Folate Augmentation of Treatment - Evaluation for Depression (FolATED): randomised trial and economic evaluation

Prevalence of depression

Depression is one of the main mental health disorders presenting in primary care. 1 The prevalence of major depression in the general population ranges from 3% to 10% with more than 150 million people at a time suffering from depression across the world. 2 In the UK prevalence of depression in 2009–10 was 11% in England, 11.5% in Northern Ireland, 8.6% in Scotland, and 7.9% in Wales. 3 Unipolar depression leads to 12.15% of years lived with disability, and is ranked as the third leading contributor to the global burden of diseases. 2 Indeed depression is currently the second cause of disability worldwide for males and females between the ages of 15 and 44 years and is predicted to reach second place for all ages by 2020. 4 Depression has become the leading cause of disability in Europe, leading to a loss of one in every 10 healthy years of life, and the leading cause of early retirement. 5 Depression and stress are now the commonest reported causes for sickness absence from work in the UK6 with over 100 million working days lost across Europe at a cost of 1% Gross Domestic Product (GDP). 7 A higher prevalence of depression is observed in women than men across the 18- to 64-year age range with women up to 2.5 times more likely to develop depression. 8

Characteristics of depression

The core symptoms of depression are low mood and loss of interest or enjoyment in usually pleasurable activities. Associated symptoms include disturbances to sleep and appetite, reduced energy and concentration, negative thoughts of guilt or worthlessness and suicidal ideation. The International Statistical Classification of Diseases & Related Health Problems (ICD-10) states that for a diagnosis of depression at least five symptoms need to be present, including at least one of the core symptoms, at an intensity that causes functional impairment and for a minimum duration of 2 weeks. Depression is classified as mild, moderate or severe according to the number of symptoms present and degree of functional impairment, and the grading of severity is of direct relevance to the treatment approaches recommended. 9 Depression is associated with increased mortality linked to suicide, alcohol and drug misuse, and increased rates of cardiovascular disease. 10 Depression thus burdens individuals, families, the NHS, and the national economy. 11 One UK study estimated the total cost of depression to the UK in 2000 at £9 billion; at that time, before the introduction of Improving Access to Psychological Therapies (also known as IAPT) in NHS England, the direct cost of treatment, mainly antidepressant medication (ADM), was £370 million; and indirect costs of 110 million working days lost to depression accounted for the vast majority of the total cost. 12 The sub-optimal treatment of depressive disorders is therefore of great public health concern.

Treatments for depression

In accordance with the joint report of the World Health Organization (WHO) and World Organization of National Colleges & Associations of Family Doctors (WONCA)1 and the National Institute for Health and Care Excellence (NICE) guidance9 the majority of people with depression are identified, treated and managed within primary care. Treatment aims to relieve symptoms, restore functioning and, in the long term, prevent relapse. The goal of treatment is complete remission which is associated with better functioning and reduced risk of relapse. 13 While there is some evidence that people with depressive symptoms improve over time without treatment,14,15 a significant proportion follow a chronic course with significant levels of depressive symptoms and functional continuing for several years. 16,17

Antidepressants are recommended as a treatment option for moderate to severe depression either in combination with psychotherapy9 or as monotherapy. 18 Owing to their greater tolerability selective serotonin reuptake inhibitors (SSRIs) are recommended as first-line treatment in primary care. 9 Patients treated with an SSRI are seven times more likely to complete a therapeutic course than those treated with tricyclic antidepressants (TCAs). 19 In clinical trials of antidepressants typically 50% of patients with depression respond to active treatment, while one-third respond to placebo20 with the placebo response appearing to increase over time in clinical trials. 21 With first-line treatments about one-third of patients achieve remission from depression, increasing to two-thirds with refinement of treatment. 22 A study of mental disorders in 14 centres worldwide found that 50% of patients continued to have a diagnosis of depression after 1 year23 with at least 10% having persistent or chronic depression. 24 Furthermore, at least 50% of people will go on to have at least one further episode of depression following their first episode of major depression. 25 The risk of further recurrences after second and third episodes rises to 70% and 90% respectively. 25 Cumulative rates of recurrence remain linear over long periods of follow-up (30–40 years), indicating a constant risk of recurrence over the lifespan. 26 Therefore recurrence rates increase with length of follow-up, and for the majority of patients depression is a recurrent condition.

Depression is a prevalent global health problem resulting in high levels of disability. While effective treatments are available outcomes remain sub-optimal with a significant proportion of patients failing to achieve remission and experiencing chronic illness, early relapse and multiple recurrences across the lifespan. There remains a pressing need for research to optimise outcomes from antidepressant treatment.

Depression and folate

Over recent years there has been a growing interest and an increasing body of evidence exploring the relationship between B vitamins, in particular folate, and depression. 27–30 Folate is a naturally occurring B vitamin and can be found in leafy green vegetables, fruits, dried beans and peas. 31 Folic acid is the synthetic form of folate, which is inexpensive and found in supplements and fortified food. 31,32

There is evidence to suggest that low folate intake is associated with symptoms of depression. 33–36 Studies report that up to one-third of patients with depressive illness have decreased serum and red cell folate levels. 37 Many people with depression have lower concentrations of folate than people with other psychiatric disorders or no psychiatric disorder. 34,38,39 Associations between folate deficiency and depressive symptoms, symptom severity and treatment outcomes in adults and the older adult population have been reported. 38,40–44 Low folate intake may also increase the risk of recurrent depression33 and depression in later life. 43 Gilbody and colleagues conducted a systematic review of observational epidemiological studies investigating the relationship between low folate status and depression. 45 They concluded that low folate status was associated with depression but could not conclude that that was a causal relationship. Though low folate may result from poor nutrition or socio-economic disadvantage, confounders are common in chronic mental illness. Other recent evidence suggests that low folate may be a consequence rather than a cause of depressive symptoms. 46 While there is weak evidence that increased folate intake may prevent depressive symptoms,35,47 this is not a consistent finding. 48 In one randomised controlled trial (RCT), for example, giving folate (2 mg) with vitamins B₁₂ and B₆ did not reduce severity of depressive symptoms. 49

The role of homocysteine

Folate is absorbed and transported in the blood in the form of 5-methyltetrahydrofolate (5-MTHF)50 and is measured in the blood as either serum folate or red cell folate. 50

Homocysteine is a highly sensitive marker of folate status51 and functional folate deficiency is indicated by elevated homocysteine. Tiemeier et al. found a significant relationship between depression and hyperhomocysteinemia, folate and B₁₂ deficiency. 52 Observational studies indicate that patients with depression have increased plasma total homocysteine concentrations. 53,54 A recent meta-analysis showed that older adults with a high total homocysteine concentration have an increased risk of depression [odds ratio (OR) = 1.70; 95% confidence interval (CI) from 1.38 to 2.08]. 55

There are important relationships between folic acid and vitamin B₁₂ that have implications for how folate can be used therapeutically. For people who are deficient in vitamin B₁₂, exposure to high levels of folate can result in subacute combined degeneration of the spinal cord, which may be linked to impaired methionine biosynthesis. 56 As methylmalonyl CoA mutase, a vitamin B₁₂-dependent enzyme, converts methylmalonyl-CoA to succinyl-CoA, blood methylmalonic acid (MMA) levels increase with suboptimal B₁₂ status. The cross-sectional study of > 10,000 participants in the US National Health and Nutrition Examination Survey (NHANES) observed a novel relationship between MMA and folate:57 in patients with low B₁₂ (< 148 pmol/l ≈ 200 ng/l) MMA increased significantly with increasing serum folate. This finding is due, either to adverse oxidative effects of unmetabolised folic acid on B₁₂ homeostasis, or inability of patients with low B₁₂ to retain intracellular folate. 58

Antidepressant treatment and folate

Folate is an essential cofactor for the biosynthesis of both serotonin (5-hydroxytryptamine or 5-HT) and noradrenaline. Thus folate deficiency leads to impaired serotonin synthesis in the human brain. 59 This may provide a theoretical model for claims that folate can play a role in the treatment and prevention of depression. 60 Virtually all antidepressants are thought to act by prolonging the activity of serotonin or noradrenaline in neurotransmission or by modulating monoamine receptor sensitivity. 61 Lower folate levels have been associated with poorer antidepressant response. 62 Further evidence suggests that baseline levels of folate within the normal range predict antidepressant response. 63

This raises the possibility of folate being used to augment antidepressants. Initial small feasibility studies seemed to confirm the potential of this augmentation strategy. 32 A Cochrane review and meta-analysis50,64 explored the role of folate augmentation in depression. Only two RCTs were identified (combined n = 151), both of which suggest possible beneficial effects of folate augmentation. 65,66 Two further small trials have given contradictory results. One study, 67 with a clinical sample of 42 patients with major depressive disorder, reported that 10 mg of escitalopram (Cipralex^®, Lundbeck) alone produced greater improvement than the combination of escitalopram and folic acid (2.5 mg/day). The other study, 68 with a clinical sample of 27, reported a greater reduction in depressive symptoms when 20 mg of fluoxetine (non-proprietary) was augmented with folic acid (10 mg/day) than when augmented with placebo. However, emerging evidence within an older depressed population suggests there may be no benefit from folate augmentation of antidepressants. 69 Despite this negative finding in the elderly there is still interest in a potential augmentation role for the B vitamins in general, with the currently recruiting B-VITAGE trial exploring B vitamin supplementation in later life. 70 These data convey mixed messages about the use of folic acid to improve ADM. To advance this debate about the clinical effectiveness of folic acid, we plan to update the Cochrane systematic review to include this and other recently reported studies of augmentation by folic acid. 50

Folate, depression and genetics

There has recently been much research aimed at identifying genetic aspects of depression and antidepressant therapy. Indeed, a number of large genome-wide association studies, and some subsequent meta-analyses, have identified genetic polymorphisms associated with both risk of depression71–73 and response to antidepressant therapy. 74,75 However, these studies have been unable to demonstrate consistent and reproducible genetic associations.

Only a few studies have described an association between risk of depression and genetic variation of genes of the one-carbon folate and methionine biosynthesis pathways. Most have focused on, and identified, an association with depression and the frequently characterised c.677C > T polymorphism (rs1801133) of the methyltetrahydrofolate reductase (MTR) gene. 76,77 This variant encodes a valine to alanine amino acid substitution at residue 222. The variant protein has reduced catalytic activity and thermolability and is associated with elevated homocysteine levels under conditions of impaired folate status. Others, in addition to MTR c.677C > T, have also described the p.D919G variant (rs1805087) of the MTR gene as a statistically significant risk factor for moderate and severe depression in postmenopausal women. 78 The MTR gene encodes the protein methyltetrahydrofolate reductase, which is a key enzyme in the biosynthesis of homocysteine to methionine.

Furthermore, it has recently been reported that the MTR c.677C > T polymorphism modifies the protective effect of folic acid against depression after pregnancy. 79 Others have demonstrated an association between c.677C > T and folate and homocysteine concentrations. 80 These observations further support the hypothesis that genetic variation in the one-carbon folate pathway may affect folic acid efficacy as an adjuvant to antidepressant therapy by altering folate bioavailability and increasing homocysteine levels.

Participants

Sample size

We originally powered FolATED to detect a difference between the two treatment groups of three points on the BDI-II at 25 weeks, judging that a clinically important difference. As we estimated the standard deviation (SD) of BDI-II scores in the trial population at 10.7, our protocol proposed a completed sample size of 400 at 25 weeks to yield 80% power to detect this difference using a significance level of 5%. As interim analysis of baseline BDI-II scores showed that their SD was about 10, we revised the target completed sample size to 358 at 25 weeks. The original protocol allowed 10% loss at each of the three follow-up assessments, thus requiring to randomise 549 to achieve 400 completers at 25 weeks. As interim analysis also showed that retention at 25 weeks was 79% rather than the 73% expected, the new target of 358 completers needed a randomised sample of 453.

Randomisation

At screening the screener took a blood sample to assess B₁₂ and folate status, and arranged a further appointment within 14 days to confirm the B₁₂ and folate results. We excluded participants who were B₁₂ or folate deficient from the main trial but offered them the opportunity to continue in the ‘comprehensive cohort’ of recruited patients.

