Objectives and methods
In this chapter we describe the relationship between depression and chest pain in primary care patients recruited into the study from the GP QOF CHD registers. The aims, objectives and methods have been presented in detail elsewhere1 and will be summarised here.
Objectives
The objectives of the cohort study were to:
determine the prevalence, incidence rate and risk factors of depressive disorders in primary care patients with CHD
explore and describe the course and prognosis of physical and depressive symptoms among primary care patients with CHD over a 3-year period
determine the effect of depression on mortality, symptom severity and pain in primary care patients with CHD.
Service use, service costs and lost employment costs will be described separately.
Methods
The cohort study methods, instruments and analysis plan have been described elsewhere.1 This was a naturalistic cohort study of primary care patients recruited from QOF CHD registers held by GP practices in south London. All participants on these registers were eligible for inclusion in the study as long as they were > 18 years of age and not temporarily registered with their GP. Eligible consenting patients were assessed at baseline and then followed up at 6-monthly intervals for a period of up to 4 years. We report on the 3-year follow-up data here. All patients were eligible for 36 months follow-up at the end of the study. A smaller population was eligible for a 4-year follow-up (and will be analysed in the future). The measures used at baseline and follow-up are shown in , previously published as table 1 in Tylee et al.1 Not all measures have been used in the analyses presented here. For more details on the instruments used, see Appendix 4.
Measures used at baseline and follow-up assessments
The reporting of chest pain by cohort members was recorded at each wave of follow-up. Chest pain is the cardinal symptom of CHD and the key to a clinical diagnosis of angina pectoris. We report here the prevalence rate of this physical symptom at baseline and the outcome in terms of cardiovascular events (e.g. MI) or continued exertional chest pain (i.e. patients reporting exertional pain on most occasions). We have analysed the relative roles of depression, anxiety and quality-of-life scores for these outcomes.
Chest pain assessment
We used the Rose questionnaire113 to assess the severity of the symptomatology (i.e. chest pain) of CHD. Devised in the 1960s for the purpose of detecting angina pectoris in field studies, it has been widely used since. The original validation criterion was agreement with CHD diagnosis in GP medical notes but since then it has been used as a useful predictor of subsequent cardiovascular events in the longer term.157–160 On the other hand, the agreement between a positive response on the Rose and concurrent clinical testing for evidence of CHD can be poor.161,162 It is used in both a long form in field studies and, as the short form of the Rose questionnaire has equivalent predictive power, the short form is often preferred. We describe how both the short form and the long form are rated below. Pain on exertion seems to be the key symptom to predict subsequent problems.163
The short form is outlined in below and the additional questions for the long form are shown in .
Rose items used for scoring the short version: specific wording and order of items given
Rose angina questionnaire items used for scoring the long version: specific wording and order of items given
To assess cardiac outcomes, all cardiology notes, investigations and interventions were collected from the GP notes by a medical doctor on the team (JP) and entered into an encrypted Microsoft Access® 2007 database (Microsoft Corporation, Redmond, WA, USA). The dates of each visit to a cardiologist, each investigation for a cardiac-related problem and each visit to a rapid access chest pain clinic, A&E or hospital admission for a cardiac investigation or intervention (e.g. bypass graft, angioplasty) were recorded. We also recorded any GP visit for cardiac-related problems or regular follow-ups. The data to be inputted and the specific definitions of cardiac investigations were decided before the hard copies were examined, with exception to the ‘rapid access chest pain clinic’ category. The GP notes were re-examined to gather this extra information and all data were entered into the Access database before any analysis of cardiac investigation data occurred. Numerous records per individual were condensed into variables that denoted: (a) the first cardiac investigation to occur within two time points would be recorded per person; and (b) the most severe cardiac investigation to occur within two time points would be recorded per person.
Thus, a cardiac outcome variable with seven hierarchical categories was created using a combination of the short form of the Rose angina questionnaire, the GP medical notes and publicly available mortality records:
no chest pain: score of 0 on the short Rose angina questionnaire
chest pain: positive response to the Rose angina questionnaire item
exertional chest pain: positive classification of grade 1 or grade 2 angina as defined by the short form Rose angina questionnaire
rapid access: any visit to the rapid access chest pain clinic, or any visit to A&E or emergency hospital admittance where cardiac-related chest pain was the diagnosis
bypass graft or angioplasty: intervention reported on GP medical notes or cardiologist/cardiology service notes
MI: occurring at any point during the follow-up period
death: any cardiovascular-related cause of death included (i.e. cardiovascular and cerebrovascular).
