Table 2. Variation in terms of statistical conclusion for Mantel-Haenszel meta-analyses*.
| RevMan method | |||||||
Fixed-effect ( ) |
Fixed-effect ( ) |
Random-effects DL ( ) |
|||||
| Counts (cell percentages) | Counts (cell percentages) | Counts (cell percentages) | |||||
| Stata method | NS | Sig | NS | Sig | NS | Sig | |
| FE (MH fixed) | NS | 8929(70.8%) | 0(0.0%) | 4311(51.7%) | 2(0.0%) | 6422(52.5%) | 51(0.4%) |
| Sig | 11(0.1%) | 3678(29.1%) | 8(0.1%) | 3983(47.7%) | 1666(13.6%) | 4067(33.3%) | |
| No comp† | 0(0.0%) | 0(0.0%) | 37(0.4%) | 3(0.0%) | 16(0.1%) | 0(0.0%) | |
| DL (MH random) | NS | 8882(70.4%) | 196(1.6%) | 4304(51.6%) | 1539(18.4%) | 8033(65.7%) | 33(0.3%) |
| Sig | 58(0.5%) | 3482(27.6%) | 15(0.2%) | 2446(29.3%) | 55(0.5%) | 4085(33.4%) | |
| No comp† | 0(0.0%) | 0(0.0%) | 37(0.4%) | 3(0.0%) | 16(0.1%) | 0(0.0%) | |
| DLb | NS | 8890(70.5%) | 336(2.7%) | 4306(51.6%) | 1676(20.1%) | 8023(65.6%) | 265(2.2%) |
| Sig | 50(0.4%) | 3342(26.5%) | 13(0.2%) | 2309(27.7%) | 65(0.5%) | 3853(31.5%) | |
| No comp† | 0(0.0%) | 0(0.0%) | 37(0.4%) | 3(0.0%) | 16(0.1%) | 0(0.0%) | |
| VC | NS | 8889(70.4%) | 282(2.2%) | 4298(51.5%) | 1601(19.2%) | 7844(64.2%) | 316(2.6%) |
| Sig | 51(0.4%) | 3396(26.9%) | 21(0.3%) | 2384(28.6%) | 244(2.0%) | 3802(31.1%) | |
| No comp† | 0(0.0%) | 0(0.0%) | 37(0.4%) | 3(0.0%) | 16(0.1%) | 0(0.0%) | |
| MVa‡ | NS | 8863(70.2%) | 787(6.2%) | 4303(51.6%) | 1882(22.6%) | 7982(65.3%) | 683(5.6%) |
| Sig | 19(0.2%) | 2876(22.8%) | 16(0.2%) | 2103(25.2%) | 84(0.7%) | 3433(28.1%) | |
| No comp† | 58(0.5%) | 15(0.1%) | 37(0.4%) | 3(0.0%) | 38(0.3%) | 2(0.0%) | |
| MVb | NS | 8917(70.7%) | 399(3.2%) | 4303(51.6%) | 1603(19.2%) | 7912(64.7%) | 374(3.1%) |
| Sig | 23(0.2%) | 3279(26.0%) | 16(0.2%) | 2382(28.5%) | 176(1.4%) | 3744(30.6%) | |
| No comp† | 0(0.0%) | 0(0.0%) | 37(0.4%) | 3(0.0%) | 16(0.1%) | 0(0.0%) | |
| BP | NS | 8933(70.8%) | 1111(8.8%) | 4305(51.6%) | 1944(23.3%) | 7908(64.7%) | 964(7.9%) |
| Sig | 7(0.1%) | 2567(20.3%) | 14(0.2%) | 2041(24.5%) | 180(1.5%) | 3154(25.8%) | |
| No comp† | 0(0.0%) | 0(0.0%) | 37(0.4%) | 3(0.0%) | 16(0.1%) | 0(0.0%) | |
| B0 | NS | 8887(70.4%) | 267(2.1%) | 4301(51.5%) | 1290(15.5%) | 7624(62.4%) | 217(1.8%) |
| Sig | 53(0.4%) | 3411(27.0%) | 18(0.2%) | 2695(32.3%) | 464(3.8%) | 3901(31.9%) | |
| No comp† | 0(0.0%) | 0(0.0%) | 37(0.4%) | 3(0.0%) | 16(0.1%) | 0(0.0%) | |
| VC2 | NS | 8887(70.4%) | 222(1.8%) | 4301(51.5%) | 1622(19.4%) | 7978(65.3%) | 202(1.7%) |
| Sig | 53(0.4%) | 3456(27.4%) | 18(0.2%) | 2363(28.3%) | 110(0.9%) | 3916(32.0%) | |
| No comp† | 0(0.0%) | 0(0.0%) | 37(0.4%) | 3(0.0%) | 16(0.1%) | 0(0.0%) | |
| DL2 | NS | 8884(70.4%) | 196(1.6%) | 4299(51.5%) | 1504(18.0%) | 7922(64.8%) | 116(0.9%) |
| Sig | 56(0.4%) | 3482(27.6%) | 20(0.2%) | 2481(29.7%) | 166(1.4%) | 4002(32.7%) | |
| No comp† | 0(0.0%) | 0(0.0%) | 37(0.4%) | 3(0.0%) | 16(0.1%) | 0(0.0%) | |
| ML | NS | 8831(70.0%) | 63(0.5%) | 4204(50.4%) | 679(8.1%) | 7113(58.2%) | 64(0.5%) |
| Sig | 42(0.3%) | 3571(28.3%) | 70(0.8%) | 3241(38.8%) | 894(7.3%) | 3985(32.6%) | |
| No conv/comp† | 67(0.5%) | 44(0.3%) | 82(1.0%) | 68(0.8%) | 97(0.8%) | 69(0.6%) | |
| REML | NS | 8629(68.4%) | 212(1.7%) | 4271(51.2%) | 1423(17.1%) | 7792(63.8%) | 98(0.8%) |
| Sig | 56(0.4%) | 3374(26.7%) | 26(0.3%) | 2538(30.4%) | 189(1.5%) | 3954(32.4%) | |
| No conv/comp† | 255(2.0%) | 92(0.7%) | 59(0.7%) | 27(0.3%) | 123(1.0%) | 66(0.5%) | |
| PL | NS | 8803(69.8%) | 846(6.7%) | 4260(51.1%) | 1679(20.1%) | 7876(64.4%) | 576(4.7%) |
| Sig | 15(0.1%) | 2766(21.9%) | 14(0.2%) | 2238(26.8%) | 114(0.9%) | 3463(28.3%) | |
| No conv/comp† | 122(1.0%) | 66(0.5%) | 82(1.0%) | 71(0.9%) | 114(0.9%) | 79(0.6%) | |
*
Following the RevMan documentation we did not compute effects for some very rare events or non-events studies (although RevMan v5.2.3 does calculate effects for them) and hence some meta-analyses were not computable; i.e. when the remaining number of eligible studies was one or zero.
†
Not computable (see above) or iterative method failed to convergence to a solution.
‡
MVa fails to produce an estimate when all effects are equal, a scenario not unheard of when meta-analysing dichotomous data.