Eligible participants completed the baseline assessments and the recruiting centre telephoned the randomisation centre at NWORTH, Bangor University. NWORTH used dynamic allocation to protect against subversion while ensuring that each arm of the trial was balanced for the stratification variables. For each participant the adaptive algorithm recalculated the likelihood of their allocation between treatment groups from the distribution of stratification variables among participants already recruited and allocated. This process keeps the balance between strata within acceptable limits of the target allocation ratio of 1 : 1 while maintaining unpredictability. 85 The selected stratification variables were:

1. centre (North East Wales, North West Wales or Swansea)

2. sex (male or female)

3. timing of ADM (new or continuing)

4. type of ADM (SSRI or other)

5. whether participant had received counselling for depression (ever or never).

Intervention

By the time participants entered the trial, we ensured they were receiving antidepressant treatment optimised to therapeutic dosages – equivalent to SSRI of at least 20 mg per day or TCA of at least 150 mg per day. Most had received an antidepressant prescription from their general practitioner (GP) before referral to the trial. For patients not on ADM, trial psychiatrists initiated ADM to meet clinical need and patient preference in accordance with routine practice. For patients on sub-therapeutic ADM, trial psychiatrists optimised the treatment regime according to the British National Formulary (BNF),86 namely citalopram dose of at least 20 mg per day or equivalent. For participants who had been receiving a therapeutic dose of ADM, we encouraged trial psychiatrists to optimise dose according to the BNF, for example by increasing citalopram dosage to 40 mg per day, or change the antidepressant, again depending on clinical need and patient preference.

Participants received a 12-week supply of either 5 mg folic acid or placebo in addition to their ADM. We selected the 5-mg dose of folic acid because that is effective for other indications,87 carries a low risk of adverse events (AEs), and is routinely used to treat folate deficiency. 31

Bilcare (formerly DHP Ltd) supplied the trial drugs and achieved the blinding needed by the trial by the process of over-encapsulation. They placed each tablet inside a size ‘1’ opaque hard gelatin capsule and added lactose BP, an ingredient of the tablet, to fill the capsule. To produce placebo for the trial they filled the same capsules with lactose. They packed capsules into high-density polyethylene bottles with tamper-evident child-resistant screw caps. They tested to confirm that the over-encapsulated tablet complied with the British Pharmacopoeia88 standard for disintegration in vitro. They also checked each batch for the presence or absence of 5 mg folic acid.

Blinding

The North Wales Organisation for Randomised Trials in Health coded the identically packaged folic acid and placebo randomly for each stratification group. Each patient’s prescription indicated his or her trial number and package serial number generated by NWORTH, thus determining the appropriate trial package. NWORTH and the local pharmacies held the key to the randomisation codes. The telephone numbers of those pharmacies were available to break codes in emergency.

This ensured that throughout recruitment treatment allocations were unknown to participants, healthcare professionals, investigators, and researchers. We broke randomisation codes for two participants. One was diagnosed with lower oesophageal cancer, and the other collapsed with agitation, breathing difficulty, raised pulse, and reduced consciousness. On both occasions the local pharmacist revealed the code from a scratch card. This ensured that we revealed only the individual allocation, only to those who needed to know.

An independent GP monitored blood results, notably to check for B₁₂ and folate deficiency at follow-up. To avoid accidental unblinding researchers engaged in clinical data collection or analysis did not have access to these blood results. Furthermore we separated pharmacogenetic and biochemistry analysis from clinical effectiveness analysis, and combined these results only when analyses were complete. Formal unblinding of the randomisation codes took place at the joint final meeting of the Trial Steering Committee (TSC) and the Data Monitoring and Ethics Committee (DMEC) on 10 October 2011.

Data collection

We collected data at screening, baseline and randomisation, and 4, 12 and 25 weeks after randomisation, using the trial case report forms (CRFs). Though we aimed to collect the data on the day due, this was not always possible. Hence there was a window for each data collection (Table 1).

We designed some questionnaires for completion by researchers or clinicians, and others by participants. The preferred mode was face to face. When that was not possible, we permitted completion over the telephone and mailed the questionnaire to the participant in advance.

Primary outcome measure

The main outcome measure was self-rated symptom severity as measured by the Beck Depression Inventory version 2 (BDI-II). 84 The BDI-II consists of 21 items, each rated on a four-point scale ranging from 0 to 3; a total score of 1–13 indicates no depression, 14–18 mild depression, 19–28 moderate depression, and 29–63 severe depression. As BDI-II scores at 25 weeks are useful in assessing participants’ medium-term recovery, that was the basis of our original power calculation (see Sample size, above). When updating that calculation, we also designated the primary outcome as the area under the curve (‘AUC’ for short) of mean BDI-II scores between randomisation and the 25-week follow-up, because this summarises participants’ recovery across the whole of that period. 84 Though there was less prior information on AUCs, notably on clinically important differences, we judged that the combination of well-behaved BDI-II scores and the mathematically robust trapezium method of estimating AUCs89 would make 358 an appropriate target sample size (see Sample size, above). In the event we were able to analyse many more than 358 trial participants. Because the actual follow-up time could vary by up to 4 weeks from the target of 25 weeks (see Table 1), we converted the area under the curve to a more meaningful ‘AUC average’, which represents a participant’s BDI-II (or other outcome) score averaged over that participant’s follow-up period.

Secondary outcome measures

Other data

We collected basic demographic information for each participant including sex, age, ethnicity, employment status, marital status and number of dependent children. We also recorded smoking and alcohol consumption, which are known to affect homocysteine levels.

Follow-up

Thus we thoroughly assessed participants at 4, 12 and 25 weeks after randomisation. Antidepressants show a delayed and variable onset of clinical improvements in depression. 97–99 Previous trials suggest that 50% of those who eventually respond to ADM start to respond within 2 weeks, 75% within 4 weeks, and almost all within 6 weeks. 100 Hence we scheduled the first assessment at week 4, 6 weeks after the start or optimisation of antidepressant treatment. Non-response at 4 weeks may lead to changes in the ADM in accordance with the BNF and NICE guidelines. Hence the second assessment at 12 weeks could measure both continuing and late responses to ADM and folate augmentation. The third assessment at 25 weeks addressed any changes in effectiveness after the end of folic acid therapy, but during ADM, since that is the minimum duration of maintenance antidepressant treatment. 18

Quality assurance

The conduct of this trial followed the principles of good clinical practice (GCP) outlined by the ICH-GCP and complied with EU directive 2001/20/EC. 101 The research also adhered to the Medical Research Council (MRC) guidelines for clinical trials102–104 and the Research Governance Frameworks for England and Wales. 105–107 In particular we anonymised all research data and stored them securely. All research team members received general training in GCP and trial-specific training in the protocol, recruiting participants, taking blood, completing CRFs, conducting assessments, and reporting AEs. We also developed a fieldworker’s manual to maintain consistency between sites.

Independent Trial monitoring

We established a TSC and a DMEC to oversee FolATED through biannual meetings or telephone conferences. The TSC comprised an independent chair, three independent members, and five members of the FolATED trial management team. The DMEC comprised an independent chair and two independent members, with the trial statistician in attendance (see Appendix 2).

Approvals

The Multicentre Research Ethics Committee (MREC) for Wales gave initial ethical approval on 6 November 2006, and the Medicines and Healthcare products Regulatory Agency (MHRA) issued the Clinical Trial Authorisation (CTA) on 21 December 2006. Appendix 3 lists the dates of approvals for individual centres.

Summary of changes to the project protocol

Appendix 4 lists all substantial changes to the protocol approved by TSC, DMEC, MHRA, MREC, and primary and secondary care R&D Departments.

Statistical analysis plan

Before starting analysis we developed our analysis plan for approval by the DMEC (see Appendix 5).

Trial populations

Imputation of missing data for ‘analysis by treatment allocated’

We excluded participants without follow-up data from the primary analysis ‘by treatment allocated’. For each variable we summarised missing data by reason (mainly participant withdrew; questionnaire not returned; page missing; item missing). Where < 10% of data were missing, we treated them as if they were missing completely at random (MCAR). 108 If > 10% of data were missing, we explored the missing data and tabulated them by the stratification variables both as reported at randomisation and as validated after quality assurance; by participant demographics; and by other important covariates. Rather than exclude participants missing some data, we chose to impute these data (see Appendix 5).

Data description and transformation

Initially we summarised data by allocated treatment and centre. Rather than test for statistical differences between allocated groups at baseline, we adjusted for any imbalance by analysis of covariance. Our analysis plan assumed that residual variation from our statistical models follows Normal distributions. This is a robust assumption in the sense that only a substantial deviation would invalidate each analysis. So we plotted and reviewed residual distributions. As none of these was substantial, we did not need to transform data to improve consistency with the assumption of Normality. Hence we present all data as collected.

Methods for analysing outcomes

All of our statistical tests were two-sided with a significance level of 5%.

Introduction

There are no economic evaluations of folic acid in managing depression. If shown to be effective, however, folic acid could represent a highly cost-effective treatment option. As it costs only 3 pence a day, the main cost drivers are likely to be hospital admissions, use of health and personal social services, ADM, and other aspects of care which might change following therapeutic benefit.

The aim of the economic analysis was therefore to assess whether the addition of 5 mg folic acid, once daily for 12 weeks, in new or existing users of antidepressants, represents a cost-effective use of healthcare resources. We limited this analysis to trial-generated estimates of costs and benefits, without modelling wider effects.

Perspective

In line with the NICE reference case,112 we adopted the costing perspective of the National Health Service (NHS) and Personal Social Services (PSS). We estimated all costs in 2009–10 prices.

Data sources

Health outcomes

The primary measure of health outcome for economic analysis was the quality-adjusted life-year (QALY), estimated from the EQ-5D questionnaire administered at baseline and 4, 12 and 25 weeks. This assesses health-related quality of life on five dimensions – mobility, self-care, usual activities, pain-discomfort and anxiety-depression. The three possible responses on each dimension are ‘no problems’, ‘moderate problems’ and ‘extreme problems’. We converted participants’ responses into a single, preference-based utility using on the UK tariff. 121

Secondary measures of health outcome for economic analysis included the EQ-VAS, the UK Short Form Health Survey – 6 Dimensions (SF-6D) and BDI-II, all completed at the same times as the EQ-5D. The EQ-VAS is a vertical 20-cm visual analogue scale for recording participants’ rating of their current health-related quality of life. The SF-6D derives an alternative preference-based utility from SF-12 responses using weights estimated from a sample of the general population by the standard gamble technique. 122

For the cost-effectiveness analysis, we calculated the number of weeks free from moderate or severe depression (defined as a BDI-II score < 13)123 by statistical inference from the observed distribution of BDI-II scores, assuming linear interpolation between time points (baseline, 4, 12 and 25 weeks).

Data analysis

We combined data on costs and outcomes in the ‘treatment allocated’ population (see Statistical methods, Imputation of missing date for analysis by treatment allocated, Missing items within a subscale, above) over 25 weeks to estimate an incremental cost-effectiveness ratio (ICER) for comparison against accepted thresholds. Primary analysis was on the data set in which we had imputed missing data in the manner described above (see Statistical methods, Trial populations).

DNA isolation

We extracted genomic DNA from 5 ml of whole blood using the Chemagic Magnetic Module (MSM) 1 system according to the manufacturer’s protocol (Chemagen Biopolymer-Technologie AG, Baesweiler, Germany). We eluted samples in 500 µl of the manufacturer’s elution buffer.

Single nucleotide polymorphism selection

We compiled a list of 25 candidate genes127 associated with either the one-carbon folate or methionine synthesis pathways. We identified 48 non-synonymous single nucleotide polymorphisms (SNPs) within these genes from the Single Nucleotide Polymorpism Database [dbSNP;134 www.ncbi.nlm.nih.gov/SNP (accessed May 2008)] and selected for analysis those with minor allele frequency greater than 5%. We included a further 100 SNPs from the HapMap (http://hapmap.ncbi.nlm.nih.gov) population of Utah residents with ancestry from northern and western Europe which, when analysed by Haploview software version 402 (www.broadinstitute.org/haploview/haploview), tagged at least one other SNP. We added a 19 base pair (bp) deletion polymorphism of intron 1 of the dihydrofolate reductase gene (DHFR) and a 28 base pair double or triple tandem repeat polymorphism of the thymidylate synthase (TYMS) gene because both have been extensively characterised in clinical studies. 127

Genotyping

We designed multiplex assays for the MALDI-TOF-based Sequenom iPLEX system (Sequenom Inc., San Diego, CA, USA) using the software at https://mysequenom.com/default.aspx. We included 140 SNPs in five-assay plexes ranging from 11 to 35 SNPs in size. We excluded five SNPs which we could not incorporate in assays because of proximal nucleotide sequence constraints and another three SNPs which we could not include at a minimum plexing level of > 10 SNPs per assay (see Appendix 6, Table 35).

