The possible benefits of blood transfusions in anemic preterm infants have never been well investigated. There is some reliable data about the outcomes of babies randomized to different goals of hemoglobin concentration (for example the PINT study Kirpalani H, Whyte RK, Andersen C, Asztalos EV, Heddle N, Blajchman MA, Peliowski A, Rios A, LaCorte M, Connelly R, Barrington K, Roberts RS. The premature infants in need of transfusion (pint) study: A randomized, controlled trial of a restrictive (low) versus liberal (high) transfusion threshold for extremely low birth weight infants. J Pediatr 2006, 149(3):301-307.) but little data about what clinical benefits may truly be expected in an anemic preterm infant if you are unsure whether to transfuse them or not.

A new observational study attempted to determine whether anemic preterm babies would suckle better after a transfusion. (Bromiker R, Kasinetz Y, Kaplan M, Hammerman C, Schimmel M, Medoff-Cooper B: Sucking improvement following blood transfusion for anemia of prematurity. Arch Pediatr Adolesc Med 2012, 166(10):1-5.) Unfortunately the analysis of their data is an object lesson in how not to analyze observational data. I will explain…

After measuring the outcome of interest, in this case it was suckling before and after a transfusion, measured with a special bottle as the number of sucks in a 5 minute period, the authors divided the data into those with lower than average suckling, and those with higher than average.What is entirely unsurprising is that the babies with lower than average suckling had an increase on the second measurement. This is a phenomenon known as the regression to the mean. It is something that happens with any data set. you can construct completely random data sets and find that the cases below average on the first measure will increase overall on the second measure. And vice versa, those above the average on the first measure will be lower overall on the second measure.

I did this below, I randomly generated 200 normally distributed numbers in column A, the mean is 0 and the SD is 1. In column B I put another randomly generated series of 200 numbers, same mean and SD. Then I joined the first number in column A to the first number in column B, and so on.

The next thing I did was to take out all those that had a value below the mean in column A, and their partner in column B.

Then if you do a paired t-test the p value is <0.001 !

Remember these are entirely random numbers! This is regression to the mean, and analyzing data in this way is a common error.

This phenomenon is also responsible for most of the placebo effect. Most demonstrations of the placebo effect are nothing to do with the healing capacity of the body, or the power of our brains to mislead us, they are just due to the simple arithmetic phenomenon I just showed. If you have more movement restriction than usual in your knee on the day you get the placebo, then the next day you will tend to be better!

Another feature of this phenomenon is that the more abnormal the initial value, the greater on average will be the change. So if I plot the data in the last graph in a different way, comparing column A to the change between A and B, you find that there is a significant correlation (r=0.56, p<0.0001).

To return to the paper about sucking after transfusion, there was NO overall difference before and after transfusion. Only when the babies with lower than average sucking before transfusion are analyzed separately was there an increase after transfusion. As you can also see in the results, those with a higher than average sucking before transfusion actually had a decrease! Which is exactly what you would expect if transfusion had no effect.

This is not the first time similar interpretations have been made regarding transfusion. A paper published in 1984 showed that babies who had poorer weight gain prior to transfusion had more weight gain after transfusion, again entirely compatible with regression to the mean. Now it is touchy to perform randomized trials of blood transfusion, which is why there are so few, I guess. But the only way to answer questions about the clinical efficacy of transfusion is to have controlled trials.

And it is obviously not just in this situation, a paper published in the New England Journal in 1983 on the effects of digoxin in infants with a VSD and circulatory congestion showed that the 6 babies with the lowest shortening fraction before digoxin had an increase in this measure, and the 15 with better pre-digoxin shortening fraction had a decrease. The appropriate interpretation should have been that digoxin had no measurable effect, but the paper was published as if there were 6 ‘responders’ to digoxin, and 15 non-responders. I don’t think there has ever been a substantial RCT of digoxin in infants with circulatory congestion, which is worrying as it is potentially toxic, and this type of paper, incorrectly interpreting the results, has the potential to mislead physicians for years to treat with an agent which may be ineffective.

The importance of Controlled trials to answer clinically important question such as these cannot be overstated.