[ad_1]
- Jeremy P Brown, doctoral researcher1,
- Jacob N Hunnicutt, director2,
- M Sanni Ali, assistant professor1,
- Krishnan Bhaskaran, professor1,
- Ashley Cole, director3,
- Sinead M Langan, professor1,
- Dorothea Nitsch, professor1,
- Christopher T Rentsch, affiliate professor1,
- Nicholas W Galwey, statistics chief4,
- Kevin Wing, assistant professor1,
- Ian J Douglas, professor1
1Division of Non-Communicable Illness Epidemiology, London Faculty of Hygiene and Tropical Drugs, London, UK
2Epidemiology, Worth Proof and Outcomes, R&D International Medical, GSK, Collegeville, PA, USA
3Actual World Analytics, Worth Proof and Outcomes, R&D International Medical, GSK, Collegeville, PA, USA
4R&D, GSK Medicines Analysis Centre, GSK, Stevenage, UK
- Correspondence to: J P Brown jeremy.brown{at}lshtm.ac.uk (or @jeremy_pbrown on X)
- Accepted 12 February 2024
Bias in epidemiological research is a serious concern. Biased research have the potential to mislead, and in consequence to negatively have an effect on medical follow and public well being. The potential for residual systematic error on account of measurement bias, confounding, or choice bias is commonly acknowledged in publications however is seldom quantified.1 Subsequently, for a lot of research it’s troublesome to evaluate the extent to which residual bias may have an effect on research findings, and the way assured we ought to be about their conclusions. More and more massive datasets with thousands and thousands of sufferers can be found for analysis, equivalent to insurance coverage claims knowledge and digital well being information. With rising dataset dimension, random error decreases however bias stays, probably resulting in incorrect conclusions.
Sensitivity analyses to quantify potential residual bias can be found.234567 Nonetheless, use of those strategies is restricted. Efficient use sometimes requires enter from a number of events (together with clinicians, epidemiologists, and statisticians) to deliver collectively medical and area space information, epidemiological experience, and a statistical understanding of the strategies. Improved consciousness of those strategies and their pitfalls will allow extra frequent and efficient implementation, in addition to vital interpretation of their utility within the medical literature.
On this article, we goal to supply an accessible introduction, description, and demonstration of three widespread approaches of quantitative bias evaluation, and to explain their potential limitations. We briefly assessment bias in epidemiological research on account of measurement error, confounding, and choice. We then introduce quantitative bias analyses, strategies to quantify the potential influence of residual bias (ie, bias that has not been accounted for by means of research design or statistical evaluation). Lastly, we talk about limitations and pitfalls within the utility and interpretation of those strategies.
Abstract factors
-
Quantitative bias evaluation strategies permit investigators to quantify potential residual bias and to objectively assess the sensitivity of research findings to this potential bias
-
Bias formulation, bounding strategies, and probabilistic bias evaluation can be utilized to evaluate sensitivity of outcomes to potential residual bias; every of those approaches has strengths and limitations
-
Quantitative bias evaluation depends on assumptions about bias parameters (eg, the power of affiliation between unmeasured confounder and final result), which could be knowledgeable by substudies, secondary research, the literature, or skilled opinion
-
When making use of, deciphering, and reporting quantitative bias evaluation, it is very important transparently report assumptions, to think about a number of biases if related, and to account for random error
Forms of bias
All medical research, each interventional and non-interventional, are probably susceptible to bias. Bias is ideally prevented or minimised by means of cautious research design and the selection of applicable statistical strategies. In non-interventional research, three main biases that may have an effect on findings are measurement bias (also referred to as info bias) on account of measurement error (known as misclassification for categorical variables), confounding, and choice bias.
Misclassification happens when a number of categorical variables (such because the publicity, final result, or covariates) are mismeasured or misreported.8 Steady variables may also be mismeasured resulting in measurement error. As one instance, misclassification happens in some research of alcohol consumption owing to misreporting by research individuals of their alcohol consumption.910 As one other instance, research utilizing digital well being information or insurance coverage claims knowledge may have final result misclassification if the result isn’t at all times reported to, or recorded by, the person’s healthcare skilled.11 Measurement error is claimed to be differential when the chance of error is dependent upon one other variable (eg, differential participant recall of publicity standing relying on the result). Errors in measurement of a number of variables might be dependent (ie, related to one another), significantly when knowledge are collected from one supply (eg, digital well being information). Measurement error can result in biased research findings in each descriptive and aetiological (ie, cause-effect) non-interventional research.12
Confounding arises in aetiological research when the affiliation between publicity and final result isn’t solely as a result of causal impact of the publicity, however slightly is partly or wholly on account of a number of different causes of the result related to the publicity. For instance, researchers have discovered that better adherence to statins is related to a discount in motorcar accidents and a rise in the usage of screening providers.13 Nonetheless, this affiliation is sort of actually not on account of a causal impact of statins on these outcomes, however extra most likely as a result of attitudes to precaution and danger which can be related to these outcomes are additionally related to adherence to statins.