We genotyped patients for these 140 SNPs according to the manufacturer’s protocol using 40 ng/reaction genomic DNA. We obtained sequence-specific polymerase chain reaction (PCR) and extension reaction oligonucleotides from Metabion GmbH (Martinsried, Germany). Table 36 of Appendix 6 defines the corresponding primer and probe sequences.

We typed the 19 bp deletion polymorphism from the dihydrofolate reductase gene (DHFR) and the 28 bp tandem repeat polymorphism from the thymidylate synthase (TYMS) gene using previously published protocols and PCR primer sequences21,22 with minor modification. Briefly the 25-µl PCR reaction consisted of 20 ng genomic DNA, 5 pmol each of primer and 18 µl 1.1× ReddyMix™ PCR mastermix (Abgene Ltd, Epsom, UK).

We resolved all PCR products with ethidium bromide staining on a 3% agarose gel. For the DHFR 19 bp deletion, a 92 bp product identified the deletion allele and a 113 bp product identified the insertion allele. For the TYMS tandem repeat a 144 bp product distinguished the triple repeat allele from the double (116 bp).

We undertook all genotyping with 10% of DNA samples duplicated as well as positive and negative controls to confirm genotype calling accuracy and concordance.

Genetics outcomes

For this genetic sub-study, the primary outcome was self-rated symptom severity on the Beck Depression Inventory (BDI-II) at baseline, and 4, 12 and 25 weeks, consistent with the trial’s primary outcome. Secondary outcomes were:

1. symptom severity rated by clinicians on the MADRS and the CGI of change (also at baseline and 4, 12 and 25 weeks)

2. mental and physical aspects of self-reported health status on the SF-12 (ditto)

3. side effects assessed by the UKU side effects scale and reported AEs (ditto), and

4. proportion of patients with self-rated moderate or severe depression (i.e. BDI-II score ≥ 19) at 25 weeks.

Genetics statistical methods

Before the analyses of association, we tested each SNP for Hardy–Weinberg Equilibrium, and excluded those found to deviate at a significance level of 0.1%. We also excluded SNPs which did not meet all our genotype quality criteria:

1. a. minor allele frequency greater than 1%

2. b. genotyping rate greater than 95% per SNP, and

3. c. samples more than 90% of SNPs called.

We fitted three mixed models to all five outcomes for each included SNP. The first (‘baseline model’) included the baseline value of the outcome, covariates representing the three time points (4, 12 and 25 weeks), three validated stratifying variables – centre, type of antidepressant, new or continuing patient – and treatment received, that is whether participants supplemented their medication with folic acid or not. We also tested non-genetic factors known to be generally associated with outcome (age, gender, marital status, employment status, number of dependents, smoking and alcohol consumption, previous counselling and treatment adherence as assessed by the Morisky scale) for univariate association with each outcome and included them in the model if the significance level was less than 10%.

The second (‘SNP’) model was identical to the first with the addition of the SNP as covariate. The third (‘interaction’) model was identical to the second with the addition of interaction between SNP and treatment received. To test for statistical significance of SNP main effects, we used likelihood ratio tests to compare the specific SNP model with the baseline model. To test for statistical significance of the SNP-folated interaction, we again used the likelihood ratio test to compare the specific interaction model with the specific SNP model. Each test tried two models – one making no assumption about the underlying mode of inheritance, the other assuming an additive mode of inheritance – and used the lower significance level for each SNP.

To take account of the multiple comparisons due to four tests on each of more than 100 SNPs, we estimated the false discovery rate (‘FDR’) for each comparison, and treated FDRs less than 5% as statistically significant associations. We used the statistical software packages: R;111 PLINK version 1.07 from http://pngu.mgh.harvard.edu/~purcell/plink/; and PASW version 18135,136 for these analyses.

Introduction

At the start of the FolATED trial understanding of the benefits of folate augmentation of ADM stemmed from a recent Cochrane systematic review. 50,64 The authors concluded that there was limited evidence that adding folate to ADM was helpful, and recommended larger trials to test this hypothesis thoroughly. That recommendation led directly to the funding of FolATED.

Method

Centre differences in recruitment patterns

North West Wales received 47% of the referrals to the trial and randomised 50% of the final sample. North East Wales and Swansea received and randomised very similar proportions of the total – 27% and 26% respectively of referrals received and 25% of the randomised sample each). Table 4 summarises these flows by centre.

Loss to follow-up

We randomised 475 participants (excluding four randomised in error): 237 to receive folic acid and 238 to receive placebo. In the folic acid group 15 people withdrew and 26 were lost to follow-up by 25 weeks. In the placebo group 18 people withdrew and 32 were lost to follow-up by 25 weeks. Table 5 shows the reasons for loss at each stage.

Participant drop-out and missing data

There were no follow-up data for 35 randomised participants; 18 withdrew before the 4-week follow-up (8 folic acid group, 10 placebo group) and 17 did not attend any appointments (6 folic acid group, 11 placebo group). Thus 14 dropped out of the folic acid group and 21 out of the placebo group. We removed these from further analysis. Therefore 440 entered the main analysis, 223 from the folic acid group and 217 from the placebo group. If these evaluable participants missed follow-up appointments, we imputed their data in accordance with Chapter 2, Statistical methods, Trial populations, above (Table 6). However we imputed no baseline measures.

Of the 440 evaluable participants 36 (8%) missed follow-up at 4 weeks, 40 (9%) at 12 weeks, and 56 (13%) at 25 weeks. Thus 10% of follow-up assessments were missing. Sixty-two participants missed one assessment: 33 at 4 weeks, 6 at 12 weeks and 23 at 25 weeks. Thirty-five participants missed two assessments: 2 at 4 and 12 weeks; 1 at 4 and 25 weeks; and 32 at 12 and 25 weeks. Thus 343 (78%) participants undertook all three assessments.

For the eight main outcome measures (BDI-II, MADRS, CGI, SF-12, EQ-5D, EQ-VAS, MINI and Morisky) a full data set over the four times would have comprised 102,520 data items. Only 2572 (2.5%) were missing, of which 2476 (2.4%) items were in missing subscales or times while 96 (0.1%) items were isolated missing values within otherwise complete subscales. Reassuringly there was no hint of significant differences between trial groups in either respect.

The missing item rate for seven of these measures is fairly consistent: BDI-II, EQ-5D, EQ-VAS, and MINI all 2.3%; MADRS 2.4%; SF-12 3.2% and CGI 3.3%. In contrast the Morisky data had a missing item rate of 10%, not explained by being collected only at 12 weeks. Predictably the pattern of missing data differed very significantly between centres: North West Wales, the best recruiting centre, missed more 4-week data than other centres, but fewer at 25 weeks. This pattern was due to heavy workload early in the trial when several 4-week appointments were missed; fortunately routine monitoring recognised and rectified the problem. Reassuringly there was no significant difference between centres in the proportion followed up.

Validation of stratification variables

The trial used five variables to balance allocation between treatment groups during dynamic randomisation: centre; sex; antidepressant type; whether the participant was new or continuing on treatment; and whether the participant had ever received counselling for depression. Not surprisingly given the speed of the recruitment process, data validation identified a few inconsistencies in these data. However we found no misclassification of centre, sex or counselling, and minimal misclassification of antidepressant type or new patient (Table 7).

Differences between centres at baseline

The management of the FolATED trial included a monthly telephone conference between the three clinical centres and NWORTH, the coordinating Clinical Trials Unit. These conferences soon identified substantial differences between centres, notably in psychiatric practice and recruitment policy. Judging that these differences would enhance the generalisability of the trial provided the conduct of the research was consistent across centres, we pursued such consistency, notably by maintaining a rigorous fieldwork handbook and arranging regular inter-centre training.

Table 9 shows that participants differed significantly between centres, notably in:

1. Mean age Those in North West Wales were on average more than 3 years younger than those in the other centres.

2. Numbers of dependent children Half of those in North West Wales had children, but only 30% of those elsewhere.

3. Employment status Swansea had many more students, while North East Wales had more unemployed.

4. Smoking rates Nearly half of those in Swansea smoked, but only 30% of those elsewhere.

However there was no significant difference in mean BDI-II scores across centres; or in alcohol consumption.

Another difference between centres identified by our monthly management conferences is that one centre did more to optimise medication than the other two, notably by using reboxetine (Edronax^®, Pfizer) to augment basic ADM. Though we plan to analyse the process and outcome of optimisation in detail, we have confirmed that this reboxetine augmentation did not differ between allocated groups; hence there was no danger of bias from differential optimisation.

Baseline demographic profile

Predictably our randomisation algorithm generated similar treatment groups (Tables 10 and 11).

Baseline clinical profile

Is folic acid clinically effective?

Primary clinical effectiveness outcomes

The primary clinical effectiveness outcome measure was self-rated symptom severity as measured by BDI-II scores over 25 weeks, as summarised by the AUC of mean BDI-II scores from randomisation till 25 weeks. This provided no evidence that folic acid was effective (Figure 3). The adjusted difference in AUC between folic acid and placebo was 1.09 (95% CI from –0.48 to 2.66; p = 0.17). The adjusted difference between folic acid and placebo at 25 weeks was 1.27 (95% CI from –0.99 to 3.54; p = 0.27). Similarly there was no significant difference in the proportion of patients who were depressed at 25 weeks (at least moderately, i.e. BDI-II ≥ 19): there were 126/223 (57%) depressed participants in the folate group and 118/217 (54%) in the placebo group. The adjusted OR of being depressed in the folate group compared with placebo group was 1.09 (95% CI from 0.75 to 1.59; p = 0.65).

FIGURE 3.

Estimated mean BDI-II scores over time adjusted for baseline score and stratification variables – by treatment allocated. Baseline score = observed population mean BDI-II score.

Secondary clinical effectiveness outcomes

Table 12 shows the unadjusted AUC results for all the main outcome measures, together with two-sample t-tests. The only significant result favoured the placebo in the SF-12 mental component. By convention generic outcome measures like EQ-5D and SF-12 show good health by high scores, while condition-specific outcome measures like BDI-II, CGI and MADRS show good health by scores that are low, if not zero. Thus five of the non-significant differences favoured placebo and seven favoured folate. Table 42 of Appendix 10 elaborates on Table 12 by showing the results of unadjusted two-sample t-tests for each variable at each time point separately.

TABLE 12 - Unadjusted AUC average (using values at baseline and 4, 12 and 25 weeks) of main outcomes by treatment allocated

After logarithmic transformation.

Estimated CGI scale scores in each group from the transformed model.

Ratio of estimated CGI scores and CI for the ratio.

After square-root transformation.

Estimated UKU scale scores in each group from the transformed model.

Difference significant at 5% level with effect size = 0.23.

Outcome variable	Folate	Placebo	Difference (folate minus placebo)
	Mean (SD)	Mean (SD)	Mean (SE)	95% CI	Significance
BDI-II	25.70 (10.83)	25.64 (11.43)	0.06 (1.06)	−2.02 to 2.15	0.953
MADRS	22.17 (7.59)	22.39 (8.34)	−0.22 (0.76)	−1.71 to 1.27	0.771
Euroqol
EQ-5D	0.573 (0.259)	0.591 (0.262)	−0.018 (0.025)	−0.067 to 0.031	0.476
EQ-VAS	54.58 (18.38)	54.34 (17.85)	0.25 (1.73)	−3.15 to 3.64	0.887
SF-12
SF-12 PCS	44.89 (11.13)	44.28 (11.48)	0.61 (1.08)	−1.51 to 2.73	0.571
SF-12 MCS *	31.99 (9.34)	34.09 (9.01)	−2.09 (0.88)	−3.81 to −0.37	0.017
CGI
CGI: Severity	3.60 (0.87)	3.61 (0.94)	−0.00 (0.09)	−0.17 to 0.17	0.980
CGI: Improvement	3.11 (0.86)	3.09 (0.95)	0.02 (0.09)	−0.15 to 0.19	0.794
CGI: Efficacya	0.29 (0.49)	0.30 (0.50)	−0.01 (0.05)	−0.10 to 0.08	0.826
Estimated CGI	1.34b (–)	1.35b (–)	0.99c (–)	0.90 to 1.08c
UKU
UKU: Psychic	8.21 (3.86)	8.16 (4.01)	0.05 (0.38)	−0.68 to 0.79	0.894
UKU: Neurologicd	0.73 (0.67)	0.79 (0.71)	−0.06 (0.07)	−0.19 to 0.07	0.371
Estimated UKU	0.53e (–)	0.62e (–)	(–) (–)	(–)
UKU: Autonomic	2.93 (2.22)	2.85 (2.51)	0.08 (0.23)	−0.37 to 0.52	0.723
UKU: Other	3.62 (2.68)	4.06 (3.11)	−0.44 (0.28)	−0.98 to 0.11	0.114

Area under the curve analysis adjusted for stratification variables and baseline score of the variable in question gave a very similar picture. Table 13 shows these results and reports all significant stratification and baseline variables. For most variables the baseline score and antidepressant type were significant covariates, but allocated treatment was not. As for the unadjusted AUCs the only outcome on which the allocated treatment had a statistically significant effect was the SF-12 mental component, again favouring the placebo.