For the purpose of the longitudinal analysis, the most severe outcome was assigned to the time point at the end of the period where the outcome occurred. MIs and deaths had to be combined for some analyses, when there were insufficient numbers to fit full models. When linking cardiac investigation and mortality data with interview data at specific time points, deaths and cardiac investigations were coded into time points at the end of the 6-month window between interviews. If a participant had a recorded cardiac investigation within the 6 months before baseline, this was recorded in the baseline time point. All cardiac investigation data as described, with dates of the investigations, were merged into the cohort data set. The aim of this time assignment was to produce data where mental health status could be clearly identified in its temporal relation to cardiac outcomes (the actual date of death was, however, used for a more detailed and specific study of mortality).
Mortality data was gathered from a number of sources, including case notes and publicly available records. The deaths of the cohort participants were tracked in the first instance via the practice manager at the different GP surgeries, where we asked for date and cause of death. In those cases in which the participant had moved away from the surgery, we contacted the primary care support services within the specific boroughs. Further to this, we purchased death certificates through the health authorities relating to the specific boroughs. Where consent had been withdrawn we recorded the fact of death if known but not the reason. The censoring date was the date of death, the final interview date or the date when the interview would have taken place for those who were lost to follow-up. For the latter group, deaths were monitored using public sources up to their censoring date. If the reason for loss to follow-up was withdrawal of consent, the reasons for death were not sought unless available as part of the public record. An intermediate censoring date was also created, based on the 36-month interview date. Only deaths that occurred before that date were used when analysing the 36-month data.
A sample size of at least 800 participants was calculated to be adequate to allow modelling of associations between physical and mental illness.1
Analyses
Descriptive statistics were used to summarise the cohort at baseline and logistic regression was used to estimate associations between outcome and predictor variables.164
The incidence of depression was described from the raw data and separately by depression at baseline. The incidence rates are minimum rates, as missing time points (non-response) were coded as non-depression incident. Mechanisms of missingness were explored and the analyses were adjusted for variables associated with missing time points (age, ethnicity and cancer). The analysis was conducted under the assumption that data were missing at random.
Data were set up for a discrete-time survival analysis as per Singer and Willett.165 Discrete-time survival analysis was chosen as the appropriate model to estimate the odds of an event and the time to an event occurring. Discrete-time (as opposed to the more common continuous time) survival analysis was chosen to reflect the sampling design of 6-monthly time intervals. For more information on discrete-time survival analyses please refer to Singer and Willett.165
The baseline hazard was first modelled as unstructured, thereby allowing the odds of depression incidence to vary freely among time points; however, we chose to adopt a growth curve modelling framework for the odds of becoming depressed over time. This was estimated using Mplus version 7 (Muthén & Muthén, Los Angeles, CA, USA) and in a latent variable framework using robust maximum likelihood. For more information please see Muthén and Masyn.166 This approach allows for both time-invariant and time-variant predictors to affect both the intercept and the change in odds with time (linear and quadratic terms). The assumption of proportional odds was tested by including time-varying effects of predictors and comparing nested models using the Satorra-Bentler scaled chi-squared difference test.167 Model development started with an unconstrained ‘full’ model, with a number of variables included on a theoretical basis and inspired by our baseline paper (see List of variables tested). Parameters were tested using chi-squared difference testing and assessment of model fit statistics (Akaike information criterion and Bayesian information criterion), and the model presented is our best fitting constrained model.
List of variables tested
Depression at baseline, history of depression, sex, age, employment status, relationship status, alcohol intake, smoking status, EQ-5D items (problems with mobility/pain, self-care, usual activities, pain/discomfort), SPQ items covering a range of social problems, diabetes, Rose angina questionnaire, cardiac events during the cohort (including MI) and cardiac operations (bypass graft or angioplasty).
