Background: Missing data are unavoidable in epidemiological research, potentially leading to bias and loss of precision. Multiple imputation (MI) is widely advocated as an improvement over complete case analysis (CCA). However, contrary to widespread belief, CCA is preferable to MI in some situations. Methods: We provide guidance on choice of analysis when data are incomplete. Using causal diagrams to depict missingness mechanisms, we describe when CCA will not be biased by missing data and compare MI and CCA, with respect to bias and efficiency, in a range of missing data situations. We illustrate selection of an appropriate method in practice. Results: For most regression models, CCA gives unbiased results when the chance of being a complete case does not depend on the outcome after taking the covariates into consideration, which includes situations where data are missing not at random. Consequently, there are situations in which CCA analyses are unbiased while MI analyses, assuming missing at random (MAR), are biased. By contrast MI, unlike CCA, is valid for all MAR situations and has the potential to use information contained in the incomplete cases and auxiliary variables to reduce bias and/or improve precision. For this reason, MI was preferred over CCA in our real data example. Conclusions: Choice of method for dealing with missing data is crucial for validity of conclusions, and should be based on careful consideration of the reasons for the missing data, missing data patterns and the availability of auxiliary information.
Schools in many countries undertake programmes for smoking prevention, but systematic reviews have shown mixed evidence of their effectiveness. Most peer-led approaches have been classroom-based, and rigorous assessments are scarce. We assessed the effectiveness of a peer-led intervention that aimed to prevent smoking uptake in secondary schools. We undertook a cluster randomised controlled trial of 10 730 students aged 12-13 years in 59 schools in England and Wales. 29 schools (5372 students) were randomly assigned by stratified block randomisation to the control group to continue their usual smoking education and 30 (5358 students) to the intervention group. The intervention (ASSIST [A Stop Smoking In Schools Trial] programme) consisted of training influential students to act as peer supporters during informal interactions outside the classroom to encourage their peers not to smoke. Follow-up was immediately after the intervention and at 1 and 2 years. Primary outcomes were smoking in the past week in both the school year group and in a group at high risk of regular smoking uptake, which was identified at baseline as occasional, experimental, or ex-smokers. Analysis was by intention to treat. This study is registered, number ISRCTN55572965. The odds ratio of being a smoker in intervention compared with control schools was 0.75 (95% CI 0.55-1.01) immediately after the intervention (n=9349 students), 0.77 (0.59-0.99) at 1-year follow-up (n=9147), and 0.85 (0.72-1.01) at 2-year follow-up (n=8756). The corresponding odds ratios for the high-risk group were 0.79 (0.55-1.13 [n=3561]), 0.75 (0.56-0.99 [n=3483]), and 0.85 (0.70-1.02 [n=3294]), respectively. In a three-tier multilevel model with data from all three follow-ups, the odds of being a smoker in intervention compared with control schools was 0.78 (0.64-0.96). The results suggest that, if implemented on a population basis, the ASSIST intervention could lead to a reduction in adolescent smoking prevalence of public-health importance.
1 June 2011 in The American Journal of Clinical Nutrition
Background: Few data are available on the long-term immunologic response to antiretroviral therapy (ART) in resource-limited settings, where ART is being rapidly scaled up using a public health approach, with a limited repertoire of drugs. Objectives: To describe immunologic response to ART among ART patients in a network of cohorts from sub-Saharan Africa, Latin America, and Asia. Study population/methods: Treatment-naive patients aged 15 and older from 27 treatment programs were eligible. Multilevel, linear mixed models were used to assess associations between predictor variables and CD4 cell count trajectories following ART initiation. Results: Of 29 175 patients initiating ART, 8933 (31%) were excluded due to insufficient follow-up time and early lost to follow-up or death. The remaining 19 967 patients contributed 39 200 person-years on ART and 71 067 CD4 cell count measurements. The median baseline CD4 cell count was 114 cells/μl, with 35% having less than 100 cells/μl. Substantial intersite variation in baseline CD4 cell count was observed (range 61–181 cells/μl). Women had higher median baseline CD4 cell counts than men (121 vs. 104 cells/μl). The median CD4 cell count increased from 114 cells/μl at ART initiation to 230 [interquartile range (IQR) 144–338] at 6 months, 263 (IQR 175–376) at 1 year, 336 (IQR 224–472) at 2 years, 372 (IQR 242–537) at 3 years, 377 (IQR 221–561) at 4 years, and 395 (IQR 240–592) at 5 years. In multivariable models, baseline CD4 cell count was the most important determinant of subsequent CD4 cell count trajectories. Conclusion: These data demonstrate robust and sustained CD4 response to ART among patients remaining on therapy. Public health and programmatic interventions leading to earlier HIV diagnosis and initiation of ART could substantially improve patient outcomes in resource-limited settings.
30 June 2020 in Statistical Methods in Medical Research
Multiple imputation has become one of the most popular approaches for handling missing data in statistical analyses. Part of this success is due to Rubin’s simple combination rules. These give frequentist valid inferences when the imputation and analysis procedures are so-called congenial and the embedding model is correctly specified, but otherwise may not. Roughly speaking, congeniality corresponds to whether the imputation and analysis models make different assumptions about the data. In practice, imputation models and analysis procedures are often not congenial, such that tests may not have the correct size, and confidence interval coverage deviates from the advertised level. We examine a number of recent proposals which combine bootstrapping with multiple imputation and determine which are valid under uncongeniality and model misspecification. Imputation followed by bootstrapping generally does not result in valid variance estimates under uncongeniality or misspecification, whereas certain bootstrap followed by imputation methods do. We recommend a particular computationally efficient variant of bootstrapping followed by imputation.