Article type
Year
Abstract
Introduction/Objectives: Methods for combining survival data have not been widely used although four methods are available: 1) iterative generalized least-squares (IGLS, Dear 94), 2) meta-analysis (MA) of failure-time data with adjustment for covariates (MFD, Hunink 94), 3) non-linear regression (NLR, Shore 90) and 4) log relative risk (LRR, Voest 89). We compared thee accuracy, applicability and ease of use of these methods using individual patient data (IPD) and published data summaries from four advanced non small cell lung cancer studies.
Methods: Data from eight single arms of the four studies with 918 patients were used. Survival estimates from IPD were considered the gold standard. Survival curves from published articles were scanned into a graphical package, coordinates along each curve identified and data elements required for each method extracted semi-automatically. We examined the max difference between MA and IPD curves (Kohnogorow-Smimoff, KS test). We evaluated the applicability of the four methods with respect to their required data elements and assumptions. The effect of survival heterogeneity between studies on the combined survival estimates was assessed using the jackknife method. The ease of use of each method was assessed by 1) the technical understanding requirements from a user's perspective, and 2) how robust each procedure's implementation was with practical problems.
Results: The semi-automatic data extraction process was reproducible with both inter- and intra-observer reliability coefficients ICC above 0.95. Three methods except MFD operated on survival proportions. Censored patients were often not extractable from published articles hence MFD, operating on numbers at risk and failed, required additional assumptions. All methods were fairly accurate within a time frame in which all study contributed data to the estimates; the max difference ranged from 1.8% for IGLS, 2.1% MFD, 4.7% NLR and 3.7% LRR (KS critical value 6.3%). Beyond this time frame, IGLS, NLR and LRR could result in survival estimates to actually rise after a trial ended. With a max censor rate in one study arm of 22%, MFD produced reasonable survival estimates ignoring censored patients. Both NLR and LRR were highly susceptible to heterogeneity. IGLS however was consistently accurate in many scenarios. LRR was the easiest method to use for homogeneous study results.
Discussion: MFD could be ideal for non-ignorable covariates if all life-table data elements were available. Overall, IGLS was possibly the preferred method.
Methods: Data from eight single arms of the four studies with 918 patients were used. Survival estimates from IPD were considered the gold standard. Survival curves from published articles were scanned into a graphical package, coordinates along each curve identified and data elements required for each method extracted semi-automatically. We examined the max difference between MA and IPD curves (Kohnogorow-Smimoff, KS test). We evaluated the applicability of the four methods with respect to their required data elements and assumptions. The effect of survival heterogeneity between studies on the combined survival estimates was assessed using the jackknife method. The ease of use of each method was assessed by 1) the technical understanding requirements from a user's perspective, and 2) how robust each procedure's implementation was with practical problems.
Results: The semi-automatic data extraction process was reproducible with both inter- and intra-observer reliability coefficients ICC above 0.95. Three methods except MFD operated on survival proportions. Censored patients were often not extractable from published articles hence MFD, operating on numbers at risk and failed, required additional assumptions. All methods were fairly accurate within a time frame in which all study contributed data to the estimates; the max difference ranged from 1.8% for IGLS, 2.1% MFD, 4.7% NLR and 3.7% LRR (KS critical value 6.3%). Beyond this time frame, IGLS, NLR and LRR could result in survival estimates to actually rise after a trial ended. With a max censor rate in one study arm of 22%, MFD produced reasonable survival estimates ignoring censored patients. Both NLR and LRR were highly susceptible to heterogeneity. IGLS however was consistently accurate in many scenarios. LRR was the easiest method to use for homogeneous study results.
Discussion: MFD could be ideal for non-ignorable covariates if all life-table data elements were available. Overall, IGLS was possibly the preferred method.