Perinatal mental health services in pregnancy and the year after birth: the ESMI research programme including RCT

Article history

The research reported in this issue of the journal was funded by PGfAR as project number RP-PG-1210-12002. The contractual start date was in September 2013. The final report began editorial review in June 2020 and was accepted for publication in June 2021. As the funder, the PGfAR programme agreed the research questions and study designs in advance with the investigators. The authors have been wholly responsible for all data collection, analysis and interpretation, and for writing up their work. The PGfAR editors and production house have tried to ensure the accuracy of the authors’ report and would like to thank the reviewers for their constructive comments on the final report document. However, they do not accept liability for damages or losses arising from material published in this report.

Copyright statement

Copyright © 2022 Howard et al. This work was produced by Howard et al. under the terms of a commissioning contract issued by the Secretary of State for Health and Social Care. This is an Open Access publication distributed under the terms of the Creative Commons Attribution CC BY 4.0 licence, which permits unrestricted use, distribution, reproduction and adaption in any medium and for any purpose provided that it is properly attributed. See: https://creativecommons.org/licenses/by/4.0/. For attribution the title, original author(s), the publication source – NIHR Journals Library, and the DOI of the publication must be cited.

2022 Howard et al.

Background and overview

Many women experience mental health problems during pregnancy and the year after birth (i.e. the perinatal period),2,3 and these are associated with adverse effects on the fetus and infant, and subsequent behavioural and emotional problems in the child and adolescent,4 with additional negative impacts for other family members. At the time we developed this programme, perinatal mental health (PMH) services in England were very fragmented, despite National Institute for Health and Care Excellence (NICE) guidance,5,6 and little was known about optimal PMH service configurations. The publication of Five Year Forward View For Mental Health,7 accompanied by NHS England’s commitment to women with PMH problems,8 meant that this research occurred during a time of considerable expansion of PMH services. A phased 5-year transformation programme, backed by £365M, was under way when we obtained approvals for our research. NHS England committed to increased access to the following by 2020/21:

. . . specialist perinatal mental health support in all areas of England, allowing at least an additional 30,000 women each year to receive evidence-based treatment, closer to home, when they need it . . . the right range of specialist community and inpatient care.

NHS England. 8

Contains public sector information licensed under the Open Government Licence v3.0

This had a positive and negative impact on our research programme.

Despite this expansion, relatively little was known about the prevalence of mental disorders in early pregnancy, or how best to identify and optimally treat disorders. By 2013, maternity services had introduced two depression screening questions at antenatal booking, and it was not clear whether or not the two questions were the optimal method of detection, which is the focus of work package (WP) 1(i).

Early identification of mental disorders is recommended to facilitate early evidence-based interventions to optimise maternal and child outcomes. The extent to which generic interventions, in comparison with specialist interventions and services, are clinically effective remains unclear. In line with stepped-care approaches for depression, guided self-help (GSH) delivered by psychological well-being practitioners (PWPs) in primary care was recommended by NICE, but, to the best of our knowledge, there have been no evaluations of GSH modified for pregnancy. Therefore, we aimed to develop such materials and carry out an exploratory trial of modified GSH [WP1(ii)].

Research into the experiences of the whole care pathway for women with PMH disorders was also important and could directly feed into the new services nationally. We investigated the experiences of women and their significant others, along with the experiences of health-care professionals (WP2).

Our National Institute for Health and Care Research (NIHR) programme development grant (PDG), RP-DG-1108-10012, explored methodological issues in evaluating services for women with severe acute postnatal disorders, including the need for modified tools [WP3(i)] and the best way to evaluate mother and baby units (MBUs) [WP3(ii)]. National guidance9 stated that women needing admission postnatally should be admitted to MBUs, but large parts of the country had little or no access to these units and women were cared for either by crisis resolution teams (CRTs) or in acute inpatients wards. This geographical inequity meant that some women cannot access MBUs. Therefore, we could not investigate the clinical effectiveness and cost-effectiveness of MBUs using a randomised controlled trial (RCT) and so we used a quasi-experimental design [WP3(ii)].

Our overarching aims were to investigate (1) the clinical effectiveness and cost-effectiveness of identification and treatment of antenatal depression and other disorders, and (2) which perinatal care pathways are optimal and cost-effective for women with common and severe mental disorders, their infants and partners/significant others.

We consider that we have achieved both these aims in broad terms through a series of inter-related mixed-methods WPs (Figure 1).

FIGURE 1.

Research pathway diagram.

Patient and public involvement

Background

For further background reading please see Howard et al. 10 and Nath et al. 11

During pregnancy, women have frequent contact with health-care professionals, but, despite this, PMH disorders are unrecognised and untreated. 5 As these contacts provide a unique opportunity to identify PMH problems, it is important to establish the optimum case identification method. NICE5 has recommended that health-care professionals consider using the Whooley questions12 to identify depressive symptoms in the perinatal period (Box 1). Research in primary care with non-pregnant populations found that answering ‘yes’ to either or both Whooley questions detected most cases of depression. 12 Inclusion of these questions at the first antenatal appointment had been rolled out in most English maternity services by 2014. However, it was unclear whether or not use of the Whooley questions is the optimal method. The 10-item self-complete Edinburgh Postnatal Depression Scale (EPDS)13 is used internationally in maternity services and could be an alternative; however, antenatal validation has been primarily in the second and third trimesters. 14 The cost-effectiveness of the different approaches to case identification was also unclear.

Aims

The aims of the WEll-being in pregNancy stuDY (WENDY) were to:

• investigate, at antenatal booking, the diagnostic accuracy and acceptability of the Whooley questions compared with the EPDS

• estimate prevalence of depression and other disorders in early pregnancy

• evaluate relative cost-effectiveness of the different tools.

Methods

Sample size

A power calculation was undertaken using simulation with bootstrap estimation of confidence intervals (CIs) for the weighted estimators of sensitivity, specificity and prevalence that corrected for the sample stratification. We assumed an overall prevalence of depression of 9%, a Whooley sensitivity of 0.95 and a Whooley specificity of 0.89. Screening 6000 women, 66% of whom consent to participate, and sampling 54% of the W+ women (i.e. n = 400) and 6% of the W– women (i.e. n = 200), would result in 600 women for interview. In this sample of women, we would expect 185 to be depressed. Assuming a sensitivity of 0.80 and a specificity of 0.71, the width of the 95% CI for the EPDS sensitivity would be 0.19 and that for specificity would be 0.13. A conservative estimate of power based on the 185 women would have > 90% power for a 0.8 and 0.65 sensitivity and specificity difference, respectively (comparing Whooley and the EPDS).

Adjustments to sampling fractions were necessary, as fewer W+ women than anticipated were recruited. The original recruitment target of 200 W– women was reached. After discussion with the Data Monitoring and Ethics Committee, we aimed for 300 W+ women and 300 W– women so that both arms could be recruited over the same period, with random sampling of W– women of 1 : 6. A total of 545 women were recruited, which was within 10% of our target.

Analysis

We used inverse probability weights to provide population estimates of mental disorders that account for bias induced by stratified sampling and missing SCID-I diagnoses. We used bootstrap resampling of the weighted estimators for calculation of most CIs and p-values. 24

We planned to investigate use of multiple imputation methods, which can yield more efficient estimates. This was not possible because of changes at the maternity unit that resulted in comprehensive background data being unavailable.

Qualitative analysis was conducted using thematic and framework approaches25–27 and has been published. 23

Economic analysis

Economic modelling was used to explore the relative cost-effectiveness of the Whooley questions, the EPDS, and the Whooley followed by the EPDS, in comparison with a hypothetical ‘no-screen instrument’ cohort. A decision tree was developed to model possible identification and treatment pathways from the first antenatal appointment to 3 months post partum. Data on sensitivity and specificity of the tools were taken from our cohort. Parameters for the no-screen option were taken from relevant literature and supplemented by expert (co-investigator) consensus (see Appendix 2).

The economic analysis took the NHS and Personal Social Services perspective preferred by NICE,28 with outcome-measured quality-adjusted life-years (QALYs). Cost-effectiveness was assessed in terms of incremental cost per true-positive case detected and incremental cost per QALY. Model parameters were entered into the model with associated probability distributions to explore uncertainty using Monte Carlo simulation. Probabilistic sensitivity analysis explored robustness of the model and the impact of alternative model assumptions. See Appendix 2 for detailed methods.

Following discussion with the funding panel about limited evidence on the optimal tool for use in this population, we additionally explored the comparative psychometric properties29 of the EQ-5D-5L22 and Short Form questionnaire-6 Dimensions (SF-6D). 20

Results

Tables 1 and 2 provide key characteristics of participants. The flow of participants can be seen in Figure 2 (see Appendix 1 for the recruitment chart).

FIGURE 2.

Flow chart of women: WENDY. DNA, did not attend.

Key findings

The weighted estimated population prevalence of mental disorders in early pregnancy are presented in Table 3.

The diagnostic accuracy of the Whooley questions was as follows: a weighted sensitivity of 0.41, a specificity of 0.95, a positive predictive value (PPV) of 0.45, a negative predictive value (NPV) of 0.93, a positive likelihood ratio of 8.2, a negative likelihood ratio of 0.62 and an area under the receiver operating characteristic curve of 0.37 (95% CI 0.34 to 0.40). For the EPDS, using a cut-off point score of 12 (out of 13), diagnostic accuracy was as follows: a weighted sensitivity of 0.59, a specificity of 0.94, a PPV of 0.52, a NPV of 0.95, a positive likelihood ratio of 9.8, a negative likelihood ratio of 0.44 and an area under the receiver operating characteristic curve of 0.89 (95% CI 0.88 to 0.90). For identification of depression, EPDS was more accurate than the Whooley questions. We also found that older age was associated with decreased diagnostic accuracy for the EPDS. 10

As the Whooley questions are the main mental health screen used by maternity services, we also examined the diagnostic accuracy of both the Whooley questions and the EPDS for ‘any mental disorder’, and found that both could be useful in detecting mental disorders, with a likelihood ratio of > 5 in each case. 10

Cost-effectiveness

In terms of detection of depression or any mental disorder, both the Whooley questions and the EPDS appeared to be cost-effective compared with a ‘no-screen’ option and the combined Whooley questions and EPDS option. The Whooley questions were the most cost-effective option at low values of willingness to pay per true-positive case detected (£0–250), and the EPDS was the most cost-effective option at higher willingness-to-pay values (> £250). The analysis for depression (but not for any mental disorder) was sensitive to assumptions and data inputs, such that the ‘no-screen’ option had the highest probability of being cost-effective at willingness-to-pay values of > £600. The Whooley questions remained the most cost-effective option at low values of willingness to pay and the EPDS was the most cost-effective option at willingness-to-pay values between £250 and £600.

In terms of cost per QALY, the ‘no-screen’ option was dominated by the other three options, and the Whooley questions, EPDS, and Whooley questions followed by the EPDS each had a probability of being cost-effective of around 30% at willingness-to-pay values of £0 to £50,000 per QALY. See Appendices 2 and 3 for detailed economic methods and results, respectively.

Psychometric comparison of the EQ-5D-5L and SF-6D

We found a lack of concordance between the EQ-5D-5L and the SF-6D. The EQ-5D-5L scores tended to be higher than SF-6D scores in individuals with better health states, whereas the SF-6D scores tended to be higher than EQ-5D-5L scores in individuals with poorer health states. Convergent and known-group validity were comparable between the two measures. Longitudinally, women who recovered showed larger increases in SF-6D utilities than those who did not recover at follow-up. With the EQ-5D-5L, this was not the case. In addition, ceiling effects were more apparent in the EQ-5D-5L. Therefore, the effectiveness of PMH interventions may be better captured by the SF-6D than by the EQ-5D-5L. See Heslin et al. 29 for further details.

Acceptability of the Whooley questions

Most women found the Whooley questions enquiry acceptable, although those with a history of or a current PMH problem and/or a history of abuse found enquiry less acceptable because of emotional responses triggered by the questions and response to disclosures. Women wanted to be asked simple questions about mental health, to have sufficient time to discuss issues and to receive normalising and well-informed responses from midwives. In addition, there were some reported concerns regarding the consequences of disclosure (e.g. information-sharing).

See Yapp et al. 23 for further details.

Strengths and limitations

We used a gold standard instrument for the diagnosis of depression and other disorders, with translated instruments and a language interpreter where required, making our study representative of the base population. Unlike most previous studies,5,30 our study established the diagnostic accuracy of the Whooley questions, as asked by midwives themselves, and the stratified sampling design enabled us to efficiently investigate the diagnostic accuracy of the tools. However, use of a single maternity site limits generalisability. We were unable to meet the recruitment target of one in every three women having a mental disorder (comparable with W– women) and, although the population was broadly representative, selection bias is likely. Limitations of the economic component include model assumptions, some parameters coming from expert clinical opinion (as no evidence existed to complete certain parameters) and questionable generalisability due to most screening tool sensitivity and specificity data coming from one study.

