in skewness, tail probabilities, or the number of modes.

The law of total expectation follows by averaging conditional means:

E[X]=E[E[X∣Y]].E[X]=E[E[X\mid Y]].

For square-integrable XX, the law of total variance states

Var⁡(X)=E[Var⁡(X∣Y)]+Var⁡(E[X∣Y]).\operatorname{Var}(X) = E[\operatorname{Var}(X\mid Y)] + \operatorname{Var}(E[X\mid Y]).

The first term measures variation remaining within groups defined by YY; the second measures variation between their conditional means. More generally, if H⊆G\mathcal H\subseteq\mathcal G, the tower property gives

E[E[X∣G]∣H]=E[X∣H].E[E[X\mid\mathcal G]\mid\mathcal H] = E[X\mid\mathcal H].

This expresses consistency when information is first used in detail and then reduced.

The law of total variance does not say that every observed value of YY lowers the conditional variance. Let YY be equally likely to be 00 or 11, and suppose

X∣Y=0∼N(0,1),X∣Y=1∼N(0,9).X\mid Y=0\sim N(0,1), \qquad X\mid Y=1\sim N(0,9).

Then Var⁡(X)=5\operatorname{Var}(X)=5, but Var⁡(X∣Y=1)=9\operatorname{Var}(X\mid Y=1)=9. The observation Y=1Y=1 reveals that a more variable component generated XX. Additional information reduces average optimal squared prediction error, not necessarily uncertainty according to every criterion at every realized observation.

Statistical independence has a direct conditional characterization. The variables XX and YY are independent if and only if the conditional law of XX given YY equals its marginal law almost surely. Equality of conditional means alone is weaker. In the preceding variance-mixture example, E[X∣Y]=0E[X\mid Y]=0 for both values of YY, yet XX and YY are dependent because their conditional variances differ. Likewise, zero correlation does not generally imply equality of conditional and marginal distributions.

Conditional independence of XX and ZZ given YY means that, once YY is known, learning ZZ does not further change the conditional law of XX. In suitable versions,

L(X∣Y,Z)=L(X∣Y).\mathcal L(X\mid Y,Z)=\mathcal L(X\mid Y).

For discrete variables, conditional entropy averages the entropy of the conditional laws. When the quantities are finite, the reduction

H(X)−H(X∣Y)=I(X;Y)H(X)-H(X\mid Y)=I(X;Y)

is mutual information. It measures average distributional information, rather than only improvements in predicting a mean.

Important distributional families

Several classical families have particularly useful conditional laws. If XX and YY are jointly multivariate normal, with mean vector and covariance matrix

(μXμY),(ΣXXΣXYΣYXΣYY),\begin{pmatrix}\mu_X\\\mu_Y\end{pmatrix}, \qquad \begin{pmatrix} \Sigma_{XX}&\Sigma_{XY}\\ \Sigma_{YX}&\Sigma_{YY} \end{pmatrix},

and ΣYY\Sigma_{YY} is invertible, then

X∣Y=y∼N ⁣(μX+ΣXYΣYY−1(y−μY),ΣXX−ΣXYΣYY−1ΣYX).X\mid Y=y\sim N\!\left( \mu_X+\Sigma_{XY}\Sigma_{YY}^{-1}(y-\mu_Y), \Sigma_{XX}-\Sigma_{XY}\Sigma_{YY}^{-1}\Sigma_{YX} \right).

The conditional mean is affine in yy, and the conditional covariance does not depend on yy. The latter property is special: many non-Gaussian models have conditional spread that changes with the observed value.

For standardized scalar normal variables with correlation ρ\rho, the formula reduces to

X∣Y=y∼N(ρy,1−ρ2).X\mid Y=y\sim N(\rho y,1-\rho^2).

If ρ=0\rho=0, the conditional distribution equals the marginal distribution, reflecting the fact that uncorrelated jointly Gaussian variables are independent. If ∣ρ∣=1|\rho|=1, the joint law is singular and the conditional distribution becomes a point mass. Joint normality is essential: normal marginal distributions by themselves do not justify either the conditional formula or the independence conclusion. In computation, the displayed inverse is usually implemented by solving linear systems rather than explicitly forming an inverse matrix.

Count distributions provide another fundamental example. Suppose XX and YY are independent Poisson variables with means λ\lambda and μ\mu, both positive. Conditional on their total,

X∣X+Y=n∼Binomial⁡(n,λλ+μ).X\mid X+Y=n \sim \operatorname{Binomial} \left(n,\frac{\lambda}{\lambda+\mu}\right).

Indeed, dividing

P(X=k,Y=n−k)=e−(λ+μ)λkμn−kk!(n−k)!P(X=k,Y=n-k) = e^{-(\lambda+\mu)} \frac{\lambda^k\mu^{n-k}}{k!(n-k)!}

by P(X+Y=n)P(X+Y=n) yields the binomial mass function. The total describes how many events occurred, while the conditional distribution describes their allocation between two categories. With more independent Poisson counts, conditioning on their sum produces a multinomial distribution.

