A diagnosis of schizophrenia indicates that a person experiences symptoms such as persistent delusions and hallucinations, the feeling that their actions or thoughts are controlled by others, and negative symptoms such as an inability to experience pleasure from activities (World Health Organization 2025). Often, long-term treatment is needed, given that relapse is a common feature of the condition (Jørgensen Reference Jørgensen, Bøg and Kabra2021).
High expressed emotion (EE) in family members has long been shown to be a risk factor for relapse in schizophrenia (Box 1). Although we can be confident about the relevance of high EE, the question of how we best address it in practice remains. National Institute for Health and Care Excellence (NICE) guidance states that families of those with schizophrenia and psychosis should be offered family intervention on the basis that it can improve relapse rates and coping skills (NICE 2015). However, evidence suggests that there is relatively low application of these guidelines clinically, with a 2015 review finding that rates of implementation for family interventions varied from 0 to 53% in practice (Ince Reference Ince, Haddock and Tai2016).
What is ‘expressed emotion’ and why does it matter?
Expressed emotion (EE) can be thought of as the ‘emotional climate’ of the relationships between someone with schizophrenia and their family (Humbeeck Reference Humbeeck, Audenhove and Hert2002). Brown & Rutter first measured this concept using the semi-structured Camberwell Family Interview (CFI), which involves interviewing those close to the patient, notably without the patient present (Brown Reference Brown and Rutter1966). Five domains are used to measure EE: criticism, hostility, warmth, positive comments and emotional over-involvement (Dodd Reference Dodd, Schoonover, Lebow, Chambers and Breunlin2017).
There has been consistent evidence that high EE in the families of individuals with diagnoses of schizophrenia is a risk factor for illness relapse (Brown Reference Brown, Monck and Carstairs1962, Reference Brown, Birley and Wing1972). Interestingly a similar phenomenon can be noted in mood and eating disorders (Butzlaff Reference Butzlaff and Hooley1998), and EE may also be relevant in dementia (Vitaliano Reference Vitaliano, Young and Russo1993).
Warmth has been shown to be protective against relapse, whereas the number of critical remarks appears to predict relapse (Ma Reference Ma, Chan and Chung2021). Since the CFI was first introduced, meta-analyses have continued to demonstrate that high EE is a robust predictor of relapse (Butzlaff Reference Butzlaff and Hooley1998; Ma Reference Ma, Chan and Chung2021).
The review in summary
This Cochrane Review (Chien Reference Chien, Ma, Bressington and Mou2024) updates a previous review (Pharoah Reference Pharoah, Mari and Rathbone2010), which suggested that family interventions could reduce relapse rates and in-patient admissions, but warned that effect sizes may have been overestimated owing to issues with methodological quality.
In the updated review, a total of 26 randomised controlled trials (RCTs) and 1985 people with schizophrenia or schizophrenia-like disorders were included, alongside 2056 family members. The review authors assessed the efficacy of a broad range of family-based interventions, varying in duration from 5 weeks to 2 years, compared with standard care. They suggested that family-based interventions resulted in both a reduction in illness relapse and caregiver burden and a shift from high to low EE in family members. Reduction in caregiver burden was the most robust finding, with other evidence rated as low-certainty, and they commented on a high risk of bias in many studies.
How was the review conducted?
The authors of the review synthesised findings from randomised controlled trials (RCTs), with the aim of assessing both beneficial and negative effects of family-based interventions for those with schizophrenia or related disorders. It is not entirely clear what proportion of included patients had a diagnosis of schizophrenia specifically.
Inclusion criteria were: adults aged 19 and over, and interventions defined as ‘family-based’ that lasted for at least five sessions. Intervention approaches were broad and included: psychoeducation, behavioural management and cognitive–behavioural therapy.
Comparisons were made with standard care, which was defined as the care that would normally have occurred in the trial location. Of note, there were no restrictions as to the clinical setting, but most care took place in psychiatric health centres or out-patient settings. A waiting-list control arm was also considered standard care.
Four primary outcomes were used: risk of relapse, caregiver burden, change in EE from high to low, and death. Multiple secondary outcomes were used, including quality of life, economic costs and adherence to medication, although these will not be discussed in detail in this article. Outcomes were assessed across three time points following the intervention, although most studies focused on follow-up at 1 month or less post-intervention.
A key element of this review is that only RCTs including both a patient- and a family member-related outcome were included.
The review authors described using standard Cochrane search methods and they included articles published up to April 2023. It is not clear whether grey literature databases were included in their search, although two clinical trial registers were searched, which can be useful sources of unpublished and ongoing trials (Box 2).
Clinical trial registers
Ultimately only around 50% of clinical trials that are conducted are published, and trials with positive results are over twice as likely to be published than those with negative or null results (Showell Reference Showell, Cole and Clarke2024). Clinical trial registers are therefore valuable places to search for trials that are unfinished or unpublished.
Examples of such registers include the US government’s website ClinicalTrials.gov and the World Health Organization’s International Clinical Trials Registry Platform (ICTRP) (https://trialsearch.who.int), both of which can be freely used online.