Standardised mortality ratios for those aged > 45 years were calculated in relation to the 2011 Office for National Statistics figures for heart disease (International Classification of Diseases, 10th edition), by age and sex.168 Time at risk was assigned to 10-year age bands with a ‘lexis’ expansion using the stsplit command in Stata. Incident rates of mortality were analysed separately using Poisson regression modelling which took into account the time at risk. Covariates were selected to be the same as those considered in the primary analysis of cardiac outcome. An analysis was also performed using the outcome of ‘non-cardiovascular deaths’ with the same covariates to explore trends for other causes of death.
Results
Recruitment and retention of the study population
Recruitment
Recruitment into the cohort study is shown in . Sixteen general practices from south London participated in the research programme, and from these practices 803 participants were recruited into the cohort study from the practice-held CHD registers. The study population represented 27% of those eligible to be included in the cohort study.164
Recruitment profile of the UPBEAT-UK cohort study.
Retention of study population
The average follow-up time for 803 participants (561 men and 242 women) from 16 practices was 2.64 years and the median was 2.94 years; 141 (17.9%) were lost to (complete) follow-up or declined to be interviewed at some point during the 36-month follow-up period. A further 22 (2.8%) died of cardiac causes, 22 (2.8%) from vascular disease and a further 28 (3.64%) died from other causes. shows the reasons for leaving the study prematurely.
Reasons for leaving the study prematurely
Baseline characteristics of the study population
The baseline characteristics of the cohort population have been published in detail.164
– show the demographic, social and clinical characteristics of participants. Essentially, this was an older population (average age 71 years), predominantly male (70%), with 87.3% classifying themselves as ‘white’. Black and ethnic minorities made up 13% of participants. As would be expected, the majority were retired (77%). The average length of time since a diagnosis of CHD was made in the GP records was 10 years, and 44% reported ongoing chest pain. Social problems were common, with relationship problems being the most common reported social problem (38%). Levels of disability were also high in this population, with 53% reporting problems with pain or discomfort and 49% with mobility problems.
Physical health status at baseline (N = 803)
Social problems and disability at baseline
Sociodemographic characteristics of the cohort at baseline
shows that our cohort study population contains approximately equal proportions of those with a completed MI and a current GP diagnosis of ischaemic heart disease with no history of MI. Similarly, equal proportions have and have not received some kind of surgical intervention.
Cardiac status of the cohort at baseline
Cardiac status at baseline is shown in .
Baseline classification of the cohort using short and long versions of the Rose angina questionnaire
Using the short Rose angina questionnaire, just over one-quarter of the cohort report exertional pain, which could indicate angina. Using the more stringent full Rose angina questionnaire criteria, 13.9% are likely to have a clinical diagnosis of angina (grade 1 or 2).
Discussion
A total of 44% of the sample stated that they currently experience chest pain. If this pain reflected just current angina pectoris, this figure would be very high, given that angina typically accounts for 4% of GP consultations and, in community samples of older people in Scotland, rates of angina pectoris using the Rose angina questionnaire are reported as 2.5% for men and 2.7% for women.169 A meta-analysis of prevalence rates of angina in community samples in 31 countries showed a weighted mean of 5.7% for men and 6.7% for women – the highest rates from any of the countries were 14.4% for men and 15.7% for women.170 Rates of angina following infarction are reported at 19%171 and there should not be any in the aftermath of successful revascularisation.
In fact, chest pain is commonly reported as a symptom in representative epidemiological studies, along with other types of pain. In one study, 12% of participants reported chest pain, compared with 41% reporting backache and 26% reporting headache.172 The most common explanatory physical diagnosis for chest pain is oesophageal reflux.173,174 There has been one epidemiological study in Australia of those with non-cardiac chest pain.175 The authors noted that (1) the most common accompanying symptom was heartburn that suggests oesophageal disorder; (2) many did not seek help for it; and (3) quality of life was reduced for those reporting it with.175,176 Non-specific chest pain is associated with an increased mortality rate.177
It has long been recognised that many patients referred for investigation of chest pain have concurrent anxiety and depression syndromes.173 Chest pain is indeed a recognised symptom for the diagnosis of a panic attack.178 Case-level anxiety or depression has been stated to be a contraindication to expensive cardiology investigation, as few patients are found to have coronary disease.179
Most of the work on the association of chest pain with psychiatric disorders has been carried out in cardiology patients. Our study population is different in that we have recruited our patients from a primary care-based register of CHD patients, some of whom are under the care of cardiologists. To our knowledge, there are no previously published studies on this topic based on GP register patients.
