Dear all:

The reason why one cannot shift groups when you have randomized a treatment is because of the 
statistical assumptions around the randomization process.  You can have "drop out", but that must 
be reported and explained in the study.  If the drop out is "biased" --- that is there is some 
reason for the drop out that coincides with the treatment effect that can negate the apparent 
findings of the study.  In some cases, particularly when one is evaluating "effectiveness of an 
intervention" rather than the "efficacy of a treatment" drop outs can be important in terms of what 
they tell you about how to apply the intervention. 

To go back to my original example of an unintended contamination - the Bolivia study had 
contamination of the control group, not from any deliberate mixing up of placebos and iodine 
capsules or anyone buying iodine capsules on the market because they realized that they were 
beneficial. The cause of the "contamination" that improved the iodine levels of all the children is 
unknown.  The endocrinologists I've talked to speculate that the levels of iodine were so high 
(most of the iodine is excreted out of the body within a few days of ingestion - the rest is retained 
in body fat stores) that iodine may have filtered into the water system through the children's urine.  
No proof of this speculation.

The Bolivia study authors did something creative with their data.  They analyzed the Change in 
iodine level with the Change in cognitive tests.  This showed an association.  Such an analysis, 
however, does not allow one to make a "probability" statement because you no longer have 
randomly assigned groups.  You can only make a statement of association, otherwise known as 
"plausibility".  While such a statement is suggestive, it cannot really determine causality.  You can 
always find spurious associations that just happen by chance and have no real relationship to each 
other. 

So, in the Schanler study, the case of the of the 17 infants(20%)  in the DM (pasteurized human 
donor milk)  most certainly could have washed out the effect of DM.  I might have handled this 
differently and considered them drop outs, BUT that would have diminished the sample size even 
further.  There are some really excellent new statistical tools that were developed when AIDS 
became a problem.  These tools are beyond my skills.  The tools were designed to anticipate a 
treatment that might actually be harmful - one that might increase the death rate even higher 
than the AIDS it was intended to treat.  So, there are rules for picking up on a significant enough 
climb in a serious risk and stopping the treatment at that point.   "Stopping a treatment"  which is 
essentially what happened in the DM group.  A significant number failed to thrive and because of 
the risks the treatment was stopped.  Had this potential problem been anticipated, a sophisticated 
statistician would have taken this into account prior to the study.  However, I think not one of us 
would have ever anticipated this as a potential outcome.

What is interesting, however, is that despite the contamination, the DM group had similar levels of 
NEC as the MM group.  We just don't have the power to figure out if this really means anything or 
if more infants had been included whether this would hold up.  

Also, one cannot make a valid statistical comparison on a "probability" level between the mother's 
milk (MM) group and the combined preterm formula (PF) and pasteurized donor milk (DM) groups 
because the ability to have enough milk to be in the MM group was NOT randomly assigned.  This 
drops the level of analysis to association and sources of confounding (eg. mother's age, 
education) should be included in the analysis.  This is a frequent occurence with breastfeeding 
studies that two randomly assigned groups such as DM and PF or two different types of formula 
are compared against a group that is not randomly assigned.  The two randomly assigned groups 
should be compared against each other.  The other comparison should be reported in a separate 
table and should never be treated as if it were a comparison of a randomized treatment.  

All-in-all I found that the study was excellently designed.  I'm impressed that Pediatrics published 
a negative study, but probably did because of the overall quality.  There is always a criticism of 
journals that don't publish negative findings.  I am unhappy about the extrapolation of the results 
beyond what can be concluded statistically.  In this case, I think the author's discussion section 
should have been far more cautious and I think they did make statements that cannot be 
supported statistically.

Susan Burger, MHS, PhD, IBCLC
PS.  Probably have bored you all to death by now.

             ***********************************************

To temporarily stop your subscription: set lactnet nomail
To start it again: set lactnet mail (or digest)
To unsubscribe: unsubscribe lactnet
All commands go to [log in to unmask]

The LACTNET mailing list is powered by L-Soft's renowned
LISTSERV(R) list management software together with L-Soft's LSMTP(R)
mailer for lightning fast mail delivery. For more information, go to:
http://www.lsoft.com/LISTSERV-powered.html