TABLE 13 - Area under curve average of main outcomes adjusted for stratification variables and their own baselines – by treatment allocated

Ratio (folate/placebo) and its CI.

Difference significant at 5% level with effect size = 0.24.

Outcome variable	Difference (folate minus placebo)			Significant covariates	Significance
	Mean (SE)	95% CI	Significance
BDI-II	1.09 (0.86)	–0.48 to 2.66	0.173	Baseline scores	< 0.001
				Type of antidepressant	0.028
				Centre	0.025
MADRS	0.71 (0.62)	–0.51 to 1.93	0.252	Baseline scores	< 0.001
				Type of antidepressant	0.002
				Centre	0.026
EQ-5D	0.00 (0.017)	–0.034 to 0.033	0.982	Baseline scores	< 0.001
EQ-VAS	–0.48 (1.36)	–3.16 to 2.20	0.726	Baseline scores	< 0.001
				Type of antidepressant	0.001
SF-12 PCS	0.40 (0.60)	–0.78 to 1.59	0.501	Baseline scores	< 0.001
				Previous treatment	0.002
SF-12 MCS *	–1.97 (0.78)	–3.49 to –0.44	0.012	Baseline scores	< 0.001
				Type of antidepressant	0.001
CGI: Severity	0.05 (0.08)	–0.11 to 0.2	0.553	Baseline scores	< 0.001
				Type of antidepressant	0.001
				Centre	< 0.001
CGI: Improvement	0.04 (0.08)	–0.13 to 0.21	0.649	Type of antidepressant	0.002
				Centre	0.009
CGI: Efficacya	0.98b (0.04)	–0.89 to 1.07	0.604	Type of antidepressant	< 0.001
				Centre	< 0.001
UKU: Psychic	0.31 (0.31)	–0.3 to 0.92	0.319	Baseline scores	< 0.001
				Type of antidepressant	0.001
				Centre	0.029
UKU: Neurologic	–0.01 (0.04)	–0.10 to 0.07	0.738	Baseline scores	< 0.001
				Centre	0.008
UKU: Autonomic	–0.11 (0.15)	–0.42 to 0.19	0.458	Baseline scores	< 0.001
UKU: Other	–0.38 (0.20)	–0.78 to 0.02	0.065	Baseline scores	< 0.001
				Centre	0.002

Figures 4–11 display the remaining adjusted analyses, except those relating to the UKU, by individual time points. Though receiving folic acid or placebo does not affect the outcome of treatment in the trial, the pattern of results over time is consistent across measures: ADM achieves major benefit over the first 4 weeks, and continuing though reducing improvement over 25 weeks. In particular Figure 11 shows how the adjusted SF-12 MCSs differ between arms, with the difference favouring placebo: the mean scores diverge by 4 weeks, achieve a substantial gap by 12 weeks, and converge a little by 25 weeks. Table 14 records the statistical analyses underpinning these nine figures, together with the corresponding analyses from the UKU side effects scale.

FIGURE 4.

Estimated mean MADRS scores over time adjusted for baseline score and stratification variables – by treatment allocated. Baseline score = observed population mean MADRS score.

FIGURE 5.

Estimated mean CGI severity scores over time adjusted for baseline score and stratification variables – by treatment allocated. Baseline score = observed population mean CGI severity score.

FIGURE 6.

Estimated mean CGI improvement scores at follow up adjusted for stratification variables – by treatment allocated.

FIGURE 7.

Estimated mean CGI efficacy scores at follow up adjusted for stratification variables – by treatment allocated.

FIGURE 8.

Estimated mean EQ-5D scores over time adjusted for baseline score and stratification variables – by treatment allocated. Baseline score = observed population mean EQ-5D score.

FIGURE 9.

Estimated mean EQ-VAS scores over time adjusted for baseline score and stratification variables – by treatment allocated. Baseline score = observed population mean EQ-VAS score.

FIGURE 10.

Estimated mean SF-12 Physical Component Scale scores over time adjusted by baseline score and stratification variables – by treatment allocated. Baseline score = observed population mean SF-12 Physical Component Scale score.

FIGURE 11.

Estimated mean SF-12 Mental Component Scale scores over time adjusted for baseline score and stratification variables – by treatment allocated. Baseline score = observed population mean SF-12 Mental Component Scale score.

TABLE 14 - Outcomes over time adjusted for significant stratification variables and their own baselines by treatment allocated

After logarithmic transformation.

After square-root transformation.

Difference significant at 5% level with effect size = 0.25 in favour of placebo.

Difference significant at 5% level with effect size = 0.44 in favour of folic acid.

Outcome variable and time point	Difference (folate minus placebo)			Significant covariates	Significance
	Mean (SE)	95% CI	Significance
BDI-II
BDI-II (4 weeks)	0.45 (0.91)	–1.34 to 2.23	0.623	Baseline scores	< 0.001
BDI-II (12 weeks)	1.52 (1.10)	–0.64 to 3.69	0.168	Baseline scores	< 0.001
				Centre	0.013
BDI-II (25 weeks)	1.27 (1.15)	–0.99 to 3.54	0.270	Baseline scores	< 0.001
				Type of antidepressant	0.004
				Centre	0.011
MADRS
MADRS (4 weeks)	1.02 (0.75)	–0.45 to 2.49	0.174	Baseline scores	< 0.001
MADRS (12 weeks)	1.14 (0.89)	–0.62 to 2.89	0.204	Baseline scores	< 0.001
				Type of antidepressant	0.012
MADRS (25 weeks)	0.02 (0.95)	–1.84 to 1.88	0.981	Baseline scores	< 0.001
				Type of antidepressant	0.003
EQ-5D
EQ-5D (4 weeks)	–0.028 (0.024)	–0.074 to 0.017	0.247	Baseline scores	< 0.001
				Previous treatment	0.005
EQ-5D (12 weeks)	0.010 (0.024)	–0.037 to 0.057	0.681	Baseline scores	< 0.001
EQ-5D (25 weeks)	0.019 (0.024)	–0.027 to 0.065	0.450	Baseline scores	< 0.001
EQ-VAS
EQ-VAS (4 weeks)	0.15 (1.74)	–3.28 to 3.58	0.932	Baseline scores	< 0.001
				Type of antidepressant	0.004
EQ-VAS (12 weeks)	–1.83 (1.90)	–5.57 to 1.91	0.337	Baseline scores	< 0.001
				Type of antidepressant	0.026
EQ-VAS (25 weeks)	1.11 (2.01)	–2.85 to 5.07	0.582	Baseline scores	< 0.001
				Type of antidepressant	0.002
				Previous treatment	0.018
SF12 PCS
SF-12 PCS (4 weeks)	1.45 (0.78)	–0.08 to 2.98	0.064	Baseline scores	< 0.001
				Previous treatment	0.008
SF-12 PCS (12 weeks)	–0.28 (0.83)	–1.91 to 1.35	0.733	Baseline scores	< 0.001
				Previous treatment	0.019
				Centre	0.063
SF-12 PCS (25 weeks)	0.76 (0.83)	–0.87 to 2.38	0.360	Baseline scores	< 0.001
				Previous counselling	0.080
				Previous treatment	0.009
SF-12 MCS
SF-12 MCS (4 weeks)	–1.28 (0.99)	–3.22 to 0.67	0.198	Baseline scores	< 0.001
SF-12 MCS (12 weeks) *	–2.95 (1.12)	–5.16 to –0.75	0.009	Baseline scores	< 0.001
				Type of antidepressant	0.002
SF-12 MCS (25 weeks)	–1.93 (1.21)	–4.30 to 0.45	0.112	Baseline scores	< 0.001
				Type of antidepressant	0.005
				Previous counselling	0.02
CGI: Severity
CGI: Severity (4 weeks)	0.06 (0.09)	–0.12 to 0.23	0.517	Baseline scores	< 0.001
				Type of antidepressant	0.008
CGI: Severity (12 weeks)	0.12 (0.11)	–0.09 to 0.32	0.276	Baseline scores	< 0.001
				Type of antidepressant	0.035
				Centre	< 0.001
CGI: Severity (25 weeks)	–0.06 (0.12)	–0.30 to 0.18	0.616	Baseline scores	< 0.001
				Type of antidepressant	< 0.001
				Centre	< 0.001
CGI: Improvement
CGI: Improvement (4 weeks)	0.14 (0.10)	–0.05 to 0.33	0.139	Type of antidepressant	0.030
CGI: Improvement (12 weeks)	0.06 (0.11)	–0.16 to 0.28	0.594	Type of antidepressant	0.052
CGI: Improvement (25 weeks)	–0.07 (0.13)	–0.32 to 0.18	0.596	Type of antidepressant	0.005
				Centre	0.012
CGI: Efficacy
CGI: Efficacy (4 weeks)a	–0.10 (0.05)	–0.20 to 0.01	0.079	Type of antidepressant	0.003
				Centre	< 0.001
CGI: Efficacy (12 weeks)a	0.04 (0.06)	–0.08 to 0.15	0.560	Type of antidepressant	0.007
				Centre	0.005
CGI: Efficacy (25 weeks)a	–0.07 (0.06)	–0.19 to 0.05	0.272	Type of antidepressant	0.004
				Centre	< 0.001
UKU: psychic
UKU: psychic (4 weeks)	0.81 (0.47)	–0.11 to 1.73	0.084	Baseline scores	< 0.001
UKU: psychic (12 weeks)	0.08 (0.47)	–0.84 to 1.00	0.869	Baseline scores	< 0.001
				Type of antidepressant	0.005
				Centre	0.003
UKU: psychic (25 weeks)	0.37 (0.46)	–0.54 to 1.28	0.426	Baseline scores	< 0.001
				Type of antidepressant	0.002
				Centre	0.013
UKU: neurologic
UKU: neurologic (4 weeks)b	–0.00 (0.06)	–0.12 to 0.12	0.989	Baseline scores	< 0.001
				Centre	< 0.001
UKU: neurologic (12 weeks)b	–0.03 (0.06)	–0.16 to 0.10	0.642	Baseline scores	< 0.001
				Centre	0.039
UKU: neurologic (25 weeks)b	–0.02 (0.07)	–0.15 to 0.12	0.821	Baseline scores	< 0.001
UKU: autonomic
UKU: autonomic (4 weeks)	–0.13 (0.22)	–0.55 to 0.30	0.565	Baseline scores	< 0.001
				Centre	0.004
UKU: autonomic (12 weeks)	–0.17 (0.23)	–0.63 to 0.28	0.460	Baseline scores	< 0.001
UKU: autonomic (25 weeks)	0.03 (0.24)	–0.43 to 0.49	0.907	Baseline scores	< 0.001
UKU: other
UKU: other (4 weeks)	–0.11 (0.28)	–0.65 to 0.44	0.694	Baseline scores	< 0.001
				Centre	< 0.001
UKU: other (12 weeks) **	–0.63 (0.30)	–1.22 to –0.04	0.037	Baseline scores	< 0.001
UKU: other (25 weeks)	–0.39 (0.30)	–0.98 to 0.20	0.195	Baseline scores	< 0.001
				Previous treatment	0.047
				Gender	0.018

The absence of the binary variable ‘Previous treatment?’ from the significant covariates in all but the SF-12 Physical Component Score of the 13 rows of Table 13, and from all but four non-psychiatric rows of the 39 rows of Table 14, confirms that this does not seem to predict the clinical outcome of ADM or of adjunctive folate.

Tables 43–46 of Appendix 10 complement Table 14, specifically the BDI-II row, by reporting the analogous logistic regression analyses for two levels of response to treatment – 50% or full improvement – both at 12 and at 25 weeks, thus illustrating the primary analysis in clinical terms.

