Background

Memantine is a moderate affinity uncompetitive antagonist of glutamate NMDA receptors. It is licensed for use in moderate and severe Alzheimer's disease (AD); in the USA, it is also widely used off‐label for mild AD.

Objectives

To determine efficacy and safety of memantine for people with dementia. To assess whether memantine adds benefit for people already taking cholinesterase inhibitors (ChEIs).

Search methods

We searchedALOIS, the Cochrane Dementia and Cognitive Improvement Group's register of trials (http://www.medicine.ox.ac.uk/alois/) up to 25 March 2018. We examined clinical trials registries, press releases and posters of memantine manufacturers; and the web sites of the FDA, EMEA and NICE. We contacted authors and companies for missing information.

Selection criteria

Double‐blind, parallel group, placebo‐controlled, randomised trials of memantine in people with dementia.

Data collection and analysis

We pooled and analysed data from four clinical domains across different aetiologies and severities of dementia and for AD with agitation. We assessed the impact of study duration, severity and concomitant use of ChEIs. Consequently, we restricted analyses to the licensed dose (20 mg/day or 28 mg extended release) and data at six to seven months duration of follow‐up, and analysed separately results for mild and moderate‐to‐severe AD.

We transformed results for efficacy outcomes into the difference in points on particular outcome scales.

Main results

Across all types of dementia, data were available from almost 10,000 participants in 44 included trials, most of which were at low or unclear risk of bias. For nearly half the studies, relevant data were obtained from unpublished sources. The majority of trials (29 in 7885 participants) were conducted in people with AD.

1. Moderate‐to‐severe AD (with or without concomitant ChEIs). High‐certainty evidence from up to 14 studies in around 3700 participants consistently shows a small clinical benefit for memantine versus placebo: clinical global rating (CGR): 0.21 CIBIC+ points (95% confidence interval (CI) 0.14 to 0.30); cognitive function (CF): 3.11 Severe Impairment Battery (SIB) points (95% CI 2.42 to 3.92); performance on activities of daily living (ADL): 1.09 ADL19 points (95% CI 0.62 to 1.64); and behaviour and mood (BM): 1.84 Neuropsychiatric Inventory (NPI) points (95% CI 1.05 to 2.76). There may be no difference in the number of people discontinuing memantine compared to placebo: risk ratio (RR) 0.93 (95% CI 0.83 to 1.04) corresponding to 13 fewer people per 1000 (95% CI 31 fewer to 7 more). Although there is moderate‐certainty evidence that fewer people taking memantine experience agitation as an adverse event: RR 0.81 (95% CI 0.66 to 0.99) (25 fewer people per 1000, 95% CI 1 to 44 fewer), there is also moderate‐certainty evidence, from three additional studies, suggesting that memantine is not beneficial as atreatment for agitation (e.g. Cohen Mansfield Agitation Inventory: clinical benefit of 0.50 CMAI points, 95% CI ‐3.71 to 4.71) .

The presence of concomitant ChEI does not impact on the difference between memantine and placebo, with the possible exceptions of the BM outcome (larger effect in people taking ChEIs) and the CF outcome (smaller effect).

2. Mild AD (Mini Mental State Examination (MMSE) 20 to 23): mainly moderate‐certainty evidence based on post‐hoc subgroups from up to four studies in around 600 participants suggests there is probably no difference between memantine and placebo for CF: 0.21 ADAS‐Cog points (95% CI ‐0.95 to 1.38); performance on ADL: ‐0.07 ADL 23 points (95% CI ‐1.80 to 1.66); and BM: ‐0.29 NPI points (95% CI ‐2.16 to 1.58). There is less certainty in the CGR evidence, which also suggests there may be no difference: 0.09 CIBIC+ points (95% CI ‐0.12 to 0.30). Memantine (compared with placebo) may increase the numbers of people discontinuing treatment because of adverse events (RR 2.12, 95% CI 1.03 to 4.39).

3. Mild‐to‐moderate vascular dementia. Moderate‐ and low‐certainty evidence from two studies in around 750 participants indicates there is probably a small clinical benefit for CF: 2.15 ADAS‐Cog points (95% CI 1.05 to 3.25); there may be a small clinical benefit for BM: 0.47 NOSGER disturbing behaviour points (95% CI 0.07 to 0.87); there is probably no difference in CGR: 0.03 CIBIC+ points (95% CI ‐0.28 to 0.34); and there may be no difference in ADL: 0.11 NOSGER II self‐care subscale points (95% CI ‐0.35 to 0.54) or in the numbers of people discontinuing treatment: RR 1.05 (95% CI 0.83 to 1.34).

There is limited, mainly low‐ or very low‐certainty efficacy evidence for other types of dementia (Parkinson's disease and dementia Lewy bodies (for which CGR may show a small clinical benefit; four studies in 319 people); frontotemporal dementia (two studies in 133 people); and AIDS‐related Dementia Complex (one study in 140 people)).

There is high‐certainty evidence showing no difference between memantine and placebo in the proportion experiencing at least one adverse event: RR 1.03 (95% CI 1.00 to 1.06); the RR does not differ between aetiologies or severities of dementia. Combining available data from all trials, there is moderate‐certainty evidence that memantine is 1.6 times more likely than placebo to result in dizziness (6.1% versus 3.9%), low‐certainty evidence of a 1.3‐fold increased risk of headache (5.5% versus 4.3%), but high‐certainty evidence of no difference in falls.

Authors' conclusions

We found important differences in the efficacy of memantine in mild AD compared to that in moderate‐to‐severe AD. There is a small clinical benefit of memantine in people with moderate‐to‐severe AD, which occurs irrespective of whether they are also taking a ChEI, but no benefit in people with mild AD.

Clinical heterogeneity in AD makes it unlikely that any single drug will have a large effect size, and means that the optimal drug treatment may involve multiple drugs, each having an effect size that may be less than the minimum clinically important difference.

A definitive long‐duration trial in mild AD is needed to establish whether starting memantine earlier would be beneficial over the long term and safe: at present the evidence is against this, despite it being common practice. A long‐duration trial in moderate‐to‐severe AD is needed to establish whether the benefit persists beyond six months.

Description of the condition

This review covers the effect of memantine in dementia of all aetiologies.

Alzheimer's disease (AD) is the commonest cause of dementia, and is found in approximately 70% of autopsies of people with dementia. The prevalence of the disease is approximately 1% to 2% at the age of 65, but doubles every five years to at least the age of 90 (Qiu 2009). The disease is progressive. For the purposes of drug trials, people with AD who have a score on the brief mini mental state examination (MMSE) of more than 20 have come to be labelled as having 'mild' AD. However, the combination of loss of memory, disorientation and frequent loss of insight which accompanies this stage means that the impact on caregivers is often very far from 'mild'. It is not uncommon for patients to need admission to care homes in this stage. 'Moderate' AD has come to be defined as those with an MMSE score of 20 to 10. In this range, patients decline more rapidly. The impairment in patients' ability to manage everyday tasks becomes marked and is obvious during even brief conversation. By the time patients have progressed to an MMSE score of 10 or less ('severe' dementia), the deficits are profound and 24‐hour supervision is required. Approximately 70% of people with AD require admission to care homes.