We are mainly reporting on the group identified as ‘exertional pain’ from the short Rose angina questionnaire, as there is adequate indication from published studies that this symptom is the most prognostically important for future events as described above.163 The distribution of the data when classified by the short Rose angina questionnaire data also allows greater power in our analyses, and there are fewer missing responses than when using the long version. Multivariate analyses have been conducted to investigate baseline associations of short Rose classification.
Exertional pain versus non-exertional pain (risk factors)
The next analysis () simultaneously compares those with chest pain with those with non-specific pain, and with those with exertional chest pain. Each paired analysis takes into account the third group.
Multinomial logistic regression of the Rose classification (short Rose angina questionnaire classification) on sociodemographic and clinical predictors (n = 626). Reference group 1: no chest pain
The variables entered were those baseline variables describing cohort members set out in the paper on baseline analyses.164 They concern sociodemographic status, cardiac variables, self-reported cardiac risk factors, physical disease comorbidity and social problems.
Variables were included in the model if they provided information to an alpha threshold of p < 0.2. Variables in the model were age, sex, ethnicity, Index of Multiple Deprivation (IMD) score, smoking status, comorbidity of COPD, depression (positive on the HADS-D scale), anxiety (positive on the HADS-A scale), total number of SPQ social problems and ‘How much bodily pain’ (i.e. non-specific pain) in the past 4 weeks? (Question 21 on the Short Form Questionnaire 36-items quality-of-life scale.) Only those variables contributing with a significance threshold of p < 0.05 are reported above.
For continuous variables (age, IMD score and total SPQ social problems): with every one unit increase in IMD deprivation, the RR of being in the chest pain group compared with the no chest pain group increases by 3% (RR 1.03; p > 1).
For the categorical variables (sex, ethnicity, depression, COPD and primary GP coded CHD diagnosis): for those classified by the HADS as having depression compared with those without HADS depression, the RR of being in the chest pain group was 2.87, and being in the exertional pain group was 2.92, both compared with the no chest pain group.
The analysis in was repeated using the full (long) Rose angina questionnaire classification variables: no pain (n = 447), pain in chest (n = 171), grade 1 and grade 2 angina (n = 111). The results, which are not reported here, were very similar.
Discussion
Notable in the results of this analysis is the absence of any of the cardiac variables or cardiac risk factors that we recorded as associations of chest pain (whether or not the chest pain was exertional) – in particular whether or not there is a history of a revascularisation procedure. The association with GP diagnosis of ischemic heart disease and exertional pain is compatible with the validation criteria of the Rose angina questionnaire. The association with COPD has clinical face validity – sufferers may experience discomfort when they exert themselves.
Depression, anxiety and social factors are the evident associations. The most potent association is the number of social problems rather than any individual social problem. For every additional social problem reported, the risk of exertional pain increases by 36%. Depression and anxiety are both associated with complaints of non-specific and exertional pain – the RR being higher for the exertional pain. Experience of panic and bodily pain remain independently associated with chest pain. Increasing report of bodily pain is associated with increased risk of being in both of the two chest pain categories.
The data from this cohort strongly suggest a psychosocial context for the complaint of chest pain whether or not it is anginal in type. As this is a cross-sectional analysis, direction of causality cannot be assumed. It is possible that having chest pain limits life to the extent that multiple social problems result, for example less social contact.
Chest pain and quality of life
Baseline analysis
The literature suggests that quality of life is impaired many years after MI for those with persisting symptoms,180 the effect being more marked in younger than older people. The same is reported for those with non-cardiac chest pain.175 The EQ-5D is used in these analyses as a measure of health-related quality of life, a score can be obtained by summing the response to the five questions asked or from the accompanying visual analogue, in which the respondent makes a judgement of their quality of life now. We have chosen the latter, as one of the five questions asks about anxiety and depression and we wish to examine quality of life separately from mental state so far as is possible.
Examination of associations of baseline European Quality of Life-5 Dimensions visual analogue scale
Quality of life was thus recorded using the EQ-5D questionnaire visual analogue scale (EQ-5D-VAS). To examine the factors associated with quality of life, we used a multiple linear regression model and selected variables from a large pool of sociodemographic and clinical factors. Variables were selected using a stepwise selection procedure that retained all predictors if evidence for their association with the visual analogue scale was lower than p = 0.2.
