Population

We will include studies that have evaluated the effectiveness of an intervention in adults (18 years and over) living with dementia or their carers.

We will include studies in all stages (including early- to late-stage dementia) and different forms of dementia (eg, Alzheimer’s disease, vascular dementia, Lewy bodies dementia and mixed dementias, as well as less common forms such as fronto-temporal dementia). We will also take into consideration risks of dementia associated with conditions such as HIV/AIDS that may be more prevalent in LMICs to ensure these are captured in our search strategy. Due to the particularly high risk of people with mild cognitive impairment (MCI) developing dementia (approximately one-third will develop dementia over a period of 3 to 10 years),17 we will also include studies focusing on this group.

Carers of people living with dementia include family members, other unpaid carers, as well as professional carers, irrespective of whether or not they are paid, and whether or not they are living with the person they care for.

Studies will be included if they were conducted in a country considered a LMIC according to the Organisation for Economic Co-operation and Development (OECD) categorisation at any time during the study period. The lists of aid recipient countries for the years 2008 to 2018 were used to identify LMICs (http://www.oecd.org/development/financing-sustainable-development/development-finance-standards/historyofdaclistsofaidrecipientcountries.htm).

We will include studies that were conducted entirely in LMICs, or (in the case of multi-country studies) where 50% or more of study participants received the intervention of interest in LMICs or where results were presented separately for participants in LMICs.

Intervention

One of the aims of this systematic review is to identify which interventions have been rigorously evaluated in LMICs. The eligibility criteria for types of interventions are therefore deliberately kept open: we will include all interventions that aim to improve the lives of people living with dementia and their carers contingent on the intervention having been subjected to a rigorous evaluation (see eligible study designs below). We will apply an ‘effectiveness’ perspective with respect to interventions, that is, interventions are eligible when defined as such in a study and improved outcomes are expected through the intervention.18 We will focus on people already living with dementia or MCI and their carers and will therefore exclude primary prevention studies (ie, prevention of dementia in people without cognitive impairment).

Dementia interventions can vary in terms of who they are targeting (person living with dementia, unpaid carers, care professionals and care systems), what the aim of the intervention is (secondary prevention, treatment, disease management, coping with caring for people living with dementia, managing the impact of dementia on the care system) and their mechanism (including diagnosis; pharmacological treatment; cognitive therapy; technological interventions; training; exercise; sleep therapy; music therapy; organisation of care, including advance care directives and case management; support for carers; financing of care; policy interventions and others).4

Comparison

All comparisons will be eligible, including active comparators, usual practice/standard care/placebo and no action.

Outcomes

Outcomes of dementia intervention studies are highly heterogeneous and may vary by who is affected (person living with dementia, their carer(s), wider society and care system) and type of outcome (clinical outcomes, such as cognitive, neurological, psychological, psychosocial; quality of life and functioning outcomes; care and delivery of care, such as the use of feeding tubes, hospitalisations, institutionalisations; economic outcomes; diagnosis rates; knowledge of the disease and ability to cope with caring for people with dementia). For example, recommendations on dementia interventions were recently made based on a review of the evidence on outcomes in domains ranging from, among others, cognition and neuropsychiatric outcomes to end-of-life care and care delivery.4

Given the broad range of potentially relevant outcomes, a search strategy that captures all of these is not feasible. Restricting inclusion criteria to certain types of outcomes would risk excluding potentially important parts of the evidence base for dementia interventions in LMICs. The impact of dementia is recognised to be experienced by different people and at different levels, including by people living with dementia, their carers, communities and the care system and wider society,19 and interventions therefore need to take all of these perspectives into account.5

Study design

The research question we aim to answer is what works in dementia in LMICs. We are therefore interested in the effect that an intervention had on outcomes in people receiving the intervention, their carers and the wider care system.20 Accordingly, we will apply a causal inference framework in this review: study designs eligible for inclusion are those that support a causal link between the intervention and observed outcomes. This includes experimental designs (RCTs and cluster-RCTs) as well as non-randomised designs suitable for causal inference, defined as ‘comparisons of ‘potential outcomes’ (…) under different treatment conditions on a common set of units’.8 We will include both randomised and non-randomised studies due to an anticipated low number of well-conducted randomised trials in LMICs, based on previous work (searches conducted for the MODEM Evidence Toolkit, and a previous review on packages of care for people living with dementia in LMICs15). Also, non-randomised studies may have greater generalisability beyond the studied environment when conducted in settings and populations closer resembling routine practice.