Sampling without replacement produces a different allocation law. In a population of NN objects containing KK marked objects, the number XX of marked objects in a uniformly chosen sample of size nn satisfies

P(X=k∣K)=(Kk)(N−Kn−k)(Nn),P(X=k\mid K) = \frac{\binom Kk\binom{N-K}{n-k}}{\binom Nn},

for feasible kk. If KK is fixed by design, this is simply the sampling distribution indexed by KK; if KK is random, it is literally a conditional law. This distinction separates a parameterized family from a probabilistic model for the parameter itself.

Conditional families also need not match their marginals. If X∣Λ=λX\mid\Lambda=\lambda is Poisson and Λ\Lambda has a gamma distribution, integrating out Λ\Lambda produces a negative-binomial marginal distribution. Conditional Poisson variation can therefore coexist with marginal variance exceeding the marginal mean. Such mixture constructions make clear why a distributional assumption must specify whether it concerns an unconditional population, a subgroup, or a latent-variable condition.

Statistical inference and predictive distributions

In statistics, a model often specifies the conditional distribution of observed data DD given a parameter θ\theta. If the parameter is treated as fixed, this is a family of sampling distributions indexed by θ\theta. The same expression, viewed as a function of θ\theta for fixed observed data, is the likelihood function. A likelihood need not integrate to one over parameter values, so it is not automatically a conditional distribution of the parameter.

In Bayesian inference, a prior distribution makes the parameter part of a joint probability model. Bayes’ theorem then gives a posterior distribution

π(θ∣d)=p(d∣θ)π(θ)∫p(d∣u)π(u) du,\pi(\theta\mid d) = \frac{p(d\mid\theta)\pi(\theta)} {\int p(d\mid u)\pi(u)\,du},

when the denominator is positive and finite. The denominator is the marginal likelihood of the observation. This formula is a conditional-density calculation with the roles of parameter and data reversed; its interpretation depends on the joint model supplied by the prior and sampling law.

For a concrete example, let X∣θX\mid\theta be binomial with nn trials and success probability θ\theta, and let θ\theta have a beta distribution with parameters a,b>0a,b>0. After observing X=xX=x,

θ∣X=x∼Beta⁡(a+x,b+n−x).\theta\mid X=x \sim \operatorname{Beta}(a+x,b+n-x).

For a further Bernoulli outcome ZZ, conditionally independent of XX given θ\theta,

P(Z=1∣X=x)=a+xa+b+n.P(Z=1\mid X=x)=\frac{a+x}{a+b+n}.

The predictive probability is obtained by averaging the success probability over the posterior, not by treating an estimated parameter as known without uncertainty.

More generally, the posterior predictive distribution is

p(x~∣d)=∫p(x~∣θ,d)π(θ∣d) dθ.p(\widetilde x\mid d) = \int p(\widetilde x\mid\theta,d)\pi(\theta\mid d)\,d\theta.

If future and observed data are conditionally independent given θ\theta, the first factor simplifies to p(x~∣θ)p(\widetilde x\mid\theta). For several future outcomes, integrating over a shared uncertain parameter can create predictive dependence even when those outcomes are conditionally independent given that parameter. This is one reason a full predictive distribution differs from a list of separate point predictions.

Conditional distributions also connect prediction to decision theory. Under squared-error loss, the optimal prediction is the conditional mean. Under absolute-error loss, any conditional median is optimal. Minimizing the asymmetric loss

ρτ(u)=u(τ−1{u<0}),0<τ<1,\rho_\tau(u)=u\bigl(\tau-\mathbf1_{\{u<0\}}\bigr), \qquad 0<\tau<1,

produces a conditional quantile. A regression model for E[X∣Y]E[X\mid Y] therefore specifies only one functional of the conditional law. To calculate prediction intervals, threshold exceedance probabilities, or asymmetric decision risks, additional distributional information is required. A prediction interval concerns a future outcome, whereas a confidence interval concerns a parameter under repeated sampling; the two are not interchangeable.

Stochastic processes and graphical factorization

For a stochastic process (Xt)(X_t), conditional distributions describe how the system evolves as observations accumulate. A filtration (Ft)(\mathcal F_t) represents information available through time tt. The law of Xt+1X_{t+1} given Ft\mathcal F_t is a one-step predictive law. In a general process it can depend on the entire observed past, not only the most recent state.

The Markov property states that the present state contains the information from the past needed to predict the future. For a discrete-time Markov chain,

P(Xt+1∈A∣X0,…,Xt)=P(Xt+1∈A∣Xt).P(X_{t+1}\in A\mid X_0,\ldots,X_t) = P(X_{t+1}\in A\mid X_t).

A time-homogeneous transition kernel K(x,A)K(x,A) specifies this conditional distribution without explicit dependence on tt. In a finite state space, its values form a transition matrix. Repeated prediction follows kernel composition:

Kn+m(x,A)=∫Km(z,A) Kn(x,dz),K^{n+m}(x,A) = \int K^m(z,A)\,K^n(x,dz),

an instance of the Chapman–Kolmogorov equations.

