CoChat
Sign Up
Cited by 810Genetic Associations and EpidemiologyAdvanced Causal Inference TechniquesGenetics and Physical Performance

Commentary: Two-sample Mendelian randomization: opportunities and challenges

Open access
Lawlor, Debbie A·2016·International Journal of Epidemiology
DOI 10.1093/ije/dyw127PMC PMC5005949
Open this paper in CoChat — free accountOpen PDF

Abstract

In this volume of the IJE , Gao and colleagues explore the causal effect of adiposity on several cancers using two-sample Mendelian randomization (MR), and find some evidence that greater adult body mass index (BMI) causally reduces the risk of breast cancer while increasing ovarian, lung and colorectal cancer. 1 The authors conclude that the study provides ‘…additional understanding of the complex relationship between adiposity and cancer risks’. Beyond the study findings themselves, this paper is interesting in its use of publicly available genome-wide association study (GWAS) summary data in a two-sample MR approach. Whereas MR has been increasingly used over the past decade since it was first proposed in the IJE , 2 two-sample MR is a relatively recent extension. 3 With the increasing availability of complete summary results from GWAS consortia that are easily accessible on the internet, the use of two-sample MR is likely to increase considerably over the next decade. 4 By complete summary data I mean results for all genetic loci with a trait or disease outcome, and not just those reaching a pre-specified P -value threshold as shown in journal publications. It is this extensive availability of results that allows Gao et al. to relate genetic variants used as instrumental variables for adiposity traits to the cancer outcomes that they are interested in; the published GWAS for the cancer outcomes would not have reported on the adiposity variants unless they reached a pre-specified GWAS significance value. MR is a form of instrumental variable analysis that uses genetic variants as instrumental variables. It has been described as nature’s or your god’s randomized controlled trial (RCT), referring to the random allocation of genetic variants at conception that mean genetic variants are less likely to violate some of the assumptions of instrumental variable analyses than non-genetic instruments. 5,6 That analogy with RCTs is also useful when teaching MR to epidemiologists who feel anxious about genetics, as illustrated in Figure 1 which provides a revision of instrumental variable analyses and the key assumptions of this approach using an example of determining the causal effect of low-density lipoprotein cholesterol (LDLc) on coronary heart disease (CHD) via an RCT ( Figure 1 a and b) and via MR ( Figure 1 c and d). In both approaches the assumption that is most likely to be violated is the ‘exclusion restriction criteria’, 7–9 which states that the instrumental variable is only related to the outcome via its effect on the risk factor. Figure 1 also shows that, if the instrumental variable assumptions are correct, the estimate of the causal effect of the risk factor on outcome can be obtained from a simple ratio of the effect of the instrumental variable on outcome divided by the effect of the instrumental variable on risk factor. In the paper by Gao and colleagues, causal effect = log odds of cancer per allele of combined adiposity genetic variants ÷ regression coefficient of adiposity measure per allele of combined adiposity genetic variants. DAG of instrumental variable analyses in an RCT and MR study exploring the effect of LDLc on CHD. These are directed acyclic graphs (DAGs), thus the absence of an arrow between any two variables (nodes) indicates we do not consider it plausible that there is a causal effect between those two. Figure shows DAGs of instrumental variable (IV) analyses to test the causal effect of low-density lipoprotein cholesterol (LDLc) on CHD. In Figure 1 a and b, the IV is randomization to receiving a statin or not (i.e this is an example of IV analyses in an RCT). In Figure 1 c and d, the IV is genetic variants that are robustly related to LDLc (i.e. this is a Mendelian randomization study). Figure 1 a and c both illustrate the three key assumptions of IV analyses: i. that the IV ‘Z’ (randomization to statins in Figure 1 a and genetic variants related to LDLc in Figure 1 c) is (or is plausibly) causally related to the risk factor (LDLc in all figures); ii. that confounding factors for the risk factor-outcome ‘X’-’Y’ association (here LDLc on CHD in all figures) are not related to the instrumental variable; iii. that the instrumental variable ‘Z’ only affects the outcome ‘Y’ (CHD) through its effect on the risk factor ‘X’ (LDLc). Two-sample MR exploits the fact that it is not necessary to obtain the effect of the instrumental variable-risk factor association (ratio denominator) and instrumental variable-outcome association (ratio numerator) from the same sample of participants. Indeed, there are some advantages to obtaining them from two different sets of participants. In particular ‘winners’ curse’, which can underestimate true causal effects in one-sample MR, 10 is unlikely to happen in two-sample MR, and unlike the impact of weak instrument bias in one-sample MR (which biases effects towards the confounded multivariable regression result), in two-sample MR weak instrument bias is

AI snapshot

Written from the paper’s full text.