Aside from age, the biggest risk factor for developing AD is possession of the ApoE4 gene, which is present in 17% to 30% of the population. Other risk factors include all vascular risks (diabetes, hypertension, high cholesterol, lack of exercise), any cerebral injury (trauma or stroke), and being female (Patterson 2007;Sibbett 2017).

Cellular and animal models, genetic, neuroimaging, clinical and postmortem brain studies have all been important in advancing understanding of the many changes which occur in the brains of people with AD. Nerve cell loss and disruption of neurotransmitter systems becomes widespread throughout many brain areas, particularly the hippocampus and cerebral cortex. Aggregation of peptide fragments of the amyloid precursor protein, which probably holds nerve cells close enough together for signals to be transmitted between them, causes malfunction of processes within cells. Large deposits of these fragments, amyloid plaques, develop outside neurones and are associated with a mild inflammatory response. However, removal of existing plaques does not appear to result in improvement in symptoms so it is unclear to what extent the amyloid plaques are a cause of AD or a consequence of some other, more fundamental, disturbance. Within the neurons, the transport of cellular components becomes disrupted because tau, which helps to keep the microtubule scaffolding together, becomes hyperphosphorylated and itself forms into paired helical filament 'tangles'. No drugs affecting this process have been proven to benefit symptoms. Cholinergic function, which mediates attention, tends to be impaired in people with AD. This can be partially corrected by cholinesterase inhibitors (ChEIs) which reduce the breakdown of acetylcholine.

Vascular dementia, in which cognitive decline is attributed to some form of vascular injury, typically ischaemic, is the second most common cause of dementia in Western societies. Developing a valid definition, distinct from AD, has been problematic. It is a heterogeneous condition and clinical manifestations differ depending on the size and location of the cerebrovascular lesions. In autopsy studies, 'mixed' AD and vascular dementia has been reported as accounting for between 0% and 55% of cases of dementia. In addition to co‐occurrence due simply to chance, AD and vascular dementia may have aetiological or pathogenetic factors in common (Kalaria 1999). In comparison with sufferers from AD, people with mixed dementia show higher frequencies of depressed mood, focal motor or sensory findings and gait disturbances, but the neuropsychological pattern is not distinctive. Using criteria which demand imaging evidence of discrete vascular lesions and a clinical event associated with cognitive decline, or evidence of marked small vessel disease, vascular dementia affects 1% to 20% of people aged 65 years or older.

Parkinson's disease dementia (PDD) and dementia with Lewy bodies (DLB) are closely related. The main clinical distinction is one of definition: those with PDD have had parkinsonism a year or more before developing dementia, whereas in DLB the onsets of dementia and any parkinsonism are closer in time. Some people with DLB show no signs of parkinsonism. Both groups are more likely to experience visual hallucinations, marked fluctuations in functional ability, and REM (rapid eye movement) sleep behaviour disorder.

Description of the intervention

Memantine was first synthesised at Eli Lilly as an agent to lower elevated blood sugar levels, but was ineffective. It has since been tested in over 100 randomised controlled trials in a wide variety of neurological and psychiatric conditions, including dementia of different sorts, depression, neuropathic pain, Parkinson's disease and autism.

In 1972, Merz applied for a German patent for memantine as a potential treatment for a wide range of cerebrovascular conditions, citing evidence of reduced degeneration and nerve cell loss following experimentally induced ischaemia in animal models. In 1975 and 1978, patents were granted in Germany and the USA, respectively (Parsons 1999). This basis for the original patent for memantine (Bormann 1991), which was due to expire in April 2010, was contested by manufacturers of generics (Forest 2007). Forest and Merz settled an agreement to provide licenses to each of Amneal, Cobalt, Dr. Reddy’s, Lupin, Orchid, Sun, Teva, Upsher−Smith, and Wockhardt that permitted these companies to launch their generic versions of 'Namenda' within three months prior to the expiration of the original patent (Forest 2010a). However, in March 2009, the U.S. Patent and Trademark Office issued a Notice of Final Determination that, after review of the regulatory timeline for approval, Namenda was entitled to a patent term extension until April 2015 and the patent finally expired in October 2015. Generic memantine is now available, but in 2010 the US Food and Drug Administration (FDA) awarded a license for an extended release, once daily preparation Namenda XR 28 mg (Forest 2010b). The patent for this expires in September 2029.

Memantine was approved in February 2002 by the European Agency for the Evaluation of Medical Products (EMEA) for the treatment of "moderately severe to severe Alzheimer's disease" (EMEA 2004) and in 2003 by the FDA for the treatment of moderate‐to‐severe AD (MMSE up to 14) (Anonymous 2003;Forest 2003). In 2006, the EMEA expanded its indication to 'moderate‐to‐severe AD' (MMSE up to 19) (EMEA 2006). In June 2008, the EMEA granted a license for a once daily 20 mg dosing schedule (EMEA 2008). Applications to the FDA and EMEA for licenses for the treatment of mild‐to‐moderate AD have been unsuccessful (Forest 2005b;Lundbeck 2005).

Memantine has not been approved for vascular dementia or earlier stages of AD in any jurisdiction.

Memantine is marketed as Axura by Merz, as Ebixa by Lundbeck in Europe, as Namenda by Forest in North America and as Mamary by Daiichi Asubio in Japan. In 2010, memantine had 34.8% share of the US market for drugs for AD (Forest 2010a). Annual global sales exceed 1 billion USD. In 2015, the UK price of memantine to the NHS decreased by 94%. Prescribing rates in England increased from approximately 100,000 items dispensed in 2011 to 784,000 (at a cost of £5.4m) in 2015 (Prescriptions England 2016).

In May 2015, Actavis launched Namzaric (a fixed‐dose combination of 28 mg extended‐release memantine and the ChEI, donepezil 10 mg) for people with moderate‐to‐severe AD, following FDA approval in December 2014.

How the intervention might work

Memantine is a low affinity antagonist of the N‐methyl‐D‐aspartate (NMDA) glutamate receptor. L‐glutamate is the main excitatory neurotransmitter in the central nervous system and is implicated in neural transmission, learning, memory processes and neuronal plasticity (Sucher 1996). Physiological glutamate activity is required for normal brain activity and so cannot be abolished completely (Kornhuber 1997). The hippocampus and other brain regions which are affected in AD, are rich in glutamate receptors of the NMDA class. Consolidation of new memories is also mediated through these receptors.

Excessive glutamate‐induced excitation, which results in excessive flow of calcium into neurons through NMDA receptors, plays a role in the pathogenesis of AD and in the damage due to an ischaemic stroke (Cacabelos 1999). The clinical actions of memantine may be mediated by preventing this excitotoxicity.

NMDA receptor‐induced responses may depend on the receptor location. Stimulation of synaptic NMDA receptors, acting primarily through nuclear Ca(2+) signalling, leads to the build‐up of a neuroprotective 'shield', whereas stimulation of extrasynaptic NMDA receptors promotes cell death (Hardingham 2010) and increases amyloid production (Bordji 2010). The differing pharmacodynamics of memantine at synaptic and extrasynaptic NMDA receptors may mean that there is a 'therapeutic window' for memantine at doses where inhibition of extrasynaptic receptors is greater than for synaptic receptors (Hardingham 2010). Inhibition of both subtypes occurs at higher doses.

