Bazo Alvarez, Juan Carlos; (2020) Practical methodology for missing data handling in interrupted time series analysis. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text UCL_PhD Thesis_JCBA.pdf - Accepted Version Download (4MB) | Preview

Abstract

Interrupted time series (ITS) is a quasi-experimental design for evaluating the effect of an intervention or treatment by comparing the outcome trajectory over time before and after initiation of the intervention. ITS became popular for evaluating interventions at the population level (e.g. policies); thus, the development of statistical methods was mainly orientated to modelling population-level data. This thesis aims to explore the issues that emerge when population-level ITS analyses are applied to incomplete individual-level data in health research, proposing alternative analysis methods. First, I performed a scoping review to demonstrate how the issues of missing data at the individual level have rarely been addressed in most recent ITS studies. Despite its limitations, complete case analysis is the most frequently used method for handling missing data. Individual-level data are usually transformed into population-level time-specific summaries before fitting ITS models. This method can lead to bias. Mixed effect models (MEM) can solve this, but my review demonstrates few studies have done so in the past. I then fitted MEM to study body weight gain induced by the initiation of antipsychotics using an ITS design on electronic health records. MEM allowed fully observed covariates to inform the implicit imputation of the outcome. ITS facilitated new clinical evidence: in the long-term, typical patients do not lose the weight they gained during the first six weeks of treatment. However, the MEM alone was not ideal for handling missing covariates (i.e. dosage). Thereafter, I used simulation studies to evaluate the performance of aggregate-level segmented regression (SR), MEM and multilevel multiple imputation (MI-JOMO) for handling data missing at random (MAR) in ITS analysis. I showed that the aggregate-level SR can over or underestimate the ITS effect. MEM is effective for handling outcomes MAR, but it should be combined with MI-JOMO when covariates are also MAR. Finally, I applied MEM with MI-JOMO to assess how dose and age modify the antipsychotic-induced weight gain. Interaction terms in MEM helped to evaluate differences in weight trajectories over time between groups by dose or age, using MI-JOMO for handling missing dose. Clinically, older people’s weight is less affected by the initiation of antipsychotic treatment than younger people’s.

Download activity - last month

Export as JSONCSVXML

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item