Types of studies

We will include RCTs using either:

• an adequate method of allocation concealment (e.g. sealed opaque envelopes);

• a quasi method of randomisation (e.g. allocation by date of birth).

Studies may be double‐blind, single‐blind or unblinded. We will exclude studies of a monotherapy design; in other words participants are randomised to treatment with a single drug throughout the study period. We will also exclude studies with an add‐on, polytherapy, transitional or withdrawal to monotherapy periods of any length.

We will include studies of parallel designs will be included. We will exclude studies of a cross‐over design as we believe this design is inappropriate for measuring our primary time‐to‐event outcome 'time to withdrawal of allocated treatment' as a withdrawal of allocated treatment in the first treatment period would mean than the participant could not cross into the second treatment period, potentially leading to a large amount of incomplete outcome data and therefore a reduction in statistical power. Furthermore, the use of cross‐over designs is no longer recommended in epilepsy due to concerns over study duration, large proportions of dropouts, unblinding of masked treatments as participants cross into the second period and potential carryover effects; a particular concern in studies of a monotherapy design which aim to assess the effect of a single treatment (Engel 2008; Wyllie 2006).

Types of participants

Children or adults with partial onset seizures (simple partial, complex partial, or secondarily generalised tonic‐clonic seizures) or generalised onset tonic‐clonic seizures (with or without other generalised seizure types). We will not include participants with other generalised seizure types alone (for example absence seizures alone without generalised tonic clonic seizures) as guidelines for the first‐line treatment of other generalised seizure types are often different from the guidelines for generalised tonic‐clonic seizures (NICE 2012), and due to documented evidence that certain drugs of interest in our review may exacerbate some generalised seizure types (How the interventions might work). We will also consider individuals with a new diagnosis of epilepsy, or who had had a relapse following antiepileptic monotherapy withdrawal.

We will exclude studies which consider antiepileptic drugs as treatment for conditions other than epilepsy.

Types of interventions

We will include the 10 antiepileptic drugs currently licensed and used as monotherapy in our network of treatments: carbamazepine, phenytoin, valproate, phenobarbitone, oxcarbazepine, lamotrigine, gabapentin, topiramate, levetiracetam, zonisamide.

Included studies must make at least one pairwise comparison between at least 2 of the 10 antiepileptic drugs included in our network. For studies with three treatment arms or more, we will include treatment arms only of the 10 antiepileptic drugs included in our network; treatment arms of drugs not included in our network will be excluded from analysis. We will not make pairwise comparisons (direct or indirect) between any antiepileptic drugs not specified above. We will make pairwise comparisons (based on direct and/or indirect evidence) between all 10 drugs (Data synthesis).

Types of outcome measures

We will investigate the following outcomes in this review (Primary outcomes; Secondary outcomes). Reporting of these outcomes in the original trial report is not an eligibility requirement for inclusion in this review.

Selection of studies

One author (SJN) will screen all titles and abstracts of all studies identified by the electronic searches as described in Search methods for identification of reviews according to the inclusion criteria specified above (Types of studies; Types of participants; Types of interventions). Subsequently, two authors (SJN and AGM) will independently assess full‐text publications according to the same inclusion criteria specified above. We will resolve disagreements by discussion or by consulting a third author (CT) where necessary. We will record the reasons for exclusion of studies at both stages of study selection. We will contact trial authors for clarification if eligibility of a study is unclear from the published information.

Data extraction and management

For all studies meeting our inclusion criteria, two authors (SJN and AGM) will send a data request form to the first and/or corresponding author of the study or to the study sponsor where appropriate (referred to as data providers herein).

Our data request form will ask the data providers if the following information is available (tick yes or no).

• Trial methods:

  • method of generation of random list;

  • method of concealment of randomisation;

  • stratification factors;

  • blinding methods.

• Participant covariates:

  • sex;

  • age;

  • seizure types;

  • epilepsy status (newly diagnosed / relapsed seizures following drug withdrawal);

  • time between first seizure and randomisation;

  • number of seizures prior to randomisation (with dates);

  • presence of neurological signs;

  • electroencephalography (EEG) results;

  • computed tomography (CT) or magnetic resonance imaging (MRI) results;

  • aetiology of seizures (if known).

• Follow‐up data:

  • treatment allocation;

  • date of randomisation;

  • dates of follow up;

  • dates of seizures post randomisation or seizure frequency data between follow‐up visits;

  • dates of treatment withdrawal and reason(s) for treatment withdrawal;

  • starting dose of treatment;

  • dates of dose changes;

  • adverse events reported.