It is possible that the effect of memantine is related to reduction of tau phosphorylation (Degerman Gunnarsson 2007) or of amyloid toxicity (Song 2008).

Memantine also preserves cerebral energy status during experimentally induced hypoglycaemia in healthy people (Willenborg 2011). Although this effect could be related to its actions in dementia, this has yet to be established.

Why it is important to do this review

First, since the last update of our review in 2006, the UK National Institute for Health and Care Excellence (NICE) has, on the basis of its 2011 appraisal (TA 217) revised an earlier recommendation that memantine was insufficiently cost‐effective to warrant prescription on the UK's National Health Service (NHS), and produced new recommendations outlining its use (NICE 2011). In a further update in 2015, NICE stated that they "identified nothing new that affects recommendations 1.1, 1.2" (those concerning memantine). Concurrent with the update of our Cochrane Review, NICE's clinical guideline on dementia (CG 42) has been partially updated, to review further the effect of two topics relevant to this memantine review: concurrent treatment with memantine and cholinesterase inhibitors (ChEIs), and memantine for non‐AD dementias. The guideline has not updated recommendations 1.1 and 1.2 (NICE 2018). The Cochrane Dementia and Cognitive Improvement review group is a stakeholder for this guideline and the authors of this review have shared data and evidence synthesis with NICE.

The UK recommendations regarding memantine monotherapy (TA 217 recommendations 1.1 and 1.2) therefore remain unchanged: that is, that memantine should be available at public expense for people who either (a) have moderate ADand are intolerant of, or have a contraindication to ChEIs or (b) have severe AD. The Appraisal Committee also concluded that "if cognitive scales are not appropriate for assessing the need for treatment, or whether to continue treatment, then clinicians should use another appropriate method of assessment".

NICE's 2011 updated decision was based on consideration of data from four studies, with emphasis on statistical significance. Their analysis showed a significant benefit on cognitive function assessed by the Severe Impairment Battery (SIB) at three months, but not at six months. Two studies provided data for assessment of activities of daily living (ADL) using the Alzheimer's Disease Cooperative Study ‐ Activities of Daily Living (ADCS‐ADL)19 scale, showing marginal significance at six months. No statistically significant effect was seen on behaviour. NICE's model suggested that memantine delayed time to institutional care by 0.8 months. Thus, NICE's 2015 decision to transfer recommendations 1.1 and 1.2 to their 'static list of technology appraisals' was taken on the basis of relatively few studies for clinical effectiveness. Memantine is now off‐patent so cost‐effectiveness will be affected.

NICE's original 2011 guidance also concluded that there was a "lack of evidence of additional clinical efficacy (of concurrent therapy with ChEIs) compared with memantine monotherapy" (NICE 2011). However, this aspect of the TA has now been updated in the newly published 2018 guideline (NICE 2018), which recommends that for people with an established diagnosis of AD who are already taking a ChEI, memantine should be additionally considered in moderate disease and additionally offered in severe disease (NICE 2018). A lack of clarity still remains because of the decision not to update the monotherapy recommendations, and there is potential confusion arising from conflating the old and new recommendations. The British Association for Psychopharmacology guideline suggests that there is "type I evidence [ie based on meta‐analysis of RCTs] for adding memantine to a cholinesterase inhibitor" (O'Brien 2017).

Second, the German Institute for Quality and Efficiency in Healthcare (IQWIG) has revised its conclusions on memantine. In 2009, they concluded that "there is no scientific proof that patients with moderate or severe Alzheimer's disease benefit from drugs containing the agent memantine". Their last search of registries for studies was in January 2009. Data from nine studies of 16 to 28 weeks duration were eventually included. IQWIG pointed to a lack of 'reliable responder analysis', which prevents an understanding of whether 'more patients in the memantine group notice a perceptible improvement in their symptoms than in the placebo group' and argued that it was "not possible to deduce a proof of a relevant effect" on clinical global because, although showing a significant benefit (standardised mean difference (SMD) 0.18, 95% confidence interval (CI) 0.05 to 0.3), the lower confidence limit fell below a threshold of 0.2. The majority of included studies collected data on the amount of care required ('resource utilisation') but this was not made available. However, in response, in 2011 Merz submitted post‐hoc findings on responder analyses which led to changes to IQWIG's conclusions (IQWIG 2011), "The data provide proof of a benefit of memantine in patients with Alzheimer’s disease with regard to the prevention of a relevant deterioration in cognitive function. For activities of daily living, taking into account the uncertain response criteria, as well as the concurrent minor size of the effect, the data provide an indication of a benefit of memantine."

Third, France's Minister for Health de‐listed all drugs routinely used for AD from August 2018 (memantine and the ChEIs, donepezil, galantamine and rivastigmine), stating, "The drugs available in early 2018 for Alzheimer's disease have only minimal and transient efficacy. They are also difficult to use because of their disproportionate adverse effects and many interactions with other drugs. None of the available drugs has been shown to slow down progression toward dependence, yet all carry a risk of life‐threatening adverse effects and severe drug interactions". De‐listing means the drugs are no longer reimbursed by the national health insurance system (Prescrire 2018). The French dementia guideline was updated at the same time and indicated that the 'medical service' was insufficient to justify national support, although noting the marketing authorisation and stating that drugs can be prescribed in line with their summary of product characteristics (HAS 2018).

Fourth, many patients in North America are taking memantine outside the terms of its license, perhaps in the belief that 'if it works for moderate AD it must work in mild AD'. In 2006, it was prescribed to 19% of people with mild AD, despite the fact that, in 2005 the FDA did not approve a supplemental new drug application (sNDA) by Forest Labs which sought specific marketing approval for a mild AD indication (McManus 2006). In the mainly academic centres that comprise the Alzheimer Disease Neuroimaging Initiative (ADNI) 45.7% of people with mild AD were receiving memantine (Schneider 2011a). Nearly 40% of US neurologists report prescribing memantine at least sometimes to people with mild cognitive impairment (Roberts 2010).

Fifth, the regulations governing publication of trial results have changed since the last update of this review in 2006. Trials conducted by US sponsors now have to be reported on clinical registries within a year of the last data being collected from the last participant in the trial. This greater transparency has led to data becoming available which is not subject to peer review, and to the possibility of identifying data which was collected but remains unavailable ('known unknowns'). The inclusion of non peer‐reviewed registry data in meta‐analysis reduces the risk of bias due to selective publication. It also helps to restore the breach of ethical contract with trial participants and their families which occurs when data remain unpublished: they usually agree to participate in the belief that the experience will contribute to publicly available knowledge.

Sixth, pharmacological strategies for managing some of the behavioural changes associated with dementia are of low benefit, often have side effects, and are overused. Considerable marketing activity, especially in the 2000s, asserted that memantine might be an alternative.

