Time to move beyond P value

Sir,

We read with great interest the perspective by Indrayan and Saini titled ‘Some newer & simpler biostatistical approaches for more credible clinical research’¹. As clinicians engaged in research, we appreciate the authors’ emphasis on aligning biostatistical methods with clinical relevance. Their focus on meaningful effect sizes, predictive performance, and patient-oriented interpretation is timely and welcome. Based on practical experience, we offer several comments to enhance the applicability of these suggestions.

The authors rightly highlighted the persistent misinterpretation of the P value. We agree that relying on a threshold, such as P<0.05, for clinical decision-making is inappropriate. As emphasised in the American Statistical Association (ASA) 2016 statement, a P value neither measures the probability that the studied hypothesis is true nor the probability that the data were produced by random chance². Equally important, it conveys neither the magnitude nor the clinical importance of the observed effect. Thus, scientific conclusions and policy decisions should not hinge solely on whether the P value crosses an arbitrary threshold.

The suggestion to adopt P<0.01 faces the same conceptual limitations as that of P<0.05. Lowering α universally increases Type II error, a concern particularly relevant in studies of rare diseases, early-phase trials, and settings where large samples are not feasible. Leading statistical guidance emphasizes that rigid cut-offs, whether 0.05 or 0.01, cannot substitute for contextual and clinical judgment^2,3. Extending this to more stringent thresholds, such as P<0.001, would further exacerbate these challenges, necessitating prohibitively large sample sizes and potentially amplifying publication bias by favouring only dramatic results while suppressing modest but clinically relevant findings⁴. Such stringent thresholds do not address the inherent limitations of the P value, including their dependence on the sample size and their inability to quantify clinical significance.

A constructive way forward is not to redefine significance thresholds but to adopt a multi-component inferential framework. Current recommendations emphasise reporting effect sizes with confidence intervals, interpreting results relative to the minimum clinically important difference, and avoiding dichotomisation into “significant” versus “non-significant”^2,3. Bayesian approaches further contribute by quantifying the probability that an effect exceeds a clinically meaningful threshold, an interpretation not possible from P value alone, facilitating nuanced and clinically relevant inferences.

In predictive modelling, the authors proposed using positive predictive value (PPV), negative predictive value (NPV), and P index as alternatives to ROC-based assessment. While the intention to emphasize clinical applicability is commendable, these measures are highly dependent on disease prevalence and lack of transportability across populations⁵. The P index also masks the asymmetric consequences of false positives and false negatives and evaluates performance at a single cutoff, making it sensitive to arbitrary threshold selection. Contemporary methodological guidelines, including TRIPOD and PROBAST, recommend a comprehensive evaluation incorporating discrimination, calibration, and decision-analytic tools, such as decision curve analysis⁵. This broader framework aligns with the authors’ goals while maintaining methodological robustness.

In summary, strengthening inferential frameworks, rather than altering P value thresholds or relying on single predictive metrics, offers a more balanced and scientifically sound path forward.