Choice bias happens when non-random choice of folks or individual time into the research ends in systematic variations between outcomes obtained within the research inhabitants and outcomes that may have been obtained within the inhabitants of curiosity.1415 This bias could be on account of choice at research entry or on account of differential loss to follow-up. For instance, in a cohort research the place the sufferers chosen are these admitted to hospital in respiratory misery, covid-19 and continual obstructive pulmonary illness may be negatively related, even when there was no affiliation within the total inhabitants, as a result of in case you shouldn’t have one situation it’s extra probably you have got the opposite situation with a purpose to be admitted.16 Choice bias can have an effect on each descriptive and aetiological non-interventional research.
Dealing with bias in follow
All three biases ought to ideally be minimised by means of research design and evaluation. For instance, misclassification could be lowered by way of a extra correct measure, confounding by means of measurement of all related potential confounders and their subsequent adjustment, and choice bias by means of applicable sampling from the inhabitants of curiosity and accounting for loss to follow-up. Different biases must also be thought of, for instance, immortal time bias by means of the suitable selection of time zero, and sparse knowledge bias by means of assortment of a pattern of adequate dimension or by way of penalised estimation.1718
Even with the perfect accessible research design and most applicable statistical evaluation, we sometimes can’t assure that residual bias will likely be absent. As an example, it’s usually not doable to completely measure all required variables, or it may be both inconceivable or impractical to gather or get hold of knowledge on each doable potential confounder. As an example, research performed utilizing knowledge collected for non-research functions, equivalent to insurance coverage claims and digital well being information, are sometimes restricted to the variables beforehand recorded. Randomly sampling from the inhabitants of curiosity may also not be virtually possible, particularly if people usually are not prepared to take part.
To disregard potential residual biases can result in deceptive outcomes and misguided conclusions. Usually the potential for residual bias is acknowledged qualitatively within the dialogue, however these qualitative arguments are sometimes subjective and sometimes downplay the influence of any bias. Heuristics are steadily relied on, however these can result in an misestimation of the potential for residual bias.19 Quantitative bias evaluation permits each authors and readers to evaluate robustness of research findings to potential residual bias rigorously and quantitatively.
Quantitative bias evaluation
When designing or appraising a research, a number of key questions associated to bias ought to be thought of (field 1).20 If, on the premise of the solutions to those questions, there’s potential for residual bias(es), then quantitative bias evaluation strategies could be thought of to estimate the robustness of findings.
Key questions associated to bias when designing and appraising non-interventional research
-
Misclassification and measurement error: Are the publicity, final result, and covariates more likely to be measured and recorded precisely?
-
Confounding: Are there potential causes of the result, or proxies for these causes, which could differ in prevalence between publicity teams? Are these potential confounders measured and managed by means of research design or evaluation?
-
Choice bias: What’s the goal inhabitants? Are people within the research consultant of this goal inhabitants?
Many strategies for quantitative bias evaluation exist, though just a few of those are often utilized in follow. On this article, we’ll introduce three easy, generally utilized, and common approaches1: bias formulation, bounding strategies, and probabilistic bias evaluation. Different strategies are additionally accessible, together with strategies for bias adjustment of linear regression with a steady final result.72122 Strategies for coping with misclassification of categorical variables are outlined on this article. Corresponding strategies for sensitivity evaluation to cope with mismeasurement of steady variables can be found and are described in depth within the literature.2324
Bias formulation
We are able to use easy mathematical formulation to estimate the bias in a research and to estimate what the outcomes could be within the absence of that bias.425262728 Generally utilized formulation, together with particulars of accessible software program to implement strategies listed, are supplied within the appendices. A few of these strategies could be utilized to the abstract outcomes (eg, danger ratio), whereas different strategies require entry to 2×2 tables or participant degree knowledge.