TL;DR

This commentary reviews the opportunities and challenges of two-sample Mendelian randomization (MR), using Gao and colleagues' study of adiposity and cancer risk as an illustrative case. It argues that while two-sample MR using publicly available GWAS summary data increases statistical power and avoids winner's curse, it introduces pitfalls around sample overlap, non-sex-specific instruments, covariate adjustment, and weak instrument bias that authors must address.

Key findings

  • Two-sample MR has key advantages over one-sample MR: winner's curse is avoided, and weak instrument bias is towards the null rather than towards the confounded multivariable regression result; the main advantage of GWAS summary data is increased statistical power, illustrated by Gao et al.'s cancer case numbers (15,748 breast, 5,100 colorectal, 12,160 lung, 4,369 ovarian, 14,160 prostate).
  • The commentary critiques Gao et al.'s conclusions: they largely base conclusions on P < 0.05 after multiple testing, and some BMI and WHR point estimates are similar (e.g. BMI–breast cancer OR 0.66 [95% CI 0.57, 0.77] concluded, but WHR 0.73 [0.53, 1.00] not; BMI–lung cancer OR 1.27 [1.09, 1.49] declared positive, but WHR–squamous lung cancer OR 1.33 [0.75, 2.36] not).
  • Potential biases in Gao et al. are identified: WHR variants were adjusted for BMI (biasing the MR estimate towards the null); the adult BMI instrument is stronger than others (R² for adult BMI 0.027, birthweight 0.008, child BMI 0.020, WHR 0.013); and sex-combined instrument–adiposity associations were used for sex-specific cancers, though 7 of 14 WHR variants were stronger in females (combined variants explained 1.34% of WHR variation in women vs 0.46% in men).
  • The commentary notes MR cannot easily distinguish effects at different life stages because 12 of 15 child BMI variants overlapped with adult BMI variants, and MR-Egger cannot differentiate adult from child BMI effects.

Methods

Not a primary study; a commentary/review. It synthesizes methodological literature on one- and two-sample MR, presents a comparison table (Table 1) of assumptions and issues across the two designs, and critically appraises Gao et al.'s two-sample MR study of adiposity traits and five cancers, including re-examination of the original GWAS papers for sex-specific effects (Table 2).

Limitations

The paper states that the extent to which bias towards the null (from weak instrument bias and WHR adjustment for BMI) is balanced by possible exaggeration from using non-sex-specific instruments is "impossible to tell." It also notes that for birthweight and child BMI, the original GWAS made no attempt to explore sex differences, so it is impossible to know whether the no-sex-difference assumption holds. Evident from the text: as a commentary, it offers no new primary data or formal analysis, and its critique of Gao et al. relies on the authors' supplementary materials and original GWAS publications rather than reanalysis.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1093/ije/dyw127

Cites

29 of 29 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 35 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Some point estimates for body mass index and waist-hip ratio in Gao et al. are not that dissimilar.

    “However, some of the point estimates for BMI and WHR are not that dissimilar.”
    Quote in the paperSubstantial claim
  • Premisecitation

    Gao et al. do not provide any information on the strength of the different instrumental variables, such as the F-statistic or R2 for the genetic instrument-adiposity trait associations.

    “Gao and colleagues do not provide any information on the strength of the different instrumental variables, such as the F-statistic or R 2 for the genetic instrument-adiposity trait associations.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1] Gao, 2016, Mendelian randomization study of adiposity-related traits and risk of breast, ovarian, prostate, lung and colorectal cancer, Int J Epidemiol source

      Not checked in the free run.

  • Premisereasoning

    Gao et al. were unable with the summary data available to test differences in effect between pre- and postmenopausal breast cancer because GWAS separated by these sub-phenotypes are not presented by the breast cancer consortia.

    “They are unable with the summary data available to test differences in effect between pre- and postmenopausal breast cancer, as GWAS separated by these sub-phenotypes are not presented by the breast cancer consortia.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1] Gao, 2016, Mendelian randomization study of adiposity-related traits and risk of breast, ovarian, prostate, lung and colorectal cancer, Int J Epidemiol source

      Not checked in the free run.

