Conditional Instrument Strength In Multivariable MR

Why Conditional Strength Matters

In multivariable Mendelian randomization (MVMR), genetic variants are used as instruments for more than one exposure. A variant can predict each exposure when examined separately, yet the variants may provide little information for distinguishing their effects from one another. In that situation, the instruments are marginally relevant but weak for estimating the exposures’ effects conditional on the other exposures.

Conditional instrument strength is therefore exposure-specific: it asks whether the genetic predictors add information about one exposure after accounting for the other exposures in the model. This is different from asking whether the exposures are correlated with one another, whether the genetic instruments satisfy the exclusion restriction, or whether a particular causal estimate is unbiased.

Nested Models and the Conditional F Statistic

For an individual-level analysis, write the full regression for exposure \(j\) as

Here \(G_k\) is the genotype for instrument \(k\), and \(X_m\) denotes exposure \(m\). The null hypothesis for the conditional-strength test is \(H_0:\delta_1=\cdots=\delta_K=0\). The reduced model omits all \(K\) genotype terms but retains the other exposures. Let \(RSS_R\) be its residual sum of squares and \(RSS_F\) the residual sum of squares from the full model. Adding the instruments should not increase the residual sum of squares, so \(RSS_R\ge RSS_F\).

With an intercept, \(K\) genotype terms, and \(L-1\) other-exposure terms in the full model, the residual degrees of freedom are \(df=n-K-L\), provided the design matrix has full column rank. The partial F statistic is:

The numerator measures the reduction in residual variation per added genotype term; the denominator estimates residual variation per residual degree of freedom in the full model. The statistic has no units. A larger value is evidence that the instruments add predictive information about \(X_j\) beyond the other exposures.

The partial \(R^2\) expresses the same nested-model comparison as a proportion of the reduced-model residual variation removed by adding the instruments:

Worked Calculations

What the Statistic Can—and Cannot—Tell You

A conditional F statistic should be examined for each exposure. A low value for one exposure means that the variants provide limited information for separating that exposure’s contribution from the other included exposures. It can be accompanied by unstable estimates or wide uncertainty intervals, but the strength statistic is not a measure of causal effect size or a test of the exclusion restriction.

The reason collinearity increases uncertainty can be seen from the regression variance. Under the classical linear model, the covariance matrix of the coefficient estimates is proportional to \((Z^\mathsf{T}Z)^{-1}\), where \(Z\) is the design matrix of predictors. When columns of \(Z\) are nearly redundant, \(Z^\mathsf{T}Z\) is close to singular, and its inverse has large diagonal entries. The coefficient estimates are then sensitive to small changes in the data. Conditional F statistics assess whether each exposure is predictably distinguished by the instruments; they complement, rather than replace, inspection of collinearity and coefficient uncertainty.

A common rule of thumb treats an F statistic near or below 10 as a warning for weak instruments. It is not a guarantee of adequate strength above 10, nor a universal decision threshold for every MVMR design. The effective strength needed depends on the setting, and weak-instrument bias and uncertainty can be complex when several exposures and correlated instruments are involved.

Diagnostics, Alternatives, and Limitations

Check that each model is full rank, that the genotype terms are jointly tested rather than judged one at a time, and that the residual degrees of freedom match the fitted model. In individual-level data, inspect residual behavior and whether a robust variance calculation is needed. In summary-data MVMR, the RSS formula is not generally applied directly to the reported SNP-exposure estimates as if they were individual observations. Use a summary-data conditional-strength method that accounts for uncertainty in those estimates and, where relevant, their covariance and linkage disequilibrium.

Marginal F statistics are useful for asking whether variants predict each exposure separately, but they can miss weak conditional prediction. A conditional partial \(R^2\) is another way to express the incremental predictive contribution; it is directly related to the F statistic through the formula above. Measures of collinearity, such as correlations among predictor columns or condition indices, address redundancy in the design. No single diagnostic replaces the others.

If conditional strength is weak, possible responses include obtaining stronger or additional instruments, reconsidering whether highly overlapping exposures can be separated in the available data, and reporting the resulting uncertainty clearly. Robust estimators aimed at other sources of bias do not automatically repair weak conditional identification. Finally, a good conditional F statistic addresses relevance in the first stage only. Causal interpretation still relies on the relevant MR assumptions, which include instrument independence and the exclusion restriction.

Common Mistakes

• Using marginal strength as a substitute. Strong separate SNP–exposure associations do not show that the instruments distinguish each exposure from the others. Calculate a conditional statistic for each exposure.
• Giving the F statistic a causal interpretation. It measures conditional prediction, not the size or direction of a causal effect and not the absence of pleiotropy.
• Using an exact F reference without its assumptions. Independent, constant-variance errors alone are insufficient for the exact finite-sample reference; conditional normality is also required in the classical model.
• Ignoring the model’s degrees of freedom. For the stated full model, the residual degrees of freedom are \(n-K-L\); a different set of predictors changes this count.
• Treating a rule of thumb as a guarantee. A value above 10 does not certify that weak-instrument concerns are absent.