Hacker News Reader

Comment by acqq | Hacker News Reader

The group receiving Calcifediol had more "no bad risk" patients:

"At least one prognostic bad risk factor(@)

Group receiving Calcifediol: 48%

Group without Calcifediol: 61.54%"

"@) Patients with at least one of the following risk factors (age >60, previous lung disease, chronic kidney disease, diabetes mellitus, hypertension, cardiovascular disease or Immunosuppressed and transplanted patients)."

That's what can be concluded from "at least one": (52% had no bad risk in D group, but only 38.46% had no bad risk in the other group). But it is also not clear where there were more patients with "multiple" factors! Or if they were those who had more problems at the end.

The whole paper contains neither raw data nor any graphs and only means and standard deviations, as far as I see? I would personally really like to see the graphs of distributions or to use raw data to check myself.

The critical question is how good the randomization was done.

View on HN

Randomization doesn't ensure perfectly balanced groups-- it just ensures that the imbalances of variables you don't measure (including things that change after randomization not related to your intervention, with blinding) are drawn from a distribution that you can apply rigorous statistical reasoning about.

Yes, we can see some things are a little unbalanced. But the effect is so massive: we might see 4/5 of the control's rate of ICU admissions if the "one prognostic bad risk factor" determined outcome entirely; instead, we see 1/50 of the rate.

> we might see 4/5 of the control's rate of ICU admissions if the "one prognostic bad risk factor" determined outcome entirely

But we still don't know if the persons with multiple bad risk factors were those who ended with bad outcomes? As far as I understand, if it was like that or not can't be seen from the paper at all, and I can imagine that it could have happened. I would really prefer the more raw data to the tables with the selected means and deviations given.

You have to get very, very unlucky on the dice rolls to get samples that are so tilted-- to pick 26 people out of 76, and somehow come up with 13/14 of those requiring ICU in the smaller group, if vitamin D has no effect. Outside of deliberate rigging... (it might even be difficult to deliberately sort and get this much of an imbalance).

Indeed, checking that the two groups look similar after randomization is completely optional. This study does an OK job of doing so.

There's both known and unknown reasons why someone might be predisposed to have a bad outcome. The reason we randomize, rather than try and make "balanced" groups, is that it addresses both unknown and known factors.

Yes, you can, by chance, get more people who are going to have a bad outcome in one group; about 5% of the time you'll get a p<0.05 finding this way. :P

Reply on news.ycombinator.com