Results of the search

We identified 267 records from the databases and search strategies outlined in Electronic searches. We found one further record by searching other resources (handsearching). We removed 98 duplicate records and screened 170 records (title and abstract) for inclusion in the review.

We excluded 148 records based on the title and abstract and assessed 22 full text articles for inclusion in the review. We excluded 9 studies from the review (see Excluded studies below) and included 13 trials in the review (see Included studies). See Figure 1 for a Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) study flow diagram.

Figure 1

---

---

Study flow diagram

Included studies

We included 13 trials in this review (Banu 2007; Bidabadi 2009; Cereghino 1974; Chen 1996; Cossu 1984; Czapinski 1997; de Silva 1996; Feksi 1991; Heller 1995; Mattson 1985; Mitchell 1987; Ogunrin 2005; Placencia 1993). Two included studies were available in abstract form only (Bidabadi 2009; Czapinski 1997), and one included study was published in Italian, which we translated into English (Cossu 1984).

Two trials recruited individuals of all ages (Feksi 1991; Placencia 1993); 5 trials recruited children only (de Silva 1996 defined children as under the age of 16, Banu 2007 and Chen 1996 defined children as under the age of 15, and Bidabadi 2009 and Mitchell 1987 defined children as under the age of 12), and the remaining 6 trials recruited adults only. Of the adults‐only trials, 3 defined adults to be individuals above the age of 18 (Cereghino 1974; Czapinski 1997; Mattson 1985), 1 trial classed adults as older than 13 years (Heller 1995), 1 trial classed adults as older than 14 years (Ogunrin 2005), and 1 trial classed adults as older than 15 years (Cossu 1984). Seven trials recruited individuals with partial onset seizures and generalised onset seizures (Banu 2007; Chen 1996; de Silva 1996; Feksi 1991; Heller 1995; Ogunrin 2005; Placencia 1993), three trials recruited individuals with partial onset seizures only (Cereghino 1974; Mattson 1985; Mitchell 1987), one trial recruited individuals with partial seizures and secondarily generalised seizures (Bidabadi 2009), one trial recruited individuals with complex partial seizures only (Czapinski 1997), and one trial recruited individuals with temporal lobe epilepsy only (Cossu 1984).

Ten trials recruited individuals with new onset seizures, or previously untreated seizures, or both (Banu 2007; Chen 1996; Cossu 1984; Czapinski 1997; de Silva 1996; Feksi 1991; Heller 1995; Mitchell 1987; Ogunrin 2005; Placencia 1993); 1 trial recruited institutionalised participants with uncontrolled seizures (Cereghino 1974); 1 trial recruited "previously untreated or under‐treated" individuals (Mattson 1985); and 1 trial reported only in abstract form provided no information regarding new onset of seizures in participants (Bidabadi 2009).

Five trials were conducted in Europe (Bidabadi 2009; Cossu 1984; Czapinski 1997; de Silva 1996; Heller 1995); three trials were conducted in the USA (Cereghino 1974; Mattson 1985; Mitchell 1987); one trial was conducted in Taiwan (Chen 1996); and four trials were conducted in rural areas or developing countries, or both: One trial was conducted in Nigeria (Ogunrin 2005), one trial was conducted in Bangladesh (Banu 2007), one trial was conducted in Kenya (Feksi 1991), and one trial was conducted in Ecuador (Placencia 1993).

We did not obtain individual participant data (IPD) for 6 trials, including a total of 317 participants, as suitable seizure data for the outcomes examined in this systematic review were not recorded (Chen 1996; Mitchell 1987), the authors no longer had a copy of the data (Cereghino 1974), or authors did not respond to our data requests (Bidabadi 2009; Cossu 1984; Czapinski 1997). A further trial, which randomised 302 participants (Feksi 1991), provided access to an IPD dataset, but this was not the final dataset used for the analysis published by the original authors. The pharmaceutical company that sponsored the trial, Ciba‐Geigy, who at that time held the product license for carbamazepine (CBZ), held the final dataset. Since the trial was undertaken, there have been a number of mergers and restructures within the industry, and the current owners of the data are Novartis. Unfortunately, Novartis were unable to locate the data for this trial. The dataset that we had for this trial contained a number of problems and inconsistencies, and we therefore decided not to include this trial in the meta‐analysis. None of these seven trials reported the specific time‐to‐event outcomes chosen for this systematic review, and we could not extract sufficient aggregate data from the trial publications in any other trial. Therefore, we could not include them in data synthesis. Table 1 contains full details of outcomes considered and summaries of results in each eligible trial for which individual participant data were not available.

Individual participant data were available for the remaining 6 trials, which recruited a total of 836 participants, representing 57% of 1455 individuals from all 13 identified eligible trials. Four trials provided computerised data directly (Banu 2007; Mattson 1985; Ogunrin 2005; Placencia 1993), and the authors of two trials, de Silva 1996; Heller 1995, supplied a combination of both computerised and hard‐copy data (although mostly computerised).

Data were available for the following participant characteristics (percentage of 836 participants with data available): sex (99%, data missing for 6 participants in de Silva 1996 and 4 participants in Mattson 1985), seizure type (100%), drug randomised (99%, data missing for 6 participants in de Silva 1996), age at randomisation (99%, data missing for 1 participant in Heller 1995, 6 participants in de Silva 1996, and 5 participants in Mattson 1985), number of seizures in 6 months prior to randomisation (98%, data missing for 5 participants from Banu 2007, 1 participant in Heller 1995, 6 participants in de Silva 1996, and 7 participants in Mattson 1985), and time since first seizure to randomisation (94%, data missing for 2 participants in Heller 1995, 6 participants in de Silva 1996, 5 participants in Mattson 1985, and all 37 participants in Ogunrin 2005).