Side effects, adverse events and suicidality

Adherence to trial medication

We assessed adherence to the trial medication at 12 weeks in four ways: scores on the Morisky Questionnaire; counting returned trial medication; serum folate levels; and homocysteine levels. We followed the published instructions for calculating respondents’ scores on the Morisky Questionnaire. We defined the biochemical criteria for adherence to folic acid treatment as: serum folate at 12 weeks greater than 15 µg/ml, and reduction of at least 15% in serum homocysteine between baseline and 12 weeks. Table 17 shows no significant difference between randomised groups in Morisky score or tablet count. Of the biochemical criteria, serum folate shows better adherence than homocysteine.

Sensitivity analyses of clinical effectiveness outcomes

Biochemistry analysis

Resource use

Table 26 presents participants’ use of NHS and PSS resources over their 25 weeks in the trial. The majority of participants, 78% and 81% in folic acid and placebo groups respectively, visited their GP on at least one occasion. The mean number of visits was 3.3 and 3.9. Many participants reported telephone contact with their GPs – 21% of those in the folic acid group and 27% in the placebo group; and visits to practice nurses – 35% and 39% respectively. Despite the NICE recommendation that people with moderate to severe depression should receive psychological therapy, only 7% of intervention participants and 6% of controls did so (0.4 times on average). However 22% of those randomised to folic acid consulted psychiatrists at hospital clinics (0.6 times on average), compared with 29% of control participants (0.8 times on average). Other hospital visits were reported by similar numbers of participants – 22% and 25% respectively; and only 6% of patients in each group were admitted to hospital. All patients received prescribed medication during the trial, with a mean of 6.5 and 7.2 antidepressant items in folic acid and placebo groups respectively. Total prescribing was comparable at 21.0 and 21.9 items respectively. Participants’ use of social services was low. Thus there were no significant differences in resource use between the two groups.

The mean percentage of responses missing from the resource use questionnaire, across all questions and time points, was 10.1% in the folic acid group and 11.0% in the control group. The level of missing data was not related to the treatment allocation at any time point (Student’s t-test: p > 0.1 for each point). The lowest rate of missing data was in response to consultations at general practices (8.1% across both treatment groups) while the highest rate of missing data related to contact with health visitors (23.5%). The high response rate is in marked contrast to the AHEAD study, which used the same questionnaire, but required patients to return their completed forms by mail, leaving 73.8% of questionnaires incomplete. 138

Costs

Table 27 presents unadjusted costs incurred during the 25 weeks of follow-up. There were no statistically significant differences in categorised or total cost between treatment groups, although there was a tendency towards lower costs in the control group for all categories. Psychiatric services in hospital or community services took more than half the total cost – £797 in the folic acid group, and £886 in the placebo control group. Participants’ attendances at hospital clinics for consultations with psychiatrists was the main driver of this. The costs of prescribed medicines were the second highest cost category at £240 and £257 in intervention and control groups respectively. Antidepressants accounted for around 30% of total medication costs.

There were differences in baseline costs, evident from the resource use questionnaire relating to the 3 months before baseline visits. Baseline costs in the folic acid group were £514 compared with £746 in the placebo group (a difference of £232, 95% CI from –£9 to £487). Although baseline costs do not contribute to the total costs, imbalances at baseline may reflect an imbalance in patient or disease characteristics. Hence they may bias cost estimates, particularly if previous use of health and social care predicts future use. The primary cost analysis therefore corrected for baseline differences by regression. 125

Adjusting for these differences at baseline reduced the mean difference in total costs from £309 (95% CI from –£155 to £764) to £48 (95% CI from –£292 to £389), still not significant.

Health outcomes

Figure 12 shows the distribution of EQ-5D scores by time, treatment group and dimension. At baseline the majority of patients described themselves as having either moderate or severe problems in relation to anxiety or depression (97% in both groups), pain or discomfort (61% and 59% for folic acid and placebo groups respectively), and usual activities (77% and 80% respectively). Improvements in the states of anxiety or depression (to 75% in both groups), and ability to perform usual activities (to 62% and 55% respectively) were evident between baseline and 25 weeks. The corresponding changes in unadjusted mean utilities from baseline to 25 weeks were 0.481 to 0.605 for folic acid and 0.514 to 0.607 for placebo.

FIGURE 12.

Participants responding to each dimension of the EQ-5D – by time and treatment allocated. n shows the number of completed responses within each treatment group. Level 3 represents the most severe problems.

Table 28 presents the numbers of gross and net QALYs gained, as measured by the EQ-5D (primary analysis), EQ-VAS and SF-6D. There were no statistically significant differences between treatment groups. Similarly there were no differences in either outcome measure in the cost-effectiveness analyses – the AUC for BDI-II scores and the number of ‘depression-free weeks’ (when participants’ BDI-II scores were less than 13). The number of participants reporting time free from depression was low – 18 (11%) in the folate group and 23 (15%) in the placebo group at 4 weeks, rising to 56 (33%) and 63 (41%), respectively, at 25 weeks.

Cost-effectiveness and uncertainty

In the primary analysis after baseline adjustment following Manca et al. ,126 folic acid is on average £48 less expensive than the placebo group, and more effective by 0.0012 QALYs. As those findings put folic acid in the south-east quadrant of the cost-effectiveness plane (Figure 13) it is therefore the dominant strategy. However there is considerable uncertainty surrounding these estimates, shown by the distribution of costs and QALYs over all four quadrants of the cost-effectiveness plane. The estimated probability of folic acid saving costs is 64%, and that of it gaining QALYs is 55% (Table 29). The resulting cost-effectiveness acceptability curve (Figure 14) shows that the probability of folic acid being cost-effective is 0.62 at the threshold of £20,000 per QALY and 0.61 at the £30,000 threshold (see Table 29).

FIGURE 13.

Cost-effectiveness plane showing joint distribution of costs and QALYs. Note: Confidence ellipses represent the 5%, 50% and 95% levels of confidence.

FIGURE 14.

Cost-effectiveness acceptability curve showing probability of folate being cost-effective – by cost-effectiveness threshold (£ per QALY). WTP, willingness to pay.

As this ICER is close to the origin of the cost-effectiveness plane, comparison with other QALY outcomes is labile. For example QALYs derived from both the EQ-VAS and SF-6D located the ICER in the south-west quadrant, where folic acid is both less effective and less expensive than placebo.

The interpretation of cost-effectiveness results suffers from lack of a benchmark for economic efficiency. The unstable direction of differences in mean effect exacerbates this. The primary clinical outcome of area under the BDI-II curve shows folic acid dominating placebo, being more effective and less costly. However the cost per depression-free week averted suggests that folic acid is less effective by an average of 6 depression-free days over 25 weeks, though less expensive. In short the proximity of all possible criteria to the origin of the cost-effectiveness plane, and associated uncertainty suggests that folic acid is no more effective, but no more expensive than placebo.

Introduction

Of the 440 patients included in the main analysis by treatment allocated, we excluded from genetic analysis: five with low genotype call rates; 14 non-Caucasians; and 38 who did not consent to the genetic study. Thus we included 383 patients in genetic analysis.

Of the 142 variants genotyped, 38 were omitted as they failed to meet minimum inclusion criteria. These included 19 SNPs owing to quality controls issues with the call rate of the specific SNP assay; 13 SNPs due to a minor allele frequency < 0.01, and six with a Hardy–Weinberg p-value of < 0.0001. In total 104 genetic variants were carried forward for analysis.

Results of the analysis of association between each SNP and each of the seven outcome measures (BDI-II, MADRS, CGI1 severity of illness, EQ-5D, EQ-5D visual analogue score (VAS), SF-12 mental, and SF-12 physical) are given in Table 30. Two associations gave a FDR < 0.05. Results of assessing for the statistical significance of a SNP-treatment group interaction term for each SNP and each outcome are given in Table 31 – for this only one SNP gave a FDR < 0.05.

Statistically significant main single nucleotide polymorphism effects

The rs11627525 SNP in the methylenetetrahydrofolate dehydrogenase (NADP⁺ dependent) 1 (MTHFD1) gene was associated with MADRS (p = 0.0004, FDR = 0.0467). This association was not replicated with any of the other six outcome measures analysed (p > 0.05; FDR > 0.05).

A plot of mean MADRS at each study time point compared with rs11627525 genotype, stratifying by genotype [CC wild-type group and a combined CT and TT group since the homozygote variant allele frequency was so low (n = 4)], is given in Figure 15. At baseline, there was no difference observed [28.3 ± 0.4 (CC) compared with 29.0 ± 0.7 (CT/TT)]. The data suggest a more dramatic reduction in MADRS between baseline and 12 weeks between CC individuals (mean difference between baseline and 12 weeks: 6.11) and the combined CT/TT group (mean difference between baseline and 12 weeks: 9.82).

FIGURE 15.

Montgomery–Åsberg Depression Rating Scale depression scores by MTHFD1 rs11627525 genotype. Note: Graphs plot mean MADRS scores at each point with vertical bars showing standard errors of means.

The rs588458 SNP in the folate hydrolase 1 (FOLH1) gene was associated with EQ-5D (p = 0.0003, FDR = 0.0337). This association was not replicated with any of the other six outcomes (FDR > 0.05), although both EQ-VAS (p = 0.0414) and SF-12 mental (p = 0.0140).

A plot of mean EQ-5D scores at baseline, 4 12 and 25 weeks, stratified into the three genotype groups, is given in Figure 16a. At baseline, TT carriers had a mean EQ-5D of 0.534 (± 0.024) – the mean scores for heterozygotes (TC carriers) were 0.049 lower, while they were 0.102 points lower for CC carriers. Similar differences were observed at 4 weeks with TC carriers being 0.078 lower than TT (0.625 ± 0.023) with CC carriers being 0.140 lower than TT. A significant difference in score was apparent at 12 weeks where the mean score (± SE) for TT carriers was 0.645 (± 0.023) compared with 0 for TC genotype (0.058 lower) and (0.140 lower than baseline) for individuals with the CC genotype. Similarly, at 25 weeks, mean EQ-5D for TT individuals (0.673 ± 0.023) was significantly higher (0.09 points) than TC which, in turn, was significantly higher (0.192 points) than for CC genotype). The overall trajectory of improvement (by EQ-5D) appears to stabilise first in the CC group (at around 4 weeks); then the TC group at around 12 weeks; whilst the wild-type TT group appear to continue to improve.

FIGURE 16.

Quality-of-life-scores by FOLH1 rs588458 genotype. Note: Graphs plot mean quality-of-life scores at each point with vertical bars showing standard errors of means.

Plots for EQ-VAS and SF-12 Mental Component Scale (MCS) against rs588458 genotype are also provided (see Figure 16b and c, respectively) since the associations with rs588458 in the main SNP effect model gave p < 0.05. For mean EQ-VAS, there was a significant divergence between genotypes at 12 weeks with TT genotype (62.2 ± 1.9) significantly higher (difference of means = 4.5) than CT (57.7 ± 1.8) and CC (52.6 ± 2.8) (difference in means = 9.6). No difference was observed at any earlier time points. Mean SF-12 did not show an obvious divergence between genotypes at any of the time points though a moderate difference was seen at 12 weeks between TT (37.1 ± 1.0) and TC (34.3 ± 1.0).

Statistically significant single nucleotide polymorphism–treatment interaction

For the model incorporating the treatment interaction term, the association of the rs17102596 SNP in the methionine adenosyltransferase I alpha (MAT1A) gene and SF-12 mental status was statistically significant (p = 0.0002, FDR = 0.0255).

A plot of mean SF-12 Mental Component Score at each study time point stratified by genotype and treatment group (folic acid or placebo) is given in Figure 17. At baseline, no difference in SF-12 is observed between TT (25.0 ± 0.8) and TC + CC (25.2 ± 1.2) groups in the folate arm of the trial. In the placebo arm, there was a modest difference at baseline between TT (25.8 ± 0.9) and TC + CC (28.2 ± 1.3). In both arms of the trial, wild-type (TT) individuals showed a similar trajectory for increase in SF-12 with similar mean SF-12 at 25 weeks (folic acid = 25.0 ± 0.8; placebo = 25.8 ± 0.9). In individuals with TC or CC genotypes in the placebo group, there was a marginally steeper trajectory of the ΔSF-12 resulting in a significantly higher mean value at 12 weeks (39.6 ± 1.7) compared with TT individuals though much of this difference could be explained by the variability in baseline measure. This is considerably higher than those individuals with TC or CC genotype receiving folic acid (31.2 ± 1.7) (difference in mean = 8.4) whose Δ mean SF-12 trajectory stabilises after 4 weeks. The SNP effect model was run on the two arms of the trial separately (data not shown) with resulting statistically significant difference in both the folate group (p = 0.02) and the placebo group (p = 0.00006).