Additional studies included:

• an examination of the effectiveness of the GAD-2 in detecting anxiety disorders31

• a NIHR Biomedical Research Council-funded PhD (Doctor of Philosophy) on migration and antenatal mental health32

• a study of PMH disorders in women aged < 25 years compared with women aged > 24 years33

• a study of the history of self-harm and mental disorders in pregnancy34

• a Nuffield Foundation-funded add-on study examining maternal personality traits, anxiety disorders, antenatal depressive symptoms and the post-partum mother–infant relationship35,36

• international individual patient data meta-analyses examining diagnostic accuracy of the EPDS. 37,38

Recommendations for future work

Replications of our studies in other services are needed, particularly within the new service structures and, specifically, they should address whether or not those who are identified and referred in early pregnancy have better mental health and fetal/infant health outcomes postnatally.

Background

In WP1(i) we found that ≈ 11% of women had depression in early pregnancy. 2,10 International guidelines5,41 for antenatal depression recommend that cognitive–behavioural therapy-based GSH should be offered as the first step in management of mild to moderate depression. GSH is usually delivered in England by PWPs in Improving Access to Psychological Therapies (IAPT) services. 42

Aims

• To establish that the trial procedures worked so that a Phase III trial could follow.

• To provide evidence on the efficacy of a GSH intervention delivered by PWPs for mild to moderate antenatal depression compared with TAU.

• To provide preliminary evidence on whether or not other outcomes improve.

• To explore if antenatal GSH is likely to be cost-effective compared with TAU.

Methods

The DAWN (Depression: an exploratory parallel-group randomised controlled trial of Antenatal guided self-help for WomeN) study was a Phase II exploratory RCT with two parallel groups and a primary end point of EPDS symptoms at 14 weeks post randomisation, initially based in one maternity unit in south-east London and then extended to five units.

Sample size calculation

Assuming a correlation of 0.4 between baseline and outcome symptom score and a two-arm parallel-group design with 52 women in each arm, the Stata^® (StataCorp LP, College Station, TX, USA) procedure sampsi gave 79% power to detect a difference of 0.5 standard deviations (SDs) using analysis of covariance and a two-tailed test using a 95% significance level. For preliminary evidence on efficacy and RCT feasibility, we aimed to recruit 110 women. 43

Economic evaluation methods

A cost-effectiveness analysis was conducted at the 3-month post-delivery follow-up point, taking an NHS/Personal Social Services perspective that is preferred by NICE. 44 QALYs were calculated using the SF-6D. 45 We calculated area-under-the-curve (AUC) values for QALYs, with linear interpolation between assessments. 43 The intervention was calculated using a micro-costing approach (see Trevillion et al. 40 for details). Data on the use of all other health and social care services were collected using the Adult Service Use Schedule (AD-SUS). 46 Cost-effectiveness was explored with the net benefit approach. Uncertainty around costs and cost-effectiveness were explored using cost-effectiveness acceptability curves (CEACs). 47

A secondary analysis was performed, substituting the SF-6D with the EQ-5D-5L measure of health-related quality of life. 22

See the study protocol39 and published results40 for full details.

Results

Fifty-three women (46.5%) were randomised, with 26 women receiving GSH and 27 women receiving TAU (Figure 3). We were unable to reach our recruitment target, despite offering home visits at evenings/weekends, attending workshops aimed at expectant parents and antenatal staff meetings to advertise the study, and extending the recruitment period and recruitment sites. During the trial recruitment period, new services led to midwives referring more women directly to IAPT. After discussion with the Trial Steering Committee and Independent Data Monitoring and Ethics Committee, we agreed that further extensions to the recruitment period would be unhelpful and costly (see Appendix 4). Demographic and clinical characteristics of participants are shown in Tables 4 and 5.

FIGURE 3.

Flow chart of women: the DAWN study.

Sixty-nine per cent (n = 18) of women attended at least four sessions of GSH. The outcome measures collected by PWPs at each session delivered to women in the GSH group are summarised for each session in Figure 4.

FIGURE 4.

Patient Health Questionnaire-9 items and EPDS medians and interquartile ranges (IQRs) for each session for the GSH arm. Horizontal lines represent depression cut-off points for the Patient Health Questionnaire-9 items and the EPDS.

At 14 weeks post randomisation, median EPDS scores in the GSH and TAU groups were 8 and 12, respectively (effect size –0.64, 95% CI –1.30 to 0.06; p = 0.066), a clinically significant difference. No statistically significant differences were observed for the secondary outcomes.

Infant birth outcomes

As sample sizes were small, further analysis was not undertaken (Table 6).

FIGURE 5.

A CEAC for GSH vs. TAU at 3 months post delivery from the health and social care perspective for SF-6D-based QALYs.

There were no adverse events in the study.

Economic findings

Costs and outcomes were similar between the groups (Table 7). CEACs showed that the probability of GSH being cost-effective compared with TAU was approximately 50% at the NICE preferred willingness-to-pay threshold of £20,000–30,000 per QALY (Figure 5). However, the results were sensitive to the assumptions made (i.e. health-care perspective, PWP indirect time and outliers and influential observations removed) and demonstrated large variation driven by the small sample size. Therefore, the cost-effectiveness of GSH remains uncertain. Please see the protocol39 and the results paper40 for full details of the economic evaluation methods and results.

Discussion

This exploratory trial suggests that GSH can be successfully delivered, is acceptable, is associated with clinically significant decreases in depressive symptoms, does not lead to harm and does not show a cost-effective disadvantage. However, as no definitive trial is possible in England, we are not able to provide definitive evidence on its clinical effectiveness or cost-effectiveness.

Recommendations for future work

Our workbook is being used in many IAPT services nationally. Ideally, we would evaluate its use through routine online IAPT data. At the time of writing, however, this was not possible, as routine IAPT data does not include a pregnancy-specific identifier. We have had excellent feedback from practitioners, including, for example, ‘Fantastic . . . really helpful for the women’.

Aims

In STAkeholders’ views and experiences of perinatal mental health CarE and services: a qualitative studY (STACEY) we aimed to explore key stakeholders’ views and experiences of support for PMH difficulties across services. We explored the perspectives of their partners and other family members (i.e. ‘significant others’), as well as practitioners.

Methods

We carried out qualitative, semistructured interviews with 52 women with babies aged 6–9 months and 32 significant others (Figure 6). Originally, the protocol proposed conducting 30 interviews with women and 20 interviews with significant others, but additional NIHR funding enabled us to capture more of the diversity of disorders and service experiences. We also carried out focus groups and interviews with 103 practitioners and commissioners. The characteristics of participating women and significant others are provided in Tables 7 and 8.

FIGURE 6.

Example timeline.

Women were recruited from 11 NHS trusts. Purposive sampling was used to obtain diversity of diagnosis, service use and sociodemographic background.

Analysis

Thematic analysis was used, with themes identified in a cyclical process of reading, coding and exploring the patterning of data. 26 Subsample data sets were created for key subgroups and separate analyses conducted (e.g. by type of service). To enhance validity, two researchers were involved in coding data for each data set.

Results

Tables 8 and 9 provide details of key characteristics of participating women and family members, respectively.

Strengths and limitations

The strengths of WP2 include the diverse participants and range of services included. However, although we sought to include varied perspectives, participants in this qualitative workstream were not representative of the wider population of service users, significant others or practitioners and, therefore, some views may not have been captured. Recruitment to the study via clinical teams may have under-represented women/significant others who are less engaged with services. Only three women with schizophrenia/schizoaffective disorder were included. Likewise, the practitioner sample may under-represent clinicians with less interest in PMH. Focus groups included a wide variety of practitioners but may limit expression of views that conflict with those of colleagues. Interpretations of the findings should be understood with these limitations in mind and further research on the key themes identified is merited.

Background and methods

See also Wykes et al. 55

A patient-reported outcome measure of perceptions of inpatient care [i.e. the perinatal VOICE (Views On Inpatient CarE) measure] was developed in the Patient Involvement in Improving Patient Care (PERCEIVE) programme. 55 A draft of the perinatal VOICE was developed and reviewed by women who had experienced acute care in the perinatal period and staff from PMH services to create the final perinatal VOICE measure. The perinatal VOICE measure contains five sections: (1) care and treatment (three items), (2) medication (two items), (3) staffing (seven items), (4) environment (five items) and (5) baby’s well-being (10 items). At the end of each section, respondents are encouraged to provide further comments.

The perinatal VOICE was included in WP3(ii) for each service experienced. A subsample of patients completed the questionnaire a second time within 6 weeks of first completion to examine test–retest reliability (see Appendix 5 for more details).

Results

A total of 267 patients completed at least one perinatal VOICE questionnaire (Table 10).

Psychometric evaluation

Descriptive statistics and test–retest

A total of 267 women provided 361 questionnaires, one for each service type accessed. Twenty-nine questionnaires were completed by mothers in the test–retest study 4–17 days after completion of the first of two assessments (mode = 7 days, mean = 8 days).

Figure 7 illustrates that response distributions for several items showed modal values at the bounds of the scale. With the exception of the generation of simple item total scores, our analyses treated the response set as ordinal categories to account for floor/ceiling effects. Figure 7 shows kappa statistics for item test–retest reliability. There was significant agreement in all cases except for the item relating to the cleanliness of the baby-changing facilities (p = 0.105). The remaining 26 items were significant. Twenty-two items showed good (> 0.6) agreement. Overall, the 27 items gave a test–retest intracluster correlation coefficient of 0.784 (95% CI 0.643 to 0.924).

FIGURE 7.

Response set and kappa for the perinatal VOICE items.

Factor analysis

Eigenvalues and goodness-of-fit measures from exploratory factor analysis (see Appendix 4) were equivocal, but suggested that two factors gave adequate fit (comparative fit index = 0.97). Those items concerned aspects of the service relating to care of the mother (and the baby) and the two factors were positively correlated at 0.49 (p < 0.0001).

Item response theory evaluation

Graded response models56 were used to examine item and test information functions. Models for single factors suggested that several items added little additional information and so a substantially shorter form might be adequate. Comparison of long and short forms suggested that the long form improved precision of estimation of the overall satisfaction dimension by ≈ 20%. Examining separately the items loading on each of the two factors suggested fewer items as contributing little information. As all items had received PPI support during development, no items were eliminated.

Construct validity

The box plot in Figure 8 shows that perinatal VOICE total scores were strongly associated with the service being received (2 degrees of freedom; p < 0.0001). Separated by factor (Figure 9), MBUs were rated as providing maternal care comparable to inpatient care, and support for caring for the baby comparable to the home treatment setting, with consistently lower scores for the inpatient service on both factors. Those in MBUs reported slightly higher scores for factor 2 (relating to care of baby) than those receiving CRT care.

FIGURE 8.

Box plot of perinatal VOICE scores by service.

FIGURE 9.

Box plot of factor scores by treatment setting: (a) factor 1 scores; (b) factor 2 scores.

Figure 10 shows the total perinatal VOICE scores related to scores from the Client Satisfaction Questionnaire (CSQ) eight-item total score (p < 0.0001). When the two perinatal VOICE factors were considered jointly, only factor 1 scores predicted the CSQ total score (Table 11), which is consistent with the CSQ focusing on satisfaction with the service directly for the woman as patient, rather than as a service for the woman as parent.

FIGURE 10.

Perinatal VOICE scores and CSQ scores.

TABLE 11 - Regression prediction of CSQ (total) from perinatal VOICE factors

SE, standard error.

Predictor	Coefficient	SE	t	p > t	95% CI
Factor 1	4.583	0.322	14.23	< 0.0001	3.949 to 5.217
Factor 2	0.154	0.346	0.45	0.656	–0.527 to 0.835

Comments provided by 139 women from 166 questionnaires respondents were analysed thematically. 57 Key themes were support networks and staff authority. In some services, women found baby support minimal or that parenting advice was too rigid. Lack of continuity of care was a particular issue in CRT support where repeated staff changes are common. Relationships with staff in all services were considered crucial, but could be compromised by understaffing and use of coercion. As in WP2, mothers reported difficulties involving their families in their care, including specific support for family members themselves. Peer support from other mothers was helpful.

Conclusions

The perinatal VOICE is an acceptable patient-reported outcome measure. Selection bias is likely and research on other samples is needed.

Background

The clinical effectiveness and cost-effectiveness of psychiatric MBUs compared with generic acute psychiatric wards or CRTs has, to the best of our knowledge, not been investigated. Our PDG demonstrated that a RCT was not possible for logistical reasons (including lack of beds and strong maternal and staff preferences for MBUs). The inequitable distribution of MBUs across England meant that a quasi-experimental observational study was possible.

Aims

In the Effectiveness and cost-effectivenesS of perinatal Mental health servIces (ESMI) MBU study we aimed to examine (1) differences in characteristics of recruited women at the point of admission to acute care, (2) clinical outcomes and (3) the cost-effectiveness of MBUs relative to TAU.