Three trials, de Silva 1996; Heller 1995; Ogunrin 2005, provided the results of neurological examinations for 220 participants (27%). Three trials provided electroencephalographic (EEG) results for 600 participants (72%) (103 participants from Banu 2007, 305 participants from Mattson 1985, and all participants from Placencia 1993). Two trials provided computerised tomography/magnetic resonance imaging (CT/MRI) results for 304 participants (36%) (26 from Banu 2007 and 278 from Mattson 1985).

See the 'Characteristics of included studies' tables for a detailed description of each study included in this systematic review.

Excluded studies

We excluded two duplicate trials (Cereghino 1973; Smith 1987), and we retained the most relevant primary reference for each trial in the review (Cereghino 1974 and Mattson 1985, respectively). We excluded five studies that were not randomised controlled trials (Bird 1966; Castro‐Gago 1998; Hansen 1980; Kuzuya 1993; Sabers 1995), and we excluded two trials that did not use CBZ and PB monotherapy (Marjerrison 1968; Meador 1990). See the 'Characteristics of excluded studies' tables for further details.

Allocation

Blinding

Incomplete outcome data

Selective reporting

We pledged study protocols in all individual participant data requests; however, protocols were not available for any of the 13 included trials, so we made a judgement of the risk of bias based on the information included in the publications or from the individual participant data we received (see the 'Characteristics of included studies' tables for more information).

Other potential sources of bias

We detected another source of bias in 6 of the 13 included trials.

Following consistency checks of individual participant data for Placencia 1993 and Banu 2007, we found some inconsistencies between the data provided and the results in the publications in terms of withdrawal and seizure recurrences, respectively, which the authors could not resolve. We performed sensitivity analysis to investigate the impact of the inconsistent data on our outcomes (see Sensitivity analysis and Effects of interventions). Furthermore, we received IPD for a seventh trial (Feksi 1991), but too many inconsistencies were present for this data to be usable (see Included studies for further details).

One trial had a cross‐over design (Cereghino 1974); such a design is unlikely to be appropriate for monotherapy treatment because of carry‐over effects from one treatment period into another (participants were also treated during washout periods with their 'regular medication'), and such a design does not allow long‐term outcomes, such as the time‐to‐event outcomes of interest in this review. For future updates of this review, we will exclude studies of a cross‐over design.

We included one trial with very small participant numbers (six participants randomised to each drug) and very short‐term follow up (three weeks), and it was unclear if this trial was adequately powered and of sufficient duration to detect differences (Cossu 1984). For future updates of this review, we will review our inclusion criteria in terms of participant numbers and trial duration.

Another trial had several potential sources of other bias (Mitchell 1987); there was evidence that the trial may have been underpowered to detect differences between the treatments, one of the tools for outcome assessment was not fully validated, and non‐randomised children from a related pilot study were included in analysis for some of the outcomes.

(1) Time to withdrawal of allocated treatment

For this outcome, a HR greater than one indicates a clinical advantage for CBZ.

Times to withdrawal of allocated treatment and reasons for withdrawal were available for 676 participants from 4 of the 6 trials providing IPD (97.8% of 691 participants from de Silva 1996; Heller 1995; Mattson 1985; and Placencia 1993 (see Included studies and Table 2) and 46.4% of the total 1455 participants from the 13 included studies). Mattson 1985 did not record follow‐up data for one participant randomised to CBZ. de Silva 1996 did not record the randomised drug for six participants, and the reason for withdrawal was not available for one participant randomised to CBZ and could not be determined from the case notes. Similarly, in Heller 1995, for one participant randomised to CBZ and three participants randomised to PB and in Placencia 1993, for one participant randomised to CBZ and two participants randomised to PB, the reason for withdrawal was not available and could not be determined from the case notes. We did not include these 15 participants with missing outcome data in the analysis of time to withdrawal of allocated treatment. All participants completed the 12‐week study in Ogunrin 2005 so could not contribute to the analysis of time to withdrawal of allocated treatment. From the IPD provided by Banu 2007, we were able to establish reasons for treatment withdrawal for all participants, but the date of withdrawal of allocated treatment was not available for all participants (see Data extraction and management for further details); therefore, we could not calculate the time to withdrawal of allocated treatment for this outcome.

Among the 784 participants for which we had reasons for treatment withdrawal from Banu 2007; de Silva 1996; Heller 1995; Mattson 1985; and Placencia 1993, 393 participants prematurely withdrew from treatment (50%): 216 out of 415 participants randomised to CBZ (52%) and 178 out of 369 participants randomised to PB (48%). (See Table 3 for reasons for premature termination of the study by treatment and how we classified these withdrawals in analysis.) We deemed 235 participants (30%) to have withdrawn for reasons related to the study drug, 125 (30%) on CBZ and 110 (30%) on PB, and we classed these withdrawals as 'events' in analysis. We classed the other 158 withdrawals to be not related to the study drug and censored these participants in analysis, in addition to those who completed the study without withdrawing.

