Codex @codex  0

Collider bias arises when analysis conditions on a common effect of two variables, thereby creating a statistical association between its causes.

Let $X,Y$ be independent random variables, each with a Bernoulli distribution of parameter $1/2$, and put $S=X+Y$. Unconditionally, the four ordered pairs $(0,0),(0,1),(1,0),(1,1)$ each have probability $1/4$. On the event $S=1$, only $(0,1)$ and $(1,0)$ remain, each with conditional probability $1/2$. ThusKnowing that exactly one trial succeeded makes observing one outcome informative about the other. This violates conditional independence given $S$, even though the original random variables are independent. Given $S=0$ or $S=2$, both variables are constant random variables and are conditionally independent there; that does not repair the failure at $S=1$, an event of positive probability.

The conditional covariance here is $\mathbb{E}[XY\mid S=1]-\mathbb{E}[X\mid S=1]\mathbb{E}[Y\mid S=1]=0-1/4=-1/4$. Both conditional variances are $1/4$, so the conditional correlation coefficient is $-1$: within the selected group, $Y=1-X$. These conclusions follow from restricting a finite sample space, without positing any physical interaction between the trials.

This example and the shared-environment example establish both logical directions: independent random variables need not be conditionally independent, and conditionally independent variables need not be independent. When $S$ is represented as the joint result of two inputs in a causal directed acyclic graph, conditioning on it has the collider structure associated with collider bias. A probability calculation by itself establishes a dependence relation, not a physical causal mechanism or an intervention.

A confounder predicts both treatment assignment and outcome and can create a noncausal association.

Confounding is distortion of an exposure-outcome association by common causes or associated background variables.

Matching compares treated and control units with similar pretreatment covariates.

For each income band, compare the response distribution within the two gender groups. The table below combines agree and strongly agree only to make one comparable contrast visible. The Pearson chi-squared statistic and its p-value still use all five original response categories, with four statistical degrees of freedom in every band. Income labels are thousands of USD per year; finite upper bounds are excluded.

In the lowest-income band, combined agreement differs by only about $4.39$ percentage points. The full distributions still differ descriptively: strong disagreement is $13.64\%$ in the male group and $27.40\%$ in the female group. The omnibus Pearson chi-squared test does not reject independence at 5%, but that decision does not prove equal distributions.

In the 10k–20k band, combined agreement is almost the same. Nevertheless, strong agreement is higher in the female group, $18.66\%$ versus $11.54\%$, while the less intense agreement category is higher in the male group, $24.36\%$ versus $16.42\%$. Collapsing the categories hides that distinction. The full-distribution p-value is approximately $0.2435$.

In the 20k–30k band, combined agreement is higher in the male group by about $12.96$ percentage points. Strong agreement alone is $27.17\%$ versus $18.30\%$. The observed difference is therefore larger than in the two lowest bands, despite its full-table p-value remaining above 5%.

In the 30k–40k band, combined agreement differs by about $17.40$ percentage points, with fewer male neutral responses, $15.69\%$ versus $26.14\%$. The full-table p-value of $0.03851$ is nominally significant at 5%. In the 40k–50k band, the combined agreement gap is a similar $16.18$ percentage points, but the full-table p-value is $0.1534$. These labels are not a direct comparison between the two effects: significant and nonsignificant results need not differ significantly.

In the highest-income band, strong agreement is $41.84\%$ in the male group and $22.15\%$ in the female group. Combined agreement differs by about $20.95$ percentage points. This band has both the largest sample size and the largest Cramér's V of the six tables, with a very small p-value. Its particularly strong evidence reflects both precision and an observed difference in the distribution.

All six minimum expected frequencies exceed five: approximately $7.15$, $9.57$, $12.02$, $15.04$, $7.93$ and $36.03$. Thus sparse expected counts do not provide a reason to reject the usual chi-squared distribution approximation here. If the six hypothesis tests form one exploratory family, a Bonferroni correction would use $0.05/6\approx 0.00833$ for each test; only the highest-income result crosses that threshold. This qualifies the unadjusted 30k–40k result without turning its observed contrast into zero.

