To the Editor:
I write to offer a comment and a suggestion on the systematic review and meta-analysis by Hussain et al. of local anesthetics used for fascial plane block. This report compared analgesia between liposomal bupivacaine and plain bupivacaine, the primary outcome being rest pain scores over the 24-to-72 h postoperative interval expressed as an area under the curve. Using data from 874 patients in 13 studies and adopting a Hartung-Knapp-Sidik-Johnkman random-effects model with the pain severity difference between treatments transformed as a standardized mean difference, the authors report in the abstract and the body that “No difference was observed between the two groups for area under the curve scores, with a standardized mean difference (95% CI) of –0.21 cm.h (–0.43 to 0.01; P = 0.058; I2 = 48%).” The authors used excellent statistical methods and software within frequentist statistical methodology; inferences were made by null hypothesis significance testing. As emphasized by statisticians and epidemiologists observing a P value > 0.05 does not demonstrate that no effect or no difference was observed or that absence of an effect was demonstrated. Neither is the P value the probability of the null hypothesis. Nor does the 95% CI have a 95% probability of including the true effect size. A correct and more cautious description for the result of a P = 0.058 for this meta-analysis would be that the null hypothesis of no difference was not rejected.
Statistical approaches using Bayesian theory, methods and software can provide an alternative for doing this meta-analysis. This random-effects meta-analysis can be expressed in the parameters μ (true mean of the population standardized mean difference) and τ2 (variance between studies). Using the 13 study values displayed in the article’s figure 3, I ran a Bayesian random-effects model comparing the standardized mean difference between liposomal and plain local anesthetic using 32,000 iterations and four chains in the Markov chain Monte Carlo algorithm of the R package brms (2.21.6). The prior distributions were Normal(0, 1) for μ and HalfCauchy(0, 0.5) for τ using guidance for choosing “weakly informative” prior distributions for a Bayesian random-effects meta-analysis. All sampling convergence characteristics of the fitted model (Gelman-Rubin convergence statistic = 1.00, large effective sample size, small Monte Carlo standard error, etc.) were satisfactory. Sensitivity analyses with other priors for both μ and τ gave very similar, almost identical estimated values and credible intervals. The model fit using the normal and half Cauchy priors was subjected to power scaling using the priorsense (1.0.1) package; there was no sensitivity of the model fit to either strengthening or weakening the prior distribution. Estimated parameter values were reported as means, standard deviations and 95% credible intervals.
The estimated parameter value for μ (standardized mean difference) was –0.21, identical to the mean value (–0.21) reported by Hussain et al. However, the 95% credible intervals (–0.43, –0.01) no longer crossed the zero line of no effect. With 98% probability, the reduction in rest pain scores was larger for liposomal bupivacaine than plain bupivacaine. Also, the estimated variability between studies (τ2) was 0.06 with the 95% credible intervals (0.01, 0.26).
The standardized mean difference used in meta-analysis is also known as Hedges’ g. It is a unitless, dimensionless metric. Heuristics for interpreting the absolute magnitude of effect has been offered with 0.2, 0.50, and 0.8 proposed as showing small, medium, and large effects. With either frequentist or Bayesian models, the summary standardized mean difference was about –0.20; this is a small, clinically unimportant effect. Using the Bayesian model, the probability that the standardized mean difference has a medium sized effect of –0.5 or better is only 1%.
Bayesian methods (1) allows direct modeling of the uncertainty of differences between studies (τ2); (2) produces the full posterior distribution of the mean and variance (μ, τ2) parameters (fig. 1); (3) permits the estimation of 95% credible intervals for μ and τ2 that have a 95% probability of containing the parameter values; and (4) allows hypothesis testing of the probability of interesting population values. All statistical models and inferences, frequentist and Bayesian, are subject to the validity of the assumptions. The use of Bayesian methods does offer some additional tools for meta-analysis.
Leave a Reply
You must be logged in to post a comment.