MERINO trial - adjusting for infection source

Hi all

You have helped me before and I come again :smiley: !

This is an interesting trial, for many reasons. It compares two antibotics, both very broad spectrum, for a set of resistant bugs (ESBL E.coli + Klebsiella). Meropenem - one of the antibiotics, was found to be significantly more efficacious than Piperacillin-Tazobactam.

23/187 died in Piperacillin arm (all cause)
7/191 died in Meropenem arm (all cause)

This is a quite significant result and has altered therapy around the world. There are a couple of caveats - firstly, no-one was adjudged to have actually died from infection (!). This is a problem in infection trials in the critically ill - to get that sick to get a nasty bug you have to be pretty sick and your underlying disease kills you. So the addendum makes for interesting reading - many had incurable cancer and perhaps the antibotics were hardly relevant.

I am intrigued about one thing - and I know this is a hated topic - but post randomisation adjustment for known variables that are associated with the outcome.

The key here is UTI. Those with UTI (and in fact, this is mentioned in the protocol- and is pre-specified) do much better with bloodstream infection, and therefore some (maybe not on this forum!) would adjust for this.

As chance would have it, the number of patients with a UTI source differed greatly between the two groups:

103/188 in the Piperacillin arm (56%)
128/193 in the Meropenem arm (67%)

So far, so good. They performed a planned multivariable regression including UTI (and Charlson’s comorbidity index) and the treatment arm and calculated adjusted and unadjusted odds ratios:

Adjustment for a urinary tract source of infection and the Charlson Comorbidity Index score resulted in little change in the findings (unadjusted odds ratio, 3.69 [1-sided 97.5% CI, 0 to 8.82]; adjusted odds ratio, 3.41 [1-sided 97.5% CI, 0 to 8.38) (eTable 5 in Supplement 2).

What I don’t understand, is how that odds ratio doesn’t change very much?

Given the small number of events, there must have been quite significant correlation between the variables. Here is the UTI data:

image

I don’t understand how the OR for UTI can be 0.34 - suggesting they are much less likely to die, but then when adjusting for treatment arm, it doesn’t affect the treatment arm OR practically at all?

Or am I really stupid and I have completely forgotten how regression and correlation between variables work? and these two results are completely plausible?

Thanks

Good question. The more extreme the imbalance or the more extreme the effect on outcome the more difference the adjustment will tend to make. This is not an extreme imbalance, so the answer may depend on how prognostically strong UTI is.

You’ve already got the best possible answer from Dr.Harrell, but this post might also apply to your question: https://thestatsgeek.com/2014/02/01/adjusting-for-baseline-covariates-in-randomized-controlled-trials/

This trial sounds like it’s had an outsized impact on practice given the very small number of outcomes of interest observed (?) Has this finding ever been replicated?

Thanks to you both.

In response -yes, it has had an oversized impact - but its a hard area to do research. Required >20 sites across multiple continents and most people excluded for one reason or another.

There are other issues which are more microbiological that are harder - some of the definitions of resistance and sensitivity are not agreed by all. I doubt it will be replicated. Which slightly stresses me out - it’s not many events to change practice worldwide (although some centres still don’t believe it).

I still cant quite work out how the OR doesn’t change, but if you and Frank Harrell think its plausible, then that’s enough for me.

Thanks again.