Women in our PAG discussed the choice of primary outcome and felt that a relapse after an episode of acute care would be devastating. Therefore, readmission to acute care was considered the most appropriate primary outcome. Women in our PAG also wanted us to measure quality of life, satisfaction with services, unmet needs and relationship with baby. Our primary objective was to test the hypothesis that women with PMH disorders who are admitted to MBUs are significantly less likely to be readmitted to acute care (i.e. MBU, CRT or a generic acute ward) in the year following discharge from acute care than women admitted to generic acute wards or CRTs. We also hypothesised that admission to MBUs would be cost-effective, compared with admission to generic wards or CRTs, for the period between index admission to 1 month post discharge.

We further hypothesised that, compared with women admitted to generic services, women admitted to a MBU in the first year after giving birth will:

• have significantly fewer unmet health and social care needs 1 month post discharge

• report significantly higher levels of service satisfaction 1 month post discharge

• have better maternal adjustment 1 month post discharge

• be significantly more sensitive and less unresponsive when interacting with their babies 1 month post discharge (and their babies will be more co-operative and less passive)

• be more likely to retain custody of their child in the year following discharge.

Methods

Women were recruited at the point of, or within 4 weeks of, discharge so that all women were interviewed at around 4 weeks (time point 1). At this point, women provided retrospective baseline information about the admission [time point 0 (t0)]. If they consented to the research team accessing their medical records, we also obtained baseline information in the records (t0). Measures are detailed in Appendix 6. Therefore, baseline data refer to the time period when women were first admitted to acute care. Long-term outcome data refer to the time period from discharge to 1 year post discharge, collected from health and social care records and a brief telephone interview.

The Research Ethics Committee (REC) declined our request for minimum data set collection using Section 251 for all women under acute care. 59

Power calculation

Pilot data were analysed using the Clinical Record Interactive Search (an anonymised case register)60 for 20 perinatal women on generic wards, 20 women in MBUs and 20 women assigned to CRTs. Generic ward patients were most likely to be readmitted (with 95% of ward patients and 35% of MBU patients readmitted during the 12-month follow-up). CRT readmission rates were similar to MBU readmission rates. Therefore, assuming similar readmission rates nationally, we could detect a doubling of risk for ward patients (with 90% power and 47 women in each group). We aimed to recruit 100 women in each group. Therefore, even if we did not manage to follow up 20% of patients in each group or needed to exclude women for being beyond the ‘region of support’ (see Primary analysis), we still anticipated being able to detect these differences.

Statistical analyses

Economic evaluation

We performed an economic evaluation at 1 month post discharge and a cost analysis at 1 year post discharge, as per the grant application. The economic evaluation took the NHS and Personal Social Services perspective preferred by NICE. 28 Service use data were combined with national published unit costs62–65 to calculate total cost of each participant over follow-up (see Appendix 13). Costs and outcomes were compared and presented as mean differences and 95% CIs obtained using bias-corrected non-parametric bootstrapping (i.e. repeat re-sampling). 66 Cost-effectiveness was assessed through the calculation of incremental cost-effectiveness ratios (ICERs)67 and were explored in terms of QALYs using the EQ-5D-5L (see Appendix 13 for an explanation of the change in primary outcome). Uncertainty around the cost and effectiveness estimates were represented by CEACs. 47 In addition, we examined readmissions rates, use of community mental health services and costs in the longer term using data collected from clinical records at 1 year post discharge. See Appendix 13 for full details on the economic methods.

Process evaluation

As service provision varies nationally, we collected detailed descriptions of the service components in participating provider organisations. We developed a structured process evaluation questionnaire, guided by the research literature and discussions within the research team and structured around service component types (e.g. interventions, facilities and staff).

Telephone contact was made with a senior member of each service type who completed the questionnaire over the telephone. The questionnaire was e-mailed ahead of the structured telephone interview so that the person could look at the forms to facilitate completion.

Significant others

We asked women to nominate a significant other who had supported them through their mental health crisis. (However, women did not have to allocate a significant other.) The nominated significant others completed a brief Involvement Evaluation Questionnaire,68 which was available online and/or in paper form at the woman’s home and includes the General Health Questionnaire-12 (GHQ-12). 69

Results

We initially recruited from 28 mental health trusts and then expanded recruitment to three Welsh health boards and 39 mental health trusts (Figure 11). A total of 279 mothers participated, of whom 108 (38.7%) received MBU care (the ‘higher’ level of care), 62 (22.2%) received generic ward (intermediate) care and 109 (39.1%) received CRT care (Figure 12) for the (index) admission to acute care after birth (see Appendix 10 for recruitment chart).

FIGURE 11.

Map of the 42 trusts/health boards recruited from in England and Wales (see Appendix 7 for further details).

FIGURE 12.

Flow chart of women participating in the ESMI MBU study.

Defining the cohort group

A total of 493 admissions occurred in 279 women recruited, with the number of admissions per participant ranging from one to seven. Women were categorised into a cohort group by highest level of care ever received. Therefore, women who spent any time in a MBU, regardless of when or for how long, were categorised as MBU. This resulted in 108 women categorised as MBU, 62 women categorised as acute ward and 109 women categorised as CRT.

Table 12 shows the number of services used, the total number of days in services and the percentage of time in the service of their cohort out of total time in all services by cohort allocation. Therefore, by our definition, 109 women (100%) in the CRT arm attended one service and spent 100% of their time under CRT services. Women in the MBU arm spent a longer time in services than women in the ward or CRT arms, which the highest level of care definition could have intensified. The total number of days in service by cohort group is presented in Figure 13 and the clinical symptoms for participants by cohort group is shown in Figure 14.

TABLE 12 - Defining the cohort group (N = 279)

IQR, interquartile range.

Variable	Service, n (%)			Total, n (%)
	CRT	Ward	MBU
Number of services (MBU/ward/CRT) used (N = 279)
1	109 (100.0)	18 (29.0)	33 (30.6)	160 (57.3)
2	0 (0.0)	44 (71.0)	50 (46.3)	94 (33.7)
3	0 (0.0)	0 (0.0)	25 (23.1)	25 (9.0)
Total number of days in services (N = 279), median (IQR)	25.0 (16.0–38.0)	34.0 (18.0–53.0)	75.5 (55.0–97.0)	42.0 (21.0–76.0)
Percentage of time in cohort service out of the total time in all services (N = 279)
< 25%	0 (0.0)	13 (21.0)	10 (9.3)	23 (8.2)
25–49%	0 (0.0)	9 (14.5)	8 (7.4)	17 (6.1)
50–74%	0 (0.0)	16 (25.8)	22 (20.4)	38 (13.6)
≥ 75%	109 (100.0)	24 (38.7)	68 (63.0)	201 (72.0)

FIGURE 13.

Histogram of total number of days in services by cohort group. (a) HTT/CRT; (b) acute ward; and (c) MBU.

FIGURE 14.

Clinical symptoms for ESMI MBU participants by cohort group.

Table 13 summarises the length of each index admission by service type.

TABLE 13 - Length of stays in each service: the ESMI MBU study (N = 493)

IQR, interquartile range.

Service	Number of participants	Number of days in admission index
		Median	IQR
Intensive care	12	25.5	12.5–46.5
Ward	125	14.0	6.0–27.0
MBU	116	49.0	34.5–74.5
CRT	237	17.0	9.0–32.0
Total	493	21.0	9.0–42.0

Key findings

Baseline measures

Table 14 shows demographics of the recruited population at baseline. Participants who attended an acute ward as their highest level of care were, on average, younger (mean 30.5 years), more likely to be white, be single, have a learning disability, have no formal qualifications/General Certificates of Secondary Education (GCSEs), be unemployed at admission, be living alone and to have been adopted/fostered as a child.

TABLE 14 - Demographics: the ESMI MBU study (N = 279)

A Level, Advanced Level; BTEC, Business and Technology Education Council; CAN-M, Camberwell Assessment of Need – Mothers; HNC, Higher National Certificate; NVQ, National Vocational Qualification.

Variable	Service			Total
	CRT	Ward	MBU
Age (years) at consent (N = 279), mean (SD)	31.1 (5.8)	30.5 (6.5)	32.5 (5.8)	31.5 (6.0)
Ethnicity (N = 279), n (%)
White	79 (72.5)	50 (80.6)	83 (76.9)	212 (76.0)
Black	5 (4.6)	4 (6.5)	11 (10.2)	20 (7.2)
Asian	14 (12.8)	3 (4.8)	8 (7.4)	25 (9.0)
Mixed/multiple ethnic groups	5 (4.6)	3 (4.8)	3 (2.8)	11 (3.9)
Other	6 (5.5)	2 (3.2)	3 (2.8)	11 (3.9)
English as first language (N = 279), n (%)
Yes	94 (86.2)	53 (85.5)	80 (74.1)	227 (81.4)
Place of birth (N = 279), n (%)
UK	94 (86.2)	49 (79.0)	71 (65.7)	214 (76.7)
Other Europe	2 (1.8)	8 (12.9)	11 (10.2)	21 (7.5)
Africa	2 (1.8)	3 (4.8)	12 (11.1)	17 (6.1)
Asia	9 (8.3)	1 (1.6)	8 (7.4)	18 (6.5)
North America/Caribbean	1 (0.9)	0 (0.0)	2 (1.9)	3 (1.1)
Central America	0 (0.0)	1 (1.6)	0 (0.0)	1 (0.4)
South America	1 (0.9)	0 (0.0)	1 (0.9)	2 (0.7)
Australasia/Oceania	0 (0.0)	0 (0.0)	3 (2.8)	3 (1.1)
Difficulty reading own language (N = 278), n (%)
Yes	12 (11.1)	7 (11.3)	8 (7.4)	27 (9.7)
Learning disability (N = 278), n (%)
Yes	2 (1.9)	5 (8.1)	4 (3.7)	11 (4.0)
Highest qualification (N = 279), n (%)
GCSE or no formal qualifications	15 (13.8)	16 (25.8)	23 (21.3)	54 (19.4)
A Level/NVQ/BTEC/HNC	48 (44.0)	28 (45.2)	43 (39.8)	119 (42.7)
Higher education/professional qualifications	46 (42.2)	18 (29.0)	42 (38.9)	106 (38.0)
Employment status prior to maternity leave (N = 279), n (%)
Working	85 (78.0)	37 (59.7)	68 (63.0)	190 (68.1)
Not working	24 (22.0)	25 (40.3)	40 (37.0)	89 (31.9)
Gross yearly household income (£) (N = 276), n (%)
0–5475	2 (1.8)	4 (6.6)	5 (4.7)	11 (4.0)
5476–14,999	18 (16.5)	15 (24.6)	22 (20.8)	55 (19.9)
15,000–30,999	30 (27.5)	14 (23.0)	31 (29.2)	75 (27.2)
31,000–45,999	17 (15.6)	10 (16.4)	15 (14.2)	42 (15.2)
46,000–60,999	15 (13.8)	5 (8.2)	6 (5.7)	26 (9.4)
≥ 61,000	19 (17.4)	5 (8.2)	22 (20.8)	46 (16.7)
Would rather not say	8 (7.3)	8 (13.1)	5 (4.7)	21 (7.6)
Current relationship status (N = 279), n (%)
Single	11 (10.1)	15 (24.2)	15 (13.9)	41 (14.7)
Partner but not cohabiting	7 (6.4)	2 (3.2)	6 (5.6)	15 (5.4)
Married/cohabiting	87 (79.8)	40 (64.5)	86 (79.6)	213 (76.3)
Separated/divorced/widowed	4 (3.7)	5 (8.1)	1 (0.9)	10 (3.6)
Current partner history of mental health problems (N = 224), n (%)
Yes	10 (10.6)	9 (23.7)	14 (15.2)	33 (14.7)
Currently living with (excluding children) (N = 279), n (%)
Alone	16 (14.7)	18 (29.0)	13 (12.0)	47 (16.8)
Spouse/partner	82 (75.2)	36 (58.1)	82 (75.9)	200 (71.7)
Parent(s)/other	11 (10.1)	8 (12.9)	13 (12.0)	32 (11.5)
Adopted/fostered as a child (N = 233), n (%)
Yes	3 (3.1)	6 (10.9)	5 (6.1)	14 (6.0)
Assigned a social worker as a child (N = 228), n (%)
Yes	10 (10.4)	8 (14.5)	5 (6.5)	23 (10.1)
Social Provision Scale total score (N = 240), mean (SD)	76.1 (12.3)	72.0 (13.6)	73.4 (11.3)	74.2 (12.3)
Number of unmet needs (CAN-M) (N = 279), mean (SD)	9.4 (4.3)	10.6 (4.7)	9.7 (4.6)	9.8 (4.5)

Table 15 and Figure 15 show clinical measures of population at baseline. Participants who attended an acute ward as their highest level of care had higher proportions of smoking at admission, substance misuse, psychotic symptoms, acts of self-injury in the 2 weeks prior to admission and a diagnosis of schizophrenia or related disorders. These women were, on average, younger at first admission and more likely to have had previous admissions in the past 2 years, and it was less likely that this was their first episode of a psychiatric disorder. In addition, they were more likely to have experienced emotional abuse, sexual abuse, emotional neglect and physical neglect as a child. These women also scored higher for the CAS, suggesting that they were experiencing higher levels of domestic abuse at point of admission.

