Types of studies

We included RCTs and quasi‐randomised trials (defined as methods of allocating people to a trial that are not random, but are intended to produce similar groups when used to allocate participants, such as those allocating by date or alternation (Higgins 2011)). One arm of the trial had to include RTT, compared against usual practice (including 'no treatment'), or an attention control group. We excluded studies where RTT was a component of both the experimental and control treatments. Examples of attention control treatments are comparable time spent receiving therapy on a different limb, or participating in an activity with no potential motor benefits. We accepted usual practice comparison groups when the intervention received by the control group was considered a normal or usual component of stroke rehabilitation practices, including neurophysiological or orthopaedic approaches. We assumed that, early after stroke, usual practice would mean that people would receive some therapy.

Types of participants

Adults (18 years and older) who have suffered a stroke. Stroke is defined by the World Health Organization (WHO) as "a syndrome of rapidly developing symptoms and signs of focal, and at times global, loss of cerebral function lasting more than 24 hours or leading to death, with no apparent cause other than that of vascular origin" (WHO 1989). We included trials starting any time after an acute stroke and in any setting. We excluded studies of participants with mixed aetiology (for example, participants with acquired brain injury) unless data were available relating to the participants with stroke only.

Types of interventions

One arm of the trial had to include an intervention where an active motor sequence was performed repetitively within a single training session, and where the practice was aimed towards a clear functional goal. Functional goals could involve complex whole tasks (e.g. picking up a cup), or pre‐task movements for a whole limb or limb segment such as grasp, grip, or movement in a trajectory to facilitate an ADL‐type activity (e.g. sit‐to‐stand). To be included, trials of repetitive activity were required to involve complex multi‐joint movement with functional measurement of outcome, rather than the exercise of a single joint or muscle group orientated to motor performance outcomes.

We included any intensity and duration of task training schedule but only included trials if the time duration or number of repetitions within a session of practice and the number of sessions delivered could be identified. We included trials that clearly used motor relearning as a whole therapy approach if we could identify the amount of task‐specific training received.

We included trials combining RTT with person‐delivered, mechanical or robotic movement assistance if the purpose of the assistance was to facilitate a task‐related repetition. We excluded studies if assisted movement was predominant, or could not easily be related to a functional goal.

We excluded trials if they combined RTT with another intervention where the influence of task repetition could not be isolated, for example electrical stimulation, virtual environments, forced use, bilateral movement, or mental rehearsal. We also excluded trials if the intervention used mechanical means simply to increase strength or endurance.

We contacted trial authors for clarification of the nature of the intervention if it was unclear whether the trial met our definition.

Types of outcome measures

Electronic searches

We searched the Cochrane Stroke Group Trials Register; this was searched by the Managing Editor on 4 March 2016. In addition, we searched the following electronic databases: the Cochrane Central Register of Controlled Trials (CENTRAL: the Cochrane Library 2016, Issue 5: 1 October 2006 to 24 June 2016; Appendix 1); MEDLINE (1 October 2006 to 8 March 2016; Appendix 2); Embase (1 October 2006 to 8 March 2016; Appendix 3); CINAHL (2006 to 23 June 2016; Appendix 4); AMED (2006 to 21 June 2016; Appendix 5); and SPORTSDiscus (2006 to 21 June 2016; Appendix 6). We developed the MEDLINE search strategy with the help of the Cochrane Stroke Group Information Specialist and adapted it for the other databases.

Searching other resources

We searched reference lists of relevant studies and contacted authors to identify missing data. In an effort to identify further published, unpublished and ongoing trials we searched the following resources using broad descriptors for stroke, rehabilitation, and physical therapy:

• ClinicalTrials.gov 15 June 2016 (http://clinicaltrials.gov/);

• World Health Organization (WHO) International Clinical Trials Registry Platform Search Portal 15 June 2016 (http://apps.who.int/trialsearch/).