These formulation require us to specify extra info, sometimes not obtainable from the research knowledge itself, within the type of bias parameters. Values for these parameters quantify the extent of bias current on account of confounding, misclassification, or choice.
Bias formulation for unmeasured confounding typically require us to specify the next bias parameters: prevalence of the unmeasured confounder within the unexposed people, prevalence of the unmeasured confounder within the uncovered people (or alternatively the affiliation between publicity and unmeasured confounder), and the affiliation between unmeasured confounder and final result.42829
These bias formulation could be utilized to the abstract outcomes (eg, danger ratios, odds ratios, danger variations, hazard ratios) and to 2×2 tables, they usually produce corrected outcomes assuming the required bias parameters are appropriate. Usually, the precise bias parameters are unknown so a spread of parameters could be entered into the components, producing a spread of doable bias adjusted outcomes underneath roughly excessive confounding situations.
Bias formulation for misclassification work in an identical method, however sometimes require us to specify optimistic predictive worth and destructive predictive worth (or sensitivity and specificity) of classification, stratified by publicity or final result. These formulation sometimes require research knowledge within the type of 2×2 tables.730
Bias formulation for choice bias are relevant to the abstract outcomes (eg, danger ratios, odds ratios) or to 2×2 tables, and usually require us to specify chances of choice into the research for various ranges of publicity and final result.25 When participant degree knowledge can be found, a common methodology of bias evaluation is to weight every particular person by the inverse of their chance of choice.31Field 2 describes an instance of the applying of bias formulation for choice bias.
Software of bias formulation for choice bias
In a cohort research of pregnant ladies investigating the affiliation between lithium use (relative to non-use) and cardiac malformations in liveborn infants, the noticed covariate adjusted danger ratio was 1.65 (95% confidence interval 1.02 to 2.68).32 Solely liveborn infants have been chosen into the research; due to this fact, there was potential for choice bias if variations within the termination chances of fetuses with cardiac malformations existed between publicity teams.
As a result of the result is uncommon, the percentages ratio approximates the chance ratio, and we will apply a bias components for the percentages ratio to the chance ratio. The bias parameters are choice chances for the unexposed group with final result S01, uncovered group with final result S11, unexposed group with out final result S00, and uncovered group with out final result S10:
ORBiasAdj = ORObs × ((S01×S10) ÷ (S00×S11))
(The place ORBiasAdj is the bias adjusted odds ratio and ORObsis the noticed odds ratio.)
For instance, if we assume that the chance of terminations is 30% among the many unexposed group (ie, pregnancies with no lithium dispensation in first trimester or three months earlier) with malformations, 35% among the many uncovered group (ie, pregnancies with lithium dispensation in first trimester) with malformations, 20% among the many unexposed group with out malformations, and 25% among the many uncovered group with out malformations, then the bias adjusted odds ratio is 1.67.
ORBiasAdj = 1.65 × ((0.7×0.75) ÷ (0.65×0.8)) = 1.67
Within the research, a spread of choice chances (stratified by publicity and final result standing) have been specified, knowledgeable by the literature. Relying on assumed choice chances, the bias adjusted estimates of the chance ratio ranged from 1.65 to 1.80 (fig 1), indicating that the estimate was sturdy to this choice bias underneath given assumptions.
Bias adjusted danger ratio for various assumed choice chances in cohort research investigating affiliation between lithium use (relative to non-use) and cardiac malformations in liveborn infants. Redrawn and tailored from reference 32 with permission from Massachusetts Medical Society. Choice chance of the unexposed group with out cardiac malformations was assumed to be 0.8 (ie, 20% chance of termination). Choice chances within the uncovered group have been outlined relative to the unexposed group by final result standing (ie, −0%, −5%, and −10%)
It’s doable to include measured covariates in these formulation, however specification then typically turns into tougher as a result of we sometimes need to specify bias parameters (such because the prevalence of the unmeasured confounder) inside stratums of measured covariates.
Though we’d not be capable of estimate these unknowns from the principle research itself, we will specify believable ranges primarily based on the printed literature, medical information, or a secondary research or substudy. Secondary research or substudies, wherein extra info from a subset of research individuals or from a consultant exterior group are collected, are significantly worthwhile as a result of they’re extra more likely to precisely seize unknown values.33 Nonetheless, relying on the actual state of affairs, they might be infeasible for a given research owing to knowledge entry limitations and useful resource constraints.