FIGURE 17.

SF-12 Mental Component Scores by MAT1A rs17102596 genotype – by treatment allocated. Note: Graphs plot mean SF-12 MCS scores at each point with vertical bars denoting standard error of the mean.

As three other outcomes measure (BDI-II, MADRS, and CGI severity of illness) gave p < 0.05 for tests of association with rs17102596-treatment interaction (see Table 31), these too were plotted (not shown). In all three instances the pattern demonstrated with SF-12 was not replicated and no clear distinction in the trajectory of the improvement determined by each outcome was seen.

Recruitment and baseline characteristics of the comprehensive cohort

Of the 635 people who originally consented to take part, we could not randomise 156 (see Figure 2). Of these 68 dropped out between screening and randomisation, and another 42 at the randomisation interview, 36 because their BDI-II scores were too low for the trial. The remaining 46 people continued as ‘residual’ members of the comprehensive cohort alongside the 475 properly randomised into the trial. Table 3 tabulates the reasons for losses by stage of recruitment; and Appendix 7 does so by centre. Table 32 compares the baseline characteristics of the ‘residual’ cohort with those of trial completers: though residual members included more economically inactive females, the most notable difference is that, while Bangor recruited most to the trial, Wrexham recruited most to the residual cohort.

Methods of analysis and results from the residual cohort

We followed the residual cohort like trial participants, but limited data collection to BDI-II at baseline, 4, 12 and 25 weeks. We also used similar methods of analysis. As the findings of fully imputed and complete case analyses were very similar, Table 33 displays the former. Though both residual and control groups show steady improvement in BDI-II scores from baseline over 25 weeks, the residual members start slower but finish with greater net improvement. We attribute this to the delayed effects of therapy to correct known deficiencies in B₁₂ or folate.

Introduction

There is general consensus in current clinical guidelines for depression that the augmentation of ADM with folate improves patient symptoms. However the supporting evidence is little more than biochemical theory and two small trials. 65,66 Indeed the authors of the current Cochrane systematic review concluded that there was limited evidence that adding folate to ADM was helpful, and recommended a trial like FolATED to test this hypothesis thoroughly. 50

Effects of interventions

Our updated search found no additional trials that met the criteria of the Cochrane review. 50 Hence our updated review analysed three trials – FolATED with 440 analysable participants, Coppen and Bailey with 100,65 and Godfrey et al. with 24. 66 As the primary analysis of the much larger FolATED trial favours the placebo (Figure 18), our updated review reverses the findings of the Cochrane review. As higher scores on both alternative outcomes – HDRS and BDI-II – show more depression, the resulting standardised mean difference of 0.05 (95% CI from –0.11 to 0.22; p = 0.52) also favours placebo. Furthermore there was no heterogeneity between studies. Hence there is no evidence to support the use of folic acid as an adjunct to ADM.

FIGURE 18.

Updated systematic review of the effectiveness of folic acid augmentation of antidepressants.

Clinical effectiveness

FolATED shows that on average routinely adding 5 mg of folic acid to antidepressants prescribed at therapeutic dosages has no clinical benefits. The lack of any mean treatment effect is highly consistent across all outcome measures and time points. The trajectory of response across time is also consistent between measures. The one exception to this is the MCS of the SF-12, which shows a significant difference favouring placebo, especially at 12 weeks. It is not immediately clear whether this single statistically significant adverse result in a secondary outcome measure has any clinical significance. As we used many secondary measures in the trial, this may be a statistical artefact of multiple testing. Given high levels of correlation between the MCS and measures of depression;139,140 however, it seems unlikely that the MCS detected a real treatment effect not detected by the other outcome measures.

To put this disappointing finding into context, we record that the biochemical outcomes confirmed that folic acid was delivered successfully, as in other studies. 141 Fortunately the few patients who were deficient in baseline red cell folate yielded consistent evidence that augmenting ADM with folic acid improved their clinical outcomes. Furthermore folic acid appeared well tolerated, resulting in no more reports of side effects, AEs or SAEs than placebo.

Cost-effectiveness

The FolATED trial showed no significant clinical effect of 5 mg folic acid once daily for 12 weeks in new or existing users of antidepressants. The economic analysis suggested that folic acid might save costs of about £48 per patient (equivalent to about two-thirds of the cost of antidepressants). According to economic theory, one might therefore conclude that folic acid has positive net benefits, and should be recommended for use. However it seems unlikely that the prescribing of an additional, ineffective medicine would result in reduced costs, since the probability of folic acid being cost saving was not high. A more appropriate conclusion, therefore, is that the economic evaluation was unable to demonstrate cost-effectiveness. We interpret results suggestive of dominance in the primary analysis with caution, as the mean difference in effect was less than half a quality-adjusted day (or six depression-free days) over the 25 weeks of the trial. Furthermore complete case analysis and alternative methods of calculating QALYs did not contradict the principal findings given the small differences in costs and effects, and their associated uncertainty.

Although it is not generally possible for an intervention to be cost-effective when clinical effectiveness has not been established, the application of a cost minimisation analysis would be inappropriate for several reasons. 142,143 First, lack of significant effects does not confirm equivalence. 144 Second, equivalence in one clinical end point does not necessarily mean equivalence in others. Third, preference-based measures of outcome may reveal differences unrelated to the effect of the intervention on depression. Finally it is arguable that hypothesis testing is arbitrary and irrelevant to decision making because the intervention with the highest net benefit should be adopted whether or not difference in benefit reach conventional levels of statistical significance. 145

We considered it inappropriate to extend the economic analysis to evaluate the cost-effectiveness of pharmacogenetic testing, as the evidence had not supported the clinical utility of testing in relation to the predicting patients’ responsiveness to folic acid.

Genetics

Our study has identified two polymorphisms within genes of the one-carbon folate and methionine metabolism pathways associated with outcome regardless of treatment arm (a main SNP effect), and one polymorphism–treatment interaction which is associated with outcome. However the well characterised c.677C > T (rs1801133) polymorphism of the MTHFR associated in previous studies with depression risk or outcome76,77,79,128 did not modify the effect of either the antidepressant therapy or the folic acid supplementation. This seems to correlate with the study finding that folic acid and modification of homocysteine levels does not influence antidepressant outcome in treatment of moderate to severe depression.

Clinical effectiveness

Our negative outcomes contrast with positive findings in two small, selective trials. Coppen and Bailey reported a significantly greater reduction in HDRS scores in females treated with fluoxetine and 400 µg of folic acid, compared with fluoxetine and placebo. 65 They suggested that males required higher doses of folic acid. In response we selected 5 mg of folic acid for FolATED to cover the entire dose–response curve They achieved very high clinical response rates – 82% in the folic acid arm and 62% in the placebo arm – surpassing those in the general literature, because their sample was much less representative than that recruited by FolATED. For example they excluded patients with continuing episodes of depression, previous poor response to fluoxetine, or comorbid substance use. One must also question the robustness of their blinding, especially in handling blood results, whereas FolATED took great care at all stages to avoid unblinding clinicians, patients and researchers. While females treated with folic acid by Coppen showed a 21% reduction in homocysteine levels, similar to the FolATED sample, the males did not. In FolATED, however, there was no gender difference in homocysteine levels following treatment with folic acid, suggesting that we covered the dose–response curve, yet without treatment effect.

In contrast Godfrey and colleagues recruited only from secondary care, specifically patients with folate deficiency (viz. red cell folate less than 200 µg) but suffering from a wide range of psychiatric disorders including depression and schizophrenia; hence they had to adapt clinical outcomes to diagnosis. 66 This makes comparison difficult as FolATED examined the routine use of folic acid augmentation in depressed patients recruited predominantly from primary care. Though Godfrey used 15 mg of the biologically active MTHF, the gross differences between studies generate no evidence for the superiority of MTHF over folic acid.

Hence, after updating the systematic review50 by including FOLATED, we found no evidence to support the previous suggestion that folic acid could improve the clinical benefits of ADM.

Cost-effectiveness

Our economic evaluation showed that folic acid, clinically ineffective, could not be a cost-effective use of resources. With no other economic evaluations of folic acid in managing depression, we assessed external validity in comparison with other studies of depression. For example the AHEAD trial pragmatically compared TCAs, SSRIs and lofepramine in UK primary care, and provided the prototype for our resource use questionnaire. Though AHEAD followed patients for 12 months, their rates of contact with healthcare professionals was comparable pro rata with FolATED findings. AHEAD trial participants visited GPs on 9.1 occasions over 12 months, compared with 3.6 over 25 weeks in FolATED; the mean length of inpatient stay was 1.05 days in AHEAD, compared with 0.58 in FolATED. With one exception we consider our participants’ use of NHS and PSS resources to be generally representative of UK practice, and typical of locations other than the recruiting sites. The exception lies in divergence from the NICE recommendation that people with moderate to severe depression receive psychological therapy, as only 6.5% of participants did so. Fortunately there is no evidence that greater use of psychotherapy, engendered for example by the English ‘Improving Access to Psychological Therapy’ programme would have changed any of our insignificant findings. As our genetic analysis found no support for pharmacogenetic testing in this field, that too could not be a cost-effective use of resources.

Genetic implications

Our study has identified three polymorphisms associated with modification of depression outcomes. For MTHFD1 rs11627525 and FOLH1 rs588458 this association was independent of treatment allocated; in particular the new link between EQ-5D and the rs588458 genotype needs further study. The MAT1A rs17102596 SNP modified the SF-12 Mental Component Scale through interaction between SNP and treatment.

In general, however, there is little evidence that any of the variants analysed can predict patients’ response to ADM or the efficacy of folic acid. Even the three SNPs showing association could not replicate this with any other outcome measure. Above all none of the polymorphisms analysed could achieve statistically significant modification of BDI-II, the primary outcome measure.

However we limited the variants analysed to genes within the one-carbon folate or methionine synthesis pathways. It is possible that other factors, genetic or otherwise, may influence patient response to ADM or folic acid. Thus it is conceivable that in future a whole-genome approach to the FolATED data set could identify markers of efficacy beyond those already analysed.

Biochemical interpretation

Folate is a naturally occurring B vitamin, needed in the brain to synthesise serotonin, noradrenaline and dopamine. In humans the biologically active form is 5-methyltetrahydrofolate (5-MTHF), which is derived from ingested folates and is the form taken up by cells and transported to the cerebrospinal fluid (CSF) via folate receptors. Folic acid is an inactive, stable, synthesised, oxidised form of folate, not naturally found in the human body; it has to undergo transformation to 5-MTHF – an inefficient process as unmetabolised folic acid remains in the blood even after an intake of only 400 µg, the recommended daily intake of folate. Folinic acid (calcium folinate) is a stable reduced form of 5-MTHF that needs no transformation before entering the CSF. 157 There is evidence that commercial preparations of folic acid can compete with 5-MTHF for the folate receptor and thus exacerbate a folate deficiency in the central nervous system (CNS). 158 Hence we wonder whether our finding that folic acid had a statistically significant negative effect on the widely used SF-12 MCS was a manifestation of such a folate deficiency rather than the type 1 error that we first supposed.

Thus better understanding of the one-carbon folate and methionine biosynthesis pathways have raised questions about the most appropriate formulation of folate to use clinically for functional folate deficiency. Stahl argues that 5-MTHF is therapeutically better than folic acid as it bypasses conversion of folic acid to the biologically active 5-MTHF, which may be difficult for some patients. 159 Furthermore high doses of inactive folic acid may compete with 5-MTHF for transport across the blood–brain barrier, potentially reducing active 5-MTHF in the brain. Other reasons why 5-MTHF synthesis may be difficult include inhibition of the enzyme by anticonvulsants. Fortunately we avoided this potential confounding factor by excluding patients on anticonvulsant medication.

Against this biomedical background our rigorous and powerful trial has established that folic acid has no general role as adjunct in antidepressant therapy. However studies in patients with cardiovascular disease have shown that higher doses of folic acid produce greater concentrations of 5-MTHF in plasma. 160 Moreover the reductions in homocysteine in FolATED participants on folic acid relative to those on placebo suggests that they were successfully metabolising folic acid to 5-MTHF. Nevertheless we suspect that the beneficial increase in 5-MTHF was masked by the excess folic acid that competed with 5-MTHF for the folate receptors and led to the negative results.