We will also request for any available, related documents such as case report forms, study protocols, clinical summaries from data providers.

In the event of no response to our IPD request, we will send a follow‐up e‐mail to the original data provider contacted. If we still receive no response for a particular study we will attempt to contact another study author or sponsor where appropriate. If a data provider is unable to make IPD available to us, we will request any aggregate data related to our outcome not reported in the publication.

If data cannot be obtained (no response to any requests or IPD is not available), two independent authors (SJN and MS) will assess whether any relevant and appropriate aggregate level data has been reported in the trial publication. We will resolve any disagreements on extracted aggregate data by discussion or by consulting a third author (CT) if necessary.

We will store all obtained data on a secure, dedicated network drive accessible only to the statisticians performing analysis (SJN, MS, CT). We will check all provided data for consistency and prepare them for analysis according to a pre‐specified procedure prepared by one author (SJN) (available on request) and piloted by two authors (SJN and MS). For each trial where IPD are supplied, we will reproduce results from trial findings where possible and we will perform the following consistency checks:

• trial details cross‐checked against any published report of the trial; original trial authors to be contacted if missing data, errors or inconsistencies are found;

• review of the chronological randomisation sequence by checking the balance of prognostic factors, taking account of factors stratified for in randomisation procedure.

We will discuss any inconsistencies in the provided data with the corresponding data providers. If large or major inconsistencies are present which cannot be resolved by data providers, we will not include the data in any analyses. If minor inconsistencies are present, we will analyse the data and conduct sensitivity analyses to test the robustness of results (Sensitivity analysis).

For the analysis of time to withdrawal of allocated treatment as a time‐to‐event outcome, we define an 'event' as either the withdrawal of the allocated treatment due to poor seizure control or adverse events or both. Non‐compliance with the treatment regimen or the addition of another antiepileptic drug will also be classed as 'events'. The outcome will be censored if treatment was withdrawn because the individual achieved a period of remission, if a participant withdrew from allocated treatment for reasons not related to the treatment (such as loss to follow‐up) or if the individual was still on allocated treatment at the end of follow‐up. Two authors (SJN and AG) will independently review reasons for treatment withdrawal for classification as events or censored observations, and we will resolve any disagreements by mutual discussion or by involving a third author (CT).

If seizure data are provided or recorded in terms of the number of seizures recorded between clinic visits rather than specific dates of seizures, to enable time‐to‐event outcomes to be calculated, we will apply linear interpolation to approximate the dates on which seizures occurred. For example, if four seizures were recorded between two visits which occurred on March 1st and May 1st (an interval of 61 days), then date of first seizure would be approximately March 13th. This will allow an estimate of the time to six‐month and 12‐month remission and the time to first seizure to be computed.

We will calculate time to six‐month and 12‐month remission from the date of randomisation to the date (or estimated date) the individual had first been free of seizures for six or 12 months respectively. If the person had one or more seizures in the titration period, a six‐month or 12‐month seizure‐free period could also occur between the estimated date of the last seizure in the titration period and the estimated date of the first seizure in the maintenance period

We will calculate time to first seizure from the date of randomisation to the date that their first seizure was estimated to have occurred. If seizure data are missing for a particular visit, these outcomes will be censored at the previous visit. These outcomes will also be censored if the individual died or if follow‐up ceased prior to the occurrence of the event of interest.

Assessment of risk of bias in included studies

Two independent authors (SJN, JW) will assess risk of bias in all included trials using the Cochrane Collaboration's tool for assessing risk of bias (Higgins 2011). The following methodological criteria are assessed according to this tool:

• selection bias (sequence generation and allocation concealment);

• performance bias (blinding of participants and personnel);

• detection bias (blinding of outcome assessment);

• attrition bias (incomplete outcome data);

• reporting bias (selective outcome reporting).

We will resolve any disagreements by discussion. In theory, a review using IPD should overcome issues of reporting biases as unpublished data can be provided and unpublished outcomes calculated. Any selective reporting bias detected could be assessed with the Outcome Reporting Bias in Trials (ORBIT) classification system (Kirkham 2010). As specified in Data extraction and management, we will ask the data providers to provide trial methods such as randomisation and blinding methods, and we will discuss any missing data and/or inconsistencies with them.

We will conduct sensitivity analyses by excluding any study judged to be at high risk of bias for any methodological aspect (Sensitivity analysis).

