Target RWE’s Post

Transcript

The simulation work will be discussing today came out of a larger comparative effectiveness study we were conducting for a client. The study compared a new treatment for a rare chronic liver condition with the standard of care. Since both the disease and the treatment were uncommon, we decided to use a sequential nested target trial emulation design. We chose this approach because it allowed us to maximize the number of patients included. In the study. However, we realized what we were doing was quite different from what we would do if we actually conducted a trial in which everyone was randomized at the beginning. This got us thinking, what exactly are we estimating with the sequential nested trial approach? What are the implications of this relatively novel design? In a treatment decision study, patients are included at the times when doctors make treatment decisions. We simulated 2 studies within this framework. The first is a single point randomized trial where patients are randomly assigned to treatment at the first decision and followed until the outcome. This is the type of trial we'd aim to conduct in real life to evaluate treatment efficacy if we had unlimited time and money. The second is a sequential nested trial using real world data where patients can be included in the untreated group at each decision point until they receive treatment or experience the outcome. Once they're treated, a patient contributes only that index and is followed until the outcome or end of follow up. The severity of the disease influence treatment assignment and how often these treatment decisions were being made. So in this study, we thought that the answers to these questions might depend on if the treatment effect differed according to other covariates. So in this simulation study, we were looking at disease severity and we considered those scenarios where the treatment effect did and didn't differ in three different study designs. So in the first one, we looked at a single point trial and we compared that to the true estimates from a single point trial and expectedly. Those were unbiased. We then considered the trial emulation, so this new methodology, but we considered a scenario where the treatment effect did not differ by disease severity. And what we found is that it's really biased unless you use the traditional methods to adjust for confounding. So in this case, this was we were emulating observational data and there was confounding by disease severity and we had to adjust for it. In the third study, we considered a scenario where that treatment effect. Offered by some other covariate, in this case disease severity. And you can kind of see a slow and attenuation of the bias over time so that crude estimates vary biased compared to a single point trial. If we use traditional confounding adjustment methodology, it's a little less biased. And then once we actually took the additional step of standardizing to the distribution of that covariate in a single point trial, we have the lease bias. And so this suggests that when you're using these this study design, you need to consider how the distribution of these modifiers differ from the distribution and the trial that you would actually want to conduct in order to get the estimate you're really looking for. In rare diseases, every patient matters. The sequential nested trial design can be incredibly useful in these settings, especially when you have a clear time point at which treatment decisions are being made.