The outcome, as well as all predictors used in this analysis, was collected by self-report questionnaire and at the same time point (baseline), and so this analysis can be seen as an examination into the factors associated with quality of life and, on an exploratory basis, an examination into the potential predictors of quality of life in our cohort. Owing to the positively skewed distribution of the EQ-5D-VAS, non-parametric bootstrapping using 200 replications was used in the analyses.
The results of this analysis are shown in –.
Description of EQ-5D-VAS at baseline
Multiple linear regression examining sociodemographic and clinical variables as predictors of EQ-5D-VAS
The data in show high correlation between anxiety and depression scores, and between EQ-5D scores and both. In the latter, high depression anxiety scores (bad) correlate inversely with high EQ-5D sores (good). shows analyses that explore, at baseline, what factors influence the EQ-5D-VAS score in this cohort.
Relationship between anxiety, depression and EQ-5D
Summary of results
The following variables associated with baseline quality of life:
Ethnicity: white ethnic group had a 4.74 unit higher score (p = 0.024). This is a small standardised effect size (0.1).
Current smokers: current smokers scored, on average, lower by 5.19 units (p = 0.022).
Exertional chest pain: those with exertional chest pain had a lower quality-of-life score by 3.90 units (p = 0.020), a small effect size (0.09).
Bodily pain: those reporting bodily pain (an item taken from the SF-12) had, on average, a 3.85-point lower score on the EQ-5D-VAS (p < 0.001). This was a moderate effect size (0.28).
Treatment for heart problem: people who had received a treatment for their heart problem in the past had a lower quality-of-life score by –2.75 units (p = 0.021). This was a small effect size (0.07).
Cancer: cancer was associated with lower quality of life by 5.63 points (p = 0.013). This was a small effect size (0.1).
Depression: HADS-D (scoring ≥ 8 on the HADS-D scale) was associated with a 9.71-unit lower quality of life (p < 0.001). This was a small effect (0.17).
Anxiety: HADS-A (score of ≥ 8) was also associated with a lower score on the EQ-5D-VAS scale by 4.8 units (p = 0.003). This was a small effect (0.11).
Depression, anxiety and chest pain all impair quality of life at baseline.
Prevalence, incidence and course of depression
Prevalence of depression
The prevalence of depressive and anxiety disorders and the risk factors associated with a depressive disorder at baseline have been presented in detail elsewhere31 and will be summarised here. The prevalence of depressive and anxiety disorders in this population at baseline is shown in .
Mental health status at baseline (N = 803)
The proportion of the population who had depression recorded as an active problem in GP records at recruitment into the cohort was 7%, whereas 4.6% of the population had an International Classification of Diseases, 10th edition, diagnosis of a depressive disorder (mild, moderate or severe) as defined by the Clinical Interview Schedule – Revised (CIS-R). According to the HADS-D scale, 12.8% of the population suffered from a depressive disorder at baseline (i.e. scored > 7 on the depression subscale).
The statistically significant associations between predictor variables and CIS-R-defined depression at baseline fell into three domains: living alone, disability and pain, and age (problems with living alone: adjusted OR for a CIS-R diagnosis of mild, moderate or severe depressive episode 5.49, 95% CI 2.11 to 13.40; p < 0.001; experiencing chest pain: adjusted OR for a CIS-R diagnosis of mild, moderate or severe depressive episode 3.27, 95% CI 1.58 to 6.76; p = 0.001; being disabled by pain or discomfort: adjusted OR for a CIS-R diagnosis of mild, moderate or severe depressive episode 3.39, 95% CI 1.42 to 8.10; p < 0.006; having problems carrying out usual activities: adjusted OR for a CIS-R diagnosis of mild, moderate or severe depressive episode 3.71, 95% CI 1.93 to 7.14; p < 0.001; and age: adjusted OR per year for a CIS-R diagnosis of mild, moderate or severe depressive episode 0.95, 95% CI 0.92 to 0.98; p < 0.001) (see Walters et al.164). Chest pain was significantly associated with depression independently of other pains and discomfort.