Selection of studies

Two review authors (from JC, LC, BF, JH, NM, LT) independently screened references identified from the searches of the electronic databases and excluded irrelevant studies. We obtained the full‐text papers of the remaining studies and the same two review authors assessed these for inclusion according to the inclusion criteria. We resolved disagreements through discussion and by referral to a third review author as necessary. We provided reasons for excluding potentially relevant studies.

Data extraction and management

Two review authors (from JC, LC, NM, LT) independently conducted data extraction using a pre‐designed data extraction form for each selected study. Data extracted included citation details, method of randomisation, study population, intervention methods and delivery, reasons for losses to follow‐up, post therapy and follow‐up outcome measures, and methodological quality. In addition, we extracted information relating to treatment monitoring, acceptability, and adherence where available. We resolved disagreements by discussion, and by referral to a third author (LT) as necessary. We contacted study authors by email to request any missing information necessary for the review.

Assessment of risk of bias in included studies

Two review authors (LT and NM) used Cochrane's 'Risk of bias' tool to independently assess the methodological quality of the included studies (Higgins 2011). The tool covers the domains of sequence generation, allocation concealment, blinding of outcome assessors, incomplete outcome data and selective reporting. We classified items as 'low risk', 'high risk' or 'unclear risk' of bias. We resolved disagreements with help from a third review author (JC).

Measures of treatment effect

For continuous outcomes using similar measurement scales, we used the mean difference (MD) with 95% confidence intervals (CIs). If similar outcomes were measured using different outcome scales, we combined results using standardised mean difference (SMD) and 95% CIs. For continuous outcomes, we extracted means and standard deviations of post‐therapy scores. We also extracted means and standard deviations of change from baseline scores where available across trials. We used the Chi² test to explore differences between subgroups.

One outcome contained both dichotomous and continuous measurement units, which we analysed using the generic inverse variance method. Four different outcome measures were used in seven trials. Three of these were continuous measures: Timed Up & Go Test (Blennerhassett 2004b; Dean 2000; Salbach 2004a); Motor Assessment Scale sit‐to‐stand (Langhammer 2000; Van Vliet 2005); sit‐to‐stand (time in seconds) (Howe 2005), the exception being 'Number of people able to stand independently and safely on two consecutive occasions' (Barreca 2004). For the six trials with continuous outcomes, we calculated the SMD and corresponding standard error in Review Manager 5 (RevMan 2014) from the SMD estimate and CI and re‐entered for the GIV‐based meta‐analysis of sit‐to stand. For Barreca 2004, we converted the log OR and its standard error (SE) to an approximate SMD scale.

Unit of analysis issues

Dealing with missing data

If data were not in a form suitable for quantitative pooling, we contacted trial authors for additional information .We attempted to obtain post therapy scores from trial authors who had reported median and inter‐quartile ranges. We presented trials reporting change scores with standard deviations in separate analyses.

Assessment of heterogeneity

We assessed the degree of heterogeneity among the trials using the I² statistic for each outcome. If less than or equal to 50%, we used a fixed‐effect meta‐analysis. If the I² statistic was greater than 50%, we explored the individual trial characteristics to identify potential sources of heterogeneity. We then performed meta‐analysis using both fixed‐effect and random‐effects modelling to assess sensitivity to the choice of modelling approach.

We addressed clinical and methodological diversity by incorporating subgroup or sensitivity analyses for type of participant (time from stroke), intervention (type and amount of intervention), and study design (comparison group, equivalence of treatment).

To test for subgroup effects we used the Chi² test with a 5% significance level.

Assessment of reporting biases

We searched clinical trial registers to assist in reducing publication bias. We also investigated selective outcome reporting through the comparison of the methods section of papers with the results reported.

Data synthesis

Where there were acceptable levels of heterogeneity, we pooled results. We used both random‐effects and fixed‐effect meta‐analysis with 95% CI using Review Manager 5 (RevMan 2014). We pooled outcomes measured with different instruments using the SMD.

