Dependent Error Misclassification in both the Response Variable and Covariate.
Date
2021-01-27
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
0000-0001-5495-5986
Type
Thesis
Degree Level
Doctoral
Abstract
Errors in Variables (EIV) are a long-standing issue in many fields, including medical
and epidemiological studies. Ignoring these errors can produce misleading inferential results.
In discrete responses, EIV are commonly termed as misclassi fication errors. Studies
on misclassifi cation have mostly focused on misclassification in only one variable. Joint misclassification in both the response variable and the covariate has been less explored. Some
literature on joint misclassification assumes the misclassification process of the response variable
is independent of the misclassification process of the covariate. However, in practice, the
dependence of misclassification errors can occur. For example both, the response variable and
covariate are obtained from a similar source as in the case of self-reported responses from a
questionnaire. The objective is to investigate (1) modeling for error-prone response variable
and error-prone covariate and (2) consequences of using an incorrect misclassification model.
In this thesis, we first introduce a model that accounts for dependent misclassification error
in a binary response variable and a binary covariate. The dependence of error is captured
through covariance-like parameters. Simulation studies are conducted to assess the consequences
of fitting an independent misclassification model to data generated from a dependent
misclassification model. The simulation experiments have several key factors to manipulate:
the amount of misclassification error (sensitivity and specificity), the dependence between
the misclassification process of the response variable, and the misclassification process of the
covariate, and the proportion of internal validation data. Further, the model is extended to a
multi-category setting and simulation study is conducted on a trinary response variable and
a trinary covariate. Results from the simulation studies indicate that ignoring dependence
of the error in misclassification can be worse than ignoring misclassification.
The proposed model is illustrated through a real data example by establishing the true
association between Trichomoniasis and Bacterial Vaginosis, using data from the HIV Epidemiology
Research Study (HERS). A likelihood-ratio test is proposed to test the independent
misclassification assumption. The test concluded that the dependent misclassification
error model fits the HERS data significantly better than the model that ignored dependence
misclassification.
Description
Keywords
misclassification, dependence
Citation
Degree
Doctor of Philosophy (Ph.D.)
Department
School of Public Health
Program
Biostatistics