Oh, and did I mention this is one of the most complex statistical analyses for a study
I've reviewed on these pages in the last 7 years? Don't worry, I won't get you too
bogged down in the details!
This study struck me as very illustrative of how researchers approach gathering information
about relatively rare diseases. Any meaningful data ideally would involve a large
number of participants, so for rare diseases that would involve multiple centers.
It's tough to coordinate that for a prospective randomized controlled trial of a treatment
intervention, but maybe a bit easier (still tough!) to gather data. That's usually
done by having national or international disease registries, where clinicians can
enter selected data about their patients, ultimately to be analyzed and reported by
a central coordinating team.
The current study pertained to "pediatric Clinically Isolated Syndrome (pCIS)," defined
as a "monofocal or multifocal clinical central nervous system event of presumed inflammatory
demyelinating cause with acute or subacute onset in the absence of encephalopathy,
not explained by fever or systemic illness" and not meeting certain criteria for other
disorders such as multiple sclerosis or acute disseminated encephalomyelitis. pCIS
tends to recur and may be a harbinger of pediatric-onset multiple sclerosis. Prior
studies had suggested that use of disease-modifying drugs (DMDs; e.g. interferon-beta,
monoclonal antibody drugs) were associated with lower risk of recurrent neurologic
Researchers used 2 different pCIS registries, one based in Italy and the other more
international, and subjects for this study were drawn from 76 different centers around
the world (none from the US). Based on their analysis of 770 patients, they found
that DMD use was associated with a lower rate of second attacks and of disability.
A few take-home points: this was a retrospective observational study, though the authors
termed it a "retrospective observational study performed on prospectively acquired
data." What I take that to mean is that clinicians entered data about their patients
prospectively, but not every patient had all the data items available. For example,
not all children had cerebrospinal fluid analysis performed. That becomes a problem
when one tries to analyze the results looking at different variables (multivariate
analysis). If some of the data points are missing, it doesn't work. A way around this
is a form of legitimate fudging called imputation or bootstrapping. We've talked about
this before, and I noted even our beloved US Census Bureau has used the technique when they are
unable to get full data with their mail and door-to-door surveys. It's a method to
"assume" what the missing data would have been, based on the rest of the dataset.
It may send a bit sketchy on the surface, but it is a validated statistical method.
As I indicated at the top, other aspects of the statistical methodology are well beyond
my expertise; 1 is of particular interest, a computerized methodology for recursive partitioning called Recursive Partitioning and Amalgamation (RECPAM) methodology. If you can fully
understand it, please contact me so I can learn too!