Why Analyze Both Directions?
Two traits can be associated because one affects the other, because their causes overlap, or because of other pathways that influence both. When the direction is uncertain, Mendelian randomization (MR) can be designed in both directions: genetic variants associated with trait \(X\) are used as instruments to estimate the effect of \(X\) on \(Y\), and variants associated with \(Y\) are used to estimate the effect of \(Y\) on \(X\). The central discipline is to treat these as two separate analyses, each with its own instruments, assumptions, and causal estimand.
A bidirectional design is useful when reciprocal causal hypotheses are scientifically plausible. It does not, by itself, demonstrate that the traits form a biological feedback loop. In particular, results in opposite directions can reflect distinct lifelong genetic perturbations, correlated pathways, or violations of the assumptions in one or both analyses. The introductory material on Mendelian randomization establishes the instrumental-variable framework; here the emphasis is on how that framework is applied and compared in both directions.
Set Up the Two Analyses Separately
For a particular direction, let \(\hat\beta_{Xj}\) denote the estimated association of variant \(j\) with the exposure \(X\), and let \(\hat\beta_{Yj}\) denote its estimated association with the outcome \(Y\). These associations must be aligned to the same effect allele. The direction-specific analysis treats the variant–exposure association as the first-stage association and the variant–outcome association as the reduced-form association.
For the \(X\to Y\) analysis, the variants should be relevant to \(X\), independent of causes of \(Y\) other than through \(X\), and should affect \(Y\) only through \(X\) (the exclusion restriction). These are the relevance, independence, and exclusion-restriction assumptions. The \(Y\to X\) analysis requires the corresponding assumptions with \(Y\) as exposure and \(X\) as outcome. Satisfying the assumptions in one direction does not imply they hold in the other.
Instrument selection is therefore direction-specific. A variant strongly associated with \(X\) may be a weak instrument for \(Y\), and vice versa. A well-planned analysis defines selection thresholds, removes or accounts for correlated variants so the chosen instruments are sufficiently independent, and harmonizes effect alleles before estimating effects. The outcome association should not be used to select instruments in a way that creates a biased estimate. If a variant has a direct pathway to the outcome that does not pass through the proposed exposure, the exclusion restriction is violated for that direction.
State what effect is being estimated for \(X\to Y\), then state the separate effect for \(Y\to X\), including the scales of both traits.
Select variants for the exposure in that direction, assess their relevance, and avoid treating the reverse-direction instrument set as interchangeable.
Align variant associations to the same alleles, estimate each direction, and report uncertainty and sensitivity diagnostics for each.
Compare the direction, magnitude, uncertainty, scales, and instrument behavior. Do not infer a feedback mechanism merely because both estimates are nonzero.
Inverse-Variance Weighted Estimation
With multiple approximately independent variants and no directional pleiotropy, a common summary-data estimator is inverse-variance weighted (IVW) regression through the origin. For the \(X\to Y\) direction, let \(s_{Yj}\) be the standard error of the variant–outcome association. Treating the exposure associations as fixed for this calculation, set \(w_j=1/s_{Yj}^2\). The IVW estimate minimizes the weighted squared residuals between the observed outcome associations and the predicted associations \(\theta\hat\beta_{Xj}\).
The standard-error expression is the usual fixed-effect approximation when the weights are based on the outcome-association standard errors and the model is appropriate. Extra heterogeneity can make this uncertainty description inadequate; it should not be hidden by reporting only a point estimate. In the reverse direction, swap the roles: variants are selected for \(Y\), their associations with \(Y\) form the exposure associations, and associations with \(X\) are the outcome associations.
Worked Direction-Specific Analyses
Worked Example: Estimating the Effect of X on Y
Suppose a hypothetical analysis uses two independent variants selected for \(X\). Both traits are standardized. The associations are \(\hat\beta_{X1}=0.12\), \(\hat\beta_{Y1}=0.03\), \(\hat\beta_{X2}=0.08\), and \(\hat\beta_{Y2}=0.01\). Suppose each outcome-association standard error is \(0.03\). The calculation assumes correct allele alignment, sufficiently strong instruments, and valid instruments for \(X\to Y\); the numerical summary alone cannot verify the independence and exclusion assumptions.
Each weight is \(w_j=1/0.03^2=1111.11\). Substitution into the IVW formula gives:
The approximate standard error is \(1/\sqrt{23.111}=0.2080\). A normal-approximation 95% confidence interval is \(0.2115\pm1.96(0.2080)\), or approximately \((-0.196,\ 0.619)\) standard deviations of \(Y\) per standard deviation of \(X\). This interval includes zero, so these data do not provide precise evidence of a nonzero \(X\to Y\) effect. The estimate is not proof of no effect, and the interval does not account for violations of the instrument assumptions.
Worked Example: Estimating the Reverse Direction
Now consider a separate hypothetical instrument set selected for \(Y\), again with both traits standardized. For two independent variants, suppose \(\hat\beta_{Y1}=0.20\), \(\hat\beta_{X1}=0.04\), \(\hat\beta_{Y2}=0.10\), and \(\hat\beta_{X2}=0.03\). Each variant–outcome association with \(X\) has standard error \(0.02\). Assume the variants are relevant to \(Y\), their alleles are aligned, and the MR assumptions are plausible for \(Y\to X\).