The printed literature could be informative if there are related printed research and the research populations within the printed research are sufficiently just like the inhabitants underneath investigation. Subjective judgments of believable values for unknowns are susceptible to the point of view of the investigator, and in consequence may not precisely replicate the true unknown values. The validity of quantitative bias evaluation relies upon critically on the validity of the assumed values. When implementing quantitative bias evaluation, or appraising quantitative bias evaluation in a broadcast research, research investigators ought to query the alternatives made for these unknowns, and report these decisions with transparency.
Bounding strategies
Bounding strategies are mathematical formulation, just like bias formulation, that we will apply to review outcomes to quantify sensitivity to bias on account of confounding, choice, and misclassification.5343536 Nonetheless, in contrast to bias formulation, they require solely a subset of the unknown values to be specified. Whereas this requirement appears advantageous, one essential drawback is that bounding strategies generate a sure on the utmost doable bias, slightly than an estimate of the affiliation adjusted for bias. When values for all unknown parameters (eg, prevalence of an unmeasured confounder) could be specified and there’s cheap confidence of their validity, bias formulation or probabilistic bias evaluation can typically be utilized and might present extra info than bounding strategies.37
One generally used bounding methodology for unmeasured confounding is the E-value.535 Through the use of E-value formulation, research investigators can calculate a sure on the bias adjusted estimate by specifying the affiliation (eg, danger ratio) between publicity and unmeasured confounder and between unmeasured confounder and final result, whereas leaving the prevalence of the unmeasured confounder unspecified. The E-value itself is the minimal worth on the chance ratio scale that the affiliation between publicity and unmeasured confounder or the affiliation between unmeasured confounder and final result should exceed to probably cut back the bias adjusted findings to the null (or alternatively to some specified worth, equivalent to a protecting danger ratio of 0.8). If the believable power of affiliation between the unmeasured confounder and each publicity and final result is smaller than the E-value, then that one confounder couldn’t totally clarify the noticed affiliation, offering assist to the research findings. If the power of affiliation between the unmeasured confounder and both publicity or final result is plausibly bigger than the E-value, then we will solely conclude that residual confounding may clarify the noticed affiliation, however it’s not doable to say whether or not such confounding is in fact adequate, as a result of we’ve not specified the prevalence of the unmeasured confounder. Field 3 illustrates the usage of bounding strategies for unmeasured confounding. Though standard, the applying of E-values has been criticised, as a result of these values have been generally misinterpreted and have been used steadily with out cautious consideration of a particular unmeasured confounder or the potential for a number of unmeasured confounders or different biases.38
Software of bounding strategies
In a cohort research investigating the affiliation between use of proton pump inhibitors (relative to H2 receptor antagonists) and all trigger mortality, investigators discovered proof that people prescribed proton pump inhibitors have been at larger danger of demise after adjusting for a number of measured covariates together with age, intercourse, and comorbidities (covariate adjusted hazard ratio 1.38, 95% confidence interval (CI) 1.33 to 1.44).39 Nonetheless, unmeasured variations in frailty between customers of H2 receptor antagonists and customers of proton pump inhibitors may bias findings. As a result of the prevalence of the unmeasured confounder within the totally different publicity teams was unclear, the E-value was calculated. As a result of the result was uncommon on the finish of follow-up, and due to this fact the chance ratio approximates the hazard ratio given proportional hazards,40 the E-value components, which applies to the chance ratio, was utilized to the hazard ratio.
E-value = RRObs + √(RRObs×(RRObs−1))
= 1.38 + √(1.38×(1.38−1))
= 2.10
(The place RRObs is the noticed danger ratio.)
The E-value for the purpose estimate of the adjusted hazard (1.38) was 2.10. Therefore both the adjusted danger ratio between publicity and unmeasured confounder, or the adjusted danger ratio between unmeasured confounder and final result, have to be better than 2.10 to probably clarify the noticed affiliation of 1.38. The E-value could be utilized to the bounds of the CI to account for random error. The calculated E-value for the decrease sure of the 95% CI (ie, covariate adjusted hazard ratio=1.33) was 1.99. We are able to plot a curve to indicate the values of danger ratios essential to probably cut back the noticed affiliation, as estimated by the purpose estimate and the decrease sure of the CI, to the null (fig 2). An unmeasured confounder with strengths of associations beneath the blue line couldn’t totally clarify the purpose estimate, and beneath the yellow line couldn’t totally clarify the decrease sure of the boldness interval.