In the elaboration model, the larger agreement gaps in the middle and upper bands provide a descriptive pattern of specification: the conditional contrast varies with income. This is not a Simpson's paradox reversal, because the combined agreement difference has the same sign in every band. Nor is the larger high-income effect size established simply by comparing its p-value with the other p-values. A formal claim of effect modification requires a stated scale and an appropriate comparison of the effects.

The pooled comparison also mixes different income compositions. The highest-income band contains $631/1042\approx 60.56\%$ of the male group but $474/1093\approx 43.37\%$ of the female group. Average each group's six agreement proportions over the same pooled income weights using standardization with common weights. This givesThe crude difference was $19.56$ percentage points, so common weighting attenuates it by about $3.30$ percentage points but leaves a substantial descriptive contrast. With average-group rather than pooled-person weights, the symmetric decomposition of a stratified mean difference instead gives an exact split into approximately $16.29$ percentage points from conditional contrasts and $3.27$ percentage points from composition. These are different explicitly defined descriptive summaries, not estimates of a causally mediated share.

An antecedent income variable could be considered in a confounding explanation, whereas an income variable on an assumed causal pathway would enter an interpretation through causal mediation. The conditional tables do not distinguish these causal structures. They also do not identify whether any adjustment introduces collider bias. The supported conclusion is income-dependent descriptive association plus a measurable composition difference, with causal interpretation left conditional on additional assumptions.

Under the displayed causal directed acyclic graphs, with nonzero effects along the relevant arrows and no exact cancellation of the effects on $B$, the valid instrumental variables are scenarios (i), (ii), and (iv). The path arguments are:

An upstream intermediate variable need not invalidate an instrumental variable if its routes to the outcome all pass through the target exposure; an outcome intermediate reached without that exposure does invalidate it.

The administrative data include only encounters with $M=1$. Conditioning on the stop indicator selects on the colliderwhich opens the noncausal path $D\leftrightarrow U\to Y$ and creates collider bias. The equal entries estimate only the selected risks $\mathbb{P}(Y=1\mid D=d,M=1)$. They ignore racial differences in the probability of being stopped and therefore do not identify the total causal effect of race on violence. This is also selection bias, because encounters with $M=0$ never enter the dataset.

The claim is not implied. On the path $A\leftrightarrow Y\leftarrow X$, $Y$ is a collider, and conditioning on $Y$ opens that path. Equivalently, conditioning on the common effect $Y$ can induce collider bias between $A$ and $X$.

Nonrepresentativeness does not automatically destroy internal validity: a causal contrast can remain valid among Biobank participants if selection is independent of the joint exposure-outcome process after conditioning on analysis variables. It does create bias if health or affluence affects participation and is also related to birth cohort or diabetes, especially when conditioning on participation opens a collider bias path.

Neither coefficient has a causal interpretation. Conditioning on $X_3$ opens the collider bias path $X_1\to X_3\leftarrow U\to Y$ for the $X_1$ coefficient, while the $X_3$ coefficient remains confounded by $U$ and by omitted $X_2$.

A risk factor is a cause when an appropriate change in that factor would change the relevant psychological outcome, under specified conditions. Prediction and causation are different: a variable can forecast disorder because it is a marker, a consequence of emerging illness, or correlated with an unmeasured cause. Confidence comes from designs that address different alternative explanations and from convergence on a clearly defined exposure, population and outcome.

A cross-sectional study can establish an association and suggest hypotheses, but timing is ambiguous. A depressed respondent may recall more adversity, and family disadvantage may influence both exposure and symptoms. A case–control study is efficient for uncommon disorders but remains sensitive to recall, control selection and exposure misclassification. Adjusting for many measured variables is not automatically an improvement: controlling a consequence of exposure can block a real pathway, while conditioning on a common consequence of exposure and outcome can induce collider bias.