Here the weights are \(1/0.02^2=2500\). With \(Y\) as exposure and \(X\) as outcome:
The approximate standard error is \(1/\sqrt{125}=0.08944\), giving a 95% normal-approximation interval of \(0.2200\pm1.96(0.08944)=(0.0447,\ 0.3953)\) standard deviations of \(X\) per standard deviation of \(Y\). This hypothetical reverse-direction estimate is more precise than the first example’s estimate. It is not valid to conclude that \(Y\to X\) is established while \(X\to Y\) is absent solely by contrasting whether each separate interval includes zero. The estimates have different uncertainty, and their validity depends on different instrument assumptions.
Worked Example: Interpreting Reciprocal Estimates
Suppose the two hypothetical analyses above are reported together. The first estimates \(0.2115\) standard deviations of \(Y\) per standard deviation of \(X\), with a 95% interval from \(-0.196\) to \(0.619\). The second estimates \(0.2200\) standard deviations of \(X\) per standard deviation of \(Y\), with an interval from \(0.0447\) to \(0.3953\).
Both point estimates are positive, but the first is imprecise and the second is more precise. Even if both had intervals excluding zero, the joint pattern would support at most two direction-specific MR findings conditional on their assumptions. It would not show that a change in \(X\) causes a later change in \(Y\), which then causes a further change in \(X\). The analyses use different genetic instruments and represent different genetically proxied perturbations; they do not measure a time-ordered cycle or identify a dynamic feedback system.
A defensible report would present the two estimates, intervals, scales, instrument counts, and direction-specific diagnostics separately, then state that reciprocal evidence does not by itself establish feedback. If the estimates conflict, possible explanations include imprecision, different instrument strength, pleiotropy, or genuinely distinct causal effects; the conflict is a reason to investigate, not to select the preferred direction by significance alone.
Diagnostics, Alternatives, and Limits
Assess instrument relevance separately for each exposure. Weak or imprecisely estimated exposure associations can destabilize ratio-based estimates, and a large number of variants does not guarantee a strong instrument set. Examine whether the variants are independent enough for the chosen estimator, whether effect alleles have been consistently aligned, and whether a small number of variants dominate the result. The Steiger directionality tests introduced earlier can provide evidence about whether variants explain more variation in the proposed exposure than outcome, but their result is not proof of causal direction; measurement error and other assumptions matter.
Heterogeneity among variant-specific estimates can signal that the simple common-effect model fits poorly. A funnel plot can reveal asymmetry, and the earlier material on MR-Egger, weighted median, weighted mode, and outlier diagnostics describes sensitivity approaches for particular forms of pleiotropy. These methods do not repair every violation: each relies on its own assumptions, and agreement across estimators is reassuring only to the extent those assumptions are plausible. For a bidirectional analysis, perform and interpret diagnostics separately in each direction rather than carrying a favorable sensitivity result from one direction over to the other.
| Approach | What it contributes | Important limitation |
|---|---|---|
| Direction-specific IVW | A weighted estimate using all selected variants under a common-effect model. | Sensitive to invalid instruments and misspecification; uncertainty depends on the model and weights. |
| Directionality evidence, including Steiger tests | Assesses whether genetic associations appear more consistent with one proposed direction. | Does not establish causality or rule out measurement error and pleiotropy. |
| Pleiotropy-robust sensitivity estimators | Test whether conclusions are stable under alternative assumptions about invalid instruments. | May be imprecise and can fail when their own assumptions are not met. |
A further limitation is that apparently similar standardized effects in the two directions are not automatically comparable causal parameters. Each estimate is conditional on its own genetic perturbations, instrument-selection process, and assumptions. If traits are measured on different scales, report the scales explicitly and avoid comparing raw numbers as though they described the same quantity. Likewise, neither a reciprocal pair of estimates nor their statistical significance identifies the biological mechanism linking the traits.
Common Mistake
A related error is to call one direction “the causal direction” merely because its \(p\)-value is below a threshold and the reverse \(p\)-value is not. A difference between “statistically significant” and “not statistically significant” is not itself evidence that the two effects differ. Compare estimates and uncertainty directly, consider their scales and assumptions, and avoid a directional conclusion stronger than the design supports.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In an \(X\to Y\) analysis, which trait should the selected genetic variants be associated with, and what does each variant’s outcome association represent?
- Why must relevance, independence, and the exclusion restriction be assessed separately for \(X\to Y\) and \(Y\to X\)?
- For one variant in an \(X\to Y\) analysis, let \(\hat\beta_X=0.10\), \(\hat\beta_Y=0.02\), and \(s_Y=0.01\). Calculate its ratio estimate and inverse-variance weight.
- Suppose well-conducted analyses estimate nonzero effects in both directions. What can be concluded, and what cannot be concluded from these results alone?
Solutions. (1) Select variants associated with \(X\), the exposure. Their associations with \(Y\) are the reduced-form, variant–outcome associations used to estimate the effect of \(X\) on \(Y\). (2) The exposure and instruments change with direction, so a variant set valid for one exposure need not be relevant or satisfy independence and exclusion for the other. (3) The single-variant ratio is \(0.02/0.10=0.20\); its weight is \(1/0.01^2=10{,}000\). The ratio is in outcome units per exposure unit. (4) The results can support two direction-specific causal estimates, conditional on the assumptions and diagnostics for each. They do not alone establish a time-ordered biological feedback loop or prove that either estimate is free of pleiotropic bias.