Seventh, there may be a therapeutic window for memantine that impacts on the timing of when memantine is prescribed. Trials in cognitive impairment in multiple sclerosis have consistently found a dose‐related worsening of neurological and psychiatric symptoms (Lovera 2010;Peyro‐Saint‐Paul 2016;Villoslada 2009), consistent with preclinical work suggesting the possibility of a therapeutic window (Hardingham 2010). In the ADNI,Schneider 2011a and colleagues found that people taking memantine and ChEIs declined faster on the MMSE and Clinical Dementia Rating (CDR) scales (but not Alzheimer's Disease Assessment Scale ‐ Cognitive subscale (ADAS‐Cog) or a functional questionnaire) than those taking ChEIs alone. Whilst this was probably because memantine was started early by clinicians faced with patients who were declining more rapidly, and the results of the two cognitive scales are not wholly consistent, the result is also consistent with an adverse effect of memantine in early AD.

It is therefore important to conduct this review to investigate these unanswered questions and inconsistencies, in order to give clear information.

Criteria for considering studies for this review

Search methods for identification of studies

Data collection and analysis

Description of studies

Studies are described in detail inCharacteristics of included studies;Characteristics of excluded studies. In the former table, we report one four‐arm study four times: as two comparisons of memantine versus placebo (with or without vitamin E in both arms), and as post‐hoc subgroups for moderate and mild Alzheimer's disease (AD) for each of these comparisons (Dysken 2014). Four other studies reported results for severity subgroups and for these studies we have also extracted data and assessed risk of bias separately for the subgroups (Bakchine 2008 (99679);Peskind 2004 (MD‐10);Porsteinsson 2008 (MD‐12);Winblad 1999 (9403)). Summary details of the participants at baseline are given inTable 21 for the AD studies. Only three studies had fewer than 50% females ((Dysken 2014 ‐ 3%;Holland 2013 ‐ 35%;Ashford 2011 (95722) ‐ 38%). Most studies had a mean age between 70 and 80 years, exceptions were older participants in theForest 2006 (MD‐22) study (mean 85 years) and theFox 2012 (MAGD) study (mean 84 years) and younger participants in theWang 2013 study (mean 65 years).

Risk of bias in included studies

A detailed analysis of the risks of bias for each study can be found inCharacteristics of included studies. 'Risk of bias' domain ratings are shown per study inFigure 2 and percentage contributions for each domain are shown inFigure 3.

Effects of interventions

See:Table 1;Table 2

We present the efficacy results according to the aetiology of the dementia in sections 1 to 3 below. We compare effects across different types of dementia in section 4. We describe safety data, including discontinuation (all‐cause and that due to adverse events) and adverse events outcomes in section 5 for all aetiologies of dementia, represented on the same forest plots, with analyses featuring separate subgroups.

We conducted the main analyses as observed case (OC), for the reasons discussed in theDealing with missing data section, and also tested the OC approach in sensitivity analyses in comparison with the last observation carried forward (LOCF) approach (Appendix 3).

We produce a 'Summary of findings' table for most sections (Table 1;Table 2;Table 22;Table 23;Table 24;Table 25;Table 26;Table 27). We generally analysed the efficacy outcomes as standardised mean difference (SMD), but back‐transformed to a common scale. Similarly, for dichotomous (safety) outcomes, we calculated the absolute risk difference (RD) for the relevant population. In doing so, we reported safety outcomes either for all trials of memantine (across all aetiologies) or for separate sub‐populations, depending on the safety outcome concerned.

Summary of main results

We discuss in this section the findings in relation to the two objectives of the review: to assess the efficacy and safety of memantine for the treatment of dementia of different aetiologies, and secondly, to assess whether memantine adds benefit for people already taking cholinesterase inhibitors (ChEIs).

Overall completeness and applicability of evidence

People with dementia who are recruited into drugs trials are often not representative of typical clinical populations. Those recruited typically lack physical comorbidities, have better psychosocial support, and are less likely to have neuropsychiatric symptoms, all of which may mitigate against decline and its functional impact.

The studies were too short and small to be expected to show any effect of memantine on life expectancy. It is possible that the drug extends the total time of deterioration without reducing the personal or social burden of the disease (Dresser 2000).The benefits of slowing Alzheimer's disease progression in the later stages can be controversial (Post 1997).

The reliability of the distinction between vascular and Alzheimer's dementia is not high: most patients with vascular dementia, especially those with severe dementia have additional Alzheimer's pathology. This limits the applicability of results from trials of mild‐to‐moderate vascular dementia to those with severe vascular dementia.

Responder analyses are not routinely presented although the data are available from all trials. However, a meta‐analysis of responders based on six trials found that 10% more placebo‐treated than memantine‐treated patients showed any clinical worsening (Wilkinson Post‐Hoc 6RCTs 2007). There was a similar difference in rates of marked clinical worsening.

Measures of executive function are difficult to assess in those with more advanced dementia and in general are not well covered in AD trials which use the Alzheimer's Disease Assessment Scale ‐ Cognitive subscale (ADAS‐Cog) and were not included in the vascular dementia trials (Roman 1999).

Comprehensive lists of adverse drug reactions are infrequently reported so there is a theoretical possibility of publication bias. Regulators require comprehensive reporting. Whilst the details of these reports to regulators remain confidential, changes in the 'Summary of Product Characteristics' are likely to be a more reliable source of information on rare adverse effects than pooled data in systematic reviews.

Quality of the evidence

The evidence for AD is mainly of high certainty (with the exception of studies in people selected for agitation), which means we can be very confident of the results, which we summarised across many studies in large numbers of participants. The evidence for mild AD was obtained from within‐trial post‐hoc subgroups, with data provided by drug companies, but this evidence is still of moderate certainty.

The evidence certainty is moderate to low for vascular dementia and mainly low for FTD. For PDD or DLB and AD participants with agitation, the evidence certainty is mainly low or very low.

The authors of this review have gone to great lengths to obtain data for studies that have not been fully published, or for which there have been delays in publication. This work has much increased the volume of evidence available, nevertheless the results of some studies are still not in the public domain and there could be publication bias. Funnel plots for the main efficacy outcomes and for all‐cause discontinuation did not appear to suggest small studies' bias. There may be some asymmetry found for the adverse events outcome. Nearly all the trials were funded by drug companies.

Potential biases in the review process

In analysing the data, we have carried out a series of subgroup analyses and stratifications. At the outset, we stratified the studies by type of dementia, but later compared across aetiologies as a check. Stratification seemed appropriate for the efficacy outcomes and some safety outcomes, but we combined the results for adverse events across all studies because we found no differences in adverse events between diagnostic subgroups. We investigated duration of study, dose of memantine and type of analysis (observed case (OC) versus last observation carried forward (LOCF)) and restricted the main analyses to a duration of six months and licensed dose, also preferring to analyse OC data. We have documented our reasoning for all these decisions, but they could be a source of bias.

We split the data by severity into mild and moderate‐to‐severe AD, on the basis of subgroup analyses and pragmatism connected with current licensing requirements. These subgroup analyses seemed to provide convincing evidence, but are still non‐randomised comparisons, and there could have been confounding by some other factor. We then split the moderate‐to‐severe dataset further in our investigation of the effect of concomitant ChEI, and this further splitting could have led to random error. For this reason (and the lack of evidence of a difference between monotherapy and dual therapy), we prefer to use results for all studies regardless of their use of ChEI.

We calculated data for mild AD from published trial data in people with mild‐to‐moderate disease and drug company‐provided subgroup data for people with moderate disease. These data are from post‐hoc subgroups and so there may be differences between intervention groups in baseline characteristics. It would have been preferable to stratify patients by severity and then randomise to treatments, but this was not done by the trialists. Having said this, there are large numbers of participants and the split into mild and moderate disease is pre‐defined, so this is probably a minor limitation.

