Comparative safety and effectiveness of cognitive enhancers for Alzheimer's dementia: a systematic review and individual patient data network meta-analysis

In the analyses, we aim to include IPD on: (1) patients, including age, sex, severity of Alzheimer's disease (eg, baseline MMSE level), presence of behavioural disturbance, comorbid conditions (eg, stroke, cardiovascular conditions, Parkinson's disease), other medications used for each patient (such as ß-blockers and other antiarrhythmic drugs), drop-outs along with reasons for drop-out, and number of participants; (2) medication, including treatment patient was allocated, dosage; (3) outcomes, including event and date of event and time taken to achieve the event for SAEs, and MMSE values and measurement dates; and (4) date and method of randomisation.The data we plan to abstract from each individual study include study characteristics (eg, year of publication), aggregated patient characteristics (eg, number of patients), outcome results (eg, MMSE, SAE) and source of funding (categorised as: funded/authored by an employee of a drug manufacturer or other commercial organisation, government-sponsored/non-profit organisations, including universities and hospitals, no funding, funding unclearly reported, and funding not reported). Two reviewers will abstract data independently, and all conflicts will be resolved through discussion. The year of publication and funding are potential effect modifiers. Therefore, these factors will be explored in a network meta-regression assuming a common fixed coefficient across treatment comparisons.As with the original review, we will appraise the risk of bias using the Cochrane Risk of Bias tool. We will draw a comparison-adjusted funnel plot for both outcomes. Two review authors will also independently assess the quality of evidence in each NMA using the GRADE approach as extended for network meta-analysis.We will perform a Bayesian hierarchical random-effects meta-analysis for each treatment comparison, as we anticipate clinical and methodological between-study heterogeneity. All IPD from included studies will be combined into a single model using a multilevel model where each study is a different cluster. We will use the odds ratio for SAE and the mean difference effect size for MMSE. In case we are able to obtain IPD for a subset of trials, then we will use a two-part model with the same between-study variance in both parts and accounting for treatment-by-covariate interactions (including for example co-morbidities such as arrhythmias in the model). The first part will entail a one-stage model using IPD only, whereas the second part will entail applying a pairwise meta-analysis with aggregate data.For a connected network of trials, the random-effects NMA model will be used. If possible, we will combine information across a network of trials using only IPD. If we are not successful in obtaining IPD for at least one study, we will combine both IPD and aggregated data in a single model. Information on patient-level covariates (eg, AD severity, sex) will be included in the model as secondary analyses. We will evaluate the consistency assumption using the design-by-treatment interaction model and the loop-specific method using aggregated data. If inconsistency is suggested, we will check the data for discrepancies and if none are identified, subgroup or meta-regression analyses will be performed.We will estimate subgroup effects (eg, age, sex) using treatment-by-covariate interaction terms within studies and combining these across studies. We will apply 3 model specifications assuming that the regression coefficients are: a) different and unrelated across comparisons, b) different but related, sharing the same distribution and c) identical across comparisons. We will compare the results of the models by evaluating the statistical significance of the regression coefficients for interactions, monitoring the reduction in the between-study variance, and using the Deviance Information Criterion to compare the overall fit and parsimony of the models. We will rank the interventions for each outcome using the surface under the cumulative ranking curve. We will conduct multiple sensitivity analyses to examine the robustness of our results. We will: 1) restrict to studies with IPD only, 2) use different priors for the between-study variance, 3) restrict to RCTs with a low risk of bias, 4) use different imputation techniques for missing outcome data. We will use a data-driven approach. More specifically, all IPD variables provided will be entered in our NMA and in the meta-regression analysis we will start by including one dependent and one independent variable. Then significant moderators will simultaneously be entered into multiple regression models as long as the minimum number of cases per independent variable is 10. Our goal is to avoid overfitting and provide reliable treatment effect estimates.All analyses will be conducted using the Bayesian software OpenBUGS. Two chains will be generated and convergence will be evaluated by their mixing, after discarding the first 10,000 iterations. We will use vague priors for all parameters of the models apart from the between-study variance for which we will use informative priors.