Participants had an increased odds of having a GP-recorded diagnosis of depression if they reported problems with close relationships (adjusted OR for a GP diagnosis of depression 2.51, 95% CI 1.40 to 4.52; p = 0.002), reported having diabetes mellitus (adjusted OR for a GP diagnosis of depression 2.01, 95% CI 1.11 to 3.63; p = 0.02), were disabled by pain and discomfort (adjusted OR for a GP diagnosis of depression 1.95, 95% CI 1.04 to 3.68, p = 0.037), or were female (adjusted OR for a GP diagnosis of depression 1.88, 95% CI 1.04 to 3.37; p = 0.035) (see Walters et al.164).
Incidence of depression and risk factors for incident depression
The incidence risk of developing a depressive disorder (as defined by scoring ≥ 8 on the HADS-D subscale) is shown in . The proportion of new cases of depression was similar at each time point for those who were not depressed at baseline. However, the proportion of incident cases of depression decreased across time points for those who were depressed at baseline. The cohort followed participants for a total 1942.3 years at risk. During this period there were 240 incident cases of depression giving an incident rate ratio of 123.6 per 1000 person-years at risk.
Risk of a new depressive disorder during each 6-month time point
The risk factors associated with the incidence of depression are shown in . The strongest predictor of incident depression was having had a documented MI at some point before the incident depression. Other predictors of incident depression were: suffering with depression at baseline or having a past history of depression, self-reported exertional chest pain, problems with disability and living alone.
Discrete-time survival model for predictors of incident depression (as defined by a HADS-D subscale score of ≥ 8)
[Exertional pain was initially considered as a possible type 3 variable. However, in the final model it was better treated as a time-invariant (baseline only) predictor because of uncertainties in the longitudinal data recording. It was not always possible to establish the time sequence of the incidence of pain and depression and there were a large number of missing values, which greatly reduced the sample size.]
The course of depression in cohort participants is described in .
Patterns of depression from baseline to 36 months
Of the 12.9% of participants depressed at baseline, 42.7% remained depressed throughout the 36-month follow-up period. Only 35% were not depressed at final follow-up. Of those not depressed at baseline, 23.6% developed depression at some point over the follow-up period.
The incidence and effect of depression and anxiety on cardiac outcomes and mortality
The incidence of cardiac outcomes is reported in two ways – first, the incidence of cardiac events over the follow-up period (and the effect of depression and anxiety) and, second, the continuing reporting of chest pain over that same period.
shows the incidence of cardiac events over the 36-month follow-up. There were 44 total deaths that had a cardiovascular cause. Sixteen patients in the cohort had a MI during the follow-up period.
Incidence of cardiac interventions, MI and cardiovascular deaths up to 36-month follow-up
shows the percentage of cardiac outcomes that were related to depression and anxiety. Although our initial analysis was done with depression, we found that for every cardiac outcome anxiety was more strongly related, especially in MI and cardiovascular death.
Percentage of cardiac outcomes in which there was a prevalence of depression or anxiety at previous follow-up
shows the predictors of cardiac outcome with ‘no chest pain’ as comparator.
Full model for predictors of cardiac outcome (relative risk ratio compared with no chest pain)
Discussion
In terms of predictors of cardiac outcome status, chest pain and exertional chest pain were not associated with baseline depression status but were statistically significantly associated with baseline anxiety. There were no associations between attendance at a rapid access cardiac clinic for a cardiac reason and either baseline depression or anxiety.
There was no evidence of an association between baseline depression and having a MI or dying from a cardiac cause, but there was a statistically significant association with baseline anxiety and having a MI or dying from a cardiac cause. Rose angina questionnaire categories for chest pain at baseline were significantly and strongly associated with subsequent chest pain and exertional chest pain, especially the latter, and in the direction that would be expected, that is, both were associated with all cardiac outcome status compared with no chest pain with exertional chest pain having a stronger association with cardiac status (apart from attendance at a rapid access cardiac clinic) than chest pain. Women have an increased risk of chest pain and exertional pain (but no evidence for higher risk of other cardiac outcomes). Age was a significant predictor only of MI/death. The effect of adding covariates to an unadjusted mode, leading to the full model reported above, is shown in .