We documented the quality of evidence for each outcome based on criteria considered within the GRADE (Grading of Recommendations Assessment, Development and Evaluation) approach (Guyatt 2008); this includes the following.

• Risk of bias due to flawed design or conduct of studies (sequence generation, allocation concealment, blinding of outcome assessors and incomplete outcome data). We re‐assessed all studies from the original review using the updated 'Risk of bias' tool (Higgins 2011).

• Imprecision (e.g. when confidence intervals for treatment effect are wide).

• Inconsistency (e.g. when point estimates vary widely, the I² is large).

• Indirectness (e.g. variations in participants, interventions, comparisons and outcomes).

• Publication bias (may be explored with the use of funnel plots and classed as not suspected, suspected, strongly suspected or very strongly suspected).

Three review authors (JC, NM and LT) assessed and documented risk of bias related to study design, imprecision, inconsistency, indirectness and publication bias for each outcome within comparisons presented.

We employed GRADE to interpret findings and to create a 'Summary of findings' table (Guyatt 2008) for the following outcomes: arm function, hand function, walking distance, walking speed, functional ambulation, lower limb functional measures and global motor function. The table provides outcome‐specific information concerning the overall quality of evidence from studies included in the comparison, the magnitude of effect of the intervention and the sum of available data on the outcomes considered. We downgraded the evidence from 'high quality' by one level for serious (or by two for very serious) study limitations (risk of bias, indirectness of evidence, serious inconsistency, imprecision of effect estimates or potential publication bias). We did not assess follow‐up outcomes using GRADE.

Subgroup analysis and investigation of heterogeneity

We undertook planned subgroup analyses for all primary outcomes separately for upper limb and lower limb function, due to the potential differential impact (Table 1). Planned subgroup analyses were as follows:

• dosage of task practice: dosage of task practice was calculated by multiplying the number of weeks, by the number of sessions per week, by the session duration in hours. Trials were divided into those providing up to and including 20 hours training, and those providing more than 20 hours training in total;

• time since stroke: mean time since stroke at recruitment was used to classify trials as within zero to six months post stroke or more than six months post stroke. As a number of trials recruited very early post stroke, a post‐hoc analysis grouping was included for trials recruiting within 14 days of stroke;

• type of intervention: trials were classified as either 1) whole therapy approaches, where rehabilitation in total was directed by a motor relearning or movement science approach, 2) mixed functional task training, where therapy included a mixed combination of functional tasks, and 3) single task training, where one task was practiced repeatedly.

We intended to consider if effect sizes were related to whether training was based on pre‐functional versus functional activities, or pre‐intervention level of disability. In the event, we excluded most pre‐functional trials because they contained a large proportion of passive or active‐assisted movement, and levels of disability proved too difficult to classify because of mixed groups of participants and unsuitable measures and data for this purpose. Therefore, we have not presented these planned subgroup analyses.

We prioritised outcomes for subgroup analyses by the study authors' primary outcome choice, or the review authors' judgement as to the most suitable measure for the intervention, for example a balance measure for trials training balance functions. If more than one measure was available, we prioritised lower limb outcomes in the following order: 1) walking speed, 2) walking distance, 3) functional ambulation, and 4) lower limb functional measures. We prioritised upper limb outcomes as 1) arm function, and 2) hand function. We omitted one trial from the subgroup and sensitivity analyses because it used a dichotomous outcome (Barreca 2004).

Sensitivity analysis

We carried out planned sensitivity analyses for allocation concealment (adequate or inadequate/unclear). In addition, we included post hoc sensitivity analyses to consider the impact of different comparison groups (attention control, usual care), equivalence of therapy time (equivalent time, additional time), and intervention delivery (individual versus group). We did not undertake planned sensitivity analyses for intervention setting (hospital versus home) because of an insufficient numbers of trials.