Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls

Sterne, Jonathan A C; White, Ian R.; Carlin, John B.; Spratt, Michael F.; Royston, Patrick; Kenward, Michael G.; Wood, Angela; Carpenter, James

doi:10.1136/bmj.b2393

Cited by 5,481 publications

(4,737 citation statements)

References 24 publications

Supporting

Mentioning

4,503

Contrasting

Unclassified

Order By: Relevance

“…These include the prospective population‐based cohort design, the relatively large sample size and our ability to adjust for multiple biological, clinical and lifestyle covariates, including HPV. In addition to this, we explored the risk of bias due to missing data by employing a multiple imputation approach 47. Results of the imputed analysis was broadly consistent with those of the complete case analysis.…”

Section: Discussionmentioning

confidence: 84%

Tobacco smoking and alcohol drinking at diagnosis of head and neck cancer and all‐cause mortality: Results from head and neck 5000, a prospective observational cohort of people with head and neck cancer

Beynon

Lang

Schimansky

et al. 2018

Intl Journal of Cancer

134

View full text Add to dashboard Cite

Tobacco smoking and alcohol consumption are well‐established risk factors for head and neck cancer. The prognostic role of smoking and alcohol intake at diagnosis have been less well studied. We analysed 1,393 people prospectively enrolled into the Head and Neck 5000 study (oral cavity cancer, n=403; oropharyngeal cancer, n=660; laryngeal cancer, n=330) and followed up for a median of 3.5 years. The primary outcome was all‐cause mortality. We used Cox proportional hazard models to derive minimally adjusted (age and gender) and fully adjusted (age, gender, ethnicity, stage, comorbidity, body mass index, HPV status, treatment, education, deprivation index, income, marital status, and either smoking or alcohol use) mortality hazard ratios (HR) for the effects of smoking status and alcohol intake at diagnosis. Models were stratified by cancer site, stage and HPV status. The fully‐adjusted HR for current versus never‐smokers was 1.7 overall (95% confidence interval [CI] 1.1, 2.6). In stratified analyses, associations of smoking with mortality were observed for oropharyngeal and laryngeal cancers (fully adjusted HRs for current smokers: 1.8 (95% CI=0.9, 3.40 and 2.3 (95% CI=0.8, 6.4)). We found no evidence that people who drank hazardous to harmful amounts of alcohol at diagnosis had a higher mortality risk compared to non‐drinkers (HR=1.2 (95% CI=0.9, 1.6)). There was no strong evidence that HPV status or tumour stage modified the association of smoking with survival. Smoking status at the time of a head and neck cancer diagnosis influenced all‐cause mortality in models adjusted for important prognostic factors.

show abstract

Section: Discussionmentioning

confidence: 84%

Tobacco smoking and alcohol drinking at diagnosis of head and neck cancer and all‐cause mortality: Results from head and neck 5000, a prospective observational cohort of people with head and neck cancer

Beynon

Lang

Schimansky

et al. 2018

Intl Journal of Cancer

134

View full text Add to dashboard Cite

show abstract

“…Interactions between the variables included in the multivariable model and type of AML were estimated within the Cox model and presented in a forest plot where age was categorized for presentational purposes. To test for bias due to baseline variables missing at random (MAR), a sensitivity analysis was performed using multiple imputation for all variables with missing data, using predictive mean matching (with the function aregImpute in R) [14,15]. Outcome and survival, was included as the Nelson-Aalen estimator in the imputation model, although not imputed itself because it contained no missing values.…”

Section: Methodsmentioning

confidence: 99%

Characterization and prognostic features of secondary acute myeloid leukemia in a population‐based setting: A report from the Swedish Acute Leukemia Registry

et al. 2015

View full text Add to dashboard Cite

Patients with secondary acute myeloid leukemia (AML) often escape inclusion in clinical trials and thus, population-based studies are crucial for its accurate characterization. In this first large population-based study on secondary AML, we studied AML with an antecedent hematological disease (AHD-AML) or therapyrelated AML (t-AML) in the population-based Swedish Acute Leukemia Registry. The study included 3,363 adult patients of which 2,474 (73.6%) had de novo AML, 630 (18.7%) AHD-AML, and 259 (7.7%) t-AML. Secondary AML differed significantly compared to de novo AML with respect to age, gender, and cytogenetic risk. Complete remission (CR) rates were significantly lower but early death rates similar in secondary AML. In a multivariable analysis, AHD-AML (HR 1.51; 95% CI 1.26-1.79) and t-AML (1.72; 1.38-2.15) were independent risk factors for poor survival. The negative impact of AHD-AML and t-AML on survival was highly age dependent with a considerable impact in younger patients, but without independent prognostic value in the elderly. Although patients with secondary leukemia did poorly with intensive treatment, early death rates and survival were significantly worse with palliative treatment. We conclude that secondary AML in a population-based setting has a striking impact on survival in younger AML patients, whereas it lacks prognostic value among the elderly patients.

show abstract

“…As there were significant predictors of missingness in our baseline data, data cannot be presumed to be missing at random, potentially biasing estimates (Sterne et al., 2009). In the subsample of participants without lifetime NSSI at age 14 (our inclusion criteria), there was minimal missing data on NSSI at age 17 (12%, see results for more details).…”

Section: Methodsmentioning

confidence: 99%

Poor family functioning mediates the link between childhood adversity and adolescent nonsuicidal self‐injury

Cassels

Harmelen

Neufeld

et al. 2018

Child Psychology Psychiatry

View full text Add to dashboard Cite

BackgroundNon‐suicidal self‐injury (NSSI) is a common harmful behavior during adolescence. Exposure to childhood family adversity (CFA) is associated with subsequent emergence of NSSI during adolescence. However, the pathways through which this early environmental risk may operate are not clear.AimsWe tested four alternative hypotheses to explain the association between CFA and adolescent‐onset NSSI.MethodsA community sample of n = 933 fourteen year olds with no history of NSSI were followed up for 3 years.ResultsPoor family functioning at age 14 mediated the association between CFA before age 5 and subsequent onset of NSSI between 14 and 17 years.ConclusionsThe findings support the cumulative suboptimal environmental hazards (proximal family relationships as a mediator) hypothesis. Improving the family environment at age 14 may mitigate the effects of CFA on adolescent onset of NSSI.

show abstract

Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls

Abstract: Most studies have some missing data. Jonathan Sterne and colleagues describe the appropriate use and reporting of the multiple imputation approach to dealing with them

Cited by 5,481 publications

References 24 publications

Tobacco smoking and alcohol drinking at diagnosis of head and neck cancer and all‐cause mortality: Results from head and neck 5000, a prospective observational cohort of people with head and neck cancer

Tobacco smoking and alcohol drinking at diagnosis of head and neck cancer and all‐cause mortality: Results from head and neck 5000, a prospective observational cohort of people with head and neck cancer

Characterization and prognostic features of secondary acute myeloid leukemia in a population‐based setting: A report from the Swedish Acute Leukemia Registry

Poor family functioning mediates the link between childhood adversity and adolescent nonsuicidal self‐injury

Contact Info

Product

Resources

About