Comments on Thavendiranathan et al. Do Blood Tests Cause Anemia in Hospitalized Patients?

Authors


Address correspondence and requests for reprints to Dr. Desbiens: Department of Medicine, University of Tennessee College of Medicine, Chattanooga Unit, 975 East Third Street Box 94, Chattanooga, TN 37403 (e-mail: desbiena@erlanger.org).

To the Editor:—I am concerned about important methodological problems in the recent article by Thavendiranathan et al.1 The exclusion criteria seem overly restrictive. For example, though blood loss is difficult to quantify clinically, the amount of hemoglobin (Hgb) transfused is not, and could have been included in the analyses as a continuous variable. In addition, several other exclusions could have been eliminated by using indicator variables. The authors report their continuous variables in Table 1 as if they were normally distributed. I suspect that they are not, especially for an important variable like mean volume of phlebotomy for which they report a very large calculated standard deviation. Interquartile ranges would have been more informative. I'm not sure why the authors prescreened their candidate variables in the model using univariate analyses. This approach overfits models and leads to confidence intervals that are not appropriate.2 The authors also do not report if they adjusted for the clustering of patients in their analyses.

The authors report that Hgb concentration (C) on admission is strongly associated with drop in Hgb C during hospitalization but this is at least partially an artifact of how they set up their regression equation. Because the change in Hgb C (the dependent portion of the regression equation) is calculated as admission Hgb C minus discharge Hgb C, then Hgb C on admission when introduced on the independent side of the equation must perforce be associated with change in Hgb C.

A method like ordinary least squares regression can indicate the importance of an independent variable such as phlebotomy volume compared with all other variables in the model by producing partial R2's for each independent variable in the model. In addition, by reporting adjusted R2 the analyst can give the reader some idea of how well the model performs. These statistics are especially important for studies of changes in Hgb C that are susceptible to confounding by indication—the sicker the patient the greater the indication for blood drawing and the greater the effect on Hgb C.3 Unfortunately, the authors do not report these statistics.

Finally, the authors suggest that the Hgb C drop they observed is clinically significant but their references to this statement are a textbook of hematology and a 1994 article that references an earlier edition of the same textbook.4,5

Ancillary