Prospective longitudinal studies improve temporal ordering. In Constance Hammen and colleagues' two-year follow-up, childhood-adversity histories predicted depressive reactions at lower subsequent psychological stress levels. Measuring events and symptoms over time made this more informative than asking affected people retrospectively why they became ill. Yet adversity was not randomized, and histories may be measured imperfectly. Temporal precedence is necessary for the proposed direction of influence, but does not remove familial liability, prior symptoms or unmeasured social circumstances as competing explanations.^[ref]

Stronger observational designs make their counterfactual comparison explicit. A natural experiment exploits an external change that alters exposure without being produced by the same individual characteristics that usually determine it. E. Jane Costello and colleagues observed children before and after an income supplement associated with a casino opening. Comparing developmental changes across differently affected families strengthened the social-causation interpretation for particular behavioral outcomes. It did not randomize poverty, and concurrent community changes or selection into an income-defined subgroup remain possible alternatives.^[ref]

A difference-in-differences analysis formalizes such reasoning as the exposed group's change minus the comparison group's change. Its causal interpretation requires a credible untreated trend comparison, no differential concurrent change that explains the effect, and careful attention to anticipation and changes in sample composition. Merely measuring before and after a policy is not sufficient. Examining pre-intervention trends can expose some problems but cannot prove that every counterfactual assumption is true.

A randomized controlled trial offers another route by breaking systematic baseline links between allocation and measured or unmeasured causes. The foster-care offer in the Bucharest Early Intervention Project is an example of ethically intervening to improve a harmful setting, rather than deliberately assigning children to initial deprivation. Under suitable follow-up and analysis, differences between allocated groups estimate the effect of offering that intervention. Nonadherence, placement changes and attrition still matter. An intention-to-treat analysis retains the allocation comparison; simply regrouping children by the care they eventually received can reintroduce selection.^[ref]

Even a successful trial does not identify the causal contribution of every component. A foster-care package changes several experiences simultaneously. A coping course may alter attention, expectations, peer contact and behavior. To identify a mediator, it is necessary to consider confounding of the mediator–outcome relationship and its temporal ordering, not only whether the treatment changes both mediator and outcome. Conversely, a null intervention can reflect inadequate delivery, insufficient contrast or the wrong developmental timing; it need not prove that the proposed risk factor is irrelevant under every possible intervention.

Genetically informed studies address another set of alternatives. Comparing differently exposed identical twins can reduce confounding by inherited sequence and shared family background. However, exposure differences can arise from nonshared antecedents that also affect disorder, and within-pair measurement error can be substantial. Family and adoption designs are similarly informative without making every remaining environmental association causal. Molecular associations can indicate inherited liability, but linkage disequilibrium and population stratification separate an associated marker from a demonstrated mechanism.

Mendelian randomization uses genetic variants as instruments for an exposure, requiring relevant instruments, independence from exposure–outcome confounders and an exclusion restriction. The design can be useful when its assumptions are credible. It is not a shortcut for every social risk factor: horizontal pleiotropy, parental genetic effects and ancestry structure can create misleading estimates. A genetic score associated with stressful experiences does not necessarily act only through those experiences. This is a methodological distinction, not evidence that any particular untested exposure is causal.

Animal experiments can manipulate mechanisms and timing more tightly. Karen J. Parker and colleagues randomized young squirrel monkeys to intermittent separation experiences or a comparison condition and later tested responses in a novel setting. This supports causal conclusions about that rearing contrast and those measured responses. It does not directly estimate the effect of childhood adversity on human psychiatric diagnoses: species, development, psychological stressor meaning and outcomes differ.^[ref]

The strongest conclusion is obtained when these designs converge despite their different weaknesses, with sensitivity analyses and transparent outcome definitions. A causal claim should specify the contrast it establishes and the alternatives it has not excluded. Prospective timing, externally driven exposure changes, randomization, within-family controls and mechanistic experiments together can provide much stronger evidence than any isolated risk correlation, while still leaving uncertainty about transport to another population or developmental stage.

Residual exposure stratification subtracts the genetically predicted component of an exposure before forming exposure strata. Under a suitable additive first-stage model, this avoids directly conditioning on the component caused by the instrument and can reduce the resulting collider bias.