TABLE 15 - Clinical measures at baseline (t0): the ESMI MBU study (N = 279)

CAS, Composite Abuse Scale; CMHT, Community Mental Health Team; CTQ, Childhood Trauma Questionnaire; HoNOS, Health of the Nation Outcome Scales.

Variable	Service			Total
	CRT	Ward	MBU
Initial help-seeker, n (%)
Patient sought help	46 (42.2)	28 (45.2)	32 (30.2)	106 (38.3)
Patient’s family, friends or neighbours sought help on their behalf	33 (30.3)	23 (37.1)	40 (37.7)	96 (34.7)
Crisis was identified during a planned contact with the patient by mental health professionals (e.g. by the CMHT)	10 (9.2)	6 (9.7)	10 (9.4)	26 (9.4)
Police or court officials identified the need for mental health intervention	0 (0.0)	2 (3.2)	3 (2.8)	5 (1.8)
Health or social care staff outside the NHS mental health services sought help for the patient	17 (15.6)	2 (3.2)	19 (17.9)	38 (13.7)
Other	3 (2.8)	1 (1.6)	2 (1.9)	6 (2.2)
Previous admissions in last 2 years (N = 279), n (%)
Yes	12 (11.0)	14 (22.6)	22 (20.4)	48 (17.2)
First episode of psychiatric disorder (N = 278), n (%)
Yes	33 (30.3)	15 (24.6)	35 (32.4)	83 (29.9)
Age (years) at first contact with mental health services (N = 270), mean (SD)	25.0 (8.2)	23.1 (7.6)	26.2 (8.1)	25.0 (8.1)
Post-partum onset of episode (vs. earlier onset) (N = 277), n (%)	61 (56.5)	35 (56.5)	62 (57.9)	158 (57.0)
Detained (N = 279), n (%)
Yes	0 (0.0)	22 (35.5)	50 (46.3)	72 (25.8)
Section type for admissions during index episode and previous 2 years61 (N = 148), n (%)
Section 2	1 (100.0)	24 (51.1)	56 (56.0)	81 (54.7)
Section 3	0 (0.0)	18 (38.3)	36 (36.0)	54 (36.5)
Section 5 (2)	0 (0.0)	3 (6.4)	5 (5.0)	8 (5.4)
Section 5 (4)	0 (0.0)	0 (0.0)	1 (1.0)	1 (0.7)
Section 136	0 (0.0)	2 (4.3)	2 (2.0)	4 (2.7)
Acts of self-injury in 2 weeks before admission (N = 272), n (%)
Yes	35 (33.0)	22 (36.1)	28 (26.7)	85 (31.3)
Total HoNOS score (N = 163), mean (SD)	12.8 (5.5)	14.8 (5.2)	14.1 (6.1)	13.8 (5.7)
Psychotic symptoms (N = 278), n (%)
Yes	53 (49.1)	50 (80.6)	80 (74.1)	183 (65.8)
Smoked at point of admission (N = 270), n (%)
Yes	18 (17.0)	28 (45.2)	24 (23.5)	70 (25.9)
Substance misuse (N = 279), n (%)
Yes	14 (12.8)	12 (19.4)	4 (3.7)	30 (10.8)
Chronic physical health condition (N = 279), n (%)
Yes	55 (50.5)	33 (53.2)	50 (46.3)	138 (49.5)
Personality disorder and related disorders diagnosis (primary or secondary) (N = 278), n (%)
Yes	17 (15.7)	18 (29.0)	14 (13.0)	49 (17.6)
Primary clinical diagnosis at admission (N = 278), n (%)
Depression and other unipolar mood disorders (ICD-10 codes F32, F33, F34, F38 and F39)	61 (56.5)	15 (24.2)	34 (31.5)	110 (39.6)
Bipolar disorder (ICD-10 codes F30 and F31), including acute psychosis (due to the psychopathology of puerperal psychosis)	17 (15.7)	18 (29.0)	38 (35.2)	73 (26.3)
Schizophrenia and related disorders (ICD-10 codes F20–29, excluding acute psychotic episode)	1 (0.9)	7 (11.3)	9 (8.3)	17 (6.1)
Anxiety disorders (ICD-10 codes F40 and F41)	20 (18.5)	8 (12.9)	11 (10.2)	39 (14.0)
Eating disorders (ICD-10 code F50)	0 (0.0)	0 (0.0)	1 (0.9)	1 (0.4)
Severe mental and behavioural disorders associated with the puerperium (ICD-10 code F53)	1 (0.9)	1 (1.6)	12 (11.1)	14 (5.0)
Mental and behavioural disorder due to multiple/psychoactive drug use/cannabis/tobacco use (ICD-10 codes F10–19)	0 (0.0)	1 (1.6)	0 (0.0)	1 (0.4)
Personality and behaviour disorders (ICD-10 codes F60–69)	6 (5.6)	11 (17.7)	3 (2.8)	20 (7.2)
No diagnosis given	2 (1.9)	1 (1.6)	0 (0.0)	3 (1.1)
CTQ − emotional abuse (N = 268), n (%)
Yes	30 (28.3)	25 (43.1)	32 (30.8)	87 (32.5)
CTQ − physical abuse (N = 266), n (%)
Yes	20 (18.7)	9 (16.4)	17 (16.3)	46 (17.3)
CTQ − sexual abuse (N = 257), n (%)
Yes	19 (18.3)	22 (40.0)	22 (22.4)	63 (24.5)
CTQ − emotional neglect (N = 265), n (%)
Yes	24 (22.4)	15 (27.3)	26 (25.2)	65 (24.5)
CTQ − physical neglect (N = 269), n (%)
Yes	22 (20.6)	19 (33.3)	25 (23.8)	66 (24.5)
Childhood maltreatment (N = 271), n (%)
Yes	46 (43.0)	39 (67.2)	52 (49.1)	137 (50.6)
CAS total score (N = 249), mean (SD)	6.0 (14.7)	9.1 (18.8)	5.3 (12.3)	6.5 (15.0)
CAS total score > 3 (N = 249), n (%)
Yes	31 (31.3)	21 (36.2)	22 (23.9)	74 (29.7)

FIGURE 15.

A plot to show the mean of the binary variables in the propensity score before and after weighting by inverse of propensity score. (a) Before weighting by inverse of propensity score; and (b) after weighting by inverse of propensity score.

a, Sections for admissions in current episode and/or previous 2 years.

Please see Appendix 11 for the threshold assessment grid ratings.

Table 16 shows obstetric measures of the population at baseline. Participants whose highest level of care was on an acute ward were more likely than the women in other cohorts to have a pregnancy that ended in miscarriage or stillbirth or to experience a premature birth or a neonatal death. Women who attended a MBU or CRT as their highest level of care were more likely than women whose highest level of service was an acute ward to have had significant problems during the birth.

TABLE 16 - Obstetric measures: the ESMI MBU study (N = 279)

Obstetric measure	Service, n (%)			Total, n (%)
	CRT	Ward	MBU
Miscarriage or stillbirth (N = 278)
Yes	38 (34.9)	16 (26.2)	35 (32.4)	89 (32.0)
Termination of pregnancy (N = 277)
Yes	28 (25.9)	14 (23.0)	29 (26.9)	71 (25.6)
History of neonatal death (N = 278)
Yes	2 (1.8)	2 (3.3)	2 (1.9)	6 (2.2)
Any child(ren) born at < 37 weeks’ gestation (N = 279)
Yes	15 (13.8)	12 (19.4)	18 (16.7)	45 (16.1)
Significant problems during the birth (N = 279)
Yes	55 (50.5)	29 (46.8)	53 (49.1)	137 (49.1)
Assisted conception (N = 276)
Yes	2 (1.9)	1 (1.6)	4 (3.7)	7 (2.5)
Feeding mode (N = 276)
Breast	52 (47.7)	29 (47.5)	43 (40.6)	124 (44.9)
Bottle	30 (27.5)	27 (44.3)	28 (26.4)	85 (30.8)
Mixed	27 (24.8)	5 (8.2)	35 (33.0)	67 (24.3)
Advised against breastfeeding because of psychiatric medication (N = 272)
Yes	18 (16.7)	21 (34.4)	36 (35.0)	75 (27.6)
Advised against breastfeeding because of other medications (N = 258)
Yes	6 (5.7)	3 (5.1)	8 (8.5)	17 (6.6)

Table 17 shows child measures in the population at baseline. Women who attended an acute ward as their highest level of care were more likely than women from the other groups to give birth prematurely or very prematurely. These women also had, on average, more children and were more likely to have had children’s social services assessment or intervention. Compared with women who accessed a ward or MBU services, women who were admitted to CRT services were more likely to be admitted when their child was aged > 100 days and less likely to be admitted before birth.

TABLE 17 - Child measures: the ESMI MBU study (N = 279)

Child measure	Service, n (%)			Total, n (%)
	CRT	Ward	MBU
Gestational age (weeks) (N = 249), mean (SD)	38.2 (1.9)	37.6 (2.3)	38.2 (2.2)	38.1 (2.1)
Term status (gestational age) (N = 249), n (%)
Post term	0 (0.0)	1 (1.9)	0 (0.0)	1 (0.4)
Full term	76 (76.0)	31 (58.5)	73 (76.0)	180 (72.3)
< 37 weeks’ gestation (premature)	23 (23.0)	20 (37.7)	21 (21.9)	64 (25.7)
< 32 weeks’ gestation (very premature)	1 (1.0)	1 (1.9)	2 (2.1)	4 (1.6)
Number of children (N = 279), mean (SD)	1.8 (1.2)	2.0 (1.1)	1.7 (1.1)	1.8 (1.1)
Children’s social services assessment or intervention at baseline (N = 278), n (%)
Yes	27 (24.8)	23 (37.1)	36 (33.6)	86 (30.9)
Age of child at date of first admission (N = 277), n (%)
Admission before birth	2 (1.9)	3 (4.8)	5 (4.7)	10 (3.6)
0–100 days	58 (53.7)	36 (58.1)	79 (73.8)	173 (62.5)
> 100 days	48 (44.4)	23 (37.1)	23 (21.5)	94 (33.9)

Process evaluation results

We collected data from staff on 42 general wards, 42 CRTs and five MBUs in the participating provider organisations in England and Wales. No service declined participation.

Propensity score

The prespecified variables in the propensity scores were explored blind to outcome data. Two out of the 23 variables were continuous: age at consent, which was normally distributed [mean 31.5 (6.0 SD) years] (see Table 14), and number of children (between 1 and 7). The number of children was transformed by taking the square root so that a more continuous distribution was present. Some data were missing for 12 of 23 prespecified variables, with the number of missing values ranging from 1 to 46. Eleven of these variables were binary and one was categorical. Iterative chained equations were used to impute missing values on these 12 variables, together with the other 11 complete variables in the model. A single imputed data set (seed 123) with a burn-in of 10 cycles, logit models to impute the binary variables and mlogit to impute a categorical variable were used. The augment option was added to perform augmented regression in the presence of perfect prediction. One prespecified binary predictor with a high level of missingness and unstable imputation over multiple seeds was omitted (mother ever adopted/fostered).

With the complete set of variables, the propensity score for MBUs compared with non-MBU services was computed using pscore, rather than the teffects command, to check the region of common support and balance between groups blind to outcome. Initial analysis suggested the exclusion of 12 beyond the region of common support and that achieving convincing balance was non-trivial.

A weight was generated from the propensity score to examine if the propensity score reduced the imbalance of key variables across groups for participants in the region of common support. Each variable was summarised by group with and without the weighting (see Figure 15). From visual inspection, number of children, detention and CAS score were still slightly imbalanced after adjusting for propensity score, suggesting that additional covariate adjustment for the effects of these variables would be appropriate.

Primary outcome

Table 18 shows the primary outcome of readmission rates at 12 months post discharge for the cohort groups used in the primary analysis, that is MBUs compared with non-MBU services (acute wards readmission rate, 32.3%; CRT readmission rate, 21.3%).

To maximise the power and robustness of the analysis, our key baseline variables (excluding and, therefore, still blind to, group allocation) were assessed for their ability to predict the primary outcome of readmission. Forward stepwise logistic regression was used with an entry criterion p-value < 0.15. This was also restricted to those in the region of common support. Significant predictors were personality disorder, ethnicity (other), learning disability, age of child at admission, partner and living alone. These six variables along with the three variables already identified as still unbalanced after propensity score weighting (i.e. number of children, section and CAS) were adjusted for in the teffects analysis model.