Measures of treatment effect

We will summarise all time‐to‐event outcomes using the hazard ratio (HR) as the measure of treatment effect. We will calculate outcomes from IPD provided where possible or extract them from published studies. We will not attempt to analyse or synthesise adverse event data; a large range of different adverse events are thought to be associated with the 10 different drugs and such data is generally collected and presented in different ways across studies. For these reasons, we believe a synthesis of adverse event data would present only selective information while a narrative description of adverse event data from IPD or extracted from published studies will be the most informative way of presenting this data.

Unit of analysis issues

We do not anticipate any unit of analysis issues. For inclusion in the review, the unit of allocation must be individual and studies of a repeated measures (longitudinal) nature or of a cross‐over design are not eligible for inclusion

Dealing with missing data

For all included studies, we will conduct an assessment of the proportion of missing outcome, demographic and covariate data and make a judgement regarding the extent and nature of missing data (e.g. missing at random, missing not at random). We will attempt to contact all study authors in order to request relevant data; we will include any information regarding missing data in such requests (Data extraction and management). If further information regarding missing data cannot be provided and we judge that an important proportion of data (particularly outcome data) is missing, we will conduct sensitivity analyses to investigate the potential impact of the missing data (for example best case scenario or worst case scenario analyses, assuming those with missing outcome data all have a favourable/unfavourable outcome, respectively).

Assessment of heterogeneity

We will use a fixed‐effect model for all pairwise and network meta‐analyses in the first instance (Data synthesis). For each pairwise comparison , we will assess the presence of heterogeneity statistically using the Q test (P value < 0.10 for significance) and the I² statistic with the following interpretation (Higgins 2003).

• 0% to 40%: might not be important;

• 30% to 60%: may represent moderate heterogeneity;

• 50% to 90%: may represent substantial heterogeneity;

• 75% to 100%: considerable heterogeneity.

We will also assess the presence of heterogeneity by visually inspecting forest plots, particularly in terms of the magnitude and direction of effects. If considerable heterogeneity is found to be present, we will perform network meta‐analysis with random effects and inspect trial and participant characteristics for sources of heterogeneity.

Assessment of reporting biases

Two authors (SJN and JW) will undertake a full risk of bias check for each eligible study, including risk of reporting biases. In theory, a review using IPD should overcome issues of reporting biases as unpublished data can be provided and unpublished outcomes calculated. We will assess potential selective reporting bias using the ORBIT classification system (Kirkham 2010). As specified in Data extraction and management, we will ask the data providers for trial methods such as randomisation and blinding methods, and we will discuss any missing data and inconsistencies with them.

Data synthesis

We will present all direct evidence visually via a treatment network diagram of the 10 antiepileptic treatments. This treatment diagram will have a maximum of direct 45 pairwise comparisons

We will take an intention‐to‐treat approach to analysis; in other words, participants will be analysed in the group to which they were randomised in an individual trial, irrespective of which treatment they actually received. Therefore, for time‐to‐event outcomes, 'time to six month remission', 'time to 12 month remission' and 'time to first seizure post randomisation' participants will not be censored if treatment was withdrawn. For the primary outcome, time to withdrawal of allocated treatment, we will consider withdrawals due to lack of efficacy (i.e. recurrent seizures), poor tolerability (i.e. adverse events) or a combination of both poor efficacy and tolerability. Other withdrawals such as losses to follow up, non‐treatment related deaths, administrative trial reasons etc. will be censored at the time of withdrawal.

For all time‐to‐event outcomes, we will investigate the relationship between the time‐to‐event and treatment effect of the anti‐epileptic drugs. We will use Cox proportional hazards regression models will be used to obtain study‐specific estimates of log(HR) or treatment effect and associated standard errors. The model assumes that ratio of hazards (risks) between the two treatment groups is constant over time (i.e. hazards are proportional). This proportional hazards assumption of the Cox regression model will be tested for each outcome of each study by testing the statistical significance of a time‐varying covariate in the model for each study. If a violation of this assumption is detected, we will perform sensitivity analyses via interval censored (piecewise) Cox models. To preserve the within‐trial randomisation, we will stratify Cox regression models by trial.

