ENBIS-18 in Nancy

2 – 25 September 2018; Ecoles des Mines, Nancy (France) Abstract submission: 20 December 2017 – 4 June 2018

Effect of Neglecting Autocorrelation in Regression CUSUM Charts to Monitor Counts Time Series

4 September 2018, 14:50 – 15:10

Abstract

Submitted by
Linda Ho
Authors
Orlando Yesid Esparza Albarracín (University of São Paulo), Airlane Pereira Alencar (University of São Paulo), Linda Lee Ho (University of São Paulo)
Abstract
In practice, it is usual to monitor count time series proposing Poisson or Negative Binomial regression models (the last to control overdispersion). The main concern is that the usual generalized linear models (GLM) assume independence and the data are autocorrelated in general time series. One possibility is to fit the generalized autoregressive and moving average (GARMA) model to model counts under the negative binomial distribution with time varying means and include lagged terms to take into account the autocorrelation.

The main contribution of our research is to measure the impact, (in terms of the average run length (ARL)), on the performance of CUSUM charts with different statistics, when the serial correlation is neglected in a regression model. This is performed simulating correlated process using GARMA, fitting independent GLM models and building the corresponding CUSUM charts. High autocorrelation leads to an increase of false alarms. This analysis may help practitioners to implement control charts taking into account the serial correlation with no extra cost to fit an appropriate model

Return to programme