Folate and depression

Our inclusive multi-disciplinary trial has shown beyond doubt that folic acid is an ineffective adjunct in antidepressant therapy. Furthermore better understanding of the one-carbon pathway offers a plausible explanation why. 158 Before we can write off folic acid entirely, however, we recommend updating the Cochrane systematic review and meta-analysis,64 and summarising the systematic review data more thoroughly than was possible in Chapter 3 (see Systematic review of the effectiveness of folate in augmenting antidepressant medication).

At the time of our initial grant proposal little information was available on the use of 5-MTHF or folinic acid in patients with depression. Now there is evidence that 5-MTHF given as adjunct or monotherapy reduces depressive symptoms in patients with low folate levels or alcoholism, and improves cognitive function and reduces depressive symptoms in elderly patients with dementia and folate deficiency. 161 Furthermore there are long-standing concerns that folate may increase cancer risk, mask B₁₂ deficiency and exacerbate depressive symptoms. As 5-MTHF may reduce some of these risks, we judge that that is now a candidate for a large multi-centre trial. There is also evidence that adjunctive 5-MTHF reduces depressive symptoms in patients who were partially responsive or non-responsive to a selective SSRI. 161

In that context we offer the design of FolATED as a proven model. With the benefit of hindsight, however, we judge that a trial recruiting for 1 year in 10 centres would yield better value for money than one recruiting for 3 years in three centres.

Recruitment

Like many trials FolATED recruited more slowly than expected. We based the original target of randomising 550 participants in three centres over 2 years – slightly less than 24 per month across all three centres – on experience with a previously successful trial in one of the three centres. Initially, however, the complex design of FolATED, designed to exclude patients suffering from folate or B₁₂ deficiency while allowing psychiatric teams to optimise antidepressant treatment in normal clinical practice (Figure 1), restricted randomisation across all three centres to eight a month. We attribute our success in eventually increasing these initial rates by 50% to 12 a month across all three centres to a lot of hard work, a combination of creative protocol changes summarised in Appendix 4, and the mutual support engendered by joint training and joint monthly telephone conferences about management and recruitment.

Late in the conduct of FolATED we used a qualitative reflective data collection tool to undertake a retrospective study of recruitment and retention issues (see Appendix 11). We sought, not to evaluate which methods were the most effective in increasing recruitment and retention, but rather to gain insights into problems we faced and methods we used to overcome recruitment and retention problems within FolATED. We found little research into participants’ preferences between different approaches. This leads us to recommend that future large trials include smaller trials or qualitative studies or both to assess the effectiveness of materials used in recruitment such as posters, leaflets, patient information sheets and newsletters to participating practitioners.

Trial design

FolATED is a three-centred, double-blind, placebo-controlled, pragmatic randomised trial of folic acid augmentation of moderate-to-severe depression. It investigates the effect of augmentation on new and continuing ADM. Assessments take place 2 weeks before randomisation (‘week –2’) to screen for eligibility and initiate antidepressant if required; 1 week before randomisation (‘week –1’) by telephone to check for tolerability of antidepressant; baseline (‘week 0’) to randomise to folate or placebo; and at weeks 4, 12 and 25 to assess outcomes. To estimate the effectiveness of folic acid in augmenting ADM, the trial uses standardised instruments to measure changes in depressive symptoms from two perspectives – clinical and participants’.

Primary outcome measures

The primary clinical effectiveness outcome measure is self-rated symptom severity as measured by the Beck Depression Inventory (BDI-II). Though BDI-II scores at 25 weeks are useful in assessing participants’ medium-term recovery, the primary outcome is the ‘AUC’ of mean BDI-II scores between randomisation and the 25-week follow-up, to aggregate participants’ paths to recovery over that timeframe. The primary economic measure is ICER, namely cost per QALY gained.

Secondary outcome measures

1. Symptom severity as measured by clinicians using the MADRS and the CGI of change.

2. Health status (mental and physical components) as measured by SF-12.

3. Health utility as measured by the EuroQoL (EQ-5D).

4. Proportion of patients with moderate depression (defined as BDI-II score ≥ 19) at week 25 – estimated by statistical inference from observed distribution of BDI-II scores rather than statistically weaker technique of counting cases.

5. Side effects as measured by the UKU side effects scale and reported AEs – serious examples include psychiatric inpatient admission, attempted or completed suicide, and other mortality.

6. Compliance and adherence of patients to take the medication as prescribed.

7. Suicidality as measured by the MINI suicidality scale.

8. Folate status as measured by homocysteine derived from blood samples taken at baseline, 12, and 25 weeks; and B₁₂ status as measured by MMA at 12 weeks only.

9. Resource use as measured by the self-completed health and social care resource use questionnaire, and

10. Genetic analysis of SNPs between the two arms.

Scope of statistical analysis plan

The statistical analysis plan focuses on clinical effectiveness as measured by its primary outcome and secondary outcomes a to g. Annexes 2 and 3 summarise the assessment timetable and outcome measures. Annex 4 outlines the methods for economic analyses; Annex 5 those for genetic analysis; and Annex 6 those for biochemical analysis.

Visit windows

Though we aim to complete questionnaires 2 weeks before randomisation and 12 and 25 weeks thereafter, we have defined an advisory window for each. We monitor reasons for collecting the data outside these windows.

Mode of data collection

Some of the trial instruments are administered by researcher or clinician, and others by participant. There is evidence from the MODE-ARTS systematic review that pragmatic changes of administrator ‘at random’ are unlikely to bias outcomes within placebo-controlled trials. 167

Allocation concealment and unblinding the data

In principle throughout recruitment to the trial we conceal treatment allocation from participants, healthcare professionals, investigators, and study team. We shall continue this during data analysis. The trial report will specify occasions when blinding was broken for specific participants.

We shall unblind the data at the joint final meeting of the TSC and the DMEC meeting on 10 October 2011. The chair of this meeting will have a sealed envelope to unblind the data when the TSC and DMEC members are content. After correcting errors detected up to, and as a consequence of, the TSC-DMEC meeting, the trial statistician (RhW) will freeze the database. No longer than a week before that meeting an independent senior statistician (CJW) will check the allocations and thus become unblinded to the treatment allocation. However the trial statisticians (RhW, YS as the assistant trial statistician, and BRC as senior trial analyst responsible for the second, validating analysis) will continue to analyse clinical effectiveness blind to allocated treatment.

Delivery of data sets to biochemistry, genetics and health economics teams

These teams will receive cleaned data from NWORTH. Annexes 4–6 describe the analysis of these data.

‘Analysed’ population

This population – the subject of the principal analyses – comprises all randomised participants with at least one post-baseline BDI-II outcome. If they fail to return any questionnaire or to complete all items for each instrument, we shall impute these data using the principles and methods described under Missing Data above.

‘Complete case’ population

This population comprises only those participants whose outcome data are complete. It will provide a useful sensitivity analysis of primary and secondary findings, especially whether they are sensitive to the presence of missing data and the methods of imputation we use to augment them.

‘Randomised’ population

At first sight it is difficult to draw inferences about this population because some contributed no data after baseline, even on the BDI-II. Because we know the baseline characteristics of all these participants, however, it is possible to reweight the analysed population so they match the characteristics of the randomised population, notably allocated treatment, stratifying variables and baseline BDI-II.

BDI-II

We shall compare the effects of adding folic acid or placebo on self-rated symptom severity by comparing the total scores of the BDI-II. 84

Scoring: This instrument collected data at: screening; baseline; 4; 12; and 25 weeks. We shall score it according to the validated manual, namely by summing the ratings of the 21 items. Each item is rated on a four-point scale ranging from 0 to 3, yielding a total possible score of 63. If a participant ticks more than one category, we shall score the highest.

MADRS

We shall compare the effects of adding folic acid or placebo on researcher-rated depression severity by comparing total MADRS scores. 90

Scoring: This instrument collected data at: baseline; 4; 12; and 25 weeks. We shall score it according to the validated manual by adding the ratings of the 10 items. Each item is rated on a seven-point scale ranging from 0 to 6, yielding a possible score of 60. If a rater selects more than one category, we shall score the highest.

CGI

We shall compare the effects of adding folic acid or placebo on the researcher-rated CGI. 91

Scoring: This instrument collected data at: baseline; 4; 12; and 25 weeks. It comprises three separate clinician-rated items: an ordinal scale of current severity of illness using a range of responses from 1 to 7; an ordinal scale of global improvement since recruitment ranging from 1 to 7; and an efficacy index ranging from 0.25 to 4.0 derived from a 4 × 4 matrix plotting therapeutic effect against side effects.

SF-12 version 2

We shall compare the effects of adding folic acid or placebo on self-rated quality of life in the form of SF-12 scores for mental and physical components. 91

Scoring: We collected SF-12 data at baseline 4, 12, and 25 weeks. The SF-12 comprises ten × 5-point items and three 3-point items. We shall score it according to the validated method, namely by applying Norm-Based Scoring (NBS) algorithms using the recommended standardisation (mean = 50, SD = 10 in the 1998 general US population) to calculate the physical and mental component scores PCS-12 and MCS-12.

Missing items: We shall use the SF-12 missing data software which uses Missing Data Estimation.

EQ-5D

We shall compare the effects of adding folic acid or placebo on self-rated health utility in the form of the EQ-5D (also known as EuroQol). This consists of a self-reported matrix and a self-rated visual analogue scale (EQ-VAS). 93 The self-reported matrix comprises five dimensions: mobility, self-care, usual activities, pain-discomfort and anxiety-depression.

Scoring: This instrument collected data at: 4, 12, and 25 weeks. The five dimensions use three-point scales ranging from 0 to 2. If a participant ticks more than one category, we shall score the highest. In the clinical effectiveness section we shall analyse this as an outcome in its own right, rather than convert it to QALYs. In contrast the health economics section will convert these data to QALYs.

UKU side-effects scale

We shall compare the effects of adding folic acid or placebo on four UKU researcher-rated system-specific side-effect scores and their total. 94

Scoring: Participants completed the instrument at telephone monitoring 1 week before baseline; baseline; 4; 12; and weeks. The instrument sums scores on 48 distinct side effects – 10 ‘psychic’, 8 ‘neurologic’, 11 ‘autonomic’ and 19 other, of which 11 are gender-free, four are male-specific and four are female-specific. It finishes with three generic items, of which we used two – global assessments by ‘patient’ and researcher – for imputation. It rates each item on a four-point scale ranging from 0 to 3, yielding possible system-specific scores of 30, 24, 33 and 45 (different questions for each sex, but same total).

Missing items: If respondents fail to answer system-specific items in the instrument, we shall consider them informative missing items and impute by assuming they did not experience those items If raters score above the range of the instrument, we shall censor at the upper limit.

Analysis: As the authors developed the UKU scale specifically for psychotropic drugs, they planned mainly to analyse specific side effects, but acknowledged the case for planned analyses of system-specific scores. 93 Believing that folate does not have independent psychotropic properties, however, we shall test whether it is safe by analysing the four system-specific scores.

Morisky compliance scale

We shall use this four-item instrument to compare the effects of adding folic acid or placebo on self-rated compliance with medication. 95

Scoring: Participants completed the instrument at 12 weeks. We shall score it as the number of ‘yes’ responses, yielding a score between 0 and 4.

MINI suicidality scale

We shall use this six-item instrument to compare the effects of folic acid and placebo on self-reported suicidality. 96

Scoring: Participants completed the instrument at all six data collection points: screening; monitoring; baseline; 4; 12 and 25 weeks. Though we selected it primarily to fulfil our duty of care to participants, the TSC and DMEC asked us to analyse it as an outcome in its own right, and the Wales MREC agreed. We shall score it by summing the scores allocated to ‘yes’ responses, which range between 0 and 10, yielding a total score between 0 and 33.

Missing items within a subscale

If a subscale comprises three or fewer items, we shall treat each as a separate subscale. In addressing missing items within a subscale thus defined, we shall take account of methodological publications about the validated instrument. In principle we seek to impute missing items to complete instruments thus:108

1. a. If < 25% of the items within a subscale are missing for a participant at a time point, we shall impute them by the weighted mean of the completed items.

2. b. If > 50% of the items within a subscale are missing for a participant at a time point, we shall treat that subscale as missing and impute it.

Missing subscales

Where between 25% and 50% of the items within a subscale are missing, we shall proceed thus:

1. If < 40% of the subscales for a participant at a time point are missing, we shall impute all missing subscales by a single application at that point of the SPSS multivariate imputation algorithm that also takes account of all validated stratification variables.

2. If > 40% of the subscales for a participant at a time point are missing, but < 20% of participants experience that problem, we shall impute all missing subscales by a single multivariate imputation that also takes account of all validated stratification variables across all time points.