The teffects-augmented inverse probability weighting approach was used, as this gave scope for double-robustness and gains in power by the inclusion of some regression covariates in addition to propensity score weighting. We included the six variables that were associated with the outcome, as described above, and the three variables for which the propensity score did not visually achieve balance (all identified without reference to the outcome variable). teffects used a logistic regression model for both the propensity score and the analysis model estimating parameters of the outcome model. As in the pscore analysis, a region of common support criterion of 0.05 was specified, which, in this setting, under maximum likelihood, identified 15 women for exclusion (the 12 women previously identified plus an additional three women). The model was then re-estimated using the wnls algorithm and, as described below, reported achieving satisfactory balance. Results are presented in Table 19.

A sensitivity analysis was performed with three different imputation model seeds (23,42,170). Estimates obtained all met the balancing criterion and yielded log-odds treatment effect estimates that varied by an odds ratio (OR) of ± 0.01. Robustness of the effect estimate to changes in estimator was also examined, by using block stratification, nearest neighbour and radius estimators (atts, attr and attn). Although yielding larger standard errors than teffects aipw, the effect estimates themselves changed little.

Our analysis showed that women who were admitted to MBUs had 0.95 times the odds of being readmitted at 12 months, compared with women who were not admitted to a MBU. If those in the MBU arm had not attended a MBU service, their readmission rate was estimated to have been 26.4% (95% CI 19.1% to 33.7%).

The analysed sample showed satisfactory overlap and a test for covariate balance after teffects inverse probability weighting gave no indication of imbalance (p = 0.9942) (see Appendix 12).

Secondary outcomes

Table 20 summarises the secondary outcome results.

Participants completed a CSQ and the perinatal VOICE for each service that they accessed. The secondary analysis associated with these measures used the CSQ/perinatal VOICE score relating to the service they are categorised under via highest level of care. For example, if a participant accessed MBU services, then they were categorised under MBU and their CSQ/perinatal VOICE score for their MBU experience was included in the analysis. Perinatal VOICE analysis included only women who accessed a MBU or an acute ward, as ‘environment themes’ on this questionnaire are not applicable to home treatment CRTs.

Table 21 shows the results for each secondary outcome. Simple linear models are used with propensity scores unless otherwise stated.

The mean maternal sensitivity scores are in the range suggesting a need for parenting intervention and support.

Economic evaluation results

See also Appendix 14 for more details.

In adjusted analyses, total costs from admission to 1 month post discharge were significantly higher in the MBU group (£60,007) than in the non-MBU group (£13,673) (mean difference £44,049, 95% CI £36,638 to £51,461; p < 0.001) (Table 22). This was because of a combination of higher unit costs for MBUs (£707/day), than for generic acute wards (£385/day) and CRTs (£199/contact), and longer lengths of stay in MBUs. QALYs were not significantly different (adjusted mean difference 0.007, 95% CI –0.013 to 0.027; p = 0.496) (see Table 22). The CEACs suggest that there is a 0% probability of MBUs being cost-effective compared with non-MBUs over a willingness-to-pay range from £0 to £50,000 per QALY in complete-case and imputed analyses and using QALYs generated from both EQ-5D-5L and SF-6D. This was because of the high costs associated with the MBU group and similar outcomes between the groups in terms of QALYs.

Use of acute care and other mental health services between discharge and 12-month follow-up was similar between the two groups. Total cost of these services was not significantly different between the groups (mean difference £632, 95% CI –£1326 to £2589; p = 0.527). Full details of the economic results are provided in Appendix 14.

Additional analyses

Strengths and limitations

Acute severe post-partum disorders are uncommon, although they are more common in some groups of women, such as women with bipolar disorder. Despite the help of Clinical Research Network staff, and frequent regular telephone calls and e-mails to services, we found it difficult to identify women, particularly on acute wards, as admissions were often short, women did not have mental capacity to consent and staff were often too busy and changed too frequently for us to always feel sure that we had not missed any eligible women. Moreover, when women were identified and recruited from an acute ward, we would occasionally find that they had spent time on a MBU, sometimes a few months earlier. Where possible, we included these women in the study, as we were still able to gather sufficient baseline data from case notes and data regarding the primary outcome to include them. However, this could mean that the secondary outcome data for the correct time period could not be included and that the study was underpowered.

Most women accessed more than one type of acute care. Therefore, we had to best classify and analyse our data to reflect this in the context of our theoretical understanding of the important components of care. We agreed that MBUs are the most specialist and ‘highest’ level of care that is focused on mother–infant relationships and maternal illness. We debated if this should be compared with other forms of acute care or if the acute care that involved the mother staying with her baby (i.e. CRT) was an equivalent exposure to the MBU. We agreed to compare specialist with non-specialist acute care, but we would also carry out sensitivity analyses in which the exposure was defined in different ways.

Our other significant challenge was the lack of data in case notes for the primary outcome because women had moved or had been discharged back to primary care. Therefore, after REC approval, we added a telephone call at 1 year, during which we asked about acute care, current living circumstances (including care of infant) and EQ-5D-5L, and this enabled further economic evaluation.

Finally, our primary outcome has its limitations, as it could be argued that the primary objective of an admission to a MBU is not prevention of relapse and so the fact that it is not any different from other services is not surprising. The primary objective of MBU admission is to treat the presenting episode of severe mental illness while simultaneously promoting mother–infant interaction. It can take up to 2 years to completely recover from an episode of post-partum psychosis, and it is at this point that mothers (including in our PPI group) report how beneficial MBUs can be in having developed the relationship with the baby. However, a serious relapse leading to acute care could result in further potential disruption to the mother–child relationship (and it is well established from other research4,71 that it is persistence of symptoms that is most likely to be associated with adverse child outcomes). Furthermore, the challenge and resource-intensive nature of assessing the mother–child relationship again at 1 year post discharge (and even more so at retaining the cohort at 2 years) meant that it was not feasible to pursue this potential aspect of long-term outcome.

Summary

Our headline estimate for the adjusted OR for MBU reducing rates of admission is 0.95 (95% CI 0.86 to 1.04), implying a modest potential benefit and a small risk of increasing rates of readmission. This estimate was based on a comparison of women balanced (matched) across an extended list of measures describing their symptomatology and social and demographic characteristics, and on the fact that the result was robust to a range of analysis methods, and was not affected by changing how cohorts were determined, namely basing this on the service to which women were first admitted or in which they spent the majority of their time in care, rather than according to service providing the highest level of care. Nonetheless, imbalance in something excluded or unmeasured remains a possibility. The IV estimator approaches the analysis in an entirely different way, making quite different assumptions, and, in theory, can account for imbalance in these excluded and unmeasured variables. The IV analysis suggested that MBUs could reduce readmission rates by 70%, which is both much larger and strongly significant (p = 0.001), although the CI was wide. As the assumptions of the IV analysis, in particular the assumption of no correlation between quality of services with distance from MBU, cannot all be confirmed as met, we regard this estimate as a sensitivity analysis indicative of potential bias from unmeasured confounders in our headline effect estimate. Although this would provide encouragement that MBUs can cut readmission rates, and certainly reduce any concern that they could increase rates, our level of uncertainty as to the level of efficacy remains high.

Across all three services, more than one in five women were readmitted to acute care in the year after discharge, implying a significant relapse. As discussed at our 2019 stakeholder event, future research should examine the reasons for readmissions and how to reduce these relapse rates, as they are occurring at a time critical for mother and infant. The new NHS Long Term Plan72 advises commissioning of PMH services up to the age of 2 years, and our data support this recommendation. It was striking that, across services, mother–infant interactions, even after 1 month post discharge, were suboptimal. This suggests that more parenting and PMH support is needed for women after discharge.

Economic evaluation did not support the cost-effectiveness of MBUs, compared with acute wards and CRTs, in WP3 over the short term (i.e. to 1 month post discharge) because of long duration of index admissions and, therefore, the high cost of admissions in this cohort. However, if the IV analysis is valid, and MBUs are able to have a much greater impact on readmission rates than is implied by the primary analysis, cost-effectiveness advantages may exist, as the high initial cost of MBUs might be offset by savings from reduced subsequent admissions. However, as noted above, caution is required in coming to any firm conclusions, given that the assumptions of the IV analysis, in particular the assumption of no correlation between quality of services with distance from MBU, cannot be confirmed as met.

Reflections on what was and what was not successful

We had very high rates of follow-up in all WPs. This was the result of very hard work by outstanding field researchers, primary outcomes chosen on the basis of importance to women themselves (which were, therefore, meaningful for all participating women) and outcomes chosen on the basis of feasible data collection (e.g. telephone vs. face to face). We are not able to examine how representative the women recruited to WP3 were compared with women who were not recruited as, for WP3, we were not given ethics approval to record a minimum data set on all women approached. However, for WP1, at least, the base population was similar to the study population and other demographics. Although participants across all studies include very disadvantaged women, with high rates of childhood and adult abuse (WP3), selection bias remains likely. Moreover, women who lose custody, sometimes repeatedly (WP2), did not benefit from current service configurations. Our sensitivity analyses examined the effect of classifying women by the service they spent most time in or attended first and this did not change the direction of results.

Low response rates from partners is a methodological challenge, but recruiting significant others is important, and with persistence and resources can be achieved, as in WP2. This is particularly important, as the NHS Long Term Plan72 recommends that significant others are assessed for mental ill health. In addition, other guidance, for example the NICE public health guideline on domestic violence and abuse,73 recommends training for health-care professionals to safely identify and respond to domestic abuse.

Recommendations for future research

• We established that the Whooley questions and EPDS can be used to identify depression or other mental disorders in pregnancy. Predictive modelling, using other risk factors, such as social support and a history of recent self-harm, routinely collected by midwives would enable more accurate identification to take place.

• We do not know if identification of mental disorders in pregnancy improves outcomes for mother and child, although the new investment assumes that this is the case. Future research should examine whether or not the introduction of specialist services has improved outcomes, and if improvements have not been seen, then examine why. Future research should also examine the impact of interventions in the pre-conception period, as this may be an even more opportune time74 to benefit women and children.

• Future research should focus on the extent of inequity of access of certain groups and how to reach out to these communities.

• Continuity has been identified as an important component of maternity care and women emphasised its importance in WPs 1, 2 and 3. Future research could examine how best to deliver continuity of care when working with women under adult generic mental health services, PMH services and maternity and universal services.

• Future research should explore how the wider system of care can optimise support for women with acute severe postnatal mental disorders.

• It is difficult to carry out research when services are undergoing transformation, but this could not have been foreseen. In future, where services are changing dramatically, an interrupted time series using routine data may be more helpful and is planned in ESMI II [see NIHR Journals Library URL: www.journalslibrary.nihr.ac.uk/programmes/hsdr/174938/#/ (accessed 12 August 2021)]. In hindsight, our continued attempts to recruit to the exploratory GSH pilot trial [WP1(ii)], including opening new recruitment sites, should have stopped earlier, as changes in policy had occurred across England and recruitment was, therefore, likely to be difficult in the additional sites we opened. In retrospect, however, we could have added additional sites earlier in parts of the UK where admissions to MBUs were not possible, such as South Wales, as this may have facilitated recruitment of more participants to the acute ward arm; however, the governance procedures were lengthy and could have delayed progress. In addition, the cost of recruiting further away, including travel expenses and time, was also a concern. Since the pandemic, we have been carrying out recruitment remotely for new studies, such as ESMI II, and this recruitment strategy could be explored for future studies. At the time of this study, however, recruiting women with severe mental illness remotely may not have been approved by the REC. However, it is now appreciated that when carried out carefully, and with appropriate safeguards, this can be achievable.

Implications for practice and any lessons learnt

• Care pathways providing perinatal tailored services were clearly associated with considerably higher levels of satisfaction across WPs, whether measured by our new perinatal VOICE measure, the CSQ or qualitative interviews. However, non-specialist staff (e.g. midwives, CRT staff, etc.) need PMH training.

• The high levels of obstetric complications and poor pregnancy outcomes seen in WP3 need to be considered in terms of the broader health needs of women with mental illness. Staff training on comorbid health problems is needed.

• The high prevalence of domestic violence and abuse experienced by participants means that PMH services need to ‘factor coercive partners into our thinking’ (stakeholder comment at our final event). Conversely, women and babies live in the context of wider family life, and services need to ensure that significant others are involved in maternal care and are supported themselves.

• In view of the suboptimal maternal sensitivity seen at 1 month post discharge, aftercare needs to include parenting support, continuity of staff (where possible), interventions for maternal illness and assistance with other needs.

Patient and Carer Advisory Group members

Amanda Grey, Sarah Spring, Ceri Rose, Eleanor O’Sullivan, Maria Bavetta, Jessie Hunt, Kathryn Grant, Dr Diana Rose, Liberty Moss, Henry Fay, Catherine Muhammad, Holly Wright and Antoaneta Sokolova.

Programme Steering Committee members

Professor Rona McCandlish (chairperson), Ms Rosie-Ann Jones, Dr Heather O’Mahen, Ms Ceri Rose, Professor Pauline Slade and Sarah Spring.

Trial Steering Committee members

Professor Cathy Creswell (chairperson), Professor Christine MacArthur, Juliette O’Donnell and Professor Vivette Glover.