We also made some post‐hoc decisions, namely to treat separately studies in AD patients selected for agitation and to exclude from the analysis two randomised arms from one study because of a study‐reported interaction of memantine with Vitamin E. In the first instance, we considered the patients selected for agitation to be sufficiently dissimilar from other AD patients that they should not be included in the same analyses, and in the second case we thought the effect modification could have introduced unwarranted heterogeneity. These decisions could have meant that our findings for AD were overestimated.

We have been transparent about the approaches taken in this review, and do not consider the potential for bias is high.

Agreements and disagreements with other studies or reviews

This updated Cochrane Review has mainly similarities, but some differences with previous work. We focus here on clinical guidelines or technology appraisals for dementia, and recent systematic reviews of memantine. We mainly focus on AD, considering the number of studies included, the approach to monotherapy and concurrent therapy, ways of dealing with missing data, the effects of severity and the impact of agitation.

We have included 32 new studies (in comparison with 2006), several of which had unpublished data provided by the study authors or drug companies. We also included results for post‐hoc subgroups, which have informed analyses for both the mild and moderate‐to‐severe AD categories. We have meta‐analysed results from considerably more AD studies than were included in the NICE technology appraisal, which had four monotherapy and two dual therapy trials (NICE 2011); and consequently our review has more precise summary statistics. For moderate‐to‐severe AD, this has meant that all the main efficacy outcomes show small benefits that are statistically significant, in contrast to the largely non‐significant findings of TA 217 ‐ which were equated with no effect (seeWhy it is important to do this review).

Our analyses have shown that the benefit obtained for monotherapy versus placebo in moderate‐to‐severe AD is very similar to that obtained in dual therapy trials, and the test for subgroup differences is not significant. Therefore, we have included all trials in the meta‐analyses, regardless of the presence or absence of ChEIs. This conclusion about the efficacy of dual therapy contrasts with the conclusion of the NICE appraisal, which stated there was a lack of evidence of additional clinical efficacy (of concurrent therapy with ChEIs) compared with memantine monotherapy (NICE 2011). Since 2011, there have been many new studies of memantine in AD, leading to more up‐to‐date systematic reviews, and NICE has now published an update to its dementia guideline in June 2018 (NICE 2018). We discuss this recent literature below.

Three systematic reviews in AD have been published in the past three years, two of memantine monotherapy (Chen 2017;Matsunaga 2015, and one of dual therapy (Matsunaga 2015b). These gave similar conclusions to our review for most outcomes, even though the authors included studies reporting participants with mild AD and studies that did not have a placebo comparator, and also used LOCF approaches for missing data. A further systematic review included three studies of dual therapy and showed significant benefits when the analysis was restricted to moderate‐to‐severe AD (Muayqil 2012). We also identified a recent review of predominantly Chinese studies, which compared concurrent therapy with donepezil, however the majority of studies did not include a placebo (Chen 2017). We note that the FDA granted a license for combination therapy in patients stabilised on 10 mg donepezil once daily in December 2014 and a combined formulation product was launched in 2015.

The effect of severity was also identified in two other systematic reviews (Di Santo 2013;Kishi 2017): one of these showed greater efficacy as severity increased in a similar way to our review, but only included six trials (Di Santo 2013). The other review of 30 AD studies (including those without a placebo comparator) reached similar conclusions to ours regarding severity, but did not probe the effect of memantine in mild AD disease (Kishi 2017).

The NICE dementia guideline has been updated concurrently with the update of this Cochrane Review (NICE 2018); we have shared data with the guideline developers and the Cochrane Dementia and Cognitive Improvement review group is a registered stakeholder for the guideline. The guideline is a major update of the original guideline and updates some aspects of TA 217. The guideline preserved the TA's original stratification for analyses of monotherapy and dual therapy, and has updated the dual therapy analyses by including additional studies; however, the monotherapy analyses have not been updated (NICE 2018). The update has revised recommendations for 'people who are already taking a ChEI', for whom clinicians should now consider adding memantine to ChEIs for people with moderate disease and offer memantine in addition to ChEI to people with severe disease (NICE 2018). The evidence in this Cochrane Review supports (and indeed has informed) recommendations to offer dual therapy, but we consider that the monotherapy recommendations should also have been examined, especially because there are many new studies and memantine is now off‐patent. We are also concerned that the conflation of the old and new recommendations in the new guideline may lead to confusion for clinicians. For example, in severe disease, the unchanged monotherapy recommendation is to offer memantine (and not ChEIs), yet the new dual therapy recommendation is for people who are 'already taking a ChEI'. Additionally, the 2018 guideline is not explicit on whether combination therapy should be offered as first line therapy for people presenting with moderate or severe disease.

A systematic review of clinical guidelines reported the recommendations from 12 moderate‐ to high‐quality guidelines (Ngo 2015). The authors noted there was disagreement between two guidelines on the benefit of memantine in mild AD, but agreement in its use in moderate‐to‐severe dementia.They noted conflicting recommendations amongst four guidelines to support combining ChEI therapy with memantine in moderate–to‐severe dementia.

France's Minister for Health has recently de‐listed memantine and the ChEIs, donepezil, galantamine and rivastigmine (HAS 2018), stating that "it is better to concentrate on helping to organise daily activities, maintain activity, support and help from those around you", This de‐listing was based on work by the Haute Autorité de Santé (HAS), which stated in 2016 following evidence review, "The new data confirms that the efficacy of drugs for the symptomatic treatment of Alzheimer's disease is, at best, modest. It is established only in the short term, mainly on cognitive disorders, in placebo controlled clinical studies whose clinical relevance and transposability in real life are not ensured. Patients in these studies are indeed younger than those who are managed in real practice, and unlike these they have no comorbidities, nor risks of drug interactions. In addition, the effects on behavioral disorders, quality of life, time to enter an institution, mortality, burden of illness for carers are still not established... However, the data accumulated since the commercialization of the drugs confirm the risk of occurrence of undesirable side effects (digestive, cardiovascular or neuropsychiatric disorders for the most notable) potentially serious, which can alter the quality of life" (HAS 2016, translation). The 2018 statement was less measured, stating that, "None of the available drugs has been shown to slow down progression toward dependence, yet all carry a risk of life‐threatening adverse effects and severe drug interactions" (Prescrire 2018) (seeWhy it is important to do this review).

The HAS based its conclusions for the efficacy of memantine on theMatsunaga 2015 review for monotherapy and an earlier review authored by some of us for dual therapy (Farrimond 2012), both of which broadly agree with the conclusions of our memantine review, but there are differences for ADL for both reviews (HAS Annexe 2016). The HAS considered the benefits to be 'clinically irrelevant' and stated there is no difference in ADL for dual therapy. These conclusions are not supported either by our current memantine review or by the review and cost‐effectiveness analysis of the NICE guideline (NICE 2018). We consider that the small incremental benefits from each of ChEI and memantine for all outcomes at six months follow‐up, each having an effect size that may be less than the minimum clinically important difference, do not equate to clinical irrelevance. As stated above, we do agree that participants recruited into drugs trials are often not representative of typical clinical populations.

