Table 4.
Sensitivity analyses.
| Scenario | Time to first flatus, hours MD (95% CI) | Time to first defecation, hours MD (95% CI) | Time to first liquid intake, hours MD (95% CI) | Length of hospital stay, days MD (95% CI) |
|---|---|---|---|---|
| Panel A. Sensitivity to the between-study variance estimator and the confidence interval method | ||||
| Primary analysis: REML with the conditional inference rule | −9.61 (−12.81 to −6.42) k = 13; τ2 = 16.29; I2 = 82.6%; P < 0.001 | −15.88 (−23.31 to −8.46) k = 9; τ2 = 78.13; I2 = 90.3%; P = 0.001 | −10.65 (−18.20 to −3.10) k = 2; τ2 = 0.00; I2 = 0.0%; P = 0.006 | −0.63 (−1.36 to 0.10) k = 7; τ2 = 0.25; I2 = 57.0%; P = 0.079 |
| REML, Wald-type confidence interval | −9.61 (−12.41 to −6.81) k = 13; τ2 = 16.29; I2 = 82.6%; P < 0.001 | −15.88 (−22.34 to −9.42) k = 9; τ2 = 78.13; I2 = 90.3%; P < 0.001 | −10.65 (−18.20 to −3.10) k = 2; τ2 = 0.00; I2 = 0.0%; P = 0.006 | −0.63 (−1.16 to −0.11) k = 7; τ2 = 0.25; I2 = 57.0%; P = 0.019 |
| REML, modified Hartung–Knapp with the ad hoc variance correction | −9.61 (−12.81 to −6.42) k = 13; τ2 = 16.29; I2 = 82.6%; P < 0.001 | −15.88 (−23.48 to −8.28) k = 9; τ2 = 78.13; I2 = 90.3%; P = 0.001 | −10.65 (−59.59 to 38.29) k = 2; τ2 = 0.00; I2 = 0.0%; P = 0.221 | −0.63 (−1.36 to 0.10) k = 7; τ2 = 0.25; I2 = 57.0%; P = 0.079 |
| DerSimonian–Laird, Wald-type confidence interval | −9.21 (−11.62 to −6.81) k = 13; τ2 = 10.63; I2 = 75.6%; P < 0.001 | −15.90 (−21.95 to −9.85) k = 9; τ2 = 66.63; I2 = 88.8%; P < 0.001 | −10.65 (−18.20 to −3.10) k = 2; τ2 = 0.00; I2 = 0.0%; P = 0.006 | −0.63 (−1.14 to −0.11) k = 7; τ2 = 0.24; I2 = 55.3%; P = 0.017 |
| Paule–Mandel, Wald-type confidence interval | −9.69 (−12.59 to −6.80) k = 13; τ2 = 17.85; I2 = 83.9%; P < 0.001 | −15.89 (−22.20 to −9.58) k = 9; τ2 = 73.78; I2 = 89.7%; P < 0.001 | −10.65 (−18.20 to −3.10) k = 2; τ2 = 0.00; I2 = 0.0%; P = 0.006 | −0.66 (−1.27 to −0.05) k = 7; τ2 = 0.41; I2 = 68.2%; P = 0.034 |
| Sidik–Jonkman, Wald-type confidence interval | −9.95 (−13.19 to −6.71) k = 13; τ2 = 24.12; I2 = 87.5%; P < 0.001 | −15.89 (−22.24 to −9.53) k = 9; τ2 = 75.08; I2 = 89.9%; P < 0.001 | −10.65 (−18.20 to −3.10) k = 2; τ2 = 0.00; I2 = 0.0%; P = 0.006 | −0.68 (−1.35 to −0.01) k = 7; τ2 = 0.55; I2 = 74.0%; P = 0.048 |
| Maximum likelihood, Wald-type confidence interval | −9.47 (−12.12 to −6.82) k = 13; τ2 = 13.95; I2 = 80.2%; P < 0.001 | −15.90 (−22.00 to −9.80) k = 9; τ2 = 67.92; I2 = 88.9%; P < 0.001 | −10.65 (−18.20 to −3.10) k = 2; τ2 = 0.00; I2 = 0.0%; P = 0.006 | −0.59 (−1.04 to −0.15) k = 7; τ2 = 0.13; I2 = 40.4%; P = 0.008 |
| Common-effect (fixed-effect) model | −6.66 (−7.57 to −5.74) k = 13; τ2 = 0 by assumption; I2 not reported; P < 0.001 | −12.72 (−14.36 to −11.07) k = 9; τ2 = 0 by assumption; I2 not reported; P < 0.001 | −10.65 (−18.20 to −3.10) k = 2; τ2 = 0 by assumption; I2 not reported; P = 0.006 | −0.52 (−0.82 to −0.21) k = 7; τ2 = 0 by assumption; I2 not reported; P < 0.001 |
| Panel B. Sensitivity to study inclusion, data interpretation and comparator choice | ||||
| Primary analysis (all eligible trials) | −9.61 (−12.81 to −6.42) k = 13; τ2 = 16.29; I2 = 82.6%; P < 0.001 95% PI −18.97 to −0.25 | −15.88 (−23.31 to −8.46) k = 9; τ2 = 78.13; I2 = 90.3%; P = 0.001 95% PI −37.58 to 5.81 | −10.65 (−18.20 to −3.10) k = 2; τ2 = 0.00; I2 = 0.0%; P = 0.006 95% PI NC | −0.63 (−1.36 to 0.10) k = 7; τ2 = 0.25; I2 = 57.0%; P = 0.079 95% PI −2.07 to 0.80 |
| Excluding results judged to be at high overall risk of bias | −11.68 (−20.15 to −3.22) k = 4; τ2 = 15.91; I2 = 66.3%; P = 0.022 95% PI NC | −15.16 (−40.79 to 10.48) k = 3; τ2 = 83.68; I2 = 79.0%; P = 0.126 95% PI NC | NE (k = 1) | −0.67 (−1.58 to 0.23) k = 6; τ2 = 0.36; I2 = 67.7%; P = 0.115 95% PI −2.46 to 1.11 |
