Perpetual Chikobvu, PhD¹, Carl Lombard, PhD², Alan J. Flisher, PhD³, Loraine Townsend, MA³, Martie Muller, MSc⁴, and Gary King, PhD⁵. (1) Biostatistics Unit, Medical Research Council, P.O. Box 19070, Tygerberg, Cape Town, 7505, South Africa, (021) 938 0548, perpetual.chikobvu@mrc.ac.za, (2) Biostatic Division/Centre for Epidemilogic Research in South Africa, Medical Research Council, P.O. Box 19070, Tygerberg, 7505, South Africa, (3) Department of Psychiatry, University of Cape Town, Groote Schuur Hospital, Observatory 7925, RSA, CapeTown, Western Cape Province, South Africa, (4) Statistics, Institute of Maritime Technology, P.O. Box 181, Simon's Town, Cape Town, 7995, South Africa, (5) Department of Biobehavioral Health, Penn State University, 315 E. Henderson Bldg, University Park, PA 16802

Outcome assessment in longitudinal studies is frequently subject to measurement errors. These errors are generally found in studies on substance use or risk behaviour indicators among school children. Ignoring these errors can yield biased covariate effect estimates. This study proposes an approach for analysing binary longitudinal data subject to measurement errors focusing on large school dropout rate and misclassification of the risk behaviour indicators between time t and time t+1. The South African Community Epidemiology Network on Drug Abuse examined high school students ever cigarette use in 1997 and 1999. Ten percent of students responded positively in 1997. However when asked ever used cigarette in 1999, 29% of those who had used indicated they never smoked a cigarette. This type of misclassification in a binary outcome measured on repeated occasions was explored for an analysis of transitions with first order binary transition probabilities of making each of the possible transitions from one visit to the next. The joint probability structure of observations made at different times was modelled via the conditional probabilities of the current time outcome given the previous time outcome. Non-response bias was taken into account. These outcomes were modelled as functions of the covariates as well as previous responses. The transitions probabilities of ever used a cigarette obtained were 0.36, 0.12 and 0.81 for incident, remittance and continuity transitions respectively among the male students for the female students the respective transition probabilities were 0.26, 0.08 and 0.92.There were no significant gender differences in probabilities.

Learning Objectives:

• Modelling transition probabilities for a binary outcome observed repeatedly
• Address the problems of missing values in longitudinal analyses
• Understanding the implications of misclassification in tobacco research in South Africa and developing countries.

Keywords: Biostatistics, Smoking

Presenting author's disclosure statement:
I do not have any significant financial interest/arrangement or affiliation with any organization/institution whose products or services are being discussed in this session.