|
|
|
|
@ -4,8 +4,8 @@
|
|
|
|
|
\begin{document}
|
|
|
|
|
|
|
|
|
|
The identification strategy centers on the fact that, in the U.S., clinical trials
|
|
|
|
|
update the publically available information on \url{ClinicalTrials.gov}, which are available
|
|
|
|
|
as historical snapshots through the website.
|
|
|
|
|
update the publically available information on \url{ClinicalTrials.gov}, which are then made
|
|
|
|
|
available as historical snapshots.
|
|
|
|
|
These updates typically include information such as additional sites conducting the study,
|
|
|
|
|
the study status, and expected or current enrollment figures.
|
|
|
|
|
By measuring enrollment and other factors prior to the conclusion of a trial, we
|
|
|
|
|
@ -13,8 +13,8 @@ can measure the effect of enrollment on trial conclusion
|
|
|
|
|
(specifically whether it is registered as completed or terminated).
|
|
|
|
|
In particular, this avoids measuring the joint determination of enrollment and conclusion
|
|
|
|
|
status arising from trials terminated early.
|
|
|
|
|
Figure \cref{Fig:CausalModel} describes the structural causal model used to justify
|
|
|
|
|
our identification claims.
|
|
|
|
|
Figure \ref{Fig:CausalModel} describes the structural causal model (SCM) used to justify
|
|
|
|
|
the causal identification
|
|
|
|
|
|
|
|
|
|
\begin{figure}[H] %use [H] to fix the figure here.
|
|
|
|
|
\tikzfig{../assets/tikzit/CausalGraph}
|
|
|
|
|
@ -22,7 +22,84 @@ our identification claims.
|
|
|
|
|
\label{Fig:CausalModel}
|
|
|
|
|
\end{figure}
|
|
|
|
|
|
|
|
|
|
The identification strategy is based on the backdoor criterion due to \cite{PEARL1995}.
|
|
|
|
|
As the backdoor criteron depends on the SCM being a Directed Acyclic Graph (DAG, the first
|
|
|
|
|
step is to justify the DAG in \cref{FIG:CausalModel}.
|
|
|
|
|
|
|
|
|
|
% The data consists of individual snapshots
|
|
|
|
|
% Describe "states"
|
|
|
|
|
% Also, snapshot states are dependent across time
|
|
|
|
|
% Define conclusion state vs snapshot state.
|
|
|
|
|
The key feature of the data is that it consists of sequences of trial snapshots for each trial.
|
|
|
|
|
Snapshots prior to the start of the trial capture expected enrollment and time to completion,
|
|
|
|
|
while snapshots during the trial record actual enrollement figures, current status,
|
|
|
|
|
and the date the snapshot was recorded.
|
|
|
|
|
Finally, after a trial concludes, snapshots list final enrollment and the date at which the
|
|
|
|
|
last participant was examined\cite{CLINICALTRIALS-data_spec}.
|
|
|
|
|
In the discussion below, I refer to a snapshot's ``state'' as the enrollment, duration, and status
|
|
|
|
|
recorded at the time of the snapshot.
|
|
|
|
|
%TODO: make sure data section discusses the normalization of enrollment and duration.
|
|
|
|
|
Additionally, I distinguish between the state at trial conclusion and state from a snapshot during the running trial as
|
|
|
|
|
``conclusion state'' and ''active snapshot state''.
|
|
|
|
|
% Describe market conditions.
|
|
|
|
|
Associated with each trial snapshot are the market conditions existing at that point in time.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
%Describe the observed and unobserved events and their supposed relationships.
|
|
|
|
|
%%%%% Relationships of interest
|
|
|
|
|
% Snapshot State -> Conclusion state
|
|
|
|
|
% Discuss how the data captures this - time dependence
|
|
|
|
|
%TODO
|
|
|
|
|
|
|
|
|
|
% Market -> Snapshot state
|
|
|
|
|
% Market -> Conclusion state
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
%%%%% Confounding relationships and controls
|
|
|
|
|
% Disease Burden -> Market Conditions, Snapshot State
|
|
|
|
|
In addition to the relationships of interest between teh active snapshot states and
|
|
|
|
|
the conclusion states, there are various biasing effects that need to be accounted for.
|
|
|
|
|
The first of these is the fact that enrollment and the drugs currently on the market are
|
|
|
|
|
both affected by the number of people who are affected by the disease under examination.
|
|
|
|
|
This biases not only the estimate of the total causal effect of market conditions
|
|
|
|
|
on conclusion state but also the direct effects of both
|
|
|
|
|
market conditions and active snapshot enrollment on conclusion state.
|
|
|
|
|
Additionally, it biases the estimation of the effect of market conditions on
|
|
|
|
|
active snapshot enrollment.
|
|
|
|
|
I plan on using the WHO's Global Disease Burden Survey
|
|
|
|
|
to control for population size. %CITE - ekaterina
|
|
|
|
|
|
|
|
|
|
% Biasing Pathways
|
|
|
|
|
|
|
|
|
|
% Compound Safety -> Current Adverse Events -> Conclusion State.
|
|
|
|
|
% Note: Compound Safety -> Current Adverse Events -/> Snapshot State.
|
|
|
|
|
% Even if it were an issue, the direct events should still be identified?
|
|
|
|
|
A second biasing effect is related to the fact that a compounds safety drives both beliefs about
|
|
|
|
|
the compound -- affecting active snapshot enrollment -- and the current adverse effects
|
|
|
|
|
which directly influcences the conclusion state by leading to terminations.
|
|
|
|
|
The backdoor criterion implies that controlling for whether or not prior trials have
|
|
|
|
|
occurred will eliminate bias.
|
|
|
|
|
%TODO: discuss how you will be conditioning on prior trials, i.e. per compound or just phase 3 etc.
|
|
|
|
|
|
|
|
|
|
% Compound Efficacy -> Measured Effectiveness -> Conclusion State
|
|
|
|
|
Similarly, the last confounding factor is that of measured effectiveness.
|
|
|
|
|
When running a trial, the sponsor will get periodic updates as to the measured effectiveness.
|
|
|
|
|
If this is lower than expected, the trial may conclude early.
|
|
|
|
|
Although this is a direct effect, the issue comes through the backdoor path through prior trials
|
|
|
|
|
and beliefs about the compound.
|
|
|
|
|
Thus controlling for prior trials eliminates this path as well.
|
|
|
|
|
% Control
|
|
|
|
|
% Compound Safety, Compound Efficacy -> Prior Trials -> Beliefs about Compound
|
|
|
|
|
%
|
|
|
|
|
|
|
|
|
|
%%%% Variance controls
|
|
|
|
|
% Sponsor Changes -> Conclusion Status
|
|
|
|
|
Finally, the last control variable is that of sponsor changes.
|
|
|
|
|
As sponsors are captured at each snapshot, it is possible to measure when a sponsor has changed.
|
|
|
|
|
Changing sponsors is a potentially disruptive event, and so it is likely to affect the probability
|
|
|
|
|
that the trial is canceled early.
|
|
|
|
|
The purpose of including this control is to reduce the variance of our estimates.
|
|
|
|
|
%Describe what causal effects are identified by the backdoor criterion.
|
|
|
|
|
|
|
|
|
|
\end{document}
|
|
|
|
|
|
youainti
4 years ago