Statistical methods of detecting change points for the trend of count data

  • Taghreed Mohammed A Jawa

Student thesis: Doctoral Thesis

Abstract

In epidemiology, controlling infection is a crucial element. Since healthcare associated infections (HAIs) are correlated with increasing costs and mortality rates, effective healthcare interventions are required. Several healthcare interventions have been implemented in Scotland and subsequently Health Protection Scotland (HPS) reported a reduction in HAIs [HPS (2015b, 2016a)].The aim of this thesis is to use statistical methods and change points analysis to detect the time when the rate of HAIs changed and determine which associated interventions may have impacted such rates.Change points are estimated from polynomial generalized linear models (GLM) and confidence intervals are constructed using bootstrap and delta methods and the two techniques are compared. Segmented regression is also used to look for change points at times when specific interventions took place. A generalization of segmented regression is known as joinpoint analysis which looks for potential change points at each time point in the data, which allows the change to have occurred at any point over time.The joinpoint model is adjusted by adding a seasonal effect to account for additional variability in the rates. Confidence intervals for joinpoints are constructed using bootstrap and profile likelihood methods and the two approaches are compared. Change points from the smoother trend of the generalized additive model (GAM) are also estimated and bootstrapping is used to construct confidence intervals.All methods were found to have similar change points. Segmented regression detects the actual point when an intervention took place. Polynomial GLM, spline GAM and joinpoint analysis models are useful when the impact of an intervention occurs after a period of time. Simulation studies are used to compare polynomial GLM, segmented regression and joinpoint analysis models for detecting change points along with their confidence intervals.
Date of Award1 Jan 2017
LanguageEnglish
Awarding Institution
  • University Of Strathclyde
SupervisorDavid Young (Supervisor) & Charles Robertson (Supervisor)

Cite this