| Excluding trials without a sham comparator | −8.92 (−14.03 to −3.81) k = 7; τ2 = 20.95; I2 = 84.6%; P = 0.005 95% PI −21.23 to 3.39 | −16.44 (−28.25 to −4.64) k = 6; τ2 = 109.21; I2 = 89.0%; P = 0.016 95% PI −45.79 to 12.90 | NE (k = 1) | −0.63 (−4.18 to 2.91) k = 3; τ2 = 1.45; I2 = 85.5%; P = 0.523 95% PI NC |
| Excluding the trial for which means and standard deviations were converted from a median and interquartile range | −9.62 (−13.17 to −6.08) k = 12; τ2 = 18.43; I2 = 84.9%; P < 0.001 95% PI −19.71 to 0.47 | — | — | −0.75 (−1.71 to 0.21) k = 6; τ2 = 0.43; I2 = 61.6%; P = 0.102 95% PI −2.69 to 1.19 |
| Conservative analysis retaining only Pei et al. (20) among the three investigator- and institution-related trials | −9.74 (−13.63 to −5.84) k = 11; τ2 = 20.36; I2 = 87.1%; P < 0.001 95% PI −20.52 to 1.05 | −16.51 (−24.61 to −8.41) k = 8; τ2 = 81.63; I2 = 91.6%; P = 0.002 95% PI −39.36 to 6.34 | — | −0.75 (−1.71 to 0.21) k = 6; τ2 = 0.43; I2 = 61.6%; P = 0.102 95% PI −2.69 to 1.19 |
| Excluding the largest trial | −9.38 (−12.85 to −5.90) k = 12; τ2 = 16.97; I2 = 83.8%; P < 0.001 95% PI −19.09 to 0.33 | −16.03 (−24.63 to −7.43) k = 8; τ2 = 90.21; I2 = 92.0%; P = 0.003 95% PI −40.08 to 8.02 | NE (k = 1) | −0.76 (−1.47 to −0.06) k = 6; τ2 = 0.16; I2 = 46.0%; P = 0.038 95% PI −2.01 to 0.48 |
| Excluding Gu et al. 2018 (19), for which the reported time unit required interpretation | −10.11 (−12.95 to −7.26) k = 12; τ2 = 8.97; I2 = 68.4%; P < 0.001 95% PI −17.29 to −2.93 | −13.36 (−19.49 to −7.22) k = 8; τ2 = 35.05; I2 = 80.1%; P = 0.001 95% PI −28.64 to 1.93 | — | — |
| Standard-care arm substituted for the sham arm in Pei et al. (20) | −11.17 (−15.30 to −7.03) k = 13; τ2 = 34.53; I2 = 90.9%; P < 0.001 95% PI −24.62 to 2.29 | −16.80 (−24.44 to −9.15) k = 9; τ2 = 82.80; I2 = 90.7%; P < 0.001 95% PI −39.13 to 5.54 | −12.49 (−19.89 to −5.10) k = 2; τ2 = 0.00; I2 = 0.0%; P < 0.001 95% PI NC | −0.64 (−1.33 to 0.05) k = 7; τ2 = 0.19; I2 = 48.6%; P = 0.062 95% PI −1.90 to 0.62 |
CI, confidence interval; MD, mean difference; NC, not calculated; NE, not estimable; PI, prediction interval.
Effect estimates are mean differences with 95% confidence intervals; negative values favor electrical acupoint stimulation. The primary analyses used restricted maximum likelihood with Hartung–Knapp–Sidik–Jonkman intervals and tests when τ2 was greater than zero and more than two trials contributed, and Wald-type inference otherwise.
The common-effect model fixes τ2 at zero by assumption; I2 is not reported for that row, to avoid implying that the observed between-study heterogeneity was absent. P values throughout are for the test of the pooled mean difference and are not the P values from Cochran's Q test of heterogeneity.
Prediction intervals are shown in Panel B for analyses to which at least five trials contributed and are otherwise marked NC. A scenario is shown only for the outcomes to which it applies; an em dash indicates that the excluded trial did not contribute to that outcome, so the scenario would reproduce the primary analysis.
Where only one trial remained, no pooled estimate could be calculated (NE).
A conservative analysis retaining only Pei et al. (20) was performed to assess robustness to the shared investigator and institutional context of the three trials (Supplementary Table S6).
Complete leave-one-out analyses are presented in Supplementary Table S9; leave-one-out analysis was not applicable to time to first liquid intake because only two trials contributed.