A key assumption made in network meta‐analysis is that treatment effect is "exchangeable" across all included trials; in other words, the indirect comparison made between two treatments is a feasible comparison to make (known as the transitivity assumption) and that the indirect evidence is consistent with the direct evidence where a comparison exists (known as the consistency assumption). Transitivity requires that all treatments are "jointly randomisable"; in other words, all 10 antiepileptic drugs could feasibly be randomised in the same trial and those which are not treatment arms in any given trial are "missing at random" (Lu 2006). This assumption can not be formally tested statistically; transitivity must be judged by careful consideration of trial settings and characteristics, treatment mechanisms and participant demographics to investigate if any differences would be expected to modify relative treatment effects. The consistency assumption can be evaluated statistically via the Bucher Method (Bucher 1997), which applies a z‐test to the difference between the direct treatment effect estimate and the indirect estimate for each loop of evidence. Given the simplicity of this test, the influence of the precision of the treatment effect estimate on the result of this test and the complexity introduced by multi‐arm trials and therefore association between treatment effects estimated from arms of the same trial, we will use a conservative significance threshold of 10% (P value < 0.1) to judge the presence of heterogeneity. We will also evaluate the presence of inconsistency statistically by comparing the fit of the consistency model (the primary analysis model) to the inconsistency model (the primary model adjusted for inconsistency factors) (Higgins 2012; Lu 2006).

We will perform all analyses using statistical software packages SAS (version 9.2) (Copyright, SAS Institute Inc. SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of SAS Institute Inc., Cary, NC, USA.), Stata software version 11.2 (StataCorp 2009), or R version 2.13 (CRAN 2011). At the time of writing, a data access system is under development to enable secure access to de‐identified pharmaceutical data from clinical trials (ClinicalStudyDataRequest.com) (https://clinicalstudydatarequest.com/Default.aspx). It is currently unclear which format participant data held within this secure data access system will in and therefore which statistical package(s) we will be able to use for analysis. We anticipate this will become clear by the time of analysis in the review.

For the 10 antiepileptic drugs included in the treatment network, we are able to make 45 possible pairwise comparisons for both partial and generalised onset seizures. For clinical interest and relevance, we will present HR estimates for each antiepileptic drug in the network compared to the current recommended first‐line treatments (carbamazepine or lamotrigine for partial onset seizures and valproate for generalised onset seizures) in the main results of this review via forest plots, and we will present the results for all 45 pairwise comparisons by seizure type as an 'Additional table' for each outcome.

Subgroup analysis and investigation of heterogeneity

There are strong clinical beliefs that certain antiepileptic drugs are more effective in certain seizure types than others, for example carbamazepine is more effective in partial onset seizures and valproate is more effective in generalised onset seizures (Marson 2000), suggesting that there is a treatment by seizure type (partial or generalised) interaction; if such an interaction exists then the key assumption of an exchangeable treatment effect across all included trials would be violated. To account for this, we will conduct all analyses separately by epilepsy type (partial onset or generalised onset) according to the classification of main seizure type at baseline. Partial seizures (simple or complex) and partial secondarily generalised seizures are classified as partial epilepsy. Primarily generalised seizures are classified as generalised epilepsy. We will then judge exchangeability of treatment effect separately by analyses of seizure type. If sufficient demographic participant level data is made available to us, we will also consider stratified or subgroup analyses according to participant covariates specified in Data extraction and management as potential magnifiers of treatment effect and as potential sources of heterogeneity and/or inconsistency. Specifically, in addition to differences between seizure types, we would also like to explore differences in age of seizure onset (paediatric onset, adult onset, elderly onset), seizure frequency before randomisation (time since first ever seizure and/or number of seizures before randomisation) and aetiology of seizures (if known according to pre‐treatment investigations such as EEG, CT and/or MRI scan).

Sensitivity analysis

As discussed above (Data extraction and management), we will discuss any inconsistencies in provided data with the corresponding data providers. If large or major inconsistencies are present which cannot be resolved by data providers, we will not include the data in any analyses. If minor inconsistencies are present, we will include the data in analyses and pursue sensitivity analyses to test the robustness of results included in this data. The exact sensitivity analyses we would perform would depend on the inconsistency present.

Given that misclassification of seizure type is a recognised problem in epilepsy (whereby some individuals with generalised seizures have been mistakenly classed as having partial onset seizures and vice versa) and such misclassification did impact upon the results of a review in our series of pairwise reviews for monotherapy in epilepsy comparing phenytoin and sodium valproate in which nearly 50% of participants analysed may have had their seizure type misclassified (Nolan 2013b), we intend to investigate the potential impact of misclassification on results in a sensitivity analysis. Given clinical evidence that individuals with generalised onset seizures are unlikely to have an 'age of onset' greater than 25 to 30 years (Malafosse 1994), we will examine the distribution of age at onset for individuals with generalised seizures. We intend to undertake two sensitivity analyses to investigate misclassification:

• re‐classification of all individuals with generalised seizure types and age at onset greater than 30 into an 'unclassified seizure type' group;

• re‐classification of all individuals with generalised seizures and age of onset greater than 30 as having partial onset seizures.