Missing Data

Is there any issue that is more persistent and more difficult to solve than missing data?

It takes a perfectly good study and layers assumptions on it. There is a clear divide in how to handle it. One option is to argue "why would you not want to use real data" and rejects the assumptions of imputation. Of course, this approach makes it's own set of strong assumptions that are often not likely to be met.

So you'd think that doing the right thing and modeling the missing data is the way to go? Well, it's an improvement but it is pretty rare that the assumptions of missing data technique are met (missing at random is just not accurate in real data).

So what do you do? Most of the time I recommend modeling (inverse probability weighting or multiple imputation) but I must confess that the lack of a solution that is actually good is rather distressing!

A few thoughts on reviews

One of the most frustrating things that we have to face as researchers are reviews. We all want people to recognize the hard work that went into developing papers and grants. None of us got to the point that we were sending material out for review except by putting in many years and a startling amount of hard work.

So it is very annoying when strange reviews come back. But I have learned that there are a few basic rules:

Should the paper be rejected:

1) Decide if the criticisms are substantive or stylistic. If substantive, then you need to either redevelop the paper or retire it. This is never a pleasant discovery but, in the long run, you'll be glad that a sharp reviewer caught an issue. In my experiences, true substantive critiques are rare.

2) If the criticism is stylistic then don't put time into it. Likely the next set of reviewers will have different preferences. Resubmit rapidly.

3) If the criticism seems to apply to another paper, entirely, then seriously consider rewriting the confusing sections for clarity. You are the subject matter expert -- it is not reasonable that reviewers and readers will necessarily follow all of the nuances.

In the same vein, as a reviewer, asking for massive redevelopment for purely stylistic reasons is often a poor choice. Ask whether the paper is methodologically valid (no point in letting mistakes into the literature) and relevant. These are the real questions that need to be considered.

Peer Review is a frustrating process but it can really improve work if you take advantage of it.

Peer Review

There is an interesting post over at Drug Monkey on the issue of bias in peer review. I think that there are two issues that really strike me as important in peer review and it is easy to confuse them. One is the idea that individual reviewers will have preferences for the types of research that they like to see done. This issue is difficult, if not impossible, to solve.

Two, and more annoying, is the issue of competence and coherence in review. I cannot enumerate the number of reviews that I have gotten that had questionable elements. I remember one journal claiming that they could not publish a paper that had an "unverifiable assumption" in it. The assumption in question, no unmeasured confounders, was a pretty standard assumption for all research. Even clinical trials have this issue with loss to follow-up not necessarily being at random.

But the reviewer is protected from strong complaints of "what were you thinking?". Now, I too have certainly done peer reviews that could have been better. I think we all can think of examples of this. So I am not claiming to be "special" and I am sure that I have been cursed by an author more than once for not "getting it".

But I think that these concerns are what gives anonymous peer review it's bad name.

NIH Challenge Grants

I am beginning to think that the NIH challenge grants are a cunningly disguised trap. The NIH is giving out $200 million for grants that can be up to $1 million apiece and do not require preliminary data. This set-up could generate a lot of applications.

I think that it might make more sense to put that effort into a very solid R01 grant proposal and try to win longer term funding under less pressure of time and with more opportunity to generate productivity.

But, of course, the siren call of "no pilot data" is certainly sounding in my ears too!

A reply to Andrew Gelman

A reply to Andrew Gelman's latest post where he links to an old post on propensity scores:

My understanding of the issue is that there was also a prevalent user problem (creating selection bias) at least partially due to time-varying risk. While this could have been found and modeled, I am unsure about how propensity scores give any advantage over a thoughtfully constructed regression model. Unless the study you are thinking of had a lot more power to estimate predictors of exposure than outcomes due to very few outcomes (but I don't believe that this was the case with the Nurse's Health Study).

I'm not saying that better statistical models shouldn't be used but I worry about overstating the benefits of propensity score analysis. It's an extremely good technique, no question about it, and I've published on one of it's variations. But I want to be very sure that we don't miss issues of study design and bias in the process.

Issues of self-selection seriously limit all observational epidemiology. The issue is serious enough that I often wonder if we should not use observational studies to estimate medication benefits (at all). It's just too misleading.

This minor point of disagreement aside, I freely admit that Andrew Gelman is one of my heroes in the statistical community. I love some of his posts. His work on statistical significance is incredibly thought provoking, very helpful in clarifying thought and a must read for any epidemiologist.,

Self Selection

In a lot of ways, I think that different forms of self selection are the biggest threat to study validity in observational epidemiology. We see it in loss to follow-up when participants select out of clinical trials. We see it in important exposures like diet, exercise and alcohol use where the exposure is likely correlated with many other health seeking behaviors. Heck, we know that being adherent to placebo therapy is associated with good outcomes.

So the trick seems to isolating the effect of a single exposure. It is the process of thinking up ways to do this isolation that allows epidemiologists to really earn their keep.

Why do observational epidemiology?

Observational epidemiology studies are often the source of highly misleading results. And yet, despite this problem, they are essential to the better understanding of human health. There are many exposures, some of them quite critical, that cannot be studied in any other way.

My personal favorite is adverse drug effects. Clinical trials are often underpowered to detect adverse events; in order to show these effects trials often need to be combined. Given the logistics involved, it is helpful to show an association between the drug in question and adverse events in real populations.

I hope to discuss the many interesting challenges and ideas in this field of research as I try to muddle towards some sort of resolution in a confusing sea of fuzzy data.