Cardiovascular Medicine Journal Club

Statistical Questions and Commentary by Frank Harrell

Randomized trial of intravenous streptokinas, oral aspirin, both, or neither among 17187 cases of suspected acute myocardial infarction: ISIS-2. The Lancet 13 Aug 1988. [28Jul09]

  • A classic design and analysis
  • Subgroup analyses were beautifully put into context by adding a analysis of patients' birth signs
  • Article could have paid more attention to analyzing the interaction effect
  • Modern sequential monitoring techniques would have triggered early stopping of the study for efficacy

Quality of care for acute coronary syndrome patients with known atherosclerotic disease. Results from the Get With the Guidelines Program. Circ 2009; 120 :560-567. [25Aug09] Brilakis et al.

  • The authors treat acute coronary syndrome (ACS) as a homogeneous condition and hence ignore ACS patient heterogeneity. Are there angina characteristics, enzymes, troponin levels, EKG quantifications, etc. that when adjusted for would better define the problem and reduce the influence of atherosclerotic disease?
  • Exclusion of patients having some missing data is inefficient (because many of their variables are measured) and biased (because incomplete cases may have a different prognosis from the rest). A standard tool for making optimum use of available data is multiple imputation.
  • Did the authors simultaneously adjust for all variables in Table 4 when assessing the prognostic impact of polyvascular atherosclerosis? And where is LVEF? At the bottom of Table 3, only a few adjustment variables are listed.
  • A paper focusing on one risk factor implies that the authors are interested in what is unique about that risk factor, i.e., its independent contribution to prognosis. Did the key analyses adjust sufficiently for other patient characteristics?
  • A multivariable model predicting, as an ordinal response, the number of vascular disease beds, would shed light on what is unique about these patients. This is a propensity score analysis.
  • The authors considered P < 0.05 as "statistically significant". When are we going to get off the P-value bandwagon?
  • Use of the standard deviation (SD) assumes the data come from a symmetric distribution. What does the authors' statement that "the mean length of stay was 5.6 +- 6.7 days" say about the distribution of length of stay?
  • Why not replace long tables (Tables 1 and 3) with dot charts or other graphics?
  • The authors used SAS for analysis. SAS syntax lulls the analyst into assuming linear effects of covariates. If the effect of age or BMI were allowed to be nonlinear (e.g., slope increase at age 70 or BMI of 30), these variables would be more thoroughly adjusted for and the effect of the risk factor of interest might lessen.

Early diagnosis of myocardial infarction with sensitive cardiac troponin assays. NEJM 2009; 361:858-67. [29Sep09] Reichlin T, Hochholzer W, Bassetti S, et al.

  • An exceptional dataset that is amazingly under-analyzed, consistent with the absence of a senior biostatistician on a highly statistical paper
  • Authors ignored the fact that they had a nice cohort study
  • For the most part analyzed the data as if it was a case-control study
    • Emphasized receiver operating characteristic (ROC) curves, sensitivity (sens), and specificity (spec)
    • sens = Prob(test + | Dz +) (probability of a positive test given the disease is present); spec = Prob(test - | Dz -); both are in backwards time order and assume that disease and test results can be dichotomized
      • conditioning on disease present when the diagnosis is at the end of the workup
      • like sampling of cases in a retrospective case-control study
    • Figure 1 is based on an arbitrary "upper limit of normal" (multiples of 99th percentiles) instead of directly modeling risk in the cohort
  • Comparison of ROC areas is low power; it is insensitive to real differences in discrimination ability
    • Compare with multiple logistic regression likelihood ratio tests
    • Comparing ROC areas is the same as comparing two Wilcoxon statistics (one for treatment A vs B and another for treatment A vs C) instead of comparing treatments B and C head to head; it is generally not a good idea to subtract one rank statistic from another
  • No formal redundancy analysis to dissect how the information in the 4 assays overlaps
  • No estimation of risk as a continuous function of the markers as was done in Ohman EM et al NEJM 335:1333; 1996 (see below) for mortality
  • No assessment of the possible value of combining two or more markers in diagnosing AMI
  • Continuous troponin level vs. risk of death from Ohman EM et al (NEJM 1996):
    ohm96car.jpg

Trends in prevalence and outcome of heart failure with preserved ejection fraction. NEJM 2006; 355:251-9. [27Oct09] Owan TE, Hodge DO, Herges RM,, et al.

  • Is it acceptable that basic data such as height and weight were missing on so many patients?
  • The authors consider preserved ejection fraction (LVEF >= 50) and reduced LVEF (< 50) to be two groups of homogeneous patients. This is less problematic for the LVEF >= 50 group, but for the LVEF < 50 group there is a huge risk spectrum across LVEF values of 10 to 49. By treating these as homogeneous groups the results are not as easy to interpret and the power to detect differences is curtailed. The power and interpretation of results is a function of, for example, how many patients are in the sample with LVEF < 30. Had the authors removed all patients with LVEF < 40 from the analysis, the results would have been different.
  • The P-value for the LVEF comparison in Table 1 is meaningless.
  • Serum creatinine has a non-monotonic risk relationship. It is not adequate to just compare means of creatinine, and this risk factor should be modeled nonlinearly in the outcome model.
  • Time trends were modeled assuming linearity in calendar time. Time could have been modeled more flexibly.
  • The r and P-value in panel A of Figure 1 are not correct; these were based on grouped data.
  • Hazard ratios in Table 2 were not computed on reasonable intervals of continuous risk factors.
  • The analyses related to Figure 3 would have been on firmer ground had it used LVEF and calendar time in a joint dose-response-type relationship (with an interaction).

Prevalence and repair of intraoperatively diagnosed patent foramen ovale and association with perioperative outcomes and long-term survival. JAMA 2009; 302(3):290-297. Krasuski RA, Hart SA, Allen D et al. [26Jan10]

  • Length of stay of patients who died should have been deemed incomplete, i.e., LOS should be a analyzed using survival time methods that count a death on day d as a LOS that is right censored at day d
  • The authors misinterpreted P-values. The abstract states that "patients with PFO demonstrated similar rates of in-hospital death (3.4% vs. 2.6%, P=.11)". The conclusion of similarity could only have been made if the least favorable confidence limit excludes a clinically important mortality increase, which is not the case here.
  • Likewise, the statement that "PFO repair was associated with no survival difference (P=.12)" is problematic.
  • Need confidence intervals for all differences of interest, and confidence bands for differences in Figure 2
  • See Absence of Evidence is Not Evidence for Absence
  • Stepwise variable selection was used; this has been shown to be unreliable and to bias statistical results
    • Top p. 292: Entry criteria of P < 0.07 and choice of variables that are selected in at least 50% of bootstrap resamples from the data are arbitrary criteria
  • There is a need to adjust for potential prognostic variables, not just confounders (the authors may have done this but I'm not clear)
Topic attachments
ISorted ascending Attachment Action Size Date Who Comment
ohm96car.jpgjpg ohm96car.jpg manage 91.6 K 27 Sep 2009 - 08:39 FrankHarrell Continuous troponin level vs. risk of death from Ohman EM et al (NEJM 1996)
final-dotplot.pdfpdf final-dotplot.pdf manage 4.0 K 25 Aug 2009 - 11:45 FrankHarrell Dot plot showing how a complex table can be represented graphically
Topic revision: r8 - 30 Jan 2010, FrankHarrell
 

This site is powered by FoswikiCopyright © 2013-2022 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding Vanderbilt Biostatistics Wiki? Send feedback