Deadline: June 30, 2022
Journal of Data Science (http://jds-online.org) invites submissions to
a special issue on "Data Science Approaches to Vaccine Effect
Analysis."
Vaccine is an essential tool for stopping the spread of COVID-19.
Evaluation of the effectiveness of the vaccines is important for the
success of vaccines. Vaccine effect analysis consists of assessing the
efficacy of vaccines in clinical trials and the effectiveness of
vaccines under real world conditions. The Centers for Disease Control
and Prevention (CDC)'s guidance on vaccine effectiveness analysis
include (1) analysis for specific subpopulations, (2) reducing the
risk of infection, (3) protection against mild COVID-19 illness, (4)
preventing hospitalization, (5) reducing the spread of illness, (6)
assessing duration of protection, (7) evaluating the impact of the
virus new variants, (8) evaluating the effectiveness of a single dose
and delayed second dose, and (9) evaluating impact of population host
factors.
The effect analysis may include two kinds of study designs, (1)
randomized clinical trials and (2) observational studies. Since the
infection and progression of COVID-19 are stochastic and dynamic, many
unknown factors may play an important role, making accurate estimation
of the vaccine effectiveness a great challenge in both randomized
clinical trials and population-based observational studies. To
stimulate discussions, we illustrate (incompletely) some challenges
and recently developed statistical and machine learning methods for
treatment effect estimation as follows.
Methods for clinical trials:
Recently, some investigators challenged the view that randomization
implies unconfoundedness and claims that randomization and
unconfoundedness are two separate concepts. We need further
validations on both theoretical and numerical analyses about the
relationships between randomization and unconfoundedness.
The Cox proportional-hazards model is widely viewed as a causal model
and used for treatment effect estimation in both randomized clinical
trials and observational studies. However, some researchers concluded
that the Cox hazard ratio is not causally interpretable and the Cox
model still may suffer selection bias.
Methods for observation data:
Estimating the vaccine effectiveness in observational data raises two
principal challenges. First, the treatment assignment mechanism is not
known a priori. Therefore, there might be confounders, affecting both
the vaccine and protection outcomes, which lead to selection bias. The
second challenge is censoring. Censoring might be informative, which
also may lead to bias.
Machine learning (ML) methods for estimating the effects of
treatments, including tree-based methods, the nonparametric Random
Survival Forest, Bayesian Additive regression trees, Gaussian
processes and, in particular, neural networks have grown rapidly. To
address the challenges of selection and censoring biases,
counterfactual survival analysis has been developed to incorporate
censoring and balanced representation for individual treatment effect
(ITE) prediction. These methods can be applied to COVID-19 vaccine
effect analysis and their merits and limitations need to be
investigated.
To overcome the bias of Cox proportional-hazards model, existing
methods for estimating treatment-specific hazard functions are able to
address the challenges due to the potential presence of multiple
sources of covariate shift: (i) non-randomized treatment assignment
(confounding), (ii) informative censoring, and (iii) event-induced
covariate shift.
The effect of the vaccine may depend on the true patient status. The
true patient status, however, is unknown and hidden. In addition, the
confounding variables that affect both the potential outcomes and
treatments are also hidden. The variational auto-encoder (VAE) that
models the hidden variables can be integrated with the Cox model for
unbiased treatment effect estimation. VAE-Cox model can be explored
for COVID-19 vaccine effect analysis.
We welcome original research articles (methods or computing), case
studies, tutorials, reviews, and perspective articles. The formats are
flexible. For example, short opinion or commentary articles are
welcome regarding misleading discoveries in the literature and
potential misuses of statistical or data science methods. Such
articles may be as impactful as full-length research articles.
Since 2003, Journal of Data Science has published research works on a
wide range of topics that involve understanding and making effective
use of field data. The journal has been reformed since July 2020 to
better serve the data science community in the era of data science.
Attractive features of the journal are completely free access, fast
review, and reproducible data science. We look forward to your
submissions.
Guest Editors:
Wenjiang Fu, Department of Mathematics, University of Houston
Annie Qu, Department of Statistics, University of California, Irvine
Momiao Xiong, Department of Biostatistics and Data Science, University
of Texas Health Science Center at Houston