3. If > 40% of the subscales for a participant at a time point had been missing, and > 20% of participants had experienced that problem, we would have used multiple multivariate imputations; in the event, however, this never happened.

Missing time points

1. If < 15% of all time points are missing, we shall impute all subscales within each time point by one iteration of the SPSS multiple imputation algorithm using all other subscales at all time points, together with age, gender, centre and group.

2. If > 15% but < 30% of all time points are missing, we shall impute all subscales within each time point by five iterations of the SPSS multiple imputation algorithm using all other subscales at all time points, together with age, gender, centre and group.

Preliminary data description

We shall summarise baseline and demographic data by both treatment allocation and centre. Conscious that our three centres differed in many respects notably psychiatric practice and recruitment policy, we shall present outcomes believed to follow a Normal distribution in the form: number of responses; mean and SD. If Normal plots show evidence of non-normality however, we shall them present them in the form: number of responses; median and first and third quartiles.

Data transformations

Our analysis plan assumes that residual variation from our statistical models will follow approximately Normal distributions. This is a robust assumption in the sense that only a substantial deviation would invalidate each analysis. Hence the trial statistician will plot all residual distributions and discuss any evidence against normality, as shown for example by Normal plots, with the CI. If necessary we shall seek an optimal transformation to improve approximation to Normality.

Continuous outcomes with baseline and more than one follow-up (for example BDI-II)

We shall use ‘AUC’ to combine outcomes over all time points to create the primary outcome. As covariates in the AUC analysis we shall use validated stratification variables – centre, gender, new or old patient, type of antidepressant and previous counselling. For individual time points we shall use analysis of covariance to adjust for the corresponding baseline score.

As a sensitivity analysis we shall use multi-level modelling with the same covariates, also known as repeated measures analysis of variance. We shall estimate parameters for three fixed factors – the three time points (4, 12 and 25 weeks), centre and treatment group. We shall also include interactions between treatment and both time point and centre. We shall summarise all effects by parameter estimate, standard error, significance level, and 95% confidence level.

Continuous outcome with no baseline and only one follow-up (Morisky scale)

We shall use analysis of covariance, with baseline depression scores and validated stratification variables as covariates, to test whether medication adherence, measured on the Morisky scale, is significantly different between the treatment groups. If so, or if there is other evidence that the Morisky score influences the main psychological outcomes, we shall test in secondary analysis whether adding it to the usual covariates improves the fit of each model and refine those models accordingly. In these circumstances we shall test whether also to add prescribed medications recorded by GPs to the usual covariates.

Dichotomous outcomes (serious adverse effects and adverse events)

We shall use logistic regression of the binary response of (S)AE or no (S)AE over each participant’s time in the trial to test whether the proportion of participants who experienced (S)AEs differs between treatment arms. Covariates will include baseline scores and validated stratification variables. We shall transform all estimated fixed effects back from their logistic form and summarise them by OR; standard error; significance level and 95% CI.

Covariates for adjustment within statistical model

We shall keep baseline depression scores and validated stratification variables as covariates throughout. We shall explore covariates of potential scientific relevance, including: demographic (e.g. age, ethnicity, marital status, number of dependants and employment status, coded in accordance with usual demographic practice); and clinical (e.g. referral source, smoking, alcohol consumption and medication adherence, measured by both Morisky scale and recorded prescriptions). We shall retain these if they achieve significance levels of 10%.

Interactions to be tested

Within each analysis we shall test for interaction between treatment and centre, not least because our three centres differed in many respects, notably psychiatric practice and recruitment policy. Any evidence of interaction between treatment and centres will lead to exploratory analysis to explain the effects, initially by covariates within the ‘treatment allocated’ population. Failing that, we shall estimate the treatment effect for each centre separately. We shall also test interactions with significant covariates and include these interactions within the model if significant at the 10% level.

Depression outcomes

General quality-of-life outcomes

Biochemistry outcomes

Side effects and compliance

Aim

To assess whether the use of folic acid supplementation is cost-effective by estimating the incremental cost–utility and cost-effectiveness ratios of ADM plus folic acid relative to ADM alone.

Data

Healthcare resources: We measure participants’ use of health and social care services by:

1. a. self-completed questionnaires (collected by research professionals at baseline, and weeks 12 and 25), which ask patients to recall their use of general practice, community services and social services over previous 3 months

2. b. general practice records of prescribed medications over 25 weeks of follow-up, and

3. c. data on hospitalisations, triggered by notified SAEs.

Unit costs: We shall derive the cost of most resources used from national sources. 114,116,119 We shall estimate specialised costs like pharmacogenetic testing from appropriate local sources. The cost year will be 2010 and the perspective will be that of the NHS and PSS.

Health outcomes: Participants completed the EQ5D instrument at: baseline, 4, 12, and 25 weeks. We shall convert their responses into a single, preference-based utility value based on the UK tariff. 168 They completed the SF12 at: baseline, 4, 12, and 25 weeks. To generate the SF6D utility score from their responses we shall apply a valuation algorithm using preference weights obtained from a sample of the general population using the standard gamble technique. 122 For the cost-effectiveness analysis, we shall estimate the number of weeks free from moderate or severe depression from participants’ responses to the BDI-II.

Analysis

Cost analysis: In principle we shall impute missing data on resource use according to the section of the statistical analysis plan on ‘Missing items within an instrument’. We shall estimate the mean cost per patient over 25 weeks across both arms of the trial together with their respective bootstrapped 95% CIs using 10,000 replicates. We shall analyse cost data by assuming that the large samples generate nearly Normal distributions of sample means, thus justifying Student’s t-test and ordinary least squares (OLS) linear regression. If quantile–quantile plots,169 and Shapiro-Wilk170 and Shapiro-Francia171 tests for normality show problems like skewness and excess zeros (at least 20% of samples),124 we shall develop an appropriate generalised linear model. To gain precision in estimating mean costs, we shall include covariates in the cost regression models. Selection of covariates will accord with the section of the statistical analysis plan on ‘Covariates for adjustment within statistical model’.

Analysis of health outcomes: We shall impute data missing from EQ5D responses according to the section of the statistical analysis plan on ‘Missing items within an instrument’. We shall present descriptive statistics of fully imputed responses to individual items within the EQ5D (mobility, self-care, usual activity, pain-discomfort and anxiety-depression) and the derived utility scores and Visual Analogue Scale scores for each time point across treatment groups. We shall estimate the number of QALYs experienced by each patient over 25 weeks as the area under the EQ5D utility curve, using the trapezoidal rule and adjusting for baseline utility score. 170 We shall derive non-parametric 95% CIs from 10,000 bootstrapped replicates. We shall analyse these QALY data by Student’s t-test and OLS linear regression, after testing by quantile-quantile plots169 and the Shapiro-Wilk170 and Shapiro-Francia171 tests that the large samples generate Normal distributions. To gain precision in estimating mean QALYs, we shall include covariates in the QALY regression models. Selection of covariates will include baseline utility score126 and accord with the section of the statistical analysis plan on ‘Covariates for adjustment within statistical model’).

Cost-effectiveness analysis: We shall estimate the number of weeks free from moderate or severe depression (when BDI-II scores are less than 13) by statistical inference from the observed distribution of BDI-II scores and assuming linear interpolation between time points – baseline and at 4, 12 and 25 weeks.

Incremental and uncertainty analysis: We shall derive ICERs by dividing differences in adjusted mean costs by differences in adjusted mean effects. We shall explore uncertainty around ICERs using non-parametric bootstrapping. We shall display results on cost-effectiveness planes and as cost-effectiveness acceptability curves showing when the results fall below given cost-effectiveness thresholds. We shall conduct sensitivity analyses to test the robustness of our findings and examine the extent to which ICERs are sensitive to key assumptions in the analysis, notably on unit costs and SF6D utilities and through complete case analysis.

Secondary analysis: Depending on the analysis of clinical data, secondary analyses will assess relationships between cost-effectiveness and potential predictors of response, notably adverse reactions and high-cost episodes by including them in generalised linear regression models. 172 These are likely to include baseline disease severity and class of prescribed ADM, along with gender, age and genetic factors. If the results of the genetic analysis suggest clinical value, we shall undertake exploratory analysis of the cost-effectiveness of pharmacogenetic testing.

Homocysteine

Folate and vitamin B₁₂ concentrations are major determinants of one-carbon metabolism, in which the essential methyl donor S-adenosylmethionine (SAM) is formed. SAM is essential for neurologic function. Low plasma folate concentrations are associated with poor response to ADM, and treatment with folic acid can improve that response. Plasma total homocysteine (tHcy) is a sensitive marker of folate (and B₁₂) status. We measure this at baseline and following treatment with folic acid or placebo primarily to assess the response to folate therapy. It will be informative to characterise how response to folate augmentation depends upon baseline homocysteine levels. We will assay homocysteine by an automated analyser (Abbott Architect) using one-step immunoassay with chemiluminescence detection (Axis-Shield). This assay is standardised using the National Institute of Standards and Technology (NIST) Standard Reference Material. It is used to assess patient samples routinely, and is subject to rigorous internal and external quality assurance assessments.

Biochemistry analysis

As patients with depression have increased homocysteine concentrations, the primary analysis of homocysteine will follow essentially the same analysis plan as for genetics in Annex 5. We shall compare the effects of folic acid and placebo on tHcy using the mixed model approach to repeated-measures analysis of variance. As potential covariates we shall use baseline Hcy, age, gender, BMI (weight/height²), co-medications, type of ADM, and new or continuing patient. We shall estimate parameters for three fixed factors – the three time points (4, 12 and 25 weeks), centre (especially important as tHcy sample handling differs between centres) and treatment group. We shall also include the interaction between treatment group and time point to test whether differences between treatments vary over time.

We shall add reported compliance with folate therapy or placebo to this model and test whether this leads to a fall in tHcy. We shall investigate how type of ADM, lifestyle factors (e.g. smoking, BMI and alcohol) and genetic polymorphisms in the folate pathway affect patients’ baseline homocysteine concentrations; and how folic acid affects BDI-II scores, the primary outcome measure. We shall also test the interaction between biochemical variables and the final set of genetic polymorphisms analysed by the genetics analysis plan (see Annex 5). We shall explore the extent to which tHcy and MMA, both baseline and subsequent, predict BDI-II, MADRS and CGI scores.

Methylmalonic acid

Methylmalonic acid is a sensitive marker of B₁₂ status. Hence the trial affords a unique opportunity to estimate the effects of folic acid supplementation on MMA concentration specifically in individuals with low-to-normal B₁₂ levels, and thus assist in elucidating the nature of the relationship, if any, between folic acid and MMA. If folic acid is clearly shown to influence MMA concentrations, this will be of considerable importance for public health decisions relating to the issue of food folate fortification in the UK and elsewhere. There is evidence from cross-sectional studies that the combination of high plasma folate and low B₁₂ status is associated with cognitive decline. Hence we shall later explore the extent to which MMA acts as a prognostic variable for other outcomes.

Thus the non-genetic data required for the biochemical component of the study are the same as those listed in Annex 2 for the genetic component of the study.

Imputing serum folate measurements censored at 20

We shall impute separately for groups A and B without using demographic variables like age or sex.

1. Step 1:

   1. a. Regress log (serum folate) on log (red cell folate – RCF) using raw data where both values are present; and derive slope and intercept.

   2. b. Replace serum folates recorded as > 20 by 20.

   3. c. Regress log (serum) on log(RCF) after including the extra data points added in step 1b, expecting the intercept to increase and the slope to decrease.

2. Step 2:

   1. a. Use slope and intercept from step 1c to impute serum folates recorded as > 20.

   2. b. Retain all imputations that yield serum folates > 20; and replace by 20 all imputations that yield serum folates < 20 in 2.1.

   3. c. Regress log(serum) on log(RCF) after including the extra data points imputed in step 2b, expecting the intercept to increase and the slope to decrease.

3. Final steps:

   1. a. Repeat steps 2a to 2c till slope changes by < 0.1 × SD of slope or no imputed folates < 20.

   2. b. Impute all serum folates recorded as > 20 using slope and intercept from final iteration of step 2c; add normally distributed residual with mean zero and SD = residual SD of final model.

   3. c. Repeat entire process after replacing RCF by B₁₂ to impute missing values when RCF is missing but B₁₂ is present; however imputation is not possible when folate, RCF and B₁₂ are all missing.

   4. d. Assess imputation process by scatter plots of serum folate against RCF and B₁₂.