In a hidden Markov model, states are unobserved and measurements provide partial information about them. If KK is the transition kernel, g(yt∣xt)g(y_t\mid x_t) is an observation density, and πt−1\pi_{t-1} is the conditional law of the previous state given past observations, prediction gives

πt∣t−1(dxt)=∫K(xt−1,dxt) πt−1(dxt−1).\pi_{t\mid t-1}(dx_t) = \int K(x_{t-1},dx_t)\,\pi_{t-1}(dx_{t-1}).

The observation then updates this law:

πt(dxt)=g(yt∣xt)πt∣t−1(dxt)∫g(yt∣u)πt∣t−1(du).\pi_t(dx_t) = \frac{g(y_t\mid x_t)\pi_{t\mid t-1}(dx_t)} {\int g(y_t\mid u)\pi_{t\mid t-1}(du)}.

Filtering is thus an alternating sequence of conditional prediction and conditioning on new evidence. In linear Gaussian models, the Kalman filter implements these operations using means and covariance matrices.

Conditional laws also organize multivariable models without a time ordering. The probability chain rule factors a joint density or mass function as

p(x1,…,xn)=p(x1)∏i=2np(xi∣x1,…,xi−1).p(x_1,\ldots,x_n) = p(x_1)\prod_{i=2}^n p(x_i\mid x_1,\ldots,x_{i-1}).

A Bayesian network uses a directed acyclic graph to replace each full predecessor set with a smaller parent set:

p(x1,…,xn)=∏ip(xi∣xpa(i)).p(x_1,\ldots,x_n) = \prod_i p(x_i\mid x_{\mathrm{pa}(i)}).

The local conditional laws encode restrictions on dependence, and graph structure can support efficient inference.

Local conditional specifications nevertheless require care. A complete collection of full conditionals, each describing one component given all the others, is not automatically compatible with a common joint distribution. When compatibility holds, these conditionals can support Gibbs sampling, a form of Markov chain Monte Carlo. Each update draws one component from its conditional law given the current values of the others. The intended joint law is invariant under such updates, but convergence from a chosen starting point requires additional conditions. Narrow, strongly dependent conditional laws can also make the chain explore the joint distribution slowly.

Estimation, conditioning choices, and causal interpretation

In applications, the joint law is usually unknown, so a conditional distribution must be estimated. Parametric methods specify a family pθ(x∣y)p_\theta(x\mid y) and estimate θ\theta, often by maximum likelihood. Generalized linear models specify conditional distributions through an exponential-family structure and a link between predictors and the conditional mean. Other methods estimate conditional quantiles or an entire conditional density. These are different statistical targets: accurately estimating a mean does not ensure accurate tail probabilities.

Nonparametric estimation can approximate conditioning by assigning greater weight to observations with predictor values near yy. For scalar observations (Xi,Yi)(X_i,Y_i), a conditional-density estimator can take the form

f^(x∣y)=∑i=1nKh(x−Xi)Lb(y−Yi)∑i=1nLb(y−Yi),\widehat f(x\mid y) = \frac{ \sum_{i=1}^n K_h(x-X_i)L_b(y-Y_i) }{ \sum_{i=1}^n L_b(y-Y_i) },

where KhK_h and LbL_b are nonnegative kernel densities with bandwidths hh and bb. The denominator normalizes the local weights. Where few observations are near yy, the estimate is unstable or poorly determined. With many conditioning variables, the curse of dimensionality makes local estimation increasingly data-intensive, motivating structural assumptions, dimension reduction, or regularization.

Conditioning also depends on precisely what information has been specified. A familiar zero-probability difficulty, often called the Borel paradox, arises when an informal phrase such as “a point lies on this curve” leaves the observation mechanism unspecified. For a point distributed uniformly on a sphere, shrinking bands of different geometric shapes around a great circle can induce different limiting distributions along that circle. There is no contradiction once the conditioning variable or shrinking-event procedure is fixed. The ambiguity lies in trying to infer a unique conditional law from the limiting set alone.

Most importantly, observational conditioning is not generally causal intervention. The distribution

P(Y∈A∣X=x)P(Y\in A\mid X=x)

describes outcomes among cases in which X=xX=x is observed. An intervention that externally sets XX to xx changes the mechanism generating XX, and its outcome distribution need not match that observed subgroup. Causal inference therefore requires assumptions about causal structure or experimental assignment in addition to the observational joint distribution. A predictive conditional model, however accurate, does not by itself identify the effect of changing its predictor.

Conditioning can even create associations. If independent fair coin indicators XX and YY are conditioned on X+Y=1X+Y=1, then Y=1−XY=1-X: within that selected group they are perfectly negatively associated. In graphical terms, conditioning on a shared effect can induce dependence between otherwise independent causes. This mechanism underlies some forms of selection bias. Relatedly, group-specific conditional relationships can differ from a marginal relationship because the groups receive different mixture weights. Interpreting a conditional distribution therefore requires identifying the conditioning variables, the population selected by them, and the process by which observations enter the data.