E-value plot for unmeasured confounding of affiliation between use of proton pump inhibitors and all trigger mortality. Curves present the values of danger ratios essential to probably cut back the noticed affiliation, as estimated by the purpose estimate and the decrease sure of the boldness interval, to the null
Given danger ratios of >2 noticed within the literature between frailty and mortality, unmeasured confounding couldn’t be dominated out as a doable clarification for noticed findings. Nonetheless, on condition that we used a bounding methodology, and didn’t specify unmeasured confounder prevalence, we couldn’t say with certainty whether or not such confounding was more likely to clarify the noticed outcome. Extra unmeasured or partially measured confounders may need additionally contributed to the noticed affiliation.
Probabilistic bias evaluation
Probabilistic bias evaluation takes a unique method to dealing with uncertainty across the unknown values. Reasonably than specifying one worth or a spread of values for an unknown, a chance distribution (eg, a standard distribution) is specified for every of the unknown portions. This distribution represents the uncertainty concerning the unknown values, and values are sampled repeatedly from this distribution earlier than making use of bias formulation utilizing the sampled values. This method could be utilized to both abstract or participant degree knowledge. The result’s a distribution of bias adjusted estimates. Resampling ought to be carried out a adequate variety of occasions (eg, 10 000 occasions), though this requirement can turn into computationally burdensome when performing corrections on the affected person report degree.41
Probabilistic bias evaluation can readily deal with many unknowns, which makes it significantly helpful for dealing with a number of biases concurrently.42 Nonetheless, it may be troublesome to specify a sensible distribution if little info on the unknowns is obtainable from printed research or from extra knowledge assortment. Generally chosen distributions embrace uniform, trapezoidal, triangular, beta, regular, and log-normal distributions.7 Sensitivity analyses could be performed by various the distribution and assessing the sensitivity of findings to distribution chosen. When performing corrections on the affected person report degree, analytical strategies equivalent to regression could be utilized after correction to regulate associations for measured covariates.43Field 4 provides an instance of probabilistic bias evaluation for misclassification.
Software of probabilistic bias evaluation
In a cohort research of pregnant ladies performed in insurance coverage claims knowledge, the noticed covariate adjusted danger ratio for the affiliation between antidepressant use and congenital cardiac defects amongst ladies with melancholy was 1.02 (95% confidence interval 0.90 to 1.15).44
Some misclassification of the result, congenital cardiac defects, was anticipated, and due to this fact probabilistic bias evaluation was performed. A validation research was performed to evaluate the accuracy of classification. On this validation research, full medical information have been obtained and used to confirm diagnoses for a subset of pregnancies with congenital cardiac defects recorded within the insurance coverage claims knowledge. Primarily based on optimistic predictive values estimated on this validation research, triangular distributions of believable values for sensitivity (fig 3) and of specificity of final result classification have been specified and have been used for probabilistic bias evaluation.
Specified distribution of values for sensitivity of final result ascertainment
Values have been sampled at random 1000 occasions from these distributions and have been used to calculate a distribution of bias adjusted estimates incorporating random error. The median bias adjusted estimate was 1.06, and the 95% simulation interval was 0.92 to 1.22.44 This discovering signifies that underneath the given assumptions, the outcomes have been sturdy to final result misclassification, as a result of the bias adjusted outcomes have been just like the preliminary estimates. Each units of estimates urged no proof of affiliation between antidepressant use and congenital cardiac defects.
Pitfalls of strategies
Incorrect assumptions
Examine investigators and readers of printed analysis ought to be conscious that the outputs of quantitative bias analyses are solely pretty much as good because the assumptions made. These assumptions embrace each assumptions concerning the values chosen for the bias parameters (desk 1), and assumptions inherent to the strategies. For instance, making use of the E-value components on to a hazard ratio slightly than a danger ratio is an approximation, and solely a superb approximation when the result is uncommon.45
Widespread bias parameters for bias formulation and probabilistic bias evaluation
Simplifying assumptions are required by many strategies of quantitative bias evaluation. For instance, it’s usually assumed that the publicity doesn’t modify the unmeasured confounder-outcome affiliation.4 If these assumptions usually are not met then the findings of quantitative bias evaluation may be inaccurate.