Independent Data Monitoring and Ethics Committee members

Dr Roch Cantwell (chairperson), Dr Stephen Bremner, Dr Liz McDonald Clifford, Dr Ron Gray and Marian Knight.

Contributions of authors

Louise M Howard (https://orcid.org/0000-0001-9942-744X) (Consultant Psychiatrist) was chief investigator for this programme and conceived the studies with the help of other authors, supervised the London-based research team and took overall responsibility for the outputs. In addition, she led the writing of this final report.

Kathryn M Abel (https://orcid.org/0000-0003-3538-8896) (Consultant Psychiatrist) supervised research staff based at the University of Manchester, conceived the ideas for the studies with Louise M Howard and was involved in writing sections of the report.

Katie H Atmore (https://orcid.org/0000-0003-3012-4584) (Research Assistant) analysed data on significant others and assisted with the writing up of these data.

Debra Bick (https://orcid.org/0000-0002-8557-7276) (Professor of Evidence-based Midwifery) was midwifery lead for WP1 and provided expertise also for WP2.

Amanda Bye (https://orcid.org/0000-0002-9808-4956) (Research Associate) was involved in data collection for WP1, conceived analyses for eating disorders and was involved in the writing up of these data for publication.

Sarah Byford (https://orcid.org/0000-0001-7084-1495) (Professor of Health Economics) led the health economic component of the programme.

Lauren E Carson (https://orcid.org/0000-0002-7027-3077) (Postdoctoral Researcher) was involved in data collection for WP3 and the writing up of findings.

Clare Dolman (https://orcid.org/0000-0001-7466-1651) (Service User Researcher) led the PPI component of the programme.

Margaret Heslin (https://orcid.org/0000-0002-3094-9255) (Junior Health Economist) led analyses under Sarah Byford’s supervision.

Myra Hunter (https://orcid.org/0000-0002-0722-6761) (Professor of Health Psychology) contributed to study design, analyses and writing up of WP1(ii).

Stacey Jennings (https://orcid.org/0000-0002-1162-6370) (Research Associate) was involved in data collection and wrote the section on process evaluation.

Sonia Johnson (https://orcid.org/0000-0002-2219-1384) (Professor of Psychiatry) led WP2 and provided expertise on CRTs and acute care for WP3.

Ian Jones (https://orcid.org/0000-0001-5821-5889) (Consultant Perinatal Psychiatrist) conceived the idea for WP3(ii) and was involved in writing sections of the report.

Billie Lever Taylor (https://orcid.org/0000-0002-6865-3425) (Clinical Psychologist Researcher) was the researcher responsible for data collection for WP2.

Rebecca McDonald (https://orcid.org/0000-0003-3373-4943) (Temporary Programme Manager) led the construction of the geographical IV.

Jeannette Milgrom (https://orcid.org/0000-0002-4082-4595) (Professor of Perinatal Psychology) provided expertise for the development of the DAWN intervention in WP1(ii).

Nicola Morant (https://orcid.org/0000-0003-4022-8133) (Associate Professor) led the qualitative analysis of WP2.

Selina Nath (https://orcid.org/0000-0002-5976-1301) (Postdoctoral Researcher) was responsible for follow-up data collection in WP1, analysed data on the GAD-2 and assisted with the writing up of these data.

Susan Pawlby (https://orcid.org/0000-0001-9109-3457) (Developmental Psychologist) led the mother–infant interaction data collection and interpretation.

Laura Potts (https://orcid.org/0000-0002-2935-6532) (Junior Statistician) collated data and results of the programme and analysed WP3 data.

Claire Powell (https://orcid.org/0000-0002-6581-0165) (Research Associate) contributed to WP3 study design, data collection, and interpretation of results.

Diana Rose (https://orcid.org/0000-0002-6393-6389) (Professor of Service User Research) conceived the study for WP3(i) and chaired the PAG.

Elizabeth Ryan (https://orcid.org/0000-0002-6802-4194) (Junior Statistician) analysed data for WP1.

Gertrude Seneviratne (https://orcid.org/0000-0003-0793-3674) (Lead Clinician for PMH Services at South London and Maudsley NHS Foundation Trust) provided expertise in interpretation of results.

Rebekah Shallcross (https://orcid.org/0000-0003-4764-0411) (Researcher) contributed to writing of study protocols, collected data for WP3 and was involved in writing sections of the report.

Nicky Stanley (https://orcid.org/0000-0002-7644-1625) (Professor of Social Work) provided expertise from a social care perspective.

Kylee Trevillion (https://orcid.org/0000-0001-5865-4726) (Programme Manager 2013–17) was responsible for supervision of London-based research staff, wrote the pan-programme standard operating procedure with respect to safeguarding and led writing up of WP1(ii).

Angelika Wieck (https://orcid.org/0000-0002-5138-4898) (Consultant Perinatal Psychiatrist) helped write the grant, assisted with recruitment in northern trusts and was involved in developing protocols.

Andrew Pickles (https://orcid.org/0000-0003-1283-0346) (Senior Statistician) contributed to the design of the programme and oversaw the analysis and reporting.

Publications

Trevillion K, Domoney J, Pickles A, Bick D, Byford S, Heslin M, et al. Depression: an exploratory parallel-group randomised controlled trial of Antenatal guided self help for WomeN (DAWN): study protocol for a randomised controlled trial. Trials 2016;17:503. https://doi.org/10.1186/s13063-016-1632-6

Howard LM, Ryan EG, Trevillion K, Anderson F, Bick D, Bye A, et al. Accuracy of the Whooley questions and the Edinburgh Postnatal Depression Scale in identifying depression and other mental disorders in early pregnancy. Br J Psychiatry 2018;212:50–6. https://doi.org/10.1192/bjp.2017.9

Lever Taylor B, Billings J, Morant N, Johnson S. How do women’s partners view perinatal mental health services? A qualitative meta-synthesis. Clinical Psychol Psychother 2018;25:112–29. https://doi.org/10.1002/cpp.2133

Millet L, Lever Taylor B, Howard LM, Bick D, Stanley M, Johnson S. Experiences of improving access to psychological therapy services for perinatal mental health difficulties: a qualitative study of women’s and therapists’ views. Behav Cogn Psychother 2018;46:421–36. https://doi.org/10.1017/S1352465817000650

Nath S, Ryan EG, Trevillion K, Bick D, Demilew J, Milgrom J, et al. Prevalence and identification of anxiety disorders in pregnancy: the diagnostic accuracy of the two-item Generalised Anxiety Disorder scale (GAD-2). BMJ Open 2018;8:e023766. https://doi.org/10.1136/bmjopen-2018-023766

Anderson FM, Hatch SL, Ryan EG, Trevillion K, Howard LM. Impact of insecure immigration status and ethnicity on antenatal mental health among migrant women. J Clin Psychiatry 2019;80:18m12563. https://doi.org/10.4088/JCP.18m12563

Gordon H, Nath S, Trevillion K, Moran P, Pawlby S, Newman L, Howard LM, Molyneaux E. Self-harm, self-harm ideation, and mother–infant interactions: a prospective cohort study. J Clin Psychiatry 2019;80:18m12708. https://doi.org/10.4088/JCP.18m12708

Griffiths J, Lever Taylor B, Morant N, Bick D, Howard LM, Seneviratne G, Johnson S. A qualitative comparison of experiences of specialist mother and baby units versus general psychiatric wards. BMC Psychiatry 2019;19:401. https://doi.org/10.1186/s12888-019-2389-8

Heslin M, Chua KC, Trevillion K, Nath S, Howard LM, Byford S. Psychometric properties of the five-level EuroQol-5 dimension and Short Form-6 dimension measures of health-related quality of life in a population of pregnant women with depression. BJPsych Open 2019;5:e88. https://doi.org/10.1192/bjo.2019.71

Lever Taylor B, Billings J, Morant N, Bick D, Johnson S. Experiences of how services supporting women with perinatal mental health difficulties work with their families: a qualitative study in England. BMJ Open 2019;9:e030208. https://doi.org/10.1136/bmjopen-2019-030208

Lever Taylor B, Mosse L, Stanley N. Experiences of social work intervention among mothers with perinatal mental health needs. Health Soc Care Community 2019;27:1586–96. https://doi.org/10.1111/hsc.12832

Lockwood Estrin G, Ryan EG, Trevillion K, Demilew J, Bick D, Pickles A, Howard LM. Young pregnant women and risk for mental disorders: finding from an early pregnancy cohort. BJPsych Open 2019;5:e21. https://doi.org/10.1192/bjo.2019.6

MacLeod Hall C, Molyneaux E, Gordon H, Trevillion K, Moran P, Howard LM. The association between a history of self-harm and mental disorders in pregnancy. J Affect Disord 2019;258:159–62. https://doi.org/10.1016/j.jad.2019.06.062

Nath S, Pearson RM, Moran P, Pawlby S, Molyneaux E, Challacombe FL, Howard LM. The association between prenatal maternal anxiety disorders and postpartum perceived and observed mother-infant relationship quality. J Affect Disord 2019;68:102148. https://doi.org/10.1016/j.janxdis.2019.102148

Trevillion K, Shallcross R, Ryan E, Heslin M, Pickles A, Byford S, et al. Protocol for a quasi-experimental study of the effectiveness and cost-effectiveness of mother and baby units compared with general psychiatric inpatient wards and crisis resolution team services (the ESMI study) in the provision of care for women in the postpartum period. BMJ Open 2019;9:e025906. https://doi.org/10.1136/bmjopen-2018-025906

Yapp E, Howard LM, Kadicheeni M, Telesia LA, Milgrom J, Trevillion K. A qualitative study of women’s views on the acceptability of being asked about mental health problems at antenatal booking appointments. Midwifery 2019;74:126–33. https://doi.org/10.1016/j.midw.2019.03.021

Bye A, Nath S, Ryan EG, Bick D, Easter A, Howard LM, Micali N. Prevalence and clinical characterisation of pregnant women with eating disorders. Eur Eat Disord Rev 2020;28:141–55. https://doi.org/10.1002/erv.2719.

Challacombe FL, Nath S, Trevillion K, Pawlby S, Howard LM. Fear of childbirth during pregnancy: associations with observed mother–infant interactions and perceived bonding. Arch Womens Ment Health 2020;24:483–92. https://doi.org/10.1007/s00737-020-01098-w

Crowley G, Molyneaux E, Nath S, Trevillion K, Moran P, Howard LM. Disordered personality traits and psychiatric morbidity in pregnancy: a population-based study. Arch Womens Ment Health 2020;23:43–52. https://doi.org/10.1007/s00737-018-0937-8

Hazell Raine K, Nath S, Howard LM, Cockshaw W, Boyce P, Sawyer E, Thorpe K. Associations between prenatal maternal mental health indices and mother–infant relationship quality 6 to 18 months’ postpartum: a systematic review. Infant Ment Health J 2020;41:24–39. https://doi.org/10.1002/imhj.21825

Lever Taylor B, Kandiah A, Johnson S, Howard LM, Morant N. A qualitative investigation of models of community mental health care for women with perinatal mental health problems [published online ahead of print January 30 2020]. J Ment Health 2020. https://doi.org/10.1080/09638237.2020.1714006

Nath S, Busuulwa P, Ryan EG, Challacombe FL, Howard LM. The characteristics and prevalence of phobias in pregnancy. Midwifery 2020;82:102590. https://doi.org/10.1016/j.midw.2019.102590

Nath S, Pearson RM, Moran P, Pawlby S, Molyneaux E, Howard L. Maternal personality traits, antenatal depressive symptoms and the postpartum mother–infant relationship: a prospective observational study. Soc Psychiatry Psychiatr Epidemiol 2020;55:621–34. https://doi.org/10.1007/s00127-019-01790-y

Trevillion K, Ryan EG, Pickles A, Heslin M, Byford S, Nath S, et al. An exploratory parallel-group randomised controlled trial of antenatal guided self-help (plus usual care) versus usual care alone for pregnant women with depression: DAWN trial. J Affect Disord 2020;261:187–97. https://doi.org/10.1016/j.jad.2019.10.013

Powell C, Bedi S, Nath S, Potts L, Trevillion K, Howard LM. Mothers’ experiences of acute perinatal mental health services in England and Wales: a qualitative analysis [published online ahead of print September 3 2020]. J Reprod Infant Psychol 2020.

