Unrecognised oesophageal intubation is a rare but potentially catastrophic event for patients and intubators alike. Videolaryngoscopy has been proposed as a potential tool in the prevention of unrecognised oesophageal intubation [1]. In a recently updated Cochrane systematic review comparing videolaryngoscopy vs. direct laryngoscopy, we reported that, collectively, videolaryngoscopes reduce the incidence of oesophageal intubation approximately twofold, but when considered by device design only, hyperangulated videolaryngoscopes were associated with a statistically significant reduction in oesophageal intubation [2].

These conclusions are based on meta-analyses using the random-effects model, as used throughout the Cochrane review. This was chosen as the included studies were variably designed and many outcomes were poorly defined. However, in the specific case of oesophageal intubation, other models may have been reasonably used to investigate the outcome.

In a frequentist meta-analysis, the most commonly used statistical models are the fixed-effect and random-effects models [3]. Briefly, the fixed-effect model of meta-analysis assumes that the true effect of an intervention is identical between studies (i.e. the sampled groups come from the same population) and there should be no unexplained statistical heterogeneity [3, 4]. Conversely, the random-effects model makes no assumptions about the true effect, seeking instead to provide a value representing the average effect of the intervention. The model accounts for statistical heterogeneity in generating a pooled estimate and confidence interval [4]. Meta-analyses of outcomes with low event rates, such as oesophageal intubation, are vulnerable to bias and confidence intervals may be imprecise [4, 5]. For outcomes with event rates below 1%, the Peto odds ratio method may have greater statistical power and be less prone to bias [5].

Oesophageal intubation has less design and outcome heterogeneity than many of our other analyses. For oesophageal intubation, the I² statistic (a test of statistical heterogeneity) was below 40% for all blade types. Strictly, the choice of analytic model should be based on the relative uniformity of the included studies [3, 4]. Given the relative ease of clinical definition for oesophageal intubation, coupled with low statistical heterogeneity and low event rates, use of either the Peto odds ratio method or even the fixed-effect model might have been credible choices.

On this basis, we undertook an exploratory analysis of the data using the fixed-effect model and Peto odds ratio method and compared the results with those generated when using the random-effects model. The results are illustrated in Fig. 1. The point estimates of all three analyses are similar, regardless of the model, reflecting a substantial reduction in the rate of oesophageal intubation when using a videolaryngoscope. However, in the case of the Macintosh-style and channelled blade analyses, both the Peto odds ratio method and the fixed-effect model generate narrower confidence intervals, with the difference in rate being statistically significant at the p < 0.05 level.

How should we process these findings? Having chosen the random-effects model for use in the main review, it would be inappropriate to retrospectively change the chosen statistical model. Equally, is it fair to conclude that Macintosh-style and channelled videolaryngoscopes do not reduce oesophageal intubation? There are good reasons to challenge this interpretation. The random-effects model is limited by a low statistical power for rare outcomes and is more prone to bias from smaller studies [5, 6], which are common in this review. Furthermore, dichotomising results as evidenced or dismissed based on a given p value are not universally accepted. As Fisher originally described, the p value was “an index measuring the strength of evidence against the null hypothesis” rather than a test of its rejection [7].

Finally, is it clinically plausible that videolaryngoscopy reduces the rate of oesophageal intubation? Videolaryngoscopy improves visualisation of the larynx, reducing the frequency of grade 3 and 4 views and increasing grade 1 and 2 views [2]; provides a wider-angle view of the airway [8]; reduces failed tracheal intubation rates and intubation difficulty based on the intubation difficulty scale [2]; and enables other team members to observe laryngoscopy and tracheal intubation while enabling an assistant to improve the view for the intubator [1]. Reduced intubation failure rates are consistent with reduced oesophageal intubation, which are a subset of this, and the improved (team) view is consistent with improved recognition of oesophageal intubation. Conversely, the risk of videolaryngoscopy increasing the risk of oesophageal intubation, or risk overall, seems vanishingly small.

In summary, we found that, collectively, videolaryngoscopes more than halve the risk of oesophageal intubation. Our data indicate that hyperangulated videolaryngoscopes reduce oesophageal intubation in a clinically meaningful and statistically significant way. The strength of evidence for Macintosh-style and channelled devices is weaker and depends on how the data are analysed and interpreted. We leave individual readers to decide how to synthesise this evidence. We argue that reduced risk of oesophageal intubation, and the likelihood of these benefits being counterbalanced by harm is low.