The overall pooled HR (for 676 participants) was 1.49 (95% confidence interval (CI) 1.15 to 1.94, P = 0.003, from fixed‐effect analysis) indicating a statistically significant advantage for CBZ; in other words, participants withdrew significantly earlier from PB than CBZ in the 4 included trials. There was moderate statistical heterogeneity between trials (Chi² test = 7.07, df = 3, P = 0.07, I² statistic = 58%, see Analysis 1.1). When we repeated the analysis using random‐effects, the pooled HR was 1.50 (95% CI 0.95 to 2.38, P = 0.07), still indicating an advantage for CBZ, but this advantage was no longer statistically significant.

We performed sensitivity analysis excluding participants from Placencia 1993 from analysis because of high risk of selection bias due to inadequate allocation concealment (see Allocation (selection bias) and Table 4). This sensitivity analysis resulted in a larger advantage for CBZ with a pooled HR of 1.66 (95% CI 1.25 to 2.20, P = 0.0005, calculated with fixed‐effect) and reduced heterogeneity (Chi² test = 3.24, df = 2, P = 0.20, I² statistic = 35%) but no change to conclusions. Further, in Placencia 1993, we also found inconsistencies (between IPD dataset and published results) in the number of participants who withdrew from allocated treatment for certain reasons, which the trial authors could not resolve. These inconsistencies were as follows.

• Results from the IPD dataset: 51 participants withdrew, 31 from CBZ and 20 from PB: 16 participants left the area (lost to follow up), 10 withdrew due to adverse effects, 22 withdrew for personal reasons or no stated reason (classed as an event), and 3 died (see Table 3).

• Results in the trial report: 53 participants withdrew, 31 from CBZ and 22 from PB: 18 participants left the area (lost to follow up), 5 withdrew because of adverse effects, 3 died, and 27 withdrew for personal reasons or no stated reason.

As the overall number of events and censored observations was similar (results from the IPD dataset: 51 withdrew, 32 events, 19 censored; and results in the trial report: 53 withdrew, 32 events, 21 censored) and as our sensitivity analysis excluding results of Placencia 1993 gave similar results and an unchanged conclusion, we feel that these inconsistencies are minor and are unlikely to have had a large impact on the overall results. In the primary analysis of Placencia 1993, we classed those who withdrew for 'no clearly articulated reason' as events in the analysis; in other words, the withdrawal was due to the study drug. However, it is also possible that these participants may have withdrawn for reasons not related to the study drug, and we therefore should have censored them in the analysis. We performed a further sensitivity analysis censoring the 19 participants who withdrew for 'no clearly articulated reason'. Again, the results of the sensitivity analysis were similar to the primary analysis, showing a slightly larger statistically significant advantage for CBZ (pooled HR 1.65, 95% CI 1.26 to 2.17, P = 0.0003), and again,heterogeneity was substantially reduced after censoring these participants (Chi² test = 3.25, df = 3, P = 0.35, I² statistic = 8%).

In Placencia 1993 (primary analysis with events and censored observations as summarised in Table 3), there was some evidence that the proportional hazards assumption of the Cox model may have been violated; the P value of the time‐varying covariate was 0.084. In sensitivity analysis under our alternative assumption regarding censoring (we censored participants who withdrew for 'no clearly articulated reason' rather than analyse them as events), there was no evidence that the proportional hazards assumption of the Cox model was violated; the P value of the time‐varying covariate was 0.824. We therefore assume that the non‐proportionality of Placencia 1993 in our primary analysis was likely to be due to our assumptions regarding censoring of participants. The proportional hazards assumption of the Cox model was satisfied for all other trials included in analysis.

For participants with generalised onset seizures (136), the pooled HR was 1.53 (95% CI 0.81 to 2.88, P = 0.19), suggesting an advantage for CBZ that was not statistically significant. There was no evidence of statistical heterogeneity between trials (Chi² test = 0.49, df = 2, P = 0.78, I² statistic = 0%, see Analysis 1.2).

For participants with partial onset seizures (520), the pooled HR was 1.49 (95% CI 1.12 to 2.00, P = 0.007), indicating a statistically significant advantage for CBZ, but a large amount of statistical heterogeneity was present between trials (Chi² test = 8.74, df = 3, P = 0.03, I² statistic = 66%). When we repeated the analysis using random‐effects, the pooled HR for participants with partial onset seizures was 1.58 (95% CI 0.82 to 3.06, P = 0.17), still indicating an advantage for CBZ, but this advantage was no longer statistically significant.

Overall, the pooled HR (adjusted for seizure type) was 1.50 (95% CI 1.15 to 1.95, P = 0.003, from fixed‐effect analysis), providing evidence of a statistically significant advantage for CBZ. When we repeated the analysis using random‐effects (Chi² test = 9.24, df = 6, P = 0.16, I² statistic = 35%), the pooled HR was 1.53 (95% CI 1.02 to 2.28, P = 0.04). In this case, the advantage of CBZ was still statistically significant. We found no interaction between treatment and seizure type (generalised versus partial onset) (Chi² test = 0.00, df = 1, P = 0.95, I² statistic = 0%).

