Why Penalize Variant-Specific Direct Effects?
Inverse-variance weighted (IVW) Mendelian randomization estimates a common causal effect by fitting variant–outcome associations to variant–exposure associations through the origin. If every variant is a valid instrument, this is a natural summary-data model. But a variant may also affect the outcome through a pathway other than the exposure. Such a direct effect can make that variant’s ratio estimate disagree with the others and pull the IVW estimate away from the causal effect.
MR-Lasso makes this issue explicit. It assigns each variant a direct-effect parameter, but penalizes those parameters so that many are estimated as exactly zero. The intended setting is one in which most variants are valid and only a relatively small number have substantial direct effects. The method is not a general correction for every form of pleiotropy: its usefulness depends on whether this sparse-outlier description is plausible.
The basic penalized estimate minimizes a weighted squared-error term plus an L1 penalty:
The tuning parameter \(\lambda\geq 0\) controls the amount of shrinkage. Larger \(\lambda\) makes nonzero direct-effect estimates harder to retain; sufficiently small fitted direct effects become exactly zero. With \(\lambda=0\), the model can fit each outcome association by assigning it a direct effect, so the causal effect is not identified by this objective alone. With very large \(\lambda\), all direct effects are forced to zero, and the fit approaches IVW. The useful range lies between those extremes.
How the Penalty Acts
For a fixed proposed causal effect \(\theta\), define the residual before adjustment as \(r_j=\hat{\beta}_{Yj}-\theta\hat{\beta}_{Xj}\). The part of the objective involving \(\alpha_j\) is a weighted squared deviation from \(r_j\), plus \(\lambda|\alpha_j|\). Its minimizer is the soft-thresholded residual:
Thus a residual whose magnitude is at most \(\lambda s_{Yj}^{2}\) is assigned a direct effect of zero. A larger residual is assigned a nonzero direct effect, but its magnitude is reduced by \(\lambda s_{Yj}^{2}\). The factor \(s_{Yj}^{2}\) matters: the same \(\lambda\) does not imply the same residual threshold for variants with different outcome-association precision. In practice, the causal effect and the direct effects are estimated jointly, so the residuals themselves change as \(\theta\) changes.
Worked Estimates and Tuning
Worked Example: An Outlying Variant Changes the IVW Estimate
Suppose four independent, harmonized variants have hypothetical associations as follows. Assume each outcome association has standard error \(0.10\), and treat the exposure associations as fixed for this illustration.
| Variant | \(\hat{\beta}_{Xj}\) | \(\hat{\beta}_{Yj}\) |
|---|---|---|
| 1 | 1.00 | 0.20 |
| 2 | 1.00 | 0.22 |
| 3 | 1.00 | 0.18 |
| 4 | 1.00 | 1.00 |
The IVW estimate is the weighted regression slope through the origin. With equal outcome standard errors, the common weights cancel:
Now set \(\lambda=4\) in the displayed MR-Lasso objective. Since \(s_{Yj}^2=0.01\), the soft-threshold is \(\lambda s_{Yj}^2=0.04\). A solution is \(\hat{\theta}_{L}=0.2133\), with \(\hat{\alpha}_{L1}=\hat{\alpha}_{L2}=\hat{\alpha}_{L3}=0\) and \(\hat{\alpha}_{L4}=0.7467\). To check the fit, the first three residuals are approximately \(-0.0133,\ 0.0067,\ -0.0333\), all within the threshold. The fourth residual before adjustment is \(1.00-0.2133=0.7867\), so soft thresholding gives \(0.7867-0.04=0.7467\).
The causal-effect estimate follows by minimizing the objective over \(\theta\) with those direct effects fixed. The first three variants contribute residuals that sum to \(-0.0400\); the fourth has a post-adjustment residual of \(0.0400\). Their total is zero, satisfying the weighted least-squares slope condition because all four \(\hat{\beta}_{Xj}\) and all weights are equal. The Lasso estimate is therefore much closer to the three mutually consistent associations than the IVW estimate. This is a hypothetical illustration, not evidence that variant 4 is actually invalid in a study.
The example also shows why the penalty must be tuned rather than chosen by inspection. A common approach is to evaluate a grid of \(\lambda\) values, identify the direct effects that are nonzero at each value, and select a model using a criterion such as the Bayesian information criterion (BIC). For a selected set of \(k\) direct effects, one BIC-style comparison refits the causal effect and those \(k\) direct effects without shrinkage, then computes \(Q+(k+1)\log J\), where \(Q\) is the weighted residual sum of squares and \(k+1\) counts the fitted causal effect and direct effects. Smaller values are preferred. This is a model-selection device, not proof that the selected variants are biologically pleiotropic. Implementations can use different details for the tuning criterion and uncertainty estimation, so the procedure used should be reported.