未被识别的食管插管是一种罕见但对患者和插管者等人来说可能是灾难性的事件。视频喉镜已被提议作为预防未被识别的食管插管的潜在工具 [ 1 ]。在最近更新的比较视频喉镜与直接喉镜的 Cochrane 系统评价中，我们报告说，视频喉镜共同将食管插管的发生率降低了大约两倍，但仅从设备设计考虑时，超角度视频喉镜与食管插管的统计学显着减少相关[ 2 ]。

这些结论基于使用随机效应模型的荟萃分析，该模型在整个 Cochrane 评价中使用。之所以选择这一点，是因为纳入的研究设计可变，许多结果定义不明确。然而，在食管插管的具体案例中，其他模型可能已被合理地用于调查结果。

在频率论元分析中，最常用的统计模型是固定效应模型和随机效应模型 [ 3 ]。简而言之，荟萃分析的固定效应模型假设干预的真实效果在研究之间是相同的（即样本组来自同一人群）并且不应存在无法解释的统计异质性 [ 3, 4 ]。相反，随机效应模型不对真实效应做出任何假设，而是寻求提供一个代表干预平均效应的值。该模型在生成合并估计和置信区间时考虑了统计异质性 [ 4]. 对食管插管等事件发生率低的结局进行的荟萃分析容易产生偏倚，而且置信区间可能不精确 [ 4, 5 ]。对于事件发生率低于 1% 的结果，Peto 比值比方法可能具有更大的统计功效并且不易出现偏倚 [ 5 ]。

与我们的许多其他分析相比，食管插管的设计和结果异质性较小。对于食管插管，所有刀片类型的 I ²统计量（统计异质性检验）均低于 40%。严格地说，分析模型的选择应基于纳入研究的相对均匀性 [ 3, 4 ]。鉴于食管插管的临床定义相对容易，加上低统计异质性和低事件发生率，使用 Peto 比值比方法甚至固定效应模型可能是可靠的选择。

在此基础上，我们采用固定效应模型和Peto比值比法对数据进行探索性分析，并将结果与​​随机效应模型的结果进行比较。结果如图 1 所示。无论模型如何，所有三项分析的点估计都相似，反映出使用视频喉镜时食管插管率显着降低。然而，在 Macintosh 风格和通道刀片分析的情况下，Peto 比值比方法和固定效应模型都会产生更窄的置信区间，率差异在 p < 0.05 水平上具有统计显着性。

我们应该如何处理这些发现？在主要审查中选择了随机效应模型后，追溯更改所选统计模型是不合适的。同样，得出 Macintosh 式和通道式视频喉镜不能减少食管插管的结论是否公平？有充分的理由质疑这种解释。随机效应模型受到罕见结果的低统计功效的限制，并且更容易受到较小研究 [ 5, 6 ] 的偏倚，这在本综述中很常见。此外，根据给定的 p 值将结果二分法作为证据或驳回并未得到普遍接受。正如 Fisher 最初描述的那样，p 值是“衡量反对零假设的证据强度的指标” 而不是对其拒绝的测试 [ 7 ]。

最后，视频喉镜检查降低食管插管率在临床上是否合理？视频喉镜改善了喉部的可视化，减少了 3 级和 4 级视图的频率并增加了 1 级和 2 级视图的频率 [ 2 ]；提供更广角的气道视图 [ 8 ]；根据插管难度等级 [ 2 ]，降低气管插管失败率和插管难度；并使其他团队成员能够观察喉镜检查和气管插管，同时使助手能够改善插管者的视野 [ 1]. 插管失败率的降低与食管插管的减少是一致的，这是其中的一个子集，而改进的（团队）观点与对食管插管的认识的提高是一致的。相反，视频喉镜检查增加食管插管风险或总体风险的风险似乎微乎其微。

总之，我们发现，总体而言，视频喉镜可将食管插管的风险降低一半以上。我们的数据表明，超角度视频喉镜以具有临床意义和统计学意义的方式减少食管插管。Macintosh 风格和通道设备的证据强度较弱，取决于数据的分析和解释方式。我们让个别读者决定如何综合这些证据。我们认为，降低食管插管的风险，以及这些益处被危害抵消的可能性很低。