Kolk TA, Nath S, Howard LM, Pawlby S, Lockwood-Estrin G, Trevillion K. The association between maternal lifetime interpersonal trauma experience and perceived mother-infant bonding. J Affect Disord 2021;294:117–27. https://doi.org/10.1016/j.jad.2021.06.069

Lever Taylor B, Howard LM, Jackson K, Johnson S, Mantovani N, Nath S, et al. Mums alone: exploring the role of isolation and loneliness in the narratives of women diagnosed with perinatal depression. J Clin Med 2021;10:2271. https://doi.org/10.3390/jcm10112271

Nath S, Lewis LN, Bick D, Demilew J, Howard LM. Mental health problems and fear of childbirth: a cohort study of women in an inner-city maternity service. Birth 2021;48:230–41. https://doi.org/10.1111/birt.12532

Perkins A, Lever Taylor B, Morant N, Johnson S. Experiences of parent-infant teams among mothers diagnosed with perinatal mental health difficulties [published online ahead of print 30 September 2021]. J Reprod Infant Psychol 2021. https://doi.org/10.1080/02646838.2021.1983920

Rubio L, Lever Taylor B, Morant N, Johnson S. Experiences of intensive home treatment for a mental health crisis during the perinatal period: a UK qualitative study. Int J Ment Health Nurs 2021;30:208–18. https://doi.org/10.1111/inm.12774

Sweeney A, Lever Taylor B. Depathologising Motherhood. Russo J, Beresford P, editors. The Routledge Handbook of Mad Studies. London: Routledge; 2021.

Zacharia A, Lever Taylor B, Sweeney A, Morant N, Howard LM, Johnson S. Mental health support in the perinatal period for women with a personality disorder diagnosis: a qualitative study of women’s experiences. J Pers Disord 2021;35. https://doi.org/10.1521/pedi_2020_34_482

Howard LM, Trevillion K, Potts L, Heslin M, Pickles A, Byford S, et al. Effectiveness and cost-effectiveness of psychiatric mother and baby units: quasi-experimental study [published online ahead of print 4 May 2022]. Br J Psychiatry 2022.

Data-sharing statement

All available data from WPs 1 and 3 can be obtained from the corresponding author. WP2 data can be obtained from Professor Sonia Johnson.

Patient data

This work uses data provided by patients and collected by the NHS as part of their care and support. Using patient data is vital to improve health and care for everyone. There is huge potential to make better use of information from people’s patient records, to understand more about disease, develop new treatments, monitor safety, and plan NHS services. Patient data should be kept safe and secure, to protect everyone’s privacy, and it’s important that there are safeguards to make sure that it is stored and used responsibly. Everyone should be able to find out about how patient data are used. #datasaveslives You can find out more about the background to this citation here: https://understandingpatientdata.org.uk/data-citation.

Disclaimers

This report presents independent research funded by the National Institute for Health and Care Research (NIHR). The views and opinions expressed by authors in this publication are those of the authors and do not necessarily reflect those of the NHS, the NIHR, CCF, PGfAR or the Department of Health and Social Care. If there are verbatim quotations included in this publication the views and opinions expressed by the interviewees are those of the interviewees and do not necessarily reflect those of the authors, those of the NHS, the NIHR, the PGfAR programme or the Department of Health and Social Care.

Aims

Economic decision-analytic modelling was used to explore the incremental cost per:

1. true-positive case of depression detected

2. QALY of depression screening.

The above aims were from the grant application, but we also added a third aim:

1. to explore the incremental cost per true-positive case of any mental disorder detected.

Methods

Model structure

Two types of decision tree model structures were developed: (1) a screening model and (2) a screening and treatment pathway model. The screening model was used to evaluate incremental cost per true-positive case of depression (model 1) and per true-positive case of all mental disorders (model 2).

The screening model pathway used for Whooley questions, EPDS and no screening is shown in Figure 16 and the Whooley questions–EPDS pathway is shown in Figure 17. The detection and treatment model was used to evaluate the incremental cost per QALY (model 3) and was based on the detection model, but additionally modelled the impact of treatment (Figure 18). Treatment pathways consisted of being allocated to GSH in the case of mild to moderate depression or high-intensity psychological intervention for moderate to severe depression (as indicated by NICE guidance5). This is followed by response to treatment or not.

FIGURE 16.

Screening model pathway for Whooley questions, EPDS and no screening.

FIGURE 17.

Screening model pathway for Whooley questions followed by EPDS.

FIGURE 18.

Screening and treatment model. HIPI, high-intensity psychological intervention.

Model 1: incremental cost per true-positive depression case detected

Base case

The results of the model 1 base case (Table 33 and Figure 19) show that the probability of detecting a true-positive case was highest for EPDS (0.059), followed by Whooley questions (0.041), no screening (0.029) and Whooley questions–EPDS (0.028). The cost of screening was highest for EPDS (£9.38), followed by no screening (£7.95), Whooley questions–EPDS (£5.37) and Whooley questions (£4.53).

TABLE 33 - Model 1 base-case costs and probabilities of detecting a true depression case

Screening approach	Probability of being true positive	Cost per screen (£)	Status	ICER (£)
EPDS	0.0594	9.38	Trade-off	267
Whooley questions	0.0412	4.53	Trade-off
No screening	0.0292	7.95	Dominated
Whooley questions–EPDS	0.0276	5.37	Dominated

FIGURE 19.

Model 1 base-case probability of detecting a true-positive depression case by cost per screen.

Using rules of dominance and extended dominance, no screening and Whooley questions–EPDS were dominated by Whooley questions, which was more clinically effective and less costly than both. A trade-off occurred between EPDS and Whooley questions, with EPDS costing more but being more clinically effective, resulting in an ICER of £267.

The cost-effectiveness plane for Whooley questions compared with EPDS (Figure 20) shows that all scatterpoints lie to the left of the x-axis where the Monte Carlo simulations represent points where Whooley questions is less clinically effective than EPDS, and the majority of scatterpoints lie below the x-axis where Whooley questions is less costly than EPDS. Figure 21 shows the associated CEACs. There is a higher probability (> 50%) of Whooley questions being cost-effective, compared with EPDS, at low values of willingness to pay (i.e. < £250) and EPDS being cost-effective at higher values (i.e. > £250). The four-way CEAC confirms these findings (Figure 22).

FIGURE 20.

Model 1 base-case cost-effectiveness plane for Whooley questions vs. EPDS in detecting a true depression case.

FIGURE 21.

Model 1 CEAC for Whooley questions vs. EPDS in detecting a true depression case.

FIGURE 22.

Model 1 base-case CEAC for all screening options in detecting a true depression case.

Sensitivity analysis

When model 1 was reanalysed using probabilities and costs from a study on the screening of depression by GPs (see Table 32), the probability of detecting a true-positive case was highest for no screening (0.100), followed by EPDS (0.059), Whooley questions (0.041) and Whooley questions–EPDS (0.028) (Table 34 and Figure 23). The cost of screening was highest for no screening because of inclusion of GP costs (£31), followed by EPDS (£9.38), Whooley questions–EPDS (£5.37) and Whooley questions (4.53).

TABLE 34 - Model 1 sensitivity analysis costs and probabilities of detecting a true depression case

Screening approach	Probability of being true positive	Cost per screen (£)	Status	ICER (£)
No screening	0.1000	31.00	Trade-off	532
EPDS	0.0594	9.38	Trade-off	267
Whooley questions	0.0412	4.53	Trade-off
Whooley questions–EPDS	0.0276	5.37	Dominated

FIGURE 23.

Model 1 sensitivity analysis probability of detecting a true-positive depression case by cost per screen.

Using the rules of dominance and extended dominance, Whooley questions–EPDS was dominated by Whooley questions, which was more clinically effective and less costly. A trade-off occurred for no screening, EPDS and Whooley questions, with each alternative costing more but having a higher probability of detecting a true-positive case than the last. This resulted in an ICER of £267 for EPDS compared with Whooley questions and of £532 for no screening compared with EPDS.

Figures 24–26 present cost-effectiveness planes for each two-way comparison involving a trade-off. In Figures 23 and 24, the majority of scatterpoints lie to the left of the x-axis where the Monte Carlo simulations represent points where Whooley questions/EPDS are less clinically effective than no screening, and the majority of the scatterpoints lie below the x-axis, where Whooley questions/EPDS are less costly than no screening. In Figure 26, all scatterpoints lie to the left of the x-axis, where Whooley questions are less clinically effective than EPDS, and the majority of scatterpoints lie below the x-axis, where Whooley questions are less costly than EPDS.

FIGURE 24.

Model 1 sensitivity analysis cost-effectiveness plane for Whooley questions vs. no screening in detecting a true depression case.

FIGURE 25.

Model 1 sensitivity analysis cost-effectiveness plane for EPDS vs. no screening in detecting a true depression case.

FIGURE 26.

Model 1 sensitivity analysis cost-effectiveness plane for Whooley questions vs. EPDS in detecting a true depression case.

Figure 27 shows the CEACs for Whooley questions, EPDS and no screening. The probability of being cost-effective is highest for Whooley questions at low levels of willingness to pay (i.e. < £250), for EPDS at willingness-to-pay levels of approximately £250–600 and for no screening at high levels of willingness to pay (i.e. > £600). The four-way CEAC confirms these findings (Figure 28).

FIGURE 27.

Model 1 sensitivity analysis CEAC for Whooley questions, EPDS and no screening in detecting a true depression case.

FIGURE 28.

Model 1 sensitivity analysis CEAC for all screening options in detecting a true depression case.

Model 2: incremental cost per true-positive case of any mental disorder detected

Base case

The results of the model 2 base case (Table 35 and Figure 29) show that the probability of detecting a true-positive case of any mental disorder was highest for EPDS (p = 0.083), followed by Whooley questions (0.063), Whooley questions–EPDS (0.038) and no screening (0.029). The cost of screening was highest for EPDS (£9.38), followed by no screening (£7.95), Whooley questions–EPDS (£5.37) and Whooley questions (£4.53).

TABLE 35 - Model 2 base-case costs and probabilities of detecting a true case of any mental disorder

Screening approach	Probability of being a true positive	Cost per screen (£)	Status	ICER (£)
EPDS	0.0832	9.38	Trade-off	241
Whooley questions	0.0631	4.53	Trade-off
Whooley questions–EPDS	0.0376	7.95	Dominated
No screening	0.0292	5.37	Dominated

FIGURE 29.

Model 2 base-case probability of detecting a true case of any mental disorder by cost per screen for each screening approach.

Using the rules of dominance and extended dominance, no screening and Whooley questions–EPDS were dominated by Whooley questions, which was more clinically effective and less costly than both. A trade-off occurred between EPDS and Whooley questions, with EPDS costing more but having a higher probability of detecting a true-positive case, resulting in an ICER of £241.

Figure 30 presents the cost-effectiveness plane for Whooley questions compared with EPDS. All scatterpoints lie to the left of the x-axis, where the Monte Carlo simulations represent points where Whooley questions are less clinically effective than EPDS, and the majority of the scatterpoints lie below the x-axis, where Whooley questions is less costly than EPDS. Figure 31 shows the CEACs for Whooley questions compared with EPDS. There is a higher probability (> 50%) of Whooley questions being cost-effective than EPDS at low values of willingness to pay (i.e. < £250) and EPDS being cost-effective at values beyond this. The four-way CEAC confirms these findings (Figure 32).

FIGURE 30.

Model 2 base-case cost-effectiveness plane for Whooley questions vs. EPDS in detecting a true case of any mental disorder.

FIGURE 31.

Model 2 base-case CEAC for Whooley questions vs. EPDS in detecting a true case of any mental disorder.

FIGURE 32.

Model 2 base-case CEAC for all screening options in detecting a true case of any mental disorder.

Sensitivity analysis

As for model 1, when the model was reanalysed using probabilities and costs from an alternative study on the screening of depression by GPs (see Table 32), the results changed (Table 36 and Figure 33). The probability of detecting a true-positive case was highest for no screening (0.1), followed by EPDS (0.083), Whooley questions (0.063) and Whooley questions–EPDS (0.038). The cost of screening was highest for no screening (£31), followed by EPDS (£9.38), Whooley questions–EPDS (£5.37) and Whooley questions (£4.53).

TABLE 36 - Model 2 sensitivity analysis costs and probabilities of detecting a true case of any mental disorder

Screening approach	Probability of being a true positive	Cost per screen (£)	Status	ICER (£)
No screening	0.1000	31.00	Trade-off	1273
EPDS	0.0832	9.38	Trade-off	255
Whooley questions	0.0631	4.53	Trade-off
Whooley questions–EPDS	0.0376	5.37	Dominated

FIGURE 33.

Model 2 sensitivity analysis probability of detecting a true case of any mental disorder by cost per screen for each screening approach.

Using the rules of dominance and extended dominance, Whooley questions–EPDS was dominated by Whooley questions, which was more clinically effective and less costly. A trade-off occurred for no screening, EPDS and Whooley questions, with each alternative costing more but having a higher probability of detecting a true-positive case than the last. This resulted in an ICER of £255 for EPDS compared with Whooley questions and of £1273 for no screening compared with EPDS.

Figures 34–36 present cost-effectiveness planes for the two-way comparisons involving a trade-off. In Figures 34 and 35, the majority of scatterpoints lie to the left of the x-axis, where the Monte Carlo simulations represent points where Whooley questions/EPDS are less effective than no screening, and the majority of scatterpoints lie below the x-axis, where Whooley questions/EPDS are less costly than no screening. In Figure 36, all scatterpoints lie to the left of the x-axis where Whooley questions is less clinically effective than EPDS, and the majority of scatterpoints lie below the x-axis, where Whooley questions is less costly than EPDS.