  • Premisecitation

    The waist-hip ratio variants used by Gao et al. were adjusted for body mass index, which the authors do not seem to acknowledge.

    “The WHR variants used by Gao and colleagues were adjusted for BMI, which the authors do not seem to acknowledge.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [19] source

      Not checked in the free run.

  • Interpretationown data

    A further potential explanation for why most emphasized Mendelian randomization results are for adult body mass index rather than other adiposity risk factors is that the genetic instrument for adult body mass index is stronger than for the other traits.

    “A further potential explanation for why most of the emphasized (based on statistical testing) MR results are seen for adult BMI, rather than any of the other adiposity risk factors, is that the genetic instrument for adult BMI is stronger than for the other traits.”
    Quote in the paperNot a substantial claim
  • May combine two claims. 14
  • Classified as probably not a substantial claim. 6

  • May combine two claims.

    Interpretationreasoning

    Gao et al. are not really able to determine effects of body mass index at different life stages because body mass index assessed at different ages is correlated and because of the nature of Mendelian randomization.

    “Gao and colleagues examine the effects of childhood BMI and adult BMI, but they are not really able to determine effects at different life stages because of the correlation between BMI assessed at different ages and because of the nature of MR.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    Greater adult body mass index, but not waist-hip ratio, is concluded by Gao et al. to decrease breast cancer and increase ovarian, lung and colorectal cancer risk.

    “Greater adult BMI, but not waist-hip ratio (WHR), is concluded to decrease breast cancer and increase ovarian, lung and colorectal cancer risk.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1] Gao, 2016, Mendelian randomization study of adiposity-related traits and risk of breast, ovarian, prostate, lung and colorectal cancer, Int J Epidemiol source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    In two-sample Mendelian randomization, weak instrument bias is towards the null, unlike one-sample Mendelian randomization where it biases effects towards the confounded multivariable regression result.

    “unlike the impact of weak instrument bias in one-sample MR (which biases effects towards the confounded multivariable regression result), in two-sample MR weak instrument bias is towards the null.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    'Winner's curse' is unlikely to happen in two-sample Mendelian randomization, whereas it can underestimate true causal effects in one-sample Mendelian randomization.

    “In particular ‘winners’ curse’, which can underestimate true causal effects in one-sample MR, 10 is unlikely to happen in two-sample MR”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [10] source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    Gao et al. analysed large numbers of cancer cases: 15,748 breast, 5,100 colorectal, 12,160 lung, 4,369 ovarian, and 14,160 prostate cancer.

    “The paper by Gao and colleagues illustrates this with the large numbers of cancer cases that they analyse: 15 748, 5100, 12 160, 4369 and 14 160, respectively, for breast, colorectal, lung, ovarian and prostate cancer.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1] Gao, 2016, Mendelian randomization study of adiposity-related traits and risk of breast, ovarian, prostate, lung and colorectal cancer, Int J Epidemiol source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Mendelian randomization generally assesses cumulative effects of a risk factor over a long period of the life course without requiring repeat risk factor assessment and with little chance of regression dilution bias or reverse causation.

    “Whereas one of the noted advantages of MR is that it generally assesses the cumulative effects of a risk factor over a long period of the life course (potentially from conception) without requiring repeat risk factor assessment and with little chance of regression dilution bias or reverse causation (confounding by prevalent disease), 7”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [7] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Of the 15 child body mass index variants, 12 overlapped with known adult body mass index variants, illustrating the difficulty of distinguishing child and adult body mass index effects.

    “From the original GWAS, 12 of the 15 child BMI overlapped with known adult BMI variants, 14 which illustrates the difficulty of distinguishing these two.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [14] source

      Not checked in the free run.

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    Gao et al. found some evidence that greater adult body mass index causally reduces the risk of breast cancer while increasing ovarian, lung and colorectal cancer.

    “Gao and colleagues explore the causal effect of adiposity on several cancers using two-sample Mendelian randomization (MR), and find some evidence that greater adult body mass index (BMI) causally reduces the risk of breast cancer while increasing ovarian, lung and colorectal cancer.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1] Gao, 2016, Mendelian randomization study of adiposity-related traits and risk of breast, ovarian, prostate, lung and colorectal cancer, Int J Epidemiol source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    MR-Egger cannot be used to differentiate effects of adult from child body mass index because it is only valid if the effect of the genetic instrument on the risk factor is independent of its effect on other phenotypes that might violate the exclusion restriction.