In order to maximise the number of eligible studies and synthesise as much robust evidence about dementia interventions in LMICs as possible, we will extend the eligibility criteria used by the Cochrane Effective Practice and Organisation of Care (EPOC) Group21 to also include other quasi-experimental study designs in addition to those typically seen in methodologically robust evaluation studies (non-randomised controlled trials, controlled before-after studies and interrupted time series studies). The label ‘quasi-experimental’ is discipline-specific and members of the Cochrane Non-Randomised Studies for Interventions Methods Group therefore developed a ‘label-free’ taxonomy of quasi-experimental studies that aims to define such studies through a series of questions in six domains.22 We will use their checklist to guide the selection of studies by study design (see table 1). Since we anticipate a shortage of methodologically robust studies of dementia interventions in LMICs, but still aim to synthesise the best available evidence that supports causal inference of treatment effects, we adapted the original checklist by removing restrictions to answer ‘yes’ to only one question in domains 2, 3 and 4. This will enable us to include several strong experimental and quasi-experimental study designs, including RCTs (including cluster-RCTs), non-randomised controlled trials, controlled before-after studies, interrupted time series studies, difference-in-differences studies, instrumental variables studies, regression discontinuity design studies and other study designs that employ methods such as propensity score matching that attempt to control for observed and unobserved confounders.

As a contingency plan in case no such studies can be identified in LMICs, we define a secondary set of inclusion criteria to still capture intervention studies in LMICs with less robust designs but that could still inform our understanding of what works in dementia in LMICs in the absence of better study designs. The secondary set of inclusion criteria will mirror the primary inclusion criteria, with the exception of study design. In the secondary set of inclusion criteria, studies will also be eligible for inclusion if they control only for observed confounders (criterion 4 in table 1).

Publication type

We will include peer-reviewed journal articles, including primary publications of intervention studies and systematic reviews of these, and grey literature describing evaluations of dementia interventions in LMICs (eg, PhD theses and reports published by care system administrative bodies and non-governmental organisations). Systematic reviews are only eligible for inclusion if their focus is on LMICs (ie, if LMICs were specified as geographical setting in their inclusion criteria) and we will use these to identify primary studies eligible for inclusion (eligible if 50% or more participants were in LMICs or results were reported separately for LMIC settings). We will also include conference abstracts, provided they contain information to assess eligibility for inclusion.

Exclusion criteria

Exclusion criteria are listed in box 1.

Search terms

Search terms were selected to reflect the key concepts making up the review question: dementia interventions that have been studied using robust research designs in LMICs. Exemplary search terms for MEDLINE via OVID for the concepts of population and intervention, study design and LMICs are provided in table 2, including the number of hits for each search as per 10 October 2018. The search for MEDLINE will be translated for the other databases.

The ‘intervention’ concept provided in table 2 (lines 22 to 98) includes both generic terms as well as a range of specific interventions that have been studied for dementia before, although usually in high-income countries. We based the list of specific interventions on a thorough review of the evidence, conducted as part of the Lancet commission,4 and through revision by the international STRiDE consortium. While this might not be a comprehensive list of all interventions that have ever been studied for dementia, we aimed to increase the likelihood of identifying relevant studies by combining generic search terms pertaining to interventions, therapies, etc with specific intervention types and names of interventions, such as drug names.

