quantitation, and physiologic variability in clinical and investigational applications. J Am Soc Echocardiogr 1991;4:203–214. 6. Caiani EG, Corsi C, Sugeng L, MacEneaney P, Weinert L, Mor-Avi V, Lang RM. Improved quantification of left ventricular mass based on endocardial and epicardial surface detection with real time three dimensional echocardiog- raphy. Heart 2006;92:213–219. doi:10.1016/j.amjcard.2006.03.072 Statin Use and Age at Death: Evidence of a Flawed Analysis We read with great interest the recently published report by Mehta et al 1 on the effects of statin use on mortality. The report addresses a very important clinical issue. It must be expected that the results might have a major impact on future treatment recom- mendations, particularly for elderly subjects with elevated lipid levels. It is therefore cru- cial that the published results rest on sound scientific method and analysis driven by clarity and rigor. In our opinion, however, the study had severe methodologic prob- lems in its design and analysis that call into question the validity and applicability of the results and conclusion. First, the study population was not clearly defined. Apparently, it consisted of all eligible veterans who at some point in time from January 1, 1996, to March 31, 2004, were associated with 10 health centers. If this is correct, it is unclear what the investigators presented in Table 3, which supposedly lists baseline character- istics for the study population. At what date, for example, was age computed? Second, the sampling scheme of the study seems not to have been addressed adequately in the analysis. To become reg- istered as a user of statins, a patient must obviously first survive into the study period and then have a prescription recorded be- fore loss to follow-up or death. Either the investigators did not consider this aspect in the analysis, or they did not describe it in sufficient detail to allow critical appraisal. To illustrate the potential effects of ignoring the sampling scheme, we conducted a sim- ulation study, in which a given treatment has no effect on mortality. As in Mehta et al’s 1 study, the mortality followed a Weibull distribution for a population with a constant birth process, and the incidence rate of (chronic) drug use increased linearly with age for ages 35 years, whereas the inci- dence rate was zero for ages 35 years. Users of the drug were assumed to present prescriptions every 2 months after the onset of treatment, and everyone resided within the capture area throughout their entire lives. Subjects were classified as treated if filled prescriptions were observed within the 8-year period. At the start of the ob- servation period (“baseline”), the differ- ence in mean age between treated and untreated subjects closely resembled the mean difference in age at death, coin- ciding with the results of Mehta et al. 1 The 2 histograms of age at death in all who died in the observation period for the 2 groups are shown in Figure 1. Not only was the mean age at death greater in treated patients, the distribution was also narrower for the treated patients, as in the study by Mehta et al. 1 It should be noted that if the incidence rate of treatment onset were to decrease with age, the mean life- times between the 2 groups would be re- versed (i.e., those who were treated would apparently now live for a shorter period). In other words, histograms based on this type of data merely reflect the difference in age distribution between treated and untreated subjects, absent other factors, and the differ- ence observed by Mehta et al 1 can therefore not be taken as evidence of improved sur- vival due to statin treatment. Although ignoring the sampling scheme invalidates the study in itself, several other problems deserve to be mentioned, the most severe being the use of forward selection algorithms. It is well known in statistical research that this method cannot but create biased estimates with inflated statistical sig- nificance, because by definition, it uses sta- tistical significance for the inclusion of co- variates. 2,3 Furthermore, as a general rule, the timing of measurements was not de- scribed, hindering the valid interpretation of results. When was comedication measured? Before, during, or after treatment with st- atins, or at any time, as with statins? What was meant by posttreatment change in un- treated individuals? For these questions and several others, Mehta et al 1 provided no useful information. It is unfortunate that Mehta et al 1 missed the opportunity to use a unique body of data to provide a scientifically sound answer to a pressing clinical ques- tion. Instead of using proved epidemio- logic methods, they seem to have relied on computer-intensive analysis of a mas- sive data set without considering how the methods fit the research questions posed. From our point of view, some variant of a nested case-control study or a retrospec- tive cohort study might well have contrib- uted definitive and valuable insight into the real benefits of treatment with statins in the elderly. Although Mehta et al 1 sug- gested that the sheer amount of data led to unbiased estimates and can compensate for analytical shortcuts, we respectfully submit that methodologic problems only become propounded when data are very numerous. We consider the study by Mehta et al 1 ex- cellent proof of this rule. Henrik Støvring, PhD, MSc Dorte Gilså Hansen, PhD, MD Lene Jarlbæk, MD, PhD Helle Wallach Kildemoes, MPH, MA Jørgen Lous, MD Morten Andersen, PhD, MD Odense, Denmark 10 January 2007 1. Mehta JL, Bursac Z, Hauer-Jensen M, Fort C, Fink LM. Comparison of mortality rates in statin users versus nonstatin users in a United States veteran population. Am J Cardiol 2006;98:923–928. 2. Miller AJ. Selection of subsets of regression variables. J R Stat Soc A 1984;147:389 – 425. Figure 1. Distribution of age at death stratified on observed treatment status within an 8-year period. Simulated data with no treatment effect, but incidence of treatment onset increases with age. 1181 Readers Comments