    “Furthermore the MR-Egger test, 9 which the authors used to test violation of the exclusion restriction criteria, cannot be used to differentiate effects of adult from child BMI, as Gao and colleagues acknowledge. This is because MR-Egger is only valid if the effect of the genetic instrument on the risk factor of interest is independent of its effect on any other phenotypes that might violate that”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [9] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning

    This study illustrates some of the pitfalls of using summary GWAS data and methods that might be used to limit these.

    “But this study does illustrate some of the pitfalls of using summary GWAS data and methods that might be used to limit these.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationcitation

    Those retrospective findings could have been identified prospectively with easy open access to complete summary data of everything tested in all randomized controlled trials.

    “Those retrospective findings could have been identified prospectively with easy open access to complete summary data of everything tested in all RCTs.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: We still lack open access to complete summary data of everything tested in all randomized controlled trials.

  • May combine two claims.

    Premisecitation

    For adult body mass index, sex differences were reported and marked differences were found for two of the 77 variants (stronger associations in women compared with men).

    “For adult BMI, sex differences were reported and marked differences were found for two of the 77 variants (stronger associations in women compared with men).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [20] source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Because the per allele effect of the genetic instrument on waist-hip ratio adjusted for body mass index is the denominator of the Mendelian randomization ratio estimate, the estimated effect of waist-hip ratio adjusted for body mass index for female cancers (breast and ovarian) may be exaggerated and those for prostate cancers underestimated.

    “As this is the denominator of the MR ratio estimate, it means that the estimated effect of WHR adjusted for BMI for female cancers (breast and ovarian) may be exaggerated and those for prostate cancers underestimated.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Although Gao et al. use the most up-to-date body mass index GWAS data, they do not do the same for waist-hip ratio, despite the most recent GWAS for waist-hip ratio adjusted for body mass index identifying 33 additional variants and being published around the same time.

    “Interestingly, although Gao et al. use the most up-to-date BMI GWAS data, 20 they do not do the same for WHR, despite the most recent GWAS for WHR adjusted for BMI identifying 33 additional variants (as well as confirming the 14 used here from the earlier GWAS) and being published around the same time as the most up-to-date BMI GWAS.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [20] source

      Not checked in the free run.

    • cites [21] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    According to Gao's Supplementary Table 1, the association of the genetic instrumental variable with each adiposity trait was taken from samples combining females and males, whereas for breast and ovarian cancer females only were included and for prostate cancer males only were included.

    “According to data presented by Gao in their Supplementary Table 1, it seems that the association of the genetic instrumental variable with each adiposity trait has been taken from samples that combine females and males, whereas for the association of the genetic instrument with breast and ovarian cancer, females only are included and with prostate cancer males only are included.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1] Gao, 2016, Mendelian randomization study of adiposity-related traits and risk of breast, ovarian, prostate, lung and colorectal cancer, Int J Epidemiol source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Seven of the 14 waist-hip ratio adjusted for body mass index variants used by Gao et al. were stronger in females compared with males, with 19 of the 44 variants in the more up-to-date GWAS being stronger in females (and one stronger in males).

    “Seven of the 14 WHR adjusted for BMI variants used by Gao and colleagues were stronger in females compared with males ( Table 2 ), with 19 of the 44 variants in the more up-to-date GWAS being stronger in females (and one stronger in males).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [19] source

      Not checked in the free run.

    • cites [21] source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The extent to which bias towards the null from weak instrument bias and adjustment of waist-hip ratio for body mass index is balanced by possible exaggeration from not using sex-specific data for the genetic instrument-waist-hip ratio association in female cancers is impossible to tell.

    “The extent to which bias towards the null as a possible result of weak instrument bias and adjustment of WHR for BMI (discussed above) is balanced by possible exaggeration of the true effect as a result of not using sex-specific data for the genetic instrument-WHR association in the female cancers, is impossible to tell.”
    Quote in the paperSubstantial claim
17 of 17

  • The paper relies on this without support.

    We still lack open access to complete summary data of everything tested in all randomized controlled trials.

    “However, we still lack such access.”

  • The authors state this as an assumption.