The authors used the GRADE tool to systematically assess for bias, and thus the certainty of evidence (Box 3).
The GRADE tool
GRADE stands for Grading of Recommendations Assessment, Development and Evaluation, and it offers a systematic framework to assess the quality, and therefore certainty, of evidence. This is the approach adopted by Cochrane for its reviews and it consists of five areas: risk of bias, inconsistency, indirectness, imprecision and publication bias.
For one particular outcome measure the GRADE approach allows evidence to be rated as either high, moderate, low or very low. When assessing studies across each of the domains, a ‘serious’ limitation leads to ‘downgrading’ of the certainty by one level, and a ‘very serious’ limitation by two.
GRADE offer a software called GRADEpro GDT that allows a useful visual summary of the evidence quality for all studies that are part of a review.
(Schünemann et al., Reference Schünemann, Higgins, Vist, Higgins, Thomas and Chandler2023)
When analysing results the authors pooled data as part of a meta-analysis; they used one of three outcome measures – mean difference (m.d.), standardised mean difference (s.m.d.) or risk ratio (RR) – and reported 95% confidence intervals (CIs). Where scales were deemed to be very similar they calculated effect size and converted this back into the units of one of the specific instruments used, thereby making it possible to pool results.
The authors prioritised converting outcomes into dichotomous groups, of ‘clinically improved’ or ‘not clinically improved’, where the data allowed this. The cut-off points used to define these groups depended on the outcome measure used. For widely used scales such as the Positive and Negative Syndrome Scale (PANSS) a 50% reduction was assumed significant. If this was not applicable, the authors deferred to the cut-off proposed by the original study authors.
A look at the results
Of the 26 RCTs included, 13 assessed outcomes at more than 6 months post-intervention. A range of countries were represented in the results, with 9 of the 26 included studies conducted in China, 3 in Italy and 2 in India, Iran, the UK and the USA.
Pooled results from four studies (229 participants) that examined relapse at short-term follow-up (≤1 month post-intervention) revealed a risk ratio of 0.66 (95% CI 0.49–0.89; I 2 = 0% – see Box 4) for family interventions compared with standard care. Although this indicated a statistically significant result, evidence was graded as low certainty, owing to the low number of participants and high risk of performance, detection and attrition bias (Box 5).
What does I 2 tell us?
I 2 is a value often used, and potentially misinterpreted, in meta-analyses.
It is helpful to understand that there is a difference between the observed effect in studies (our results) and the true effect, which would be the result we would see if the accuracy of our each of our measurements was 100% (which is not possible in practice).
The value of I 2 ranges from 0 to 100% and helps to quantify what proportion of the observed effects in trials is due to variation in true effects, and what proportion of variation results from sampling error. Sampling error is the inherent error that results from taking a sample, given that we cannot accurately know every value of the population being sampled.
An I 2 of 30% would mean that 30% of the variance between studies in a meta-analysis would still be present even if we were able to see the true effects. A value of 0% would indicate that all of the variance would no longer be present if we knew the true effects, meaning that any differences result from sampling error. It follows that a value of 100% tells us that all of the variance would remain even if we knew the true effects.
What I 2 is not able to signify is how much the effects vary between studies, and it therefore cannot be considered as an absolute measure of heterogeneity (the variability in effects between studies), despite often being used as such.
I 2 tells us, when looking at the results of a meta-analysis, what proportion of the dispersion would remain if sampling error could be removed, and it should be used together with the observed effect sizes, to give us a sense of the true effects.
(Borenstein Reference Borenstein, Higgins and Hedges2017)
Performance, detection and attrition bias
Bias is a systematic error that results in difference between study results and the true values, not caused by chance (Krishna Reference Krishna, Maithreyi and Surapaneni2010). Types of bias are categorised by which part of the study process or design has resulted in the error.
This article specifically notes three types of bias:
-
performance bias can refer to either unequal treatment or a change in participant response between study groups, based on the fact that patients and those providing the treatment are not masked to the treatment being given (Phillips Reference Phillips, Kaiser and Thabane2021);
-
detection bias occurs when either participants, healthcare providers or investigators know which treatment was received and this influences how outcomes are measured; it is also referred to as observer, ascertainment or assessment bias (Mansournia Reference Mansournia, Higgins and Sterne2017);
-
attrition bias occurs when those who are lost from a study – also known as ‘drop-outs’ – have systematic differences from those who continued, which can influence the observed effect (Nunan Reference Nunan, Aronson and Bankhead2018).
Meta-analysis of the five studies (309 participants) exploring risk of relapse at more than 6 months post-intervention produced a risk ratio of 0.91 (95% CI 0.76–1.10; I 2 = 25%) for family interventions compared with standard care. This result was not statistically significant and was also graded as low-certainty evidence.
Pooled results from eight studies (563 participants) showed a reduction in caregiver burden at short-term follow-up, with a mean difference of −5.84 on the Family Burden Interview Schedule (FBIS) (95% CI −6.77 to −4.92; I 2 = 45%). Here the authors rated the evidence as moderate for certainty, which makes this the most certain primary outcome result. There was some evidence that the reduction in caregiver burden extended to medium- and long-term follow-up, but this was rated as being of very low certainty.