The sensitivity analysis including only the 20 participants randomised in de Silva 1996 before the withdrawal of the PB arm gave similar results with a pooled HR adjusted for seizure type for 633 participants of 1.42 (95% CI 1.08 to 1.86, P = 0.01) and heterogeneity between trials was reduced to 0 in this analysis (Chi² test = 5.66, df = 3, P = 0.14, I² statistic = 0%). Results within each seizure group were also similar in this sensitivity analysis, with a pooled HR for 115 participants with generalised seizures of 1.37 (95% CI 0.69 to 2.73, P = 0.37, I² statistic = 0%) and a pooled HR of 1.43 (95% CI 1.06 to 1.92, P = 0.02, I² statistic = 46%) for 498 participants with partial seizures (see Table 4 for further details).

Following reclassification of the 65 participants aged 30 or older with new onset generalised seizures in Heller 1995; Ogunrin 2005; and Placencia 1993 (see Sensitivity analysis), results were very similar and conclusions were unchanged (results available from review authors).

Inadequate allocation concealment in Placencia 1993 may have influenced withdrawal rates if participants, or personnel, or both, were aware of which drug the participants had been assigned; from the data we received, 19% of participants withdrew from the CBZ arm, and 15% of participants withdrew from the PB arm while the other 3 studies included in the analysis showed more participants withdrawing from the PB arm than the CBZ arm. Furthermore, inconsistencies between published data and data provided to us and unclear definitions for reason of withdrawal (participants withdrew for 'no clearly articulated reason') was likely to have influenced the results of our analysis. These factors in the Placencia 1993 trial in addition to the continuation of the CBZ arm in de Silva 1996 after the withdrawal of the PB arm are all factors that are likely to have contributed to the heterogeneity in Analysis 1.1 and Analysis 1.2. These factors may have confounded the results of our primary analyses in this review.

(2) Time to achieve 12‐month remission

For this outcome, a HR greater than one indicates a clinical advantage for PB.

Data for 683 participants from 4 trials were available for analyses of time to 12‐month remission and time to 6‐month remission (98.8% of 691 participants from de Silva 1996; Heller 1995; Mattson 1985; and Placencia 1993 (see Included studies and Table 2) and 46.9% of the total 1455 participants from the 13 included studies). Mattson 1985 recorded no follow‐up data for one participant randomised to CBZ. de Silva 1996 did not record the randomised drug for six participants, and in Placencia 1993, seizure data after occurrence of first seizure were not available for one participant randomised to PB, so we did not include this participant in the analyses. The study duration of Ogunrin 2005 was 12 weeks, so 12‐month remission was not possible among participants in this trial. Banu 2007 recorded the date of first seizure after randomisation, but all dates of subsequent seizures were not available; therefore, we could calculate 'Time to first seizure' but not 'Time to six‐month remission' and 'Time to 12‐month remission'.

Two hundred and eighty out of 683 participants (41%) achieved 12‐month remission, 163 out of 384 (45%) on CBZ and 117 out of 319 (37%) on PB. The overall pooled HR (for 683 participants) was 0.93 (95% CI 0.72 to 1.19, P = 0.57), suggesting no advantage for either drug. There was no evidence of statistical heterogeneity between trials (Chi² test = 3.54, df = 3, P = 0.32, I² statistic = 15%, see Analysis 1.3).

We performed sensitivity analysis excluding participants from Placencia 1993 from the analysis because of high risk of selection bias due to inadequate allocation concealment (see Allocation (selection bias) and Table 4). This sensitivity analysis resulted in a pooled HR of 0.82 (95% CI 0.61 to 1.09, P = 0.17), suggesting an advantage for CBZ that was not statistically significant. Again, there was no evidence of statistical heterogeneity between trials (Chi² test = 0.33, df = 2, P = 0.85, I² statistic = 0%). Our conclusion did not change following the sensitivity analysis.

In Placencia 1993, there was evidence that the proportional hazards assumption of the Cox model may have been violated; the P value of the time‐varying covariate was < 0.001. On closer inspection of the participants in Placencia 1993, all 60 participants who achieved 12‐month remission achieved immediate remission (i.e., did not have any seizures at all in the first 12 months of follow up). The trial followed up a further 42 participants for more than 365 days (up to 548 days); however, none of these participants achieved a 12‐month period of seizure freedom during the trial, so we censored them all at their last follow‐up date (after 365 days). This observation would explain the apparent change in treatment effect over time in Placencia 1993 and therefore the violation of the proportional hazards assumption. When we analysed separately those who achieved immediate 12‐month remission, the proportional hazards assumption was satisfied (P value of time‐varying covariate was 0.872). The proportional hazards assumption of the Cox model was satisfied for all other trials included in the analysis.

For participants with generalised onset seizures (158), the pooled HR was 0.64 (95% CI 0.41 to 1.01, P = 0.05), suggesting a borderline statistically significant advantage for CBZ. There was no evidence of statistical heterogeneity between studies for participants with generalised seizures (Chi² test = 0.61, df = 2, P = 0.74, I² statistic = 0%) For participants with partial onset seizures (525), the pooled HR was 1.11 (95% CI 0.81 to 1.59, P = 0.52), suggesting an advantage for PB that was not statistically significant. A considerable amount of statistical heterogeneity is present between studies for participants with partial onset seizures (Chi² test = 9.06, df = 3, P = 0.03, I² statistic = 67%). When we repeated the analysis with random‐effects, the result for participants with generalised seizures was unchanged, and for participants with partial onset seizures, the pooled HR was 1.24 (95% CI 0.69 to 2.22, P = 0.47), showing a larger advantage for PB that was not statistically significant. Overall, the pooled HR (adjusted for seizure type for 683 participants, fixed‐effect) was 0.93 (95% CI 0.72 to 1.20, P = 0.57), suggesting no clear overall advantage for either drug, but a considerable amount of heterogeneity was present between studies (Chi² test = 13.48, df = 6, P = 0.04, I² statistic = 55%). When we repeated the analysis with random‐effects, results were similar and conclusions unchanged. We found a statistically significant interaction between treatment and seizure type (generalised versus partial onset) (Chi² test = 3.81, df = 1, P = 0.05, I² statistic = 73.8%, see Analysis 1.4, calculated with fixed‐effect).

