Meta-analysis of median survival times with inverse-variance weighting

Sean McGrath, Jonathan Kimmelman, Ömer Öztürk, Russell Steele, Andrea Benedetti, Benedetti, AndreaarXiv 2025 · DOI · arXiv

This paper, recently uploaded to arXiv, provides a practical solution for researchers trying to synthesize survival data when the standard "Hazard Ratio" isn't enough.

TL;DR

The authors validate a simple "Wald approximation" method to back-calculate standard errors from reported confidence intervals of median survival times. This allows researchers to use standard inverse-variance weighting to meta-analyze medians, differences in medians, and ratios of medians—even when the original studies don't report standard errors.

What Problem Does This Paper Address?

In clinical trials, "survival" (time-to-event) data is the gold standard. Usually, meta-analyses of these trials use the Hazard Ratio (HR). However, HRs come with baggage: they rely on the "proportional hazards" assumption (the idea that the relative risk between two groups stays constant over time), which is frequently violated in real-world data.

Median survival time is a much more intuitive and robust alternative. It tells you exactly when half the patients in a study have reached the event. While primary studies almost always report the median and its 95% confidence interval (CI), they rarely report the standard error (SE). Without the SE, you cannot perform a standard inverse-variance weighted meta-analysis, which is the most common way to "pool" results from multiple studies.

The challenge is that the CIs for medians are usually calculated using nonparametric methods (like Brookmeyer-Crowley) that are asymmetric and don't technically follow a Wald distribution. This paper asks: Can we just pretend they are Wald intervals to back-calculate the SE, and does that actually work?

Methods

The authors investigate the Wald approximation-based approach. The core idea is to estimate the standard error ($\widehat{SE}$) using the reported lower ($l$) and upper ($u$) limits of a $100(1-\alpha)\%$ confidence interval:

$$\widehat{SE} = \frac{u - l}{2z_{1-\alpha/2}}$$

The Novel Contribution: 1. Theoretical Proof: The authors provide a mathematical proof (Proposition 1) showing that this approximation is "consistent" for the widely used Brookmeyer-Crowley confidence interval. Essentially, as the sample size grows, the width of that nonparametric interval behaves enough like a Wald interval to make the math work. 2. Meta-Analytic Framework: They describe how to apply this to three types of outcomes: the pooled median, the difference of medians, and the ratio of medians (using a log-transformation for the latter).

Datasets

The paper evaluates the method using two distinct approaches:

  • Extensive Simulations: They generated thousands of datasets using Exponential, Weibull, and complex "Weibull mixture" distributions to represent different survival shapes. They varied sample sizes ($N=20$ to $1000$) and censoring rates (26% to 65%).
  • Real-World Application: They applied the method to a dataset of 28 randomized trials for non-small cell lung cancer (NSCLC) therapies, comparing experimental arms to comparator arms.

Key Insights and Ideas

  • The "Asymmetry" Myth: Researchers often worry that because median CIs are asymmetric, the Wald formula (which assumes symmetry) will fail. The authors show that while asymmetry causes a slight overestimation of the SE, the bias is negligible unless the tail probabilities are extremely skewed (near zero).
  • Consistency is Key: The "aha moment" is the proof that the empirical process terms in the Brookmeyer-Crowley interval cancel out when calculating the interval width, leaving a deterministic term that directly relates to the true asymptotic variance.
  • Robustness to Heterogeneity: The method holds up well even if the meta-analyst misspecifies the form of between-study heterogeneity (e.g., assuming additive differences when the true effect is multiplicative).

Key Results

  • High Accuracy for Moderate Samples: When within-study sample sizes were above 50, the Wald approximation was nearly unbiased (relative bias < 10%). At $N=1000$, the bias dropped below 2%.
  • Comparable to "True" SEs: In meta-analysis simulations, the Wald approximation performed almost identically to a "benchmark" approach that used the true (but usually unknown) standard errors.
  • NSCLC Findings: In the lung cancer application, the method successfully pooled data to show a small but significant survival advantage for experimental treatments: a difference of 1.24 months (95% CI: 0.22, 2.26) and a ratio of 1.11 (95% CI: 1.04, 1.20).

Limitations and Future Work

  • Small Sample Bias: The method struggles when effective sample sizes are very small ($N < 50$). In these cases, it tends to underestimate the within-study variance, which can lead to overconfident results (poor CI coverage) in the meta-analysis.
  • High Censoring: Performance degrades as the percentage of "censored" individuals (those who haven't had the event by the end of the study) increases.
  • Future Directions: The authors suggest that for very small studies, researchers might consider using a jackknife or bootstrap approach to estimate the pooled median, rather than relying on the Wald approximation. They also point to the metamedian R package as a tool that now implements these findings.

Summary updated

Abstract

We consider the problem of meta-analyzing outcome measures based on median survival times. Primary studies with time-to-event outcomes often report estimates of median survival times and confidence intervals based on the Kaplan-Meier estimator. However, outcome measures based on median survival are rarely meta-analyzed, as standard inverse-variance weighted methods require within-study standard errors that are typically not reported. In this article, we consider an inverse-variance weighted approach to meta-analyze median survival times that estimates the within-study standard errors from the reported confidence intervals. We show that this method consistently estimates the standard error of median survival when applied to confidence intervals constructed by the Brookmeyer-Crowley method. We conduct a series of simulation studies evaluating the performance of this approach at the study level (i.e., for estimating the standard error of median survival) and the meta-analytic level (i.e., for estimating the pooled median, difference of medians, and ratio of medians) for commonly used confidence intervals for median survival, including the Brookmeyer-Crowley method and nonparametric bootstrap. We find that this approach often performs comparably to a benchmark approach that uses the true within-study standard errors for meta-analyzing median-based outcome measures when within-study sample sizes are moderately large (e.g., above 50). However, when the effective sample sizes are small, the method can yield biased estimates of within-study standard errors. We illustrate an application of this approach in a meta-analysis evaluating survival benefits of being assigned to experimental arms versus comparator arms in randomized trials for non-small cell lung cancer therapies.
BibTeX
@misc{mcgrath2025meta,
  author = {Sean McGrath and Jonathan Kimmelman and Ömer Öztürk and Russell Steele and Andrea Benedetti and Benedetti, Andrea},
  title = {{Meta-analysis of median survival times with inverse-variance weighting}},
  year = {2025},
  eprint = {2503.03065},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.03065},
}
Figure