We will use established study design and geographical location search filters to narrow down the number of results and will make adaptations as necessary to reflect our inclusion criteria. If such filters are not available for some databases, we will consider adapting the search filters used for MEDLINE.

The search filters for eligible study designs are provided in table 2, rows 101 to 108 (for RCTs), 110 to 121 (for cluster-RCTs) and 123 to 147 (for other quasi-experimental studies), respectively. In order to identify RCTs, we will use the sensitivity-maximising version of the highly sensitive filter for RCTs recommended in the Cochrane Handbook.23 For identification of cluster-RCTs, we will use the sensitivity-maximising filter for cluster-RCTs developed by Taljaard et al.24 This filter was validated, among others, for a sample of cluster-RCTs in sub-Saharan Africa from before 2001 and showed good sensitivity (94%), with overall higher sensitivity for more recent studies. For identification of non-randomised studies that support causal inference of the relationship between exposure and outcome, we created an additional set of search terms pertaining to commonly used labels for quasi-experimental studies. We enhanced these search terms with the ‘aetiology’ filter developed by the Hedges Group.25 The methodological criteria used for developing the filter were found to be a good match for our inclusion criteria (including the use of a clearly identified comparison group, with examples given by the authors including ‘randomised controlled trials (RCT), quasi-randomised controlled trials, non-randomised controlled trials, cohort studies with case-by-case matching or statistical adjustment to create comparable groups or nested case–control studies’25).

Finally, rows 149 to 157 of table 2 show the exemplary search terms for LMICs, based on a filter for LMICs that is available from the Cochrane EPOC website (http://epoc.cochrane.org/lmic-filters). The search terms were expanded and updated to reflect the list of LMICs from 2008 to 2018.

Filters for LMICs and study designs were tested in MEDLINE for a sample of 10 studies (six selected for being conducted in LMICs,26–31 four selected for using quasi-experimental study designs32–35). The sensitivity of the search strategy for this sample was 100% (10/10 studies) for searches with study design filters, and 100% (6/6 studies) when adding the LMIC filter.

Quantitative synthesis

We will explore the feasibility of conducting quantitative synthesis of treatment effects through traditional pairwise and network meta-analysis. Feasibility of quantitative synthesis will be assessed for each intervention studied in our sample of included studies. We will assess the similarity of the specific intervention in each study with other candidate interventions for a meta-analysis, the participants in the studies where this intervention was evaluated and whether the same outcome was used. Quantitative synthesis can only be conducted for studies reporting the same outcome, although individual studies may use different instruments to measure this (eg, the Mini-Mental State Examination (MMSE) and Alzheimer’s Disease Assessment Scale–cognitive subscale (ADAS-cog) for cognition, and the Texas Functional Living Scale and Barthel scale for activities of daily living).

We will assess the feasibility to conduct both traditional pairwise and network meta-analysis. Pooled estimates are calculated as weighted averages of the treatment effects in included studies, where weights are assigned to each study based on its precision.43 Network meta-analysis extends pairwise meta-analysis by incorporating both direct evidence about the relative effectiveness of interventions that have been compared with each other in a primary study and indirect evidence about the relative effectiveness of interventions that were never directly compared with each other, but are connected through a network of other interventions for which direct comparisons exist. In addition, network meta-analysis can be a valuable tool in delineating the effects of individual components of complex interventions, as recently shown in a network meta-analysis of multi-component interventions for dementia caregiver depression.44 The validity of treatment effect estimates obtained through network meta-analysis depends on the similarity of the various direct comparisons in the network with respect to relevant treatment effect modifiers (eg, severity of disease, age of participants, risk of bias due to study design and implementation) and consistency of treatment effects obtained from direct and indirect comparisons.45

