Anders Gorst-Rasmussen

When dealing with missing data in clinical trials, it is often convenient to work under simplifying assumptions, such as missing at random (MAR), and follow up with sensitivity analyses to address unverifiable missing data assumptions. One such sensitivity analysis, routinely requested by regulatory agencies, is the so-called tipping point analysis, in which the treatment effect is re-evaluated after adding a successively more extreme shift parameter to the predicted values among subjects with missing data. If the shift parameter needed to overturn the conclusion is so extreme that it is considered clinically implausible, then this indicates robustness to missing data assumptions. Tipping point analyses are frequently used in the context of continuous outcome data under multiple imputation. While simple to implement, computation can be cumbersome in the two-way setting where both comparator and active arms are shifted, essentially requiring the evaluation of a two-dimensional grid of models. We describe a computationally efficient approach to performing two-way tipping point analysis in the setting of continuous outcome data with multiple imputation. We show how geometric properties can lead to further simplification when exploring the impact of missing data. Lastly, we propose a novel extension to a multi-way setting which yields simple and general sufficient conditions for robustness to missing data assumptions.

Background: In the onset 5 trial, fast-acting insulin aspart (faster aspart) was noninferior to insulin aspart (IAsp) for change from baseline glycated hemoglobin at 16 weeks, when used in continuous subcutaneous insulin infusion by participants with type 1 diabetes. The aim of this post hoc analysis was to investigate whether infusion set wear-time was associated with changes in sensor glucose, measured using continuous glucose monitoring (CGM). Materials and Methods: This was a post hoc analysis of onset 5 data. Mean infusion set wear-time and duration of CGM-wearing period were assessed. Mean CGM sensor glucose 24 h before and 24 h after were used to calculate the before–after difference (CGM sensor glucose drift). Results: Mean infusion set wear-time was 2.9 and 3.0 days in the faster aspart and IAsp arms, respectively. At 16 weeks, the average duration of the CGM wearing period was 13.7 and 13.8 days, respectively. Mean CGM sensor glucose before versus after an infusion set change, at week 16, was 10.14 versus 9.39 mmol/L with faster aspart and 9.48 versus 9.47 mmol/L with IAsp. The estimated treatment difference in CGM sensor glucose drift at 16 weeks for faster aspart versus IAsp was +0.72 mmol/L (95% confidence interval: 0.48–0.96, P < 0.001). Conclusions: Mean infusion set wear-time and duration of CGM-wearing period were similar for faster aspart and IAsp. A significantly greater upward drift in CGM sensor glucose values measured during an infusion set wearing period was observed with faster aspart versus IAsp. Clinical trial registration: NCT02825251.

The use of multiple imputation (MI) to address missing data has improved the quality of clinical trial reporting but has also introduced the nuisance of results that are subject to sampling variability introduced by using a finite number of imputations. There is a tradeoff between ensuring that enough imputations are used to obtain accurate results for decision making while at the same time keeping the number of imputations low enough to avoid computational bottlenecks. We propose a rule for evaluating this tradeoff that extends the familiar idea of exact reproducibility of statistical results: enough imputations are used if an independent round of MI yields exactly the same result when rounded to the chosen reporting precision, with high probability. We explore how this reproducibility rule can be used derive simple criteria for choosing the number of imputations at the planning stage of a clinical trial.