Upon visual inspection of forest plots in Analysis 1.4, it was clear that Placencia 1993 was the main source of the heterogeneity between studies in the subgroup of participants with partial onset seizures. The other 3 studies showed moderate, non‐significant effect sizes while Placencia 1993 showed a large, significant effect size in favour of PB (HR 2.43, 95% CI 1.27 to 4.65). This effect was not shown in the subgroup of participants with generalised onset seizures in participants in Placencia 1993 (HR 0.48, 95% CI 0.19 to 1.18). Repeating our sensitivity analysis from above, excluding Placencia 1993 from analysis due to inadequate allocation concealment, heterogeneity reduced to 0 (I² statistic = 0%) in all analyses, and there was no longer evidence of an interaction between treatment and seizure type. Results were also changed for participants with generalised onset seizures (101) (pooled HR 0.71, 95% CI 0.42 to 1.19, P = 0.19), showing an advantage for CBZ that was no longer statistically significant; for participants with partial onset seizures (394), a pooled HR of 0.88 (95% CI 0.62 to 1.25, P = 0.47) showed a change in direction of effect, now indicating an advantage for CBZ that was not statistically significant. And overall, the pooled HR (adjusted for seizure type for 495 participants) was 0.82 (95% CI 0.61 to 1.10, P = 0.18), suggesting an advantage for CBZ that was not statistically significant.

The sensitivity analysis excluding participants randomised to CBZ following withdrawal of the PB arm in the de Silva 1996 trial gave similar results, with an estimated pooled hazard ratio of 0.90 (95% CI 0.69 to 1.17, P = 0.42). Results within each seizure group were also similar, with a pooled HR of 0.59 (95% CI 0.37 to 0.95, P = 0.03) for participants with generalised seizures (137) and a pooled HR of 1.09 (95% CI 0.79 to 1.49, P = 0.61) for participants with partial seizures (503), resulting in no changes in conclusions (see Table 4 for further details).

Following reclassification of the 65 participants aged 30 or older with new onset generalised seizures in Heller 1995; Ogunrin 2005; and Placencia 1993 (see Sensitivity analysis), results were very similar and conclusions were unchanged (results available from review authors).

As in the analysis of our primary outcome, Placencia 1993 seemed to be contributing the majority of the variability between trial results. This could have been a knock–on effect of the inadequate allocation concealment in this trial, which was likely to have influenced the withdrawal rates in this study and in turn the number of participants remaining in the trial who could achieve 12‐month remission. Again, we conclude that the inclusion of this study may have confounded the results of this outcome.

(3) Time to achieve 6‐month remission

For this outcome, a HR greater than 1 indicates a clinical advantage for PB. See 'Time to 12‐month remission' for details of participants included in the analyses of time to 6‐month remission.

Three hundred and eighty‐seven out of 683 participants (57%) achieved 6‐month remission, 213 out of 384 (59%) on CBZ and 117 out of 319 (55%) on PB. The overall pooled HR (for 683 participants) was 1.02 (95% CI 0.83 to 1.26, P = 0.86), suggesting no advantage for either drug. There was no evidence of statistical heterogeneity between trials (Chi² test = 3.63, df = 3, P = 0.30, I² statistic = 17%, see Analysis 1.5).

We performed sensitivity analysis excluding participants from Placencia 1993 from the analysis because of high risk of selection bias due to inadequate allocation concealment (see Allocation (selection bias) and Table 4). This sensitivity analysis resulted in a pooled HR of 0.88 (95% CI 0.68 to 1.14, P = 0.34), suggesting an advantage for CBZ that was not statistically significant. Again, there was no evidence of statistical heterogeneity between trials (Chi² test = 0.14, df = 2, P = 0.93, I² statistic = 0%). Our conclusion did not change following the sensitivity analysis.

In Mattson 1985, there was an indication that the proportional hazards assumption may have been violated (see Data synthesis); the P value of the time‐varying covariate was 0.054, and visual inspection of the cumulative incidence plot (Figure 12) showed crossing of the curves at around 300 days. In other words, up to 300 days, participants on PB seemed to be achieving 6‐month remission quicker than those on CBZ, but this changed after 300 days. However, participant numbers were reduced by 300 days (83 participants at risk out of 308 randomised), so small changes may have been magnified at this time.

Figure 12

---

---

Time to six‐month remission ‐ Mattson 1985

As a sensitivity analysis, we fitted a piecewise Cox regression model to investigate any change in treatment effect over time assuming proportional hazards within each interval. From the visual inspection of Figure 12, we split the follow‐up period of Mattson 1985 into 3 intervals: 0 to 182.5 days (immediate 6‐month remission), 182.5 to 300 days, and over 300 days (maximum follow up: 1616 days). We estimated separate hazard ratios for each interval.

