1) "Signal detection" --> find AEs that differ between arms.
2) Estimate true probability of an AE of interest, P(AE).
I will focus on 2) below.
3/n Assume we are interested in P(AE) - this is our estimand.
What do proportions in above table estimate? The incidence proportion:
IP = P(AE happening in [0, t] and that this AE is observed).
4/n
Assumptions:
- Interpretable only if all patients have ~ same follow-up.
- If we have (administratively) censored patients then IP *underestimates* true AE probability.
Challenges for estimation of P(AE): 1) Unequal follow-up in both arms. 2) What is AE hazard?
5/n
3) Competing events (CE) such as death before AE or treatment discontinuation leading to end of AE recording.
Simply censoring CEs gives biased estimates of AE probabilities P(AE).
Incidence proportion accounts for 2) and 3).
What are other estimators?
6/n
See published SAP of SAVVY project (Survival analysis for AdVerse events with VarYing follow-up times) for list of alternative estimators:
Their behaviour w.r.t. to above 3 criteria are indicated in Table 1 in this paper: arxiv.org/abs/2008.07883
7/n
In absence of CEs (1 - Kaplan-Meier) is unbiased.
However, in presence of CEs (1 – KM) *overestimates* true P(AE).
Aalen-Johansen is only estimator of P(AE) that accounts for censoring, CEs, and makes no constant hazard assumption --> can be considered gold standard.
8/n
But how large can bias of other estimators become, notably IP and (1 - KM) with simple censoring of CEs?
SAVVY: academia + industry collaboration exploring bias on data of 17 RCTs from various indications.
- Comparison of P(AE) in 2-arm RCT: arxiv.org/abs/2008.07881
--> IP underestimates P(AE) up to factor 3!
--> (1 - KM) overestimates P(AE) up to factor 5!
Papers have many more analyses, e.g. discrepancies of estimators when categorizing AE risk acc. to @iqwig guidelines.
10/n
Empirical quantification illustrates how large bias can be.
Goal of SAVVY: raise awareness that when estimating P(AE):
- Start with scientific question: Signal detection or estimation of P(AE)?
- Censoring? Competing events?
- Appropriate estimator --> Aalen-Johansen.
11/n
Aalen-Johansen is available in any standard software.
Clinical trials in patients with hematological malignancies often present unique challenges for trial design due to complexity of tmt options and existence of potential curative but highly risky procedures, for example, SCT or tmt sequence across different phases.
The paper illustrates how to apply the estimand framework in hematological clinical trials and how the estimand framework can address...
3/n
... potential difficulties in trial result interpretation. Three phase 3 RCTs are used to illustrate different scientific questions and the consequences of the estimand choice for
- trial design,
- data collection,
- analysis,
- and interpretation.
I give a list of activities that were kicked off by the ICH E9 estimands addendum (I believe) and that I am aware of (and/or involved in). I am sure there is much more, so no ambition at all to be comprehensive.
Very interesting paper on carbon footprint of pharma industry. Some takeaways:
1) Pharma significantly *more* emission intensive than automotive industry.
2) Highest and lowest intensity emitter in 2015 differ by factor of 5.5!
2/3
3) From 2012-2015 "...the leader of the pack,
namely @Roche, achieved both highest increase in revenues as well as highest reduction in emissions, suggesting, albeit on anecdotal basis, that financial and environmental performances are far from being exclusive."
3/3
4) Roche, J&J, and Amgen are already today below the 2025 target emission according to the Paris climate agreement.
There are many reasons to work for Roche, but this is definitively one more, and not the least important for me!
1/n I was asked to give an industry statistician's view on A) below. Disclaimer: I do not know about the exact regulations (which might also be region-specific). What I offer is a 1st hand experience of what happens around a trial stopping early. A thread.
Planned efficacy interim after ~245 PFS events. 1) Data with sponsor, except for tmt assignment (=rando codes). 2) Rando codes with IxRS vendor. 3) Indep. Stat. Reporting Group (ISRG) coordinates.
Registration still open for EFSPI regulatory statistics webinars. Currently we have 627 and 534 registrations for the two webinars. Webex can handle 1000 people dialing in, so go ahead and register! 😉