Thread by @AmJEpi

Hello #epitwitter! Time for an @epiellie @AmJEpi tweetorial.

Today’s topic is the Target Trial Framework for #causalinference and how to apply it to improving observational studies.

#epiellie

So, what is the #targettrial framework?

Well it’s not a new method! Instead think of it as pedagogical device that provides a structured way to build your research question and study design for observational studies and minimizes the potential for bias.

What does that mean?

To design an observational study, we first think about what the ideal hypothetical randomized trial (target trial) is that would let us answer our research question.

Then, we try to match our observational study as closely as possible to that trial design.

The target trial framework is an alternative to the PICO (population, intervention, comparison, outcome) approach to describing an observational study.

The target trial protocol should include all the PICO components by *also* clarify the start & end of follow-up, causal contrast, and statistical analysis plan.

Plus, the target trial needs two protocols — the trial you wish you could do, and the observational study you can do!

Let’s take a look at an example!

Our August issue has this great paper by @MissLodi, @_MiguelHernan, et al. which uses an actual trial as a starting point for the target trial, & then tries to emulate this as closely as possible from observational data.

academic.oup.com/aje/article/18…

Why might we want to use an existing trial as a target trial?

In order to ensure that we are “Comparing apples to apples”!

Randomized trials often differ in important ways from observational studies because of:

•different inclusion/exclusion criteria
•different target populations
•different treatment strategies
•different causal estimands!

Even if we could do a perfect observational study with no unmeasured confounding, no measurement error, etc, we can’t expect to get the same answer as a perfect randomized trial when those☝🏼 differences exist, because different questions should have different answers!

This is where the target trial framework comes in. It can help us make sure that we are aiming to answer the same question in our trial and our observational study.

So let’s dive into the paper!

(expand for poll)

The authors wanted to know why randomized trials and observational studies find different strength of effects when assessing the benefit of immediate treatment initiation for people living with HIV.

An RCT (INSIGHT-START) reported an intention-to-treat hazard ratio estimate of 0.43 (95% CI: 0.3,0.6), and a per-protocol HR estimate of 0.34 (0.2,0.5).

An observational study using the HIV-CAUSAL database reported a 7-year (per-protocol) risk ratio of 0.66 (0.6,0.8).

To compare these better, @MissLodi and coauthors used a three step process:

1) harmonize study protocols using the actual trial as the guide for the target trial
2) harmonize data analysis to estimate same measures & target parameters
3) sensitivity analyses for mismatches

For example:

START required 2 measurements of CD4 cell count above 500 cells/mm3 at least 14 days apart & within 60 days of randomization.

HIV-CAUSAL mostly had measurements every 90-180 days.

The harmonized protocol required 2 measures within 90 days.

Another example of harmonization:

START used a composite outcome of serious AIDS and non-AIDS or death, but HIV-CAUSAL had used a composite of AIDS onset or death.

The harmonized outcome was “earliest of AIDS or death”.

The causal contrast is an important harmonization point:
Randomized trials can potentially allow estimation of intention-to-treat and per-protocol effects. But intention-to-treat effects don’t make a lot of sense in observational studies, and we mostly want per-protocol effects.

After harmonizing all the components of the target trial, the next step is to emulated it with both the RCT data and the observational data.

This required re-analyzing START and HIV-CAUSAL based on the new harmonized protocol.

What do we see in the new analysis?

The estimated per-protocol hazard ratio in the re-analyzed START (actual trial) still suggests a *stronger* harm from delaying treatment compared to the HR from the HIV-CAUSAL (emulated trial).

So why might this be? Well, there were several components of the target trial which couldn’t be harmonized between the actual studies.

This is where stage 3 (sensitivity analyses) comes in. The paper explains these in lots of detail & is worth reading in full! Sneak peek here👇🏼

So there you have it. The target trial framework helps us think through our study design, and is a nice complement to using causal graphs to choose our confounding variables.

Have you ever used it or do you think you’ll try? (expand for poll)