FIGURE 34.

Model 2 sensitivity analysis cost-effectiveness plane for Whooley questions vs. no screening in detecting a true case of any mental disorder.

FIGURE 35.

Model 2 sensitivity analysis cost-effectiveness plane for EPDS vs. no screening in detecting a true case of any mental disorder.

FIGURE 36.

Model 2 sensitivity analysis cost-effectiveness plane for Whooley questions vs. EPDS in detecting a true case of any mental disorder.

Figure 37 shows the CEAC for Whooley questions, EPDS and no screening. The probability of being cost-effective is highest for Whooley questions at low levels of willingness to pay (i.e. < £250) and for EPDS at higher levels of willingness to pay (i.e. > £250). The four-way CEAC confirms these findings (Figure 38).

FIGURE 37.

Model 2 sensitivity analysis CEAC for Whooley questions, EPDS and no screening in detecting a true case of any mental disorder.

FIGURE 38.

Model 2 sensitivity analysis CEAC for all screening options in detecting a true case of any mental disorder.

Model 3: incremental cost per quality-adjusted life-year

Base case

The results of the model 3 base case are presented in Table 37 and Figure 39. The mean QALY gain per person was highest for EPDS (0.7304), followed by Whooley questions (0.7302), Whooley questions–EPDS (0.7301) and no screening (0.7255). Total cost per person was highest for EPDS (£1799), followed by Whooley questions (£1772), no screening (£1765) and Whooley questions–EPDS (£1748).

TABLE 37 - Model 3 base-case costs and QALYs for each screening option

Screening approach	QALYs	Costs (£)	Status	ICER (£)
EPDS	0.7304	1799	Trade-off	108,419
Whooley questions	0.7302	1772	Trade-off	312,181
Whooley questions–EPDS	0.7301	1748	Trade-off
No screening	0.7255	1765	Dominated

FIGURE 39.

Model 3 base-case costs and QALYs for each screening option.

Using the rules of dominance and extended dominance, no screening was dominated by Whooley questions–EPDS, which was more clinically effective and less costly. A trade-off occurred for EPDS, Whooley questions and Whooley questions–EPDS, with EPDS costing more but producing more QALYs than the other strategies. Whooley questions–EPDS had the lowest cost of the remaining options, but also produced the lowest QALYs. The ICER was £108,419 per QALY for EPDS and £312,181 per QALY for Whooley compared with Whooley questions–EPDS.

Figure 40 presents the cost-effectiveness plane for Whooley questions compared with EPDS. Figure 40 shows that the scatterpoints are approximately equal in each quadrant, suggesting no advantage for either option in terms of costs or effects. This was similar for Whooley questions compared with Whooley questions–EPDS (Figure 41) and for EPDS compared with Whooley questions–EPDS (Figure 42).

FIGURE 40.

Model 3 base-case cost-effectiveness plane for Whooley questions vs. EPDS for cost per QALY.

FIGURE 41.

Model 3 base-case cost-effectiveness plane for Whooley questions vs. Whooley questions–EPDS for cost per QALY.

FIGURE 42.

Model 3 base-case cost-effectiveness plane for EPDS vs. Whooley questions–EPDS for cost per QALY.

The four-way CEAC (Figure 43) indicates that, at a willingness to pay of £0, all options have a similar probability of being cost-effective. However, as willingness to pay increases, the probability of no screening being cost-effective falls, whereas the probability for all other options remains similar.

FIGURE 43.

Model 3 base-case CEAC for all screening options for cost per QALY.

Sensitivity analysis

The results of the sensitivity analysis were similar to the base case, with no screening being dominated and the other screening options involving a trade-off. The four-way CEAC (Figure 44) confirms that Whooley questions, EPDS and Whooley questions–EPDS all have a similar probability of being cost-effective.

FIGURE 44.

Model 3 sensitivity analysis CEAC for all screening options for cost per QALY.

Strengths and limitations

This study included data from a cross-sectional survey specifically designed to compare the accuracy of alternative approaches to detecting depression in pregnant women at the first antenatal appointment. This is the earliest opportunity to systematically detect depression in pregnancy. Furthermore, this study assessed the accuracy of the Whooley questions when asked by midwives at a routine maternity contact rather than assessing responses to researchers and, therefore, the results are of relevance to usual clinical practice. Other strengths include the use of a robust diagnostic interview, an efficient and well-powered study design and a diverse study population.

A number of limitations that could have influenced the results should be considered. Although the Whooley questions were asked by midwives in clinical practice, the EPDS was administered by researchers. Therefore, the diagnostic accuracy of the EPDS may not reflect accuracy in clinical practice; however, as it is a self-complete instrument, its administration by researchers is unlikely to change its diagnostic accuracy. Furthermore, there was a 2- to 3-week delay in administering the EPDS and the SCID-I after the first antenatal appointment when the Whooley questions were asked and so changes in mental health state over this time period are possible.

The model is also limited by the need to make certain assumptions where data were unavailable. Key assumptions included that all women are screened, no women are receiving IAPT prior to presentation, all women who screen positive are referred to IAPT and no one who screened negative becomes depressed at a later point. However, assumptions are necessary in economic modelling, as models are a simplification of reality. Furthermore, these assumptions are consistent with related models. 5,30 Additionally, the resources, and therefore the costs of identifying depressed women in the no-screening option, were estimated based on the clinical opinion of a single consultant midwife (Jill Demilew). However, this midwife has over 40 years’ clinical experience and 20 years’ experience as a consultant midwife. In addition, this estimate was varied in sensitivity analyses to check the robustness of the results to this assumption.

The generalisability of the model must also be considered, as most data on the sensitivity and specificity of the screening tools came from one study based on one inner-city area,10 and screening data were available for only 33% of all eligible women. However, this is the first study to examine the cost-effectiveness of detecting and treating depression early in pregnancy informed by real-world data on screening tool accuracy, and there is flexibility in economic models to update the model parameters as additional data become available. Additionally, the timescale of this evaluation is limited to 3 months post birth and any longer-lasting impacts of screening and treatment of depression are therefore not captured.

Psychometric characteristics of questionnaire

The response set over the 27 items of Box 2 was a six-category Likert scale, with response frequencies as shown in Figure 7. Response distributions for several items showed modal values at the bounds of the scale. Except in the generation of simple-item total scores, our analyses have treated the response set as ordinal categories throughout and, therefore, account for these item floor/ceiling effects. Item test–retest reliability was assessed using kappa statistics (quadratic weights) and intraclass correlation for the item total. Item covariance structure was investigated using exploratory factor analysis. The scope for scale abbreviation for estimation of scale and subscale totals was assessed by examination of the item and test information curves, estimated from an ordinal graded response model based on item response theory. 56 Validity of the scale and subscale scores was assessed by their association with the contemporaneous satisfaction scores from the CSQ. All analyses were undertaken in Stata 15, except for the exploratory factor analysis undertaken in Mplus (Muthén & Muthén, Los Angeles, CA, USA).

For examination of dimensional properties, participant response profiles with complete data or with up to six missing items were included in the analyses, the latter under an assumption of missing at random. This included missing responses for items where the response was ‘not applicable’, notably questions about a patient’s caring of her baby when treatment had required their separation.

Test–retest subsample

Kappa agreement statistics for each item obtained from the test–retest subsample are shown in Figure 7. All items except item 24, relating to the cleanliness of the baby-changing facilities (p = 0.105), gave significant agreement. Although all significant, the item relating to worry about harming baby also gave low agreement, and items 20, 23 and 25 gave modest agreement. All other items showed good (> 0.6) agreement. The overall item total over the 27 items gave a test–retest ICC of 0.784 (95% CI 0.643 to 0.924).

Factor analysis

In the light of the non-independence of repeated observations from the women reporting on multiple services, the exploratory factor analysis and item response theory models analysed the 267 first-completion responses (with non-applicable items treated as missing and assumed missing at random). We report the exploratory factor analysis using the goemin non-orthogonal rotation. Eigenvalues and goodness-of-it measures are shown in Table 38 and standardised factor loadings in Table 39.

Items 12 and 13 (social services involvement reassuring and family/friend contact easy, respectively) and possibly item 22 (worried about harming baby) were not strongly related to any factor. Although the eigenvalue criterion could be used to justify up to six factors, the step-down in eigenvalues, the comparative fit index and the simple division and spread of items loading on factors 1 and 2 all suggested the two-factor solution as parsimonious. Although closer-fitting models could be achieved with more factors, the adequate fit of the two-factor model, together with the broad consistency of the standardised factor loadings shown in Table 39 with those found for the non-perinatal population from whom this version has been adapted, suggested it for further investigation in addition to the single-factor/questionnaire total score. Face validity interpretation of the loading items suggest that factor 1 was concerned with women’s views about their own treatment, whereas factor 2 was concerned with women’s satisfaction in relation to baby care.

All items as a single dimension: informative items

Consistent with the factor loadings, the item information curves showed items 6, 7, 9, 10 and 11 as dominant, and items 2, 3 and 8 also exceed 1 (Figure 45).

FIGURE 45.

Item information curves for a unidimensional 27-item-graded response model.

Forming a scale from just these eight items with higher information (i.e. items 2, 3 and 6–11) gave factor scores that correlated (0.96) with those from the full 27 items. However, the test information curves for all 27 items and for the reduced set of eight items nonetheless suggested that the full item set is valuable in achieving additional precision. The relative size of the standard error from the short and long forms is shown in Figure 46, with the standard error of the estimated factor being 20% smaller throughout much of its useful range (–2 to +2) for the full as compared with the short version.

FIGURE 46.

Short form (eight items) and long form (27 items) relative standard errors.

Two-dimensional analysis: informative items

When items are separated by factor, the item information curves for the 11 items loading on factor 1 appear more consistently informative, with only one item (item 4) having an information peak of < 1. Of the 10 items contributing to factor 2, there are three less informative items (items 18, 20 and 23) (Figure 47).

FIGURE 47.

Item information curves. The two factors are positively correlated 0.49 (p < 0.0001). (a) Factor 1; and (b) factor 2.

Construct validity

The distribution of factor scores and construct validity were assessed in Figure 9.

Geographical methods

For each study participant, the driving distance from their home to the nearest MBU (i.e. all MBUs open in England and Wales at the time of recruitment) was determined. This calculation was conducted in two steps: (1) the as-the-crow-flies distance, using the haversine formula (see below) and (2) the driving distance (as per as-the-crow-flies distance).

This involved the following:

• The geolocation (latitude, longitude) was determined for each participant’s residence and each MBU using UK Postcode Geo data for the outward codes. 96

• The ‘as-the-crow-flies’ distance in miles for each participant and each MBU was determined using Haversine formula, which is used to generate the distance between two points on a sphere based on their latitude and longitude):(1)Distance=ACOS(COS(RADIANS(90 – Lat1))×COS(RADIANS(90–Lat2))+SIN(RADIANS(90 – Lat1))×SIN(RADIANS(90–Lat2))×COS(RADIANS(Long1–Long2)))×6371.

• For each participant, the shortest distance (i.e. ‘as the crow flies’) and corresponding MBU was subsequently identified by determining the minimum distance across MBUs per subject.

The driving distance was manually calculated for the MBU closest to each participant’s home residence. 96

Sensitivity analyses

Sensitivity analyses were performed on alternative cohort definitions. Women who used both MBU and acute wards were redefined using the following definitions:

• majority number of days

• first service accessed.

Forty-five women were admitted to both a MBU and a ward during their index admission and were categorised as belonging to the MBU cohort (i.e. according to their highest level of care), of whom 20 used only MBU and ward services and 25 used all three services. CRT will remain under the definition of lowest level of care, and so this cohort includes only those women who used only CRT services.

Twelve of the 45 women who were admitted to both a MBU and a ward during the index admission spent more time in a ward. Four of these women were admitted to an ICU during their index admission period. Thirty-three women who were admitted to both a MBU and a ward during the index admission accessed a ward first. Seven women changed cohort service in both alternative definitions. These seven women spent more time in a ward and accessed the ward first, and two of these women were admitted to an ICU. One woman who attended a ward first but spent the majority of days in a MBU was admitted to an ICU.

Other

Aims

Economic evaluation aimed to assess the cost-effectiveness of MBUs compared with generic acute wards and CRTs for the treatment of women with severe mental illness following birth within a quasi-experimental cohort study. The research questions were as follows:

• Are MBUs cost-effective in the short term (from index admission to 1 month post discharge) in the treatment of women with severe mental illness following birth compared with generic acute wards and CRTs?

• Are MBUs associated with a reduction in (1) readmission rates, (2) use of community mental health services and (3) costs in the year following discharge compared with generic acute wards and CRTs?

The planned comparisons were two two-way comparisons (MBUs vs. generic acute wards and MBUs vs. CRT services) plus a three-way comparison. However, owing to small sample sizes, comparison of MBUs with both other groups combined were conducted in line with clinical analyses.

Methods