Worked Example: Comparing Two BIC-Style Fits
Use the four-variant data above to compare an all-zero-direct-effect fit with a model allowing a direct effect for variant 4. For the first fit, the IVW slope is \(0.400\). Its residuals are \(-0.200,-0.180,-0.220,\) and \(0.600\), so
There is one fitted parameter, \(\theta\), so this fit has BIC-style value \(48.08+\log(4)=49.47\). For the model with variant 4’s direct effect, refit \(\theta\) to the first three variants: \((0.20+0.22+0.18)/3=0.200\). Set \(\alpha_4=1.00-0.200=0.800\). The fitted residuals are zero, so \(Q=0\); there are two fitted parameters, giving \(0+2\log(4)=2.773\). By this criterion the second model is strongly preferred for these hypothetical data. The calculation compares specified candidate supports; in real data, the support arises from the penalty path and the BIC value does not establish the instrument-validity assumptions.
Interpretation, Diagnostics, and Limitations
A nonzero \(\hat{\alpha}_{Lj}\) means that the selected penalized model assigns variant \(j\) a direct effect after accounting for its exposure association. It is not, by itself, a definitive test of a biological pathway or proof that the variant violates the instrumental-variable assumptions. Conversely, a zero estimate does not prove validity: shrinkage can conceal modest direct effects, particularly when the tuning parameter is large.
Check whether the result is stable across reasonable tuning choices, whether the selected variants have unusually large residuals, and whether the estimated causal effect changes sharply when the tuning parameter changes. Also examine whether a few variants dominate the result, whether exposure associations are sufficiently strong, and whether variant dependence or allele alignment has been handled correctly. A basic Lasso analysis that ignores correlated variants or substantial uncertainty in \(\hat{\beta}_{Xj}\) may give misleading estimates. In particular, treating noisy exposure associations as fixed can create regression dilution or otherwise distort the fitted relationship.
Worked Example: When Pleiotropy Is Not Sparse
Suppose three hypothetical variants have exposure associations \(0.10,\ 0.20,\) and \(0.30\). Let the true causal effect be \(0.50\), but suppose every variant also has the same direct effect \(0.10\). Their expected outcome associations are therefore \(0.15,\ 0.20,\) and \(0.25\), respectively. The IVW estimate is
The discrepancy from \(0.50\) arises because all variants carry direct effects, so the sparse-outlier premise fails. An MR-Lasso fit may select some direct effects or adjust the slope, but the data do not contain a majority of valid variants to anchor the causal estimate in the intended way. A penalty cannot create that information. This example is a warning about assumptions, not a claim that the Lasso must return exactly the IVW estimate.
Compared with IVW, MR-Lasso trades the simpler all-valid-instruments model for a sparse direct-effect model and can reduce the influence of a small number of outlying variants. Radial MR and MR-PRESSO provide complementary ways to identify or assess outlying contributions; they are diagnostics rather than interchangeable versions of the same penalized estimator. Unlike a procedure that merely flags observations, MR-Lasso estimates direct effects jointly with the causal effect. Its limitations include dependence on tuning, potential instability when many variants have similar evidence of invalidity, and uncertainty after selecting which direct effects are nonzero. Standard errors that ignore this selection may understate uncertainty; uncertainty should be calculated using a method appropriate to the fitted procedure and the study design.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If \(\lambda\) increases, what generally happens to the number and magnitude of estimated direct effects?
- For a variant with \(s_{Yj}=0.20\) and \(\lambda=3\), calculate the soft-threshold for its residual.
- Why does a zero Lasso estimate of \(\alpha_j\) not prove that variant \(j\) is valid?
- State the sparse-pleiotropy assumption in words and explain why a shared direct effect across all variants can violate it.
Solutions. (1) Direct effects are more strongly shrunk, so fewer will generally remain nonzero and the retained effects will generally be smaller; the jointly estimated \(\theta\) can also change. (2) The threshold is \(\lambda s_{Yj}^2=3(0.20^2)=0.120\) in outcome-association units. (3) The estimate may be zero because the penalty shrank a modest effect to zero, or because the data lack power to distinguish it from zero; validity is not established by selection alone. (4) The assumption is that only a minority of the instruments have nonzero direct effects, leaving enough valid variants to identify the causal relationship. If every variant shares a direct effect, invalidity is not confined to a sparse subset, so the intended anchor of mostly valid variants is absent.