Effect of adjusting for confounders on association between depression and cardiac outcomes (relative risk ratios compared with no chest pain)
Sensitivity analyses
General practitioner practice was also included in the model as a sensitivity analysis as the practices were situated in areas with different sociodemographic characteristics and also had varying levels of loss to follow-up. The results did not differ significantly from the original model, apart from the exertional chest pain outcome, in which depression and anxiety had similar relative risk ratios (RRRs), both of which were significant (depression: RRR 1.702, 95% CI 1.025 to 2.827; p = 0.040; anxiety: RRR 1.573, 95% CI 1.027 to 2.411; p = 0.037).
The baseline number of comorbid medical illnesses and social problems as measured by the SPQ was added to the model. Depression was now no longer significant for chest pain (RRR 1.177, 95% CI 0.650 to 2.130; p = 0.591), although anxiety remained significant (RRR 1.730, 95% CI 1.132 to 2.643; p = 0.011). Total number of social problems was also related to chest pain (p = 0.001), but not the number of comorbidities. In terms of exertional chest pain, in this model, depression was no longer statistically significant (RRR 1.424, 95% CI 0.560 to 2.373; p = 0.174) although, again, anxiety remained statistically significant (RRR 1.600, 95% CI 1.059 to 2.413; p = 0.026). There was no change in associations for rapid access chest clinic or for cardiac interventions (neither significant), but social problems were significant for chest pain and exertional chest pain (p = 0.001 for both) and for cardiac interventions (p = 0.005), but not for other outcomes.
Two alternative definitions of depression were analysed:
Depression in the 6 months before the index cardiac event. The results of this analysis were similar to the main model, except that both depression and anxiety were statistically significantly associated with exertional chest pain (depression: RRR 1.85, 95% CI 1.291 to 2.663; p = 0.001; anxiety: RRR 1.867, 95% CI 1.349 to 2.585; p < 0.001). For other outcomes, the model could not be fitted owing to small sample sizes.
Cumulative burden of depression using the sum of follow-up periods, occurring before the index cardiac outcome, in which the patient had been depressed. Chest pain and exertional chest pain were both associated with depression and anxiety (depression: RRR 1.217, 95% CI 1.000 to 1.482; p = 0.05; anxiety: RRR 1.239, 95% CI 1.041 to 1.474; p = 0.016). However, other outcomes were similar to the main model except that cumulative anxiety was associated with attendance at a rapid access chest clinic (RRR 1.350, 95% CI 1.059 to 1.722; p = 0.015).
Redefining the cardiac outcome with rapid access chest clinic combined with any other interventions did not show any association with baseline depression or anxiety.
Mortality
There were 72 deaths altogether, of which 22 were from cardiac disease, 22 from vascular disease, 16 from cancer and eight were attributable to other causes; in four cases, the cause of death was not ascertainable. Of the 22 cardiac deaths, 14 were of men (2.5% of 561) and eight of women (3.3% of 242). Incidence rates of cardiac death were 12.7 per 1000 person-years for men and 9.3 per 1000 person-years for women. Standardised mortality ratios compared with the general population were 1.13 (95% CI 0.62 to 1.90) and 1.87 (95% CI 0.81 to 3.70), for men and women, respectively.
shows a Poisson regression model for the rates of cardiovascular deaths. Only risk ratios for increasing age and a higher baseline anxiety were significant at p = 0.05. A sensitivity analysis using the CIS-R definition of depression and anxiety was similar, except that the risk ratio for depression was increased to 1.260 (95% CI 0.274 to 5.796; p = 0.766) and that for anxiety decreased (risk ratio 2.001, 95% CI 0.694 to 5.774, p = 0.199) and was no longer significant.
Baseline associations with cardiovascular mortality
shows a Poisson regression model for the rates of non-cardiovascular deaths. Only RRs for increasing age, male sex and higher baseline total comorbidities were significant at p = 0.05.
Baseline associations with non-cardiovascular mortality
The types and course of chest pain, and predictors of chest pain as an outcome
The effect of depression on continuing chest pain
This next series of analyses use data from the five-wave follow-up. We report on the frequency of cardiac events, the persistence of chest pain reported by respondents and the predictive power of depression and other psychosocial factors.
Chest pain type and frequency of report
As shown in , the amount of chest pain is reported, ranging from 0 to 100%, of the occasions when patients were interviewed using the Rose angina questionnaire. These proportions became the outcome variable to contrast with baseline predictors between those with no chest pain and those with non-exertional pain only.