For 'interval 0 to 182.5 days' (74 events from 308 at participants at risk), the HR was 1.06 (95% CI 0.63 to 1.77, P = 0.83), indicating no clear advantage of either drug. For 'interval 182.5 to 300 days' (22 events from 83 participants at risk), the HR was 0.65 (95% CI 0.37 to 1.15, P = 0.14), suggesting an advantage for CBZ that was not statistically significant. For 'interval over 300 days' (20 events from 41 participants at risk), the HR was 0.92 (95% CI 0.64 to 1.33, P = 0.65), suggesting no clear advantage of either drug.

These results suggest some indication of a change in treatment effect over time, with no clear advantage between the two drugs in the early stages of the trial for immediate remission; an advantage for CBZ emerged after the initial six months, which was no longer present by the end of the study. However, the confidence intervals of estimates were wide, particularly for later times in the trial due to small numbers of events and participants and risk, so we do not have statistically significant evidence to support the hypothesis of a change in treatment effect over time for Mattson 1985. Thus, we conclude that the observed difference in treatment effect around 180 to 300 days compared with the rest of the study follow up was likely to be due to chance that more participants on CBZ achieved 6‐month remission than those on PB at this time (16 participants on CBZ compared with 6 on PB in this time interval) while the numbers of participants achieving 6‐month remission were more comparable at other time points. The proportional hazards assumption of the Cox model was satisfied for all other trials included in the analysis.

For participants with generalised onset seizures (158), the pooled HR was 0.69 (95% CI 0.47 to 1.01, P = 0.06), suggesting a borderline statistically significant advantage for CBZ. There was no evidence of statistical heterogeneity between studies for participants with generalised seizures (Chi² test = 1.25, df = 2, P = 0.54, I² statistic = 0%). For participants with partial onset seizures (525), the pooled HR of 1.17 (95% CI 0.90 to 1.50, P = 0.24) suggested an advantage for PB that was not statistically significant. A considerable amount of statistical heterogeneity was present between studies for participants with partial onset seizures (Chi² test = 7.99, df = 3, P = 0.05, I² statistic = 62%). When we repeated the analysis with random‐effects, the result for participants with generalised seizures was unchanged, and for participants with partial onset seizures, the pooled HR of 1.15 (95% CI 0.73 to 1.82, P = 0.54) still showed an advantage for PB that was not statistically significant. Overall, the pooled HR (adjusted for seizure type for 683 participants, fixed‐effect) was 0.99 (95% CI 0.80 to 1.23, P = 0.95), suggesting no clear overall advantage for either drug, but a considerable amount of heterogeneity was present between studies (Chi² test = 14.24, df = 6, P = 0.03, I² statistic = 58%). When we repeated the analysis with random‐effects, results were similar and conclusions unchanged. We found a statistically significant interaction between treatment and seizure type (generalised versus partial onset) (Chi² test = 5.00, df = 1, P = 0.03, I² statistic = 80.0%, see Analysis 1.6, calculated with fixed‐effect).

As in Analysis 1.4, from visual inspection of forest plots in Analysis 1.6, it was clear that Placencia 1993 was the main source of the heterogeneity between studies in the subgroup of participants with partial onset seizures. The other three studies showed moderate, non‐significant effect sizes, while Placencia 1993 showed a large, significant effect size in favour of PB (HR 1.95, 95% CI 1.25 to 3.04). Again, this effect was not shown in the subgroup of participants with generalised onset seizures in participants in Placencia 1993 (HR 0.52, 95% CI 0.27 to 0.98). Repeating our sensitivity analysis from above as in the analysis of 'Time to 12‐month remission', excluding Placencia 1993 from analysis because of inadequate allocation concealment, reduced heterogeneity to 0 (I² statistic = 0%) in all analyses, and there was no longer evidence of an interaction between treatment and seizure type. Results were also changed for participants with generalised onset seizures (101), with a pooled HR of 0.81 (95% CI 0.50 to 1.32, P = 0.40) showing an advantage for CBZ that was not statistically significant; for participants with partial onset seizures (394), a pooled HR of 0.91 (95% CI 0.67 to 1.24, P = 0.56) showed a change in direction of effect, again indicating an advantage for CBZ that was not statistically significant. And overall, the pooled HR (adjusted for seizure type for 495 participants) was 0.88 (95% CI 0.68 to 1.14, P = 0.34), suggesting an advantage for CBZ that was not statistically significant.

The sensitivity analysis excluding participants randomised to CBZ following the withdrawal of the PB arm in the de Silva 1996 trial gave similar results, with an estimated pooled hazard ratio of 0.97 (95% CI 0.78 to 1.21, P = 0.79). Results within each seizure group were also similar, with a pooled HR of 0.66 (95% CI 0.45 to 0.98, P = 0.04) for participants with generalised seizures (137) and a pooled HR of 1.14 (95% CI 0.88 to 1.48, P = 0.31) for participants with partial seizures (503), resulting in no changes in conclusions (see Table 4 for further details).

Following reclassification of the 65 participants aged 30 or older with new onset generalised seizures in Heller 1995; Ogunrin 2005; and Placencia 1993 (see Sensitivity analysis), results were very similar and conclusions were unchanged (results available from review authors).

As in the analysis of our outcomes 'Time to withdrawal of allocated treatment' and 'Time to 12‐month remission', Placencia 1993 seemed to be contributing the majority of the variability between trial results (see the above outcomes for discussion). Again, we conclude that the inclusion of this study may have confounded the results of this outcome.