Second, the French authorities examined adverse effect data in detail for the ChEIs and memantine, using meta‐analyses of clinical trials, summary of products characteristics (SPC) and (for uncommon AEs) observational studies and analyses of pharmacovigilance databases (HAS Annexe 2016). The HAS reported that in placebo‐controlled clinical trials, more people discontinued donepezil 10 mg than placebo due to adverse effects, but the HAS did not report this for other ChEIs or memantine. In our current memantine review, however, we have noted a significant subgroup difference for this outcome between people with mild disease (greater discontinuation on memantine) and those with moderate‐to‐severe disease (no difference compared with placebo) (see section A above). The HAS report included some observational studies, one of which compared memantine and the three ChEIs (but had no data on untreated patients) (Fosbøl 2012). This very large study in the USA and Denmark reported for the Danish cohort a greater risk (in adjusted analyses) of myocardial infarction and cardiac death for memantine (compared with donepezil) and a smaller risk for syncope and atrioventricular block, but no differences in the USA cohort on hospitalisation for cardiac events. In both cohorts, all‐cause mortality was greater for patients receiving memantine. The authors concluded that the greater risk of cardiovascular events in the Danish cohort in users of memantine and dual therapy is probably related to selection of sicker participants, because these therapies are reserved for individuals with more severe dementia in Denmark. They also noted the lack of comparative data (with placebo/no treatment) and stated that "no clinical studies has found cardiovascular signals of clinical concern" (Fosbøl 2012). The HAS also gives pharmacovigilance data (as case reports) and indicates changes made to the SPCs of the various drugs: data on donepezil versus placebo was inconsistent for mortality in people with vascular dementia; there may be an increased risk of QT interval prolongation for galantamine and there may be an increased risk of myocardial infarction, stroke and torsade de pointes for rivastigmine. A very large pharmacovigilance study of the WHO database, VigiBase, over 58 countries, investigated all ChEIs and noted that serious cardiovascular events were frequently reported, suggesting that their significance has probably been previously underestimated, and encouraging caution when prescribing these drugs, especially as patients with Alzheimer's disease are frequently frail and receive other drugs (Krӧger 2015). The HAS has little to say about memantine pharmacovigilance except that hepatitis has been added to the SPC, but does agree with our review on the results for adverse effects from clinical trials. Instead of treating memantine separately from ChEIs, the HAS statements on safety apply to the dementia drugs as a whole (HAS Annexe 2016). Overall, we consider there is insufficient evidence to support the strong statement on adverse events (HAS 2016) and think that memantine, which has a different mode of action from ChEIs, should be considered separately.

The HAS recommends care and support for the individualinstead of the drug treatments, rather than including drug treatments as a part of a general care package. They do not appear to have reviewed the evidence on this.

Overall, our view is that the evidence in our review and our examination of the French adverse effects data raises questions about the appropriateness of the de‐listing policy taken by the French government. We note that de‐listing means the drugs are no longer reimbursed by the national health insurance system (Prescrire 2018), but that the drugs can be prescribed in line with their summary of product characteristics (HAS 2018). This may mean confusion for clinicians and potential for inequalities in the healthcare system.

Our review of the evidence on the effect on memantine in people with moderate‐to‐severe AD selected for agitation suggested that memantine may be ineffective in this population, but this is low‐ or very low‐certainty evidence. Our findings contrast with two post‐hoc analyses (Gaultier 2005;Wilcock Post‐Hoc 3RCTs 2008); the latter analysed individual participant data (IPD) from three randomised controlled trials (RCTs) in our review (Reisberg 2003 (9605);Tariot 2004 (MD‐02);van Dyck 2007 (MD‐01)); selecting only those from a 'behaviourally disturbed population', defined as a score > 0 on any of the three NPI symptoms of agitation–aggression, delusion, and hallucinations at baseline. Data were integrated in a single dataset and then analysed; meta‐analysis was not reported and the data were said to be 'pooled', indicating that the randomisation was not maintained. The study found a significant benefit for memantine for each of the efficacy outcomes at six months, but its reliability is unclear. A second review (Gaultier 2005); reported separately data from two RCTs in our review (Reisberg 2003 (9605);Tariot 2004 (MD‐02)), and, for one of the studies with memantine monotherapy (Reisberg 2003 (9605)) gave results for post‐hoc subgroups of participants with agitation and no agitation at baseline. Results were not reported for the other study (which had dual therapy). The review stated that for people with agitation at baseline (defined as inWilcock Post‐Hoc 3RCTs 2008, and comprising about 60% of the population); there was a significant improvement in agitation symptoms for memantine compared with placebo, which does not agree with our review's findings, but we note this is a different outcome to that in our review (improvement versus worsening of symptoms). In people without agitation at baseline, there were stated to be significantly fewer emergent agitation symptoms for memantine. Overall, there may be some differences with earlier findings, but all the efficacy evidence is of low or very low certainty and there is a need for further research of alternative therapies in this important patient group.

Finally, we note some differences between this Cochrane Review and the 2018 NICE guideline for types of dementia other than AD (NICE 2018). For people with vascular dementia, the NICE guideline recommends that ChEIs or memantine should be considered only if they have comorbid AD or PDD/DLB. However, this Cochrane Review has identified small clinical benefits in cognitive function and in behaviour and mood in people diagnosed with vascular dementia, and there is a post‐hoc analysis indicating there may be greater benefits in people with more severe disease. For people with FTD, the guideline recommends that ChEIs and memantine should not be offered (a strong recommendation), stating that there is not usually a cholinergic deficit in people with FTD, that there was no evidence of benefit and also citing the potential for adverse effects, whilst noting no difference between memantine and placebo. It may be that the guideline has conflated the findings for ChEIs with those for memantine, and we note that memantine has a non‐cholinergic mechanism of action and may have potential for benefit, albeit from low‐ or very low‐certainty evidence. For people with PDD or DLB, the guideline reviewed the evidence for PDD and DLB separately, and appeared to draw conclusions on the basis of a lack of statistical significance. The guideline recommended that memantine should be considered only if ChEIs are not tolerated or contraindicated, which, while consistent with our review findings of low‐ or very lo‐ certainty evidence in this patient group, has a strength which goes beyond the evidence.

Implications for practice.

Implications for research.

Appendix 1. Search: September 2011, February 2016, March 2017 and March 2018

Appendix 2. Outcome measures (brief description)

The following range of outcome measures was used in the trials.Table 20 summarises their use in the included studies for AD.

1. Global rating scales

• Clinician's Interview‐Based Impression of Change scale (CIBIC‐Plus) provides a global rating of function in four areas, general, cognitive, behaviour and activities of daily living. All patients are scored as 4 at baseline and subsequent assessments on a scale of 1 to 7 are relative to baseline, with 1 showing marked improvement and 7 marked worsening. Information is obtained from the caregiver and the patient. There are different versions: the Alzheimer's Disease Cooperative Study format (Schneider 1997) and the New York version (Reisberg 1997).

• The Clinical Global Impression of Change (CGIC) is a global rating of all domains of a patient's current condition in comparison with baseline (Guy 1976). It is a seven‐point scale, from 1 (very much improved) to 7 (very much worse), 4 indicating no change. The assessment is conducted by the same clinician at both time points with input from relatives or carers.