Ideally, assumptions could be primarily based on supplemental knowledge collected in a subset of the research inhabitants (eg, inner validation research to estimate predictive values of misclassification) or, within the case of choice bias, within the supply inhabitants from which the pattern was chosen, however extra knowledge assortment isn’t at all times possible.7 Validation research could be an essential supply of proof on misclassification, though correct design is essential to acquire legitimate estimates.33
A number of biases
If the outcomes are sturdy to at least one supply of bias, it’s a mistake to imagine that they need to essentially replicate the causal impact. Relying on the actual research, a number of residual biases may exist, and collectively quantifying the influence of all of those biases is important to correctly assess robustness of outcomes.34 Bias formulation and probabilistic bias analyses could be utilized for a number of biases, however specification is extra difficult, and the biases ought to sometimes be accounted for within the reverse order from which they come up (appendices 2 and three present an utilized instance).74647 Bounding strategies can be found for a number of biases.34
Prespecification
Prespecification of quantitative bias evaluation within the research protocol is effective in order that selection of unknown values and option to report bias evaluation isn’t influenced by whether or not the outcomes of bias evaluation are according to the investigators expectations. Clearly a wide variety of analyses is feasible, though we might encourage even handed utility of those strategies to cope with biases judged to be of particular significance given the constraints of the precise research being performed.
Accounting for random and systematic error
Each systematic errors, equivalent to bias on account of misclassification and random error on account of sampling, have an effect on research outcomes. To precisely replicate this difficulty, quantitative bias evaluation ought to collectively account for random error in addition to systematic bias.48 Bias formulation, bounding strategies, and probabilistic bias evaluation approaches could be tailored to account for random error (appendix 1).
Reporting
Deficiencies within the reporting of quantitative bias evaluation have been beforehand famous.1484950 When reporting quantitative bias evaluation, research investigators ought to state:
-
The tactic used and the way it has been applied
-
Particulars of the residual bias anticipated (eg, which particular potential confounder was unmeasured)
-
Any estimates for unknown values which have been used, with justification for the chosen values or distribution for these unknowns
-
Which simplifying assumptions (if any) have been made
Quantitative bias evaluation is a worthwhile addition to a research, however as with all side of a research, ought to be interpreted critically and reported in adequate element to permit for vital interpretation.
Different strategies
Generally utilized and broadly relevant strategies have been described on this article. Different strategies can be found and embrace modified probability and predictive worth weighting with regression analyses,515253 propensity rating calibration utilizing validation knowledge,5455 a number of imputation utilizing validation knowledge,56 strategies for matched research,3 and bayesian bias evaluation if a completely bayesian method is desired.5758
Conclusions
Quantitative bias strategies present a method to quantitatively and rigorously assess the potential for residual bias in non-interventional research. Growing the suitable use, understanding, and reporting of those strategies has the potential to enhance the robustness of medical epidemiological analysis and cut back the probability of misguided conclusions.
Footnotes
-
Contributors: This text is the product of a working group on quantitative bias evaluation between the London Faculty of Hygiene and Tropical Drugs and GSK. An iterative technique of on-line workshops and e-mail correspondence was used to resolve by consensus the content material of the manuscript. Primarily based on these selections, a manuscript was drafted by JPB earlier than additional remark and reviewed by all group members. JPB and IJD are the guarantors. The corresponding creator attests that each one listed authors meet authorship standards and that no others assembly the standards have been omitted.
-
Funding: No particular funding was given for this work. JPB was supported by a GSK PhD studentship.
-
Competing pursuits: All authors have accomplished the ICMJE uniform disclosure kind at https://www.icmje.org/disclosure-of-interest/ and declare: AC, NWG, and JNH have been paid staff of GSK on the time of the submitted work; AC, IJD, NWG, and JNH personal shares in GSK; AC is presently a paid worker of McKesson Company in a job unrelated to the submitted work; JNH is presently a paid worker of Boehringer Ingelheim in a job unrelated to this work; DN is UK Kidney Affiliation Director of Informatics Analysis; JPB was funded by a GSK studentship obtained by IJD and reviews unrelated consultancy work for WHO Europe and CorEvitas; SML has obtained unrelated grants with business collaborators from IMI Horizon, however no direct business funding; all authors report no different relationships or actions that would seem to have influenced the submitted work.
-
Provenance and peer assessment: Not commissioned; externally peer reviewed.
[ad_2]