4) Time to first seizure postrandomisation

For this outcome, a HR greater than one indicates a clinical advantage for CBZ.

We had data for 822 participants from 6 trials (98.3% of 836 participants from Banu 2007; de Silva 1996; Heller 1995; Mattson 1985; Ogunrin 2005; and Placencia 1993 (see Included studies and Table 2). de Silva 1996 did not record the randomised drug for 6 participants, and dates of seizure recurrence were not available for 8 participants (4 randomised to CBZ and 4 to PB) in Mattson 1985; therefore, we did not include these 14 participants in the analysis.

Four hundred and fifty‐three out of 822 participants (55%) experienced seizure recurrence, 264 out of 434 (61%) on CBZ and 189 out of 388 (49%) on PB. The overall pooled HR (for 822 participants) was 0.86 (95% CI 0.71 to 1.04, P = 0.12), suggesting an advantage for PB that was not statistically significant. There was no evidence of statistical heterogeneity between trials (Chi² test = 6.26, df = 5, P = 0.28, I² statistic = 20%, see Analysis 1.7).

We performed sensitivity analysis excluding participants from Placencia 1993 from analysis because of high risk of selection bias due to inadequate allocation concealment (see Allocation (selection bias) and Table 4). This sensitivity analysis resulted in a pooled HR of 0.87 (95% CI 0.71 to 1.08, P = 0.22), still suggesting an advantage for PB that was not statistically significant. Again, there was no significant evidence of statistical heterogeneity between trials (Chi² test = 6.04, df = 4, P = 0.20, I² statistic = 34%). Our conclusion did not change following the sensitivity analysis.

In Banu 2007, we found inconsistencies (between the IPD dataset and published results), which the study authors could not resolve; the publication reported that only 7 participants had experienced no seizures from the start of treatment (3 randomised to PB and 4 randomised to CBZ); however, from IPD provided, 21 participants did not experience seizures from the start of treatment (12 randomised to PB and 9 randomised to CBZ). Given these inconsistencies and the limited data available on seizure recurrence, we performed sensitivity analysis excluding the participants from Banu 2007 from Analysis 1.7. This sensitivity analysis resulted in a pooled HR of 0.82 (95% CI 0.66 to 1.01, P = 0.06), suggesting a slightly larger advantage to PB, which is now borderline statistically significant. Again, there was no evidence of statistical heterogeneity between trials (Chi² test = 5.11, df = 4, P = 0.28, I² statistic = 22%). This sensitivity analysis showed that Banu 2007, a trial which showed a small, non‐significant advantage for CBZ, may have confounded the results of our analysis; without the inclusion of this trial, our results indicated a larger, borderline statistically significant advantage for PB for the outcome of time to first seizure.

For participants with generalised onset seizures (238), the pooled HR was 1.23 (95% CI 0.86 to 1.77, P = 0.27), suggesting an advantage for CBZ that was not statistically significant. A considerable amount of statistical heterogeneity was present between studies for participants with generalised onset seizures (Chi² test = 8.65, df = 4, P = 0.07, I² statistic = 54%). For participants with partial onset seizures (584), the pooled HR of 0.76 (95% CI 0.60 to 0.96, P = 0.02) suggested a statistically significant advantage for PB. There was no evidence of statistical heterogeneity between studies for participants with partial onset seizures (Chi² test = 4.55, df = 5, P = 0.47, I² statistic = 0%). When we repeated the analysis with random‐effects, the result for participants with partial onset seizures was unchanged, and for participants with generalised onset seizures, the pooled HR of 1.15 (95% CI 0.66 to 2.02, P = 0.62) still showed an advantage for CBZ that was not statistically significant. Overall, the pooled HR (adjusted for seizure type for 822 participants, fixed‐effect) was 0.87 (95% CI 0.72 to 1.06, P = 0.18), suggesting an advantage for PB that was not statistically significant. A considerable amount of heterogeneity was present between studies (Chi² test = 17.98, df = 10, P = 0.06, I² statistic = 44%). When we repeated the analysis with random‐effects, the results were similar and conclusions unchanged. We found a statistically significant interaction between treatment and seizure type (generalised versus partial onset) (Chi² test = 4.78, df = 1, P = 0.03, I² statistic = 79.1%, see Analysis 1.8, calculated with fixed‐effect).

From visual inspection of forest plots in Analysis 1.8, it was clear that Ogunrin 2005 was the main source of the heterogeneity between studies in the subgroup of participants with generalised onset seizures. The other 4 studies showed non‐significant advantages of CBZ, while Ogunrin 2005 showed a large, significant effect size in favour of PB (HR 0.21, 95% CI 0.06 to 0.76). The subgroup of participants with partial onset seizures in participants in Ogunrin 2005 did not show this effect (HR 1.42, 95% CI 0.26 to 7.80). Reclassification of the 65 participants aged 30 or older with new onset generalised seizures in Heller 1995; Ogunrin 2005; and Placencia 1993 (see Sensitivity analysis) into an uncertain seizure type group (see Analysis 1.9) reduced heterogeneity between studies for the remaining 757 participants to 0 (I² statistic = 0%); the results among participants with partial onset seizures were unchanged. For participants with generalised onset seizures (173), a pooled HR of 1.39 (95% CI 0.86 to 1.77, P = 0.13) indicated a larger advantage of CBZ that still does not reach statistical significance. (We note that we could not calculate the HR for Ogunrin 2005 as following reclassification, only a single participant remained in the PB group and did not experience seizure recurrence.) Among the group of participants with 'uncertain' seizure type (65), the pooled HR of 1.22 (95% CI 0.59 to 2.51, P = 0.59) suggested an advantage of CBZ that was not statistically significant. A considerable amount of heterogeneity was present in the analysis of reclassified participants (Chi² test = 4.78, df = 2, P = 0.09, I² statistic = 58%), which was perhaps unsurprising as this relatively small group was made up of participants with 'uncertain' and likely different seizure types. Following reclassification, a statistically significant interaction between treatment and seizure type (generalised versus partial onset) still existed (Chi² test = 6.64, df = 2, P = 0.04, I² statistic = 69.9%, see Analysis 1.9), indicating an advantage for PB for participants with partial onset seizures and an advantage for CBZ for participants with generalised onset seizures.