    The instrumental variable is only related to the outcome via its effect on the risk factor (the exclusion restriction criterion).

    “the assumption that is most likely to be violated is the ‘exclusion restriction criteria’, 7–9 which states that the instrumental variable is only related to the outcome via its effect on the risk factor.”

  • The authors state this as an assumption.

    Confounding factors for the risk factor-outcome association are not related to the instrumental variable.

    “ii. that confounding factors for the risk factor-outcome ‘X’-’Y’ association (here LDLc on CHD in all figures) are not related to the instrumental variable;”

  • The authors state this as an assumption.

    The instrumental variable is causally related to the risk factor.

    “i. that the IV ‘Z’ (randomization to statins in Figure 1 a and genetic variants related to LDLc in Figure 1 c) is (or is plausibly) causally related to the risk factor (LDLc in all figures);”
4 of 4

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Two methods (MR-Egger and the weighted median estimator) are likely to become increasingly used in MR sensitivity analyses as two-sample MR using summary GWAS data becomes more common.

    “Two of these methods in particular (MR-Egger36 and the weighted median estimator37) are likely to become increasingly used in MR sensitivity analyses, as two-sample MR using summary genome-wide association studies (GWAS) data becomes more common.38”

    From Triangulation in aetiological epidemiology (2017) · cites this paper as [38]

  • MR-Egger and weighted median estimator could be used as a sensitivity analysis to assess violation of the exclusion restriction criteria when several different treatment IVs are used in RCTs.

    “Although these approaches have been developed for use in MR, and in particular in the case of MR-Egger for two-sample MR,36,38 they could clearly be used as a sensitivity analysis to assess violation of the exclusion restriction criteria when several different treatment IVs are used in RCTs.”

    From Triangulation in aetiological epidemiology (2017) · cites this paper as [38]

  • Many genome-wide association consortia have now made their aggregate results fully available, with massive potential for these to increase aetiological understanding, including mechanistic insights, through two-sample MR.

    “Many genome-wide association consortia have now made their aggregate results fully available (i.e. all genetic associations for up to several million genetic variants with an ever-increasing number of outcomes from molecular to behavioural phenotypes, irrespective of P-values), with massive potential for these to increase aetiological understanding, including mechanistic insights, through two”

    From Triangulation in aetiological epidemiology (2017) · cites this paper as [38]

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Smith ADAC Hardy R Heron J . Model selection of life course hypotheses involving continuous exposures. Int J Epidemiol 2016; in review.

Cited by

Sorted by citations

Related

Authors
Lawlor, D. A.
Affiliations
University of Bristol (GB)
Corresponding author
Lawlor, Debbie A
ORCID
Lawlor, Debbie A
Venue
International Journal of Epidemiology
Publisher
Oxford University Press (OUP)
Published
2016-06
Volume
45
Issue
3
Pages
908-915
ISSN
  • 0300-5771 (ISSN-L)
  • 1464-3685
DOI
10.1093/ije/dyw127
PMC
PMC5005949
OpenAlex
W2490528887
Cited by
810
References
29
Citations by year
2026: 42 · 2025: 97 · 2024: 172 · 2023: 170 · 2022: 81 · 2021: 83 · 2020: 48 · 2019: 59 · 2018: 38 · 2017: 14 · 2016: 6
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • International Journal of Epidemiology (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Adelaide Research & Scholarship (AR&S) (University of Adelaide) (submittedVersion, cc-by)
  • HighWire Press Open Archive (submittedVersion)
Funders
  • Universidade Federal de Pelotas
  • National Institute for Health and Care Research (NF-SI-0611-10196, SI-0611-10196)
  • University of Exeter
  • University of Bristol (NF-SI-0611-10196, MC_UU_12013, MC_UU_12013/5)
  • Medical Research Council (MC_UU_12013_5, MC_UU_12013, NF-SI-0611-10196)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Genetic Associations and Epidemiology
  • Physical Sciences › Mathematics › Statistics and Probability › Advanced Causal Inference Techniques
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Genetics and Physical Performance
Keywords
two-sample Mendelian randomization, adiposity, body mass index, cancer risk, breast cancer, ovarian cancer, lung cancer, colorectal cancer, exclusion restriction, GWAS summary data, causal inference, genetic variants, causal effect estimation, confounding
MeSH terms
Genotype, Humans, Random Allocation, Mendelian Randomization Analysis

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-10.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-10.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review