I know that organized randomized studies are what is really needed, but there is a poor-man's substitute : Simply gather the data systematically, and data-mine anonymized records.
Of course, there will be unwanted correlations in the results (Dr X always does steroids), but those can be weeded out statistically (similarly for Hospital Y kills patients). But if the patients are receiving/not-receiving steroids at anything approaching random (as claimed in the article) then good data collection, and good statistics should be able to tease out results.
Many important findings in medicine are due to meta-analysis : whereby multiple trials are aggregated, and answers gathered across the meta-trial (and the questions being answered may even be different from what the trials were originally seeking to answer).
Hmm, I wonder if it's any harder to get the results of these meta-analyses "generally accepted" than if you did your own (equally statistically valid) trials?
In any case, weeding out the extra correlations requires more data, yada yada, people die while you're collecting it.
Of course, there will be unwanted correlations in the results (Dr X always does steroids), but those can be weeded out statistically (similarly for Hospital Y kills patients). But if the patients are receiving/not-receiving steroids at anything approaching random (as claimed in the article) then good data collection, and good statistics should be able to tease out results.
Many important findings in medicine are due to meta-analysis : whereby multiple trials are aggregated, and answers gathered across the meta-trial (and the questions being answered may even be different from what the trials were originally seeking to answer).