1. The Need for Competing Risks
Differential Analysis & Reporting of Longitudinal Event Data in Type 2 Diabetic Nephropathic Study Populations: A Systematic Review
The Need for Competing Risks
Recent clinical guidelines state the need to move away from traditional interval-based risk calculators in favour of those estimating lifetime-risk1; the rationale being that being able to treat patients earlier may be more beneficial to their survival.
This shift in timescale has important implications for our assumptions regarding ‘loss to follow-up’, a phenomenon whereby we can no longer observe some members of the study population. Standard survival models assume that patients lost to follow-up have the same chance of experiencing the main outcome as those still in the study population. This is called ‘non-informative censoring’.
A problem arising from the use of lifetime risk is that all censoring shifts towards becoming failure from a competing cause. As such, the assumption of non-informative censoring is violated. (It’s not possible for someone to die from heart failure if they’ve already died from a stroke).
It follows that the use of standard models in such scenarios could introduce bias leading to potential overestimation of the risk. This could become particularly evident in studies where follow-up is long or the competing outcomes are common.
Background
Type 2 (T2) diabetes is one of the largest health challenges facing the UK today. Over 2.9 million new diagnoses were made 2012 with yearly incidence projected to exceed 5 million by
The disease is slow-progressing, with prolonged poor glucose regulation resulting in systemic endothelial damage. This can lead to eventual mortality from any one of several causes, including vascular diseases and neoplasms.
Diabetic nephropathy (DN) is one of the complications associated with long-standing diabetes in which damage to glomeruli leads to impaired filtration and leakage of protein into the urine (proteinuria). The presence and degree of proteinuria can be used to determine the extent of endothelial damage in sufferers of type 2 diabetes.
Given that DN can predict a range of serous health outcomes, analysis using competing risks (CRs) methods seems appropriate, yet whether or not this is being done is not currently known.
MEDLINE
N = 2825
EMBASE
N = 2756
CINAHL
N = 332
2329 duplicates excluded
3886 abstracts screened
179 full-texts assessed for eligibility
3707 records excluded
16 studies included in review
10 reported composite outcomes
1 reported competing risks
5 reported individual outcomes
Search Strategy
The CINAHL database, and Ovid versions of the MEDLINE and EMBASE databases were searched for both published & unpublished prognostic studies using keyword and MeSH-term search strategies. A degree of specificity was induced in the searches through:
limiting them to the title and abstract.
requiring keywords relating to ‘original research’ be present.
the use of “AND” Boolean operators stipulating that the article contain terms pertaining to population, study design and outcome.
Exclusion Criteria
Articles were excluded if:
The study was not of trial or cohort design.
The study population consisted of less than 1,000 individuals with T2 DN at baseline.
The study had an average follow-up less than five years.
The study did not investigate incidence, relative risk, hazard ratio, cumulative incidence or odds ratio metrics of microvascular, macrovascular or neoplasm events.
The study was not available in English.
The date of publication was prior to 1980.
Results
After removing duplicate records, the initial search identified 3,886 abstracts, published up to 21st March From these, 179 studies were eligible for full-text review, of which 141 have currently been screened. Of the 16 papers identified for data extraction to date:
one implemented a CRs model.
10 used one or more individual outcome models or general incidence statistics.
five reported composite outcome models.
Conclusions
Preliminary results suggest that most studies do not account for CRs in their analyses, and instead, use either single or composite outcome models. In a CRs scenario, both of these techniques can lead to bias through misspecifying the population at risk. In the case of single outcome models, this can be a result of failing to acknowledge other potential outcomes, or to jointly evaluate them in the presence of the main outcome. Conversely, the use of composite outcome models imposes a strong and often untested assumption of risk uniformity between the grouped outcomes. Furthermore, the inconsistency of the groupings between papers can result in them having poor external validity. By simultaneously evaluating the risk of developing multiple outcomes, the use of CRs models goes some way towards addressing these issues. Through these techniques, it may be possible to develop more targeted intervention strategies tailored to an individual’s risk-outcome profile.
Future Research
There is on-going simulation work to explore the above biases and their potential effect on the risk measures. However, further work is needed to assess the relevance of baseline data to eventual failure in lifetime-risk scenarios, and ways to more efficiently use routinely collected clinical data in prediction models.
Feakins BG, Stevens RJ, McFadden E, Farmer A
References
1 Anekwe L ~ 2011 ~ JBS to Rip Up Rules on Cardiovascular Risk ~ ~ [Accessed ]
2 Diabetes UK ~ 2012 ~ Diabetes in the UK: Key Statistics on Diabetes ~
Objectives
{i} To identify longitudinal studies examining the occurrence of vascular and neoplastic events in individuals with T2 DN.
{ii} To assess the extent to which these studies account for competing risks in their methodology and reporting.