| dc.description.abstract | In order to evaluate the impact of outcome misclassification on the possible causal association
between Adverse Events of Special Interest (AESIs) and COVID-19 vaccination, we conducted
a literature review and a simulation study. The literature study aimed to obtain a plausible range
of outcome misclassification indices of the International Classification of Diseases (ICD) coding
systems used in electronic healthcare databases.
We used logistic regression to contrast a naïve estimator that disregards misclassification
with a misclassification-integrating Maximum Likelihood Estimation (MLE) model to explore
the relationship between vaccine exposure and the occurrence of AESIs. The MLE model
employed marginal probability from a Bernoulli distribution to account for misclassification,
facilitating a comparison of log of odds ratios and relative risks between the models.
In our simulation study, we generated data which incorporated a fixed vaccination rate,
varying sample sizes, regression coefficients, and misclassification rates to evaluate bias and
mean squared error (MSE) of the log of odds ratio and the relative risk. The analyses showed that
the MLE model exhibited reduced bias under high prevalence and specificity conditions when
examining the relative risk bias. Despite its great variability, the MLE model outperformed the
naïve estimator in specific simulations with increased association strengths, supporting our
hypothesis. Our findings reveal a need for rigorous methodologies to address misclassification in
vaccine safety assessments and indicate that the degree of misclassification may be significant,
depending on the specific ICD code.
Our research highlighted the importance of thoroughly recognising misclassified data and
outlined critical areas for future research in epidemiology. Our observations demonstrated that
the traditional approach, which disregards the subsequent bias brought on by misclassification –
is not fundamentally flawed, emphasising the need to investigate how and when these errors can
be significant. | |