The sensitivity analysis excluding participants randomised to CBZ following withdrawal of the PB arm in the de Silva 1996 trial gave similar results, with an estimated pooled hazard ratio of 0.87 (95% CI 0.71 to 1.06, P = 0.10). Results within each seizure group were also similar, with a pooled HR of 1.20 (95% CI 0.82 to 1.75) for participants with generalised seizures (217) and a pooled HR of 0.77 (95% CI 0.61 to 0.97, P = 0.007) for participants with partial seizures (562) (see Table 4 for further details).

In de Silva 1996, there was an indication that the proportional hazards assumption may have been violated (see Data synthesis); the P value of the time‐varying covariate was 0.08, and visual inspection of the cumulative incidence plot (Figure 13) showed crossing of the curves at around 100 days. In other words, up to 100 days, more participants on CBZ seemed to be experiencing first seizure recurrence earlier than those on PB, but this changed after 100 days. However, participant numbers were reduced by 100 days (26 participants at risk out of 64 randomised), so small changes may have been magnified at this time. Furthermore, curves also seemed to cross at around 800 days, when even fewer participants remained at risk of first seizure in the trial (11 participants at risk out of 64 randomised).

Figure 13

---

---

Time to first seizure ‐ de Silva 1996

As a sensitivity analysis, we fitted a piecewise Cox regression model to investigate any change in treatment effect over time assuming proportional hazards within each interval. From the visual inspection of Figure 13, we split the follow‐up period of de Silva 1996 into 3 intervals: 0 to 100 days, 100 to 800 days, and over 800 days (maximum follow up 4163 days). We estimated separate hazard ratios for each interval.

For 'interval 0 to 100 days' (38 events from 64 participants at risk), the HR was 0.92 (95% CI 0.36 to 2.34, P = 0.83), indicating no clear advantage of either drug. For 'interval 100 to 800 days' (14 events from 26 participants at risk), the HR was 1.06 (95% CI 0.55 to 2.01, P = 0.86), again, suggesting no clear advantage of either drug. Over 800 days, 11 participants remained at risk; however, neither of the 2 remaining participants randomised to PB experienced an event (shown by the flattening of the curve at around 700 to 800 days in Figure 13); therefore, the HR of first seizure recurrence was undefined over this time period. Furthermore, in sensitivity analysis excluding participants randomised to CBZ following withdrawal of the PB arm in the de Silva 1996 trial, there was no longer evidence that the proportional hazards assumption had been violated; the P value of the time‐varying covariate was 0.316 among these 20 participants.

We did not find any statistically significant evidence to support a change in treatment effect over time in de Silva 1996 for the outcome of time to first seizure. We conclude that the imbalance in participant numbers in the 2 randomised groups (54 randomised to CBZ and 10 randomised to PB) magnified the apparent crossing of the survival plots over time and the majority of participants experiencing an event (60 participants experienced a seizure while only 4 were censored in this analysis) was also likely to be an influence. The proportional hazards assumption of the Cox model was satisfied for all other trials included in the analysis.

We conclude from this analysis that there was likely to be a difference in efficacy of the drugs (in terms of time to first seizure recurrence after randomisation) by seizure type, that participants with generalised seizures experience seizure recurrence later on CBZ than PB, and that participants with partial onset seizures experience seizure recurrence later on PB than CBZ. The overall trend towards an advantage for PB for all included participants reflects that the majority of participants included in this analysis had partial onset seizures (71% of 822 included participants). It was possible that inconsistencies in data provided to us (Banu 2007) and misclassification of seizure type in participants over the age of 30 (Heller 1995;Ogunrin 2005; Placencia 1993) may have confounded the results of this analysis. However, in a sensitivity analysis to take account of these confounding factors, the association between treatment and seizure type still existed and therefore could be a true association.

5) Adverse events

We extracted all reported information related to adverse events from the study publications. Cossu 1984 did not report any findings related to adverse events, and we are uncertain if these data were collected without access to protocols (see Selective reporting (reporting bias)). (See Table 5 for details of all adverse event data provided in the other 12 studies included in this review.) Two studies reported only numbers of withdrawals due to adverse events (Chen 1996; Czapinski 1997), and two reported the rate of adverse events/number of participants reporting adverse events (Bidabadi 2009; Placencia 1993); these four studies did not report specific adverse events. For the eight studies that did report specific adverse events, in summary, the adverse events reported by two or more studies in this review are as follows.