• Physician's global impression. This unvalidated four‐point rating of the dementia syndrome and patient's general health status was used in theDitzler 1991 study.

• Clinical Global Impression (CGI). This seven‐point rating of severity of illness was used in theGortelmeyer 1992 study to assess whether patients improved, remained unchanged or worsened.

• The Sandoz Clinical Assessment Geriatric Scale (SCAG) is a physician rating (Shader 1974). It consists of 18 items and an overall impression (item 19), all rated on a seven‐point format. There are five sub‐scores: cognitive disturbances, disturbances in social behaviour, lack of drive, affective disturbances, somatic disturbances. Item 19 was used in thePantev 1993 study.

2. Cognitive Tests

• The cognitive part of the Alzheimer's Disease Assessment Scale (ADAS‐Cog) comprises 11 individual tests, spoken language ability (0 to 5), comprehension of spoken language (0 to 5), recall of test instructions (0 to 5), word finding difficulty (0 to 5), following commands (0 to 5), naming object (0 to 5), construction drawing (0 to 5), ideational praxis (0 to 5), orientation (0 to 8), word recall (0 to 10) and word recognition (0 to 12) (Rosen 1984). The total score ranges from 0 to 70, the high score indicating greater impairment.

• Syndrom‐Kurztest determines patient attention and memory disturbances (Kim 1993).

• Severe Impairment Battery (SIB) evaluates cognitive performance in advanced Alzheimer's Disease (Schmitt 1997). It is a 51‐item scale which assesses social interaction, memory, language, visuospatial ability, attention, praxis and construction. The scores range from 0 (greatest impairment) to 100.

3. Activities of Daily Living

• Alzheimer's Disease Cooperative Study‐Activities of Daily Living (ADCS‐ADL) was specifically designed to assess functional capacity over a broad range of severity in patients with Alzheimer's disease (Galasko 1997). The 19‐item ADCS‐ADLsev19 (ADCS‐ADL19) has 54 points. The 23‐item ADCS‐ADL23 has 78 points. High score indicates greater functional capacity.

• Activities of daily living test. This scale evaluates patients' abilities to cope with five instrumental task under the guidance of a psychologist.

• Behavioural Rating Scale for Geriatric Patients (BGP) is a 45‐item observer‐rated scale for the assessment of functional and behavioural disturbances of geriatric patients, performed by nursing staff (Van der Kam 1989). The BGP contains several subscales: care dependence, aggressiveness, physical, mental, disability and depressiveness, and inactivity. The "care dependence" scale consists of 23 items, measures activities of daily living and has the highest reliability and validity. High score indicates lower functional capacity.

• Bristol Activities of Daily Living (BADL) is a caregiver rated scale. Scores range from 0 to 60, with higher scores indicating greater impairment (i.e. a negative outcome).

• Nurse's Observational Scale for Geriatric Patients (NOSGER) contains 30 items of behaviour, each rated in a five‐point scale according to frequency of occurrence (Spiegel 1991). Item scores are summarised into six dimension scores (memory, instrumental activities of daily life, self‐care activities of daily living, mood, social behaviour, and disturbance behaviour).

• Disability Assessement Daily (DAD) (Gélinas 1994) contains 40 items, 17 related to self‐care and 23 items related to instrumental activities of daily living. Scores range from 0% to100%, with higher scores indicating greater impairment.

4. Mood and Behavioural measures

• Neuropsychiatric Inventory (NPI) assesses the frequency and the severity of behavioral and neuropsychiatric symptoms in patients with dementia based on an interview with the caregiver (Cummings 1994). There are 12 items with a total score ranging from 0 to 144. Higher scores indicate greater neuropsychiatric impairment.

• Cohen Mansfiedl Agitation Inventory (CMAI) assesses the frequency of behaviour on a seven‐point scale (1: never; 2: less than once a week; 3: once or twice a week; 4: a few times a week; 5: once or twice a day; 6: a few times a day; 7: a few times an hour). There are 29 items (Cohen Mansfield 1989).

• Nurses Observation Scale for Inpatient Evaluation (NOSIE) assesses behaviour of psychiatric patients (Honigfeld 1974). It comprises 30 items of behaviour and the frequency of their occurrence. NOSIE subscale scores are for social competence, social interest, personal neatness, irritability, manifest psychosis, retardation, depression. Increasing values of the NOSIE index are indicative of improvement.

• Sandoz Clinical Assessment Geriatric Scale (SCAG) is a physician rating scale (Shader 1974). It consists of 18 items and an overall impression (item 19), all rated on a seven‐point format. There are also five sub‐scores: cognitive disturbances, disturbances in social behaviour, lack of drive, affective disturbances, somatic disturbances.

• Behavioural Rating Scale for Geriatric Patients (BGP) is a 45‐item observer‐rated scale for the assessment of functional and behavioural disturbances of geriatric patients, performed by nursing staff (Van der Kam 1989). The BGP contains several sub scales: care dependence, aggressiveness, physical, mental, disability and depressiveness, and inactivity. The "care dependence" scale consists of 23 items, measures activities of daily living and has the highest reliability and validity

• Nurse's Observational Scale for Geriatric Patients (NOSGER) contains 30 items of behaviour, each rated in a five‐point scale according to frequency of occurrence (Spiegel 1991). Item scores are summarised into six dimension scores (memory, instrumental activities of daily life, self‐care activities of daily living, mood, social behaviour, and disturbance behaviour).

• The BEHAVE‐AD scale has 25 items, is based on caregiver interview and is more specific for psychotic disorders in people with dementia. There are seven subscales: paranoid and delusional ideation, hallucinations, disturbances of daily activities, aggression, sleep disturbances and circadian rhythm disorder, affective disorder, and anxiety or phobias. Higher scores indicate greater neuropsychiatric impairment.

5. Combination Scales

Several scales used in earlier studies combine elements that are now more commonly assessed using separate instruments. Such scales have been used as a different method from 'overall clinical impression' for the systematic global assessment of dementia. These combination scales typically have several subscales, the results of which are sometimes presented separately and can be included in meta‐analyses.

• Gottfries‐Brane‐Steen Scale (GBS) is a 26‐item, physician‐assessed observer scale based on caregiver's information and an interview with the patient (Gottfries 1982). It comprises three subscales: motor performance, intellectual and emotional capacity, and a group of six symptoms commonly observed in dementia.

• Other combination scales used in studies of memantine are the SCAG, NOSIE, NOSGER and BGP which are detailed above.

6. Cost of resource utilisation

• Resource Utilization in Dementia questionnaire (RUD) is a structured interview of the patient's caregiver consisting of a baseline questionnaire (assessing basic patient demographic data and specific events that occurred during the past one month) and follow‐up questionnaire (assessing specific events that occurred during the past three months) (Wimo 1988). The specific events included time spent caring for the patient, changes in the caregiver's work status, healthcare resource utilisation by the caregiver, healthcare resource utilisation by the patient and the patient's residential status.

