Course 6: Doing and Critiquing a Two-Sample MR Study
Course06-DoingAndCritiquingATwoSampleMRStudy.RmdWhat you’ll learn
By the end of this chapter, you should be able to:
- describe an end-to-end two-sample MR study workflow,
- identify major design and interpretation checkpoints,
- use Medusa outputs to critique an analysis,
- explain when an MR estimate should be treated cautiously.
Why this matters
By this point, you know the logic of MR and the mechanics of estimation. The last step is judgment: can you design a credible study, read one critically, and know when the output supports a cautious causal interpretation?
A practical study checklist
Before running an MR study, ask:
- Is the exposure well defined and biologically coherent?
- Are the instruments strong enough?
- Are the outcome data appropriate for the question?
- Is allele harmonization trustworthy?
- Are the sensitivity analyses informative?
- Are the limitations stated honestly?
This checklist is often more important than the final p-value.
A small end-to-end synthetic analysis
library(Medusa)
set.seed(1006)
beta_grid <- seq(-1.5, 1.5, by = 0.05)
sim_data <- simulateMRData(
n = 5000,
nSnps = 6,
trueEffect = 0.30,
seed = 1006
)
profile <- fitOutcomeModel(
cohortData = sim_data$data,
covariateData = NULL,
instrumentTable = sim_data$instrumentTable,
betaGrid = beta_grid,
siteId = "site_A",
analysisType = "perSNP"
)
#> Fitting outcome model at site 'site_A' (3141 cases, 1859 controls)...
#> Site 'site_A': beta_ZY_hat = 0.7014 (SE = 0.1192).
combined <- poolLikelihoodProfiles(list(site_A = profile))
#> Pooling profile likelihoods from 1 site(s)...
#> Pooling complete: 1 sites, 3141 total cases, 1859 total controls.
estimate <- computeMREstimate(combined, sim_data$instrumentTable)
#> MR estimate: beta = 1.9712 (95% CI: 1.4080, 2.5344), p = 3.52e-08
#> Odds ratio: 7.179 (95% CI: 4.088, 12.609)
sensitivity <- runSensitivityAnalyses(
profile$perSnpEstimates,
methods = c("IVW", "MREgger", "WeightedMedian", "LeaveOneOut"),
engine = "internal"
)
#> Running sensitivity analyses with 6 SNPs...
#> Engine: internal
#> IVW...
#> MR-Egger...
#> Weighted Median...
#> Leave-One-Out...
#> Sensitivity analyses complete.How to critique the analysis
1. Instrument quality
Check:
- approximate F-statistics,
- whether instruments come from a biologically sensible region,
- whether the number of SNPs is enough to support meaningful sensitivity analyses.
2. Harmonization and data integrity
Check:
- allele labels,
- palindromic SNP handling,
- allele-frequency consistency,
- genotype missingness.
3. Estimator agreement
Check:
- whether IVW, weighted median, and MR-Egger point in the same general direction,
- whether one SNP dominates leave-one-out results,
- whether uncertainty is very large.
4. Interpretation discipline
Even a well-conducted MR result should be described carefully.
Good language:
- “consistent with a causal effect,”
- “supports a likely causal role,”
- “evidence is compatible with…”
Overstated language:
- “proves causation,”
- “settles the question,”
- “guarantees a drug will work.”
Key takeaway: the output of MR is evidence, not certainty.
Common reasons not to trust an MR estimate
- weak instruments,
- obvious horizontal pleiotropy,
- severe disagreement across methods,
- poor harmonization,
- implausible biology,
- strong chance of sample selection artifacts,
- outcome definitions that do not match the scientific question.
Exercise set 1
Exercise 1
If IVW is positive, weighted median is near zero, and leave-one-out changes sign when one SNP is removed, what should you conclude?
Show solution
You should conclude that the result is unstable and should not be treated as robust causal evidence without deeper investigation. One SNP may be overly influential, or the methods may be reacting to invalid instruments.
Exercise set 2: critique a mock abstract
Mock claim:
We found a statistically significant MR estimate linking biomarker X to disease Y, therefore biomarker X is definitively causal and should be targeted therapeutically.
What is missing from this claim?
Show solution
The claim ignores instrument validity, pleiotropy, sensitivity analyses, harmonization quality, uncertainty, and the difference between causal evidence and therapeutic success. A responsible abstract would discuss the assumptions and supporting diagnostics rather than making a definitive statement.
Where to go next
You now have the conceptual foundation to:
- read ordinary two-sample MR papers critically,
- understand why sensitivity analyses matter,
- understand how Medusa’s federated workflow connects to textbook MR logic.
For next steps inside the package:
- Getting Started with Medusa for a runnable workflow,
- IL-6 Signaling and Colorectal Cancer for a scientific walkthrough,
- Federated Analysis Guide for network-level deployment.
Chapter summary
- A good MR study is judged by design quality and robustness, not only by the headline estimate.
- You should inspect instrument strength, harmonization, diagnostics, and sensitivity analyses together.
- Medusa provides tools that support this critical workflow, but the final judgment remains scientific and epidemiologic.
References
- Davies, N.M., Holmes, M.V., & Davey Smith, G. (2018). Reading Mendelian randomisation studies: a guide for clinicians. BMJ, 362, k601.
- Burgess, S., Davey Smith, G., Davies, N.M., et al. (2019). Guidelines for performing Mendelian randomization investigations. Wellcome Open Research, 4, 186.
- Skrivankova, V.W., Richmond, R.C., Woolf, B.A.R., et al. (2021). Strengthening the reporting of observational studies in epidemiology using Mendelian randomization: the STROBE-MR statement. JAMA, 326(16), 1614–1621.