Background: Nearly one in 5 patients with ischemic stroke will invariably experience a second stroke within 5 years. Stroke risk stratification schemes based solely on clinical variables perform only modestly in non-atrial fibrillation (AF) patients and improvement of these schemes will enhance their clinical utility. Cerebral white matter hyperintensities are associated with an increased risk of incident ischemic stroke in the general population, whereas their association with the risk of ischemic stroke recurrence is more ambiguous. In a non-AF stroke cohort, we investigated the association between cerebral white matter hyperintensities and the risk of recurrent ischemic stroke, and we evaluated the predictive performance of the CHA2DS2VASc score and the Essen Stroke Risk Score (clinical scores) when augmented with information on white matter hyperintensities. Methods: In a registry-based, observational cohort study, we included 832 patients (mean age 59.6 (SD 13.9); 42.0% females) with incident ischemic stroke and no AF. We assessed the severity of white matter hyperintensities using MRI. Hazard ratios stratified by the white matter hyperintensities score and adjusted for the components of the CHA2DS2VASc score were calculated based on the Cox proportional hazards analysis. Recalibrated clinical scores were calculated by adding one point to the score for the presence of moderate to severe white matter hyperintensities. The discriminatory performance of the scores was assessed with the C-statistic. Results: White matter hyperintensities were significantly associated with the risk of recurrent ischemic stroke after adjusting for clinical risk factors. The hazard ratios ranged from 1.65 (95% CI 0.70-3.86) for mild changes to 5.28 (95% CI 1.98-14.07) for the most severe changes. C-statistics for the prediction of recurrent ischemic stroke were 0.59 (95% CI 0.51-0.65) for the CHA2DS2VASc score and 0.60 (95% CI 0.53-0.68) for the Essen Stroke Risk Score. The recalibrated clinical scores showed improved C-statistics: the recalibrated CHA2DS2VASc score 0.62 (95% CI 0.54-0.70; p = 0.024) and the recalibrated Essen Stroke Risk Score 0.63 (95% CI 0.56-0.71; p = 0.031). C-statistics of the white matter hyperintensities score were 0.62 (95% CI 0.52-0.68) to 0.65 (95% CI 0.58-0.73). Conclusions: An increasing burden of white matter hyperintensities was independently associated with recurrent ischemic stroke in a cohort of non-AF ischemic stroke patients. Recalibration of the CHA2DS2VASc score and the Essen Stroke Risk Score with one point for the presence of moderate to severe white matter hyperintensities led to improved discriminatory performance in ischemic stroke recurrence prediction. Risk scores based on white matter hyperintensities alone were at least as accurate as the established clinical risk scores in the prediction of ischemic stroke recurrence.

Purpose: To evaluate effectiveness and safety of rivaroxaban versus warfarin or dabigatran etexilate in a prospective cohort of routine care non‐valvular atrial fibrillation (AF) patients during February 2012 to August 2014.Methods: We identified in nationwide health registries a cohort of AF patients who were new‐users of rivaroxaban 15 mg (R15) or 20 mg (R20); dabigatran 110 mg (D110) or 150 mg (D150); or warfarin. Propensity‐adjusted Cox regression was used to compare outcome rates in four settings: ‘R15 vs. warfarin’; ‘R15 vs. D110’; ‘R20 vs. warfarin’; and ‘R20 vs. D150’.Results: Rivaroxaban users (R15: n = 776; R20: n = 1629) were older and with more comorbidities than warfarin (n = 11 045) and dabigatran users (D110: n = 3588; D150: n = 5320). Rivaroxaban 15‐mg users had the overall highest crude mortality rate. After propensity adjustment, rivaroxaban had lower stroke rates vs. warfarin (R15: hazard ratio [HR] 0.46, 95% confidence interval [CI]: 0.26–0.82; R20 HR: 0.72, 95%CI: 0.51–1.01), and similar stroke rates vs. dabigatran. The bleeding rate was similar to warfarin and moderately higher vs. dabigatran (R15 vs. D110 HR: 1.28, 95%CI: 0.82–2.01; R20 vs. D150 HR: 1.81, 95%CI: 1.25–2.62). The mortality rate was higher vs. dabigatran (R15 vs. D110 HR: 1.43, 95%CI: 1.13–1.81; R20 vs. D150 HR: 1.52, 95%CI: 1.06–2.19).Conclusions: Rivaroxaban was associated with similar or lower stroke rates, but higher bleeding and mortality rates. Channeling of rivaroxaban towards elderly and less healthy patients may have generated residual confounding. In particular, our findings cannot stand alone when deciding which oral anticoagulant to prescribe. Copyright © 2016 John Wiley & Sons, Ltd.

The FARIMA models, which have long-range-dependence (LRD), are widely used in many areas. Through deriving a precise characterisation of the spectrum, autocovariance function, and variance time function, we show that this family is very atypical among LRD processes, being extremely close to the fractional Gaussian noise in a precise sense. Furthermore, we show that this closeness property is not robust to additive noise. We argue that the use of FARIMA, and more generally fractionally differenced time series, should be reassessed in some contexts, in particular when convergence rate under rescaling is important and noise is expected.

The treelet transform is a recent data reduction technique from the field of machine learning. Sharing many similarities with principal component analysis, the treelet transform can reduce a multidimensional dataset to the projections on a small number of directions or components that account for much of the variation in the original data. However, in contrast to principal component analysis, the treelet transform produces sparse components. This can greatly simplify interpretation. I describe the tt Stata add-on for performing the treelet transform. The addon includes a Mata implementation of the treelet transform algorithm alongside other functionality to aid in the practical application of the treelet transform. I demonstrate an example of a basic exploratory data analysis using the tt add-on.