7. Safety and tolerability were assessed by the frequency of reported adverse effects

Appendix 3. Last Observation carried forward (LOCF) versus observed case (OC) assumptions for missing data

Two sets of analyses were compared, one using results based on an OC assumption, and the other with a LOCF assumption. One study reported raw data for both a per protocol analysis and a retrieved dropout analysis (Howard 2012 (DOMINO‐AD)). One reported a per protocol analysis at 52 weeks and LOCF at the end of the study (assumed to be 52 weeks) (Peters 2015 (MEGACOMBI2)).

Firstly, within‐trial comparisons are reported inAnalysis 10.1 andAnalysis 10.2 for the outcomes cognitive function and decline in ADL. There is generally little difference between the results for OC and LOCF, with the possible exceptions of two studies for the decline in activities of daily living (ADL) (Peters 2015 (MEGACOMBI2);Reisberg 2003 (9605)).

Secondly, we conducted meta‐analyses of studies in patients with moderate‐to‐severe dementia for the above outcomes, comparing analyses restricted to OC and LOCF data (Analysis 10.3;Analysis 10.4). Again, there appeared to be little difference between the results of two analyses, although there was heterogeneity within each analysis. We decided to report OC analyses in the main part of the review wherever possible: if studies did not report OC data, we used the data reported by the authors. The degree of missing data was assessed and is reported in the 'Risk of bias' section of theCharacteristics of included studies table.

Appendix 4. Sensitivity and subgroup analyses to investigate heterogeneity in efficacy outcomes

Initially, we included data from all 24 trials in AD in the same forest plots (Analysis 11.1;Analysis 11.2;Analysis 11.3;Analysis 11.4) using the OC approach to missing data, wherever possible. Two studies had no results for any efficacy outcome (Lundbeck 2006 (99817);Merz 2003 (MRZ‐9104). For most outcomes, there was too much heterogeneity to make overall statements of effect ‐ this is seen in the variation of point estimates in the forest plots and in the statistics: for the cognitive function outcome (20/24 studies), I² was 41% and P = 0.03. For the decline in ADL (14/24), I² was 47% and the P value was 0.03. For the behaviour and mood outcome (15/24), I² was 38% and the P value was 0.07. However, for the clinical global rating (14/24 studies), I² was 0% and the P value for heterogeneity was 0.59. Therefore, we investigated the heterogeneity for these outcomes in sensitivity and subgroup analyses, which cover risk of bias, severity of AD and the presence or absence of concomitant ChEIs.

Appendix 5. Conclusions from subgroup analyses

After reducing the dataset to studies with the licensed dose (20 mg/day) and a duration of follow‐up (six to seven months), we conducted subgroup analyses investigating the presence or absence of concomitant ChEI and two main approaches to severity of disease (mild–to–moderate versus moderate–to–severe, and mild versus moderate–to–severe) in 17 studies in 5813 randomised participants.

There is not much effect of concomitant ChEI as an effect modifier, but there are large differences in the test for subgroup differences related to severity. The exception to this may be for the behaviour and mood outcome, which also shows a small non‐significant difference for the effect of concomitant ChEI.

We investigated effect modification further by conducting meta‐regressions in STATA (StataCorp version 13). For the cognitive function outcome the following results were found:

• for the potential effect modifier of concomitant ChEI (versus monotherapy), the meta‐regression coefficient was 0.60 (95% CI ‐1.06 to 2.26) and the proportion of between‐study variance explained (adjusted R²) was negative, suggesting a poor fit;

• for the potential effect modifier of severity (mild‐to‐moderate versus moderate‐to‐severe), the meta‐regression coefficient was ‐1.61 (95% CI ‐2.94 to ‐0.27) and the proportion of between‐study variance explained was 48.3%;

• Inclusion of both variables gave coefficients of ‐1.70 (95% CI ‐3.06 to ‐0.34) for severity and 0.66 (95% CI ‐0.70 to 2.02) for ChEI, with an adjusted R² of 42.0% and an overall model significance of 0.046.

The meta‐regression for behaviour and mood gave negative adjusted R² values and suggested one study may have been an outlier (Nakamura 2016). In its absence, the meta‐regressions that included both severity and ChEI gave coefficients.

• Behaviour and mood: ‐2.21 (95% CI ‐3.89 to ‐0.52) for severity and ‐1.81 (95% CI ‐3.49 to ‐0.13) for ChEI, with an adjusted R² of 100% and an overall model significance of 0.02.

• Cognitive function: ‐1.94 (95% CI ‐3.29 to ‐0.60) for severity and 0.37 (95% CI ‐0.98 to 1.72) for ChEI, with an adjusted R² of 59.1% and an overall model significance of 0.025 for cognitive function.

These meta‐regressions are underpowered, but give an indication that much of the variance may be explained by the severity factor. We note that the coefficient for severity is negative, which means that there may be a bigger effect in the more severe populations.

All these findings suggest that it is more important to stratify the studies by severity first and then investigate the effect of ChEI to address our second objective.

On the basis of the evidence from these subgroup analyses, we decided to report, in the main text, the results separately for mild and moderate‐to‐severe AD, but not to report the results for patients with mild‐to‐moderate AD, because this is not a licensed indication. Thus, the main analyses reported in the text for AD are:

• moderate‐to‐severe AD; six to seven months duration;

• mild AD; six to seven months months duration.

We note that the medians of the standardised change from baseline for the placebo group varied between the mild, moderate and severe AD subgroups. In the moderate and severe disease subgroups, the cognitive function, decline in ADL, behaviour and mood, and, to a lesser extent, global scores, all tended to worsen over six months. By comparison, for the mild severity subgroup, the scores for the outcomes cognitive function, and behaviour and mood tended to improve with time.

Appendix 6. Subgroup analysis by concomitant cholinesterase inhibitor in people with moderate to severe AD

We conducted subgroup analyses to investigate any differences in the effect of memantine versus placebo as monotherapy or with concomitant ChEI in the moderate‐to‐severe AD population.Analysis 2.1;Analysis 2.2;Analysis 2.3 andAnalysis 2.4 are between‐trial subgroup analyses andAnalysis 2.5;Analysis 2.6 andAnalysis 2.7 show the within‐trial, per protocol (PP) analyses for cognitive function (MMSE), decline in ADL (BADL) and neuropsychiatric inventory (NPI) outcomes (Howard 2012 (DOMINO‐AD)).

Comparison 1. Memantine 20 mg or equivalent vs placebo for moderate‐to‐severe Alzheimer's disease 24‐ to 30‐week data. OC.

Comparison 2. Memantine 20 mg or equivalent vs placebo in patients taking versus not taking cholinesterase inhibitors with moderate to severe Alzheimer's disease 24‐ to 30‐week data.

Comparison 3. Memantine 20 mg or equivalent vs placebo for moderate‐to‐severe AD with agitation.

Comparison 4. Memantine 20 mg vs placebo for mild AD (MMSE 20‐23) OC six‐month studies.

Comparison 5. Memantine 20 mg vs placebo for mild‐to‐moderate vascular dementia. Six‐month studies.

Comparison 6. Memantine vs placebo for Parkinson's disease dementia (PDD) and dementia with Lewy bodies (DLB). ITT‐LOCF Data.

Comparison 7. Memantine vs placebo for frontotemporal dementia (FTD).

Comparison 8. SUMMARY: memantine 20 mg/day or equivalent vs placebo at six to seven months, by dementia type and by severity.