It is anticipated that in our sample of included studies, there will be underlying differences between individual studies of the same intervention, in particular with respect to the details of the intervention and the setting. For example, a training intervention in one study is unlikely to be exactly the same intervention as in another study, even if a protocol is used to standardise the intervention, because, for example, the individual delivering the training changes. Such differences do not preclude pooling study results, but the heterogeneity in treatment effects resulting from underlying differences needs to be taken into account when assigning weights to studies for a pooled estimate. Our default will therefore be to use a random-effects model, which takes into account between-study variation and assumes that included studies come from different populations, with unique details of the intervention and conditions of the study. The true treatment effect can therefore vary from study to study.46

Subgroup analysis

While a random-effects model takes into account between-study variation, it does not explain heterogeneity. We will aim to use subgroup analysis and meta-regression to identify any study-level characteristics that might explain variation in treatment effects. Potential candidate covariates for subgroup analysis and meta-regression are study design and quality (within RCTs: risk of bias; within non-RCTs: analytical method employed, for example, instrumental variable, risk of bias), intervention details (eg, drug dose, intensity of training, intensity of stimulation therapy) and setting (by country, by bracket of gross national income, by rural vs urban setting). However, subgroup analysis will only produce meaningful insights when enough studies are included in each subgroup to detect any difference by the selected covariate (as opposed to a chance finding) and meta-regression should not be conducted for samples of less than 10 studies.47 Given that we do not expect to identify a large number of robust dementia intervention studies in LMICs, it is unlikely that there will be sufficient studies of a given intervention with a specific outcome to allow meaningful subgroup analysis and meta-regression.

In cases where interventions were evaluated both by RCTs and non-randomised studies, we will obtain separate pooled effects for the two study types.18 We will extract information on study design characteristics that will allow us to examine these as potential sources of heterogeneity.38

Exploring meta-bias

For each intervention-outcome pair, we will first assess in a funnel plot whether asymmetry exists with respect to expected random variation of treatment effects around the pooled effect estimate with decreasing study precision. Funnel plots are used to detect possibly ‘missing’ smaller studies with larger or smaller treatment effects than what would be expected by chance alone, for which publication bias is one possible explanation.48 We will then use contour-enhanced funnel plots to assess whether any observed asymmetry is likely due to publication bias favouring the publication of statistically significant results.49

We anticipate identifying only a small number of studies eligible for inclusion in our quantitative review (meeting the criteria of being conducted in a LMIC and using a robust study design). In a scenario where not more than one LMIC study exists for any combination of intervention and outcome, pairwise meta-analysis will become impossible. For network meta-analysis, the existence of only a handful of LMIC studies, even if a common comparator (eg, standard care) exists, makes it difficult to assess whether the assumptions of similarity (with respect to treatment effect modifiers) and consistency (agreement between direct and indirect estimates of treatment effect) hold. In case of a lack of eligible studies meeting our primary inclusion criteria, we will use our secondary set of inclusion criteria for studies using less robust research designs and will explore the feasibility of quantitative synthesis of these.

Given the reliance of past efforts to develop recommendations for dementia interventions in LMICs on evidence from high-income countries,7 15 our primary aim is to advance our understanding of what works in LMICs and provide policymakers, people living with dementia and their carers, health and care professionals and others with evidence that is of immediate relevance to their setting. However, should quantitative synthesis of LMIC studies not be feasible due to a lack of studies reporting the same outcome, we will pursue an alternative way of obtaining pooled estimates of the comparative effectiveness of dementia interventions relevant for LMICs, as outlined below. In a scenario where we are unable to identify enough primary studies of dementia interventions in LMICs, we will quantitatively synthesise available evidence from methodologically robust studies from high-income countries, as identified in the MODEM Toolkit (www.modem-dementia.org.uk). We will discuss existing interventions with dementia experts from the STRiDE partner countries to identify those that have the highest relevance for LMICs. We will then extract relevant information from studies meeting our study design inclusion criteria in the MODEM database and obtain pooled estimates for the effectiveness and comparative effectiveness of these interventions. We will develop a separate protocol for this approach should it become necessary (ie, in the case of not identifying eligible studies in LMICs).