The FBIS comprises 24 questions, each rated on a Likert scale ranging from 0 (no burden) to 2 (serious burden). It could be argued that a change of 5.84 is therefore significant, given that it equates to a change from serious burden to no burden over three questions. However, a relatively recent study suggested a cut-off score of 23 to identify to family members who may require an intervention (Yu Reference Yu, Liu and Zhou2019). Four of the eight studies included in Chien et al’s meta-analysis had mean FBIS scores >23 in both the intervention and the control group post-intervention. This might indicate that although caregiver burden has decreased, caregivers are still experiencing stress levels indicative of the need for further intervention.
Eight studies examined post-intervention change in EE in family members. Two of these (72 participants) were combined for analysis, although the time point for outcome measures actually varied between 1 and 10 months after the intervention had finished. Results suggested that family-based interventions may be more likely than standard care to result in a shift in EE in family members from high to low (RR = 3.90, 95% 1.11–13.71; I 2 = 0%), with the evidence rated as low certainty.
In terms of adverse outcomes, six studies (304 participants) reported death rates – which in five of the six was defined as death by suicide – as being no different between the two arms (RR = 0.48, 95% CI 0.18–1.32), with the evidence graded as low certainty.
All or nothing
There are several important considerations when drawing conclusions from this review. It is interesting that the authors chose to use only RCTs that included both patient- and family-member outcomes. They justify this choice by citing a review that showed family psychoeducation alone was the most efficacious method in preventing relapse (McFarlane Reference McFarlane2016).
Although demonstrating benefit for family members is important, we might wonder how well this correlates with patient outcomes. Furthermore, the decision to exclude studies that examined either patient or family member outcomes alone limits the ability of this review to make full use of available literature. This is especially true given that they state that most available reviews, and therefore presumably most of the existing evidence, primarily focus on patient outcomes.
The matter of heterogeneity
In general, the combined study sample used demonstrates a high degree of heterogeneity. Although this is a strength with regard to the wider applicability and external validity of the results, it may limit their specificity.
The use of standard care as the comparator introduces considerable variation between the controls used in these studies.
Waiting-list controls, used in three of the included studies (Posner Reference Posner, Wilson and Kral1992; Buchkremer Reference Buchkremer, Mönking and Holle1995; Magliano Reference Magliano, Fiorillo and Malangone2006), were considered as standard care. This decision allowed more studies to be included but their results may be less generalisable to situations where, for example, patients are already receiving care in the community or on in-patient wards.
Moreover, if additional forms of treatment were received by both the intervention and the control group, these results were still included, but we cannot necessarily assume the additional treatments had an equal effect on both arms of the studies.
The use of the umbrella term ‘family-based interventions’ as an inclusion criterion introduces considerable variety in terms of type of intervention, number of sessions (from 6 to 72), duration of intervention (5 weeks to 2 years) and role of the professionals delivering the intervention. The most common intervention, used in 13 of the RCTs, was family psychoeducation.
The large variation in intervention type makes it more difficult to conclude which specific family intervention is likely to be helpful for patients. However, it might be clinically useful to know that our conclusions from this review may well be applicable to a range of strategies for our patients as long as they involve their family. Perhaps, the intervention being family based is of more importance that the specific type of intervention delivered.
Expressed emotion in practice
As detailed in Box 1, EE is an interesting concept, and it would be useful to consider, if it were possible, how EE may vary across cultures and whether it can be reliably measured in the same way in each location. Research involving EE may assume that it is a viable transcultural concept; however, it seems inevitable that family systems will vary broadly across cultures.
It is important to note that conclusions of this review cannot be applied to patients who are socially isolated and who lack a family system. This is highly relevant given that social fragmentation may be associated with higher rates of psychosis and that there appears to be a high prevalence of loneliness in people with schizophrenia (Compton Reference Compton, Druss, Walker and Ku2022; Suman Reference Suman, Nehra and Sahoo2022).
In addition, the authors’ decision to convert results to dichotomous outcomes where possible introduces the potential of representing the data too simplistically. As regards EE, the results – specifically, the combined time-point post-interventions results – may not indicate what proportion of family members moved from low to high EE, nor indicate whether there was no change in EE. Overall, it is disappointing, as the authors stated, that there was not a clearer focus on EE in the available studies.
Conclusion
This review suggests that family-based interventions such as family psychoeducation, probably reduce caregiver burden and might help family members of those with schizophrenia shift from high to low EE. There is some evidence that such interventions may also lead to a reduction in relapse rates for patients.
However, the high degree of heterogeneity within the combined study sample and notable risk of bias across many studies means we must interpret the results with caution. The exclusion of studies without both a family member and a patient outcome also likely resulted in the omission of relevant studies.
Funding
This work received no specific grant from any funding agency, commercial or not-for-profit sectors.
Declaration of interest
None.
eLetters
No eLetters have been published for this article.