Skip to main content

Table 3 Results, Overall Survival, KRAS identifiable population

From: A cautionary tale: an evaluation of the performance of treatment switching adjustment methods in a real world case study

Analysis*

Estimand 1 (all KRAS assessable patients) hazard ratio (95% CI)

Estimand 2 (KRAS WT) hazard ratio (95% CI)

Expected ‘truth’§

0.76–0.77 (0.60–0.98)

0.55–0.73 (0.41–0.93)

ITT

0.97 (0.79–1.18)

0.99 (0.75–1.29)

IPCW 1 (primary analysis, inclusive model, stabilised weights)

0.75 (0.59–0.95)

0.55 (0.28–1.09)

IPCW 5 (secondary analysis, inclusive model, stabilised weights)

0.75 (0.59–0.96)

0.65 (0.50–0.85)

TSEsimp 1 (primary analysis, inclusive model, with re-censoring)

0.64 (0.46–0.98)

0.44 (0.24–0.95)

TSEsimp 9 (secondary analyses, inclusive model, with re-censoring)

0.70 (0.54–0.90)

0.59 (0.43–0.78)

TSEgest 1 (primary analysis, inclusive model, with re-censoring)

0.54 (0.30–1.51)

0.31 (0.13–2.12)

TSEgest 9 (secondary analyses, inclusive model, with re-censoring)

0.79 (0.57–0.99)

0.67 (0.46–0.86)

RPSFTM 1 (with re-censoring, ever treated, interval bisection)

0.91 (0.49–1.68)

0.87 (0.02–35.88)

  1. Notes: CI - Confidence Interval; IPCW - Inverse Probability of Censoring Weights; ITT - Intention to treat; KRAS - Kirsten Rat Sarcoma Virus; RPSFTM - Rank preserving structural failure time model; TSEgest - Two-stage estimation with g-estimation; TSEsimp - Two-stage estimation with simple regression; WT - Wild Type. Confidence intervals for TSE analyses were calculated by bootstrapping the entire adjustment analyses, with 5,000 samples. For TSEgest 1 the g-estimation process did not converge in 17% of samples. For TSEgest 9 the g-estimation process did not converge in 4% of samples
  2. *For IPCW, primary analyses only included patients with WT KRAS in models used to obtain weights. Secondary analyses added BSC MT KRAS patients to the weighting models, under the assumption that KRAS was not prognostic for survival in patients receiving BSC. For TSE, primary analyses compared BSC WT KRAS switchers to BSC WT KRAS non-switchers to estimate the effect of switching. Secondary analyses compared BSC WT KRAS switchers to BSC WT KRAS non-switchers and BSC MT KRAS patients (irrespective of switch status) to estimate the effect of switching. ‘Inclusive’ models included all variables considered potentially important, and ‘reduced’ models included only those variables considered most important. The analysis number refers to the way in which the method was applied - see supplementary materials for further details
  3. §The expected truths are taken from Table 1, with the point-estimate range representing the range of point-estimates observed in the relevant analyses from Study 20020408, 20100007 and CO.17, and the confidence interval range representing the minimum and maximum 95% confidence intervals reported for these point-estimates
Back to article page

BMC Medical Research Methodology

ISSN: 1471-2288

Contact us