Frequency of chest pain (both exertional and non-exertional) reported across the follow-up period
In contrast to , which takes into account the whole cohort, subsequent analyses report on those who remained under follow-up who did not experience any of the events mentioned above. In addition, only those who had, at most, two missing chest pain responses out of the five possible over the follow-up period were included. Those with more than two missing responses were excluded. (Sensitivity analysis was carried out restricting it to one or fewer missing responses, and no differences were found, thus we selected this analysis, which includes a larger portion of the sample.) These two exclusions reduced the sample for analysis to n = 568. Baseline depression and anxiety scores were examined separately, and then quality of life, as measured by the EQ-5D-VAS, was entered into the analysis.
Predictors of non-exertional pain
Females and those with asthma comorbidity were most at risk of reporting non-exertional chest pain ( and ). Shorter amount of time on the GP CHD register is also predictive. Anxiety and depression, counterintuitively, do not have an effect. The addition of the quality-of-life measure does produce a significant finding, with a lower score being predictive of more pain report.
Effect of anxiety alone on non-exertional chest pain and with inclusion of EQ-5D-VAS in patients with two or fewer missing time points
Effect of depression alone on non-exertional chest pain and with inclusion of EQ-5D-VAS in patients with two or fewer missing time points
Exertional pain
The situation with exertional pain is quite different ( and ). Exertional pain reported at baseline had a massive effect in predicting subsequent reports of chest pain, as did, to a lesser extent, non-exertional pain. Depression and anxiety each had an additional effect on baseline pain. The length of time on the register was predictive, but in the opposite direction to non-exertional pain, meaning the longer time spent on the register, the more likely they were to report exertional pain. Older age was a weak risk factor. Addition of the quality-of-life measure eliminated the effect of depression, and weakened that of anxiety.
Effect of anxiety alone and with inclusion of EQ-5D-VAS, patients with two or fewer missing time points
Effect of depression alone and with inclusion of EQ-5D-VAS in patients with two or fewer missing time points
Economic analysis of the UPBEAT-UK programme
Objectives
To:
assess the long-term costs of care for a sample of patients with CHD and comorbid (a) baseline, and (b) subsequent depression compared with those without comorbid depression
identify predictors of these costs.
Methods
Health-care utilisation and costs
The primary measures for the economic analysis were health-care utilisation and total health-care and societal costs. Medical and informal (i.e. family care) resource utilisation was assessed using a questionnaire, the Client Service Receipt Inventory, developed by members of the Centre for the Economics of Mental and Physical Health.127 The questionnaire is frequently used in mental and physical care evaluations. Medical utilisation included the number of contacts with hospital and community services and (where appropriate) the length of these contacts and the duration of stay for inpatient care. Hospital service utilisation consisted of outpatient, inpatient, A&E and day hospital visits. Community services mainly included contacts with GPs, practice and district nurses, other medical professionals (psychiatrist, other community-based doctor, community mental health nurse, health visitor, other nurse, psychologist, counsellor, occupational therapist, physiotherapist, ‘alternative’ medicine or therapy, other therapist) and care professionals (social worker, housing worker, home help/home care worker, care attendant, community support worker, voluntary worker, day centre/drop-in/social club, any other community-based service). Informal care involved assistance in daily activities by family members, relatives or friends (i.e. personal or child care, help in/around and outside the house) and was measured as the average weekly hours spent with the patient specifically because of their health problems.
Data were collected every 6 months for a period of 3 years between September 2007 and September 2012.
Total costs were calculated from both a health-care payer’s and a wider societal perspective, by multiplying the estimated resource use by the corresponding unit cost. Unit costs were applied to service-use data using the NHS Reference Costs for 2011–12181 and the 2012 Unit Costs of Health and Social Care.129 The unit cost of a home care worker was used as a proxy for costing informal care. Wherever necessary, costs were inflated to 2012 prices using the inflation indices for hospital and community health services.129 All costs were expressed in pounds sterling (GBP).
Indirect costs of productivity loss due to CHD and comorbid depression were not included in the analysis because of incomplete data on sickness absence from work, occupation and status or mode of employment (i.e. full time or part time). However, the average age was 71 years and most participants were retired. Therefore, productivity loss is not likely to be a major consideration.
Medication use was self-reported and was recorded for baseline only (for the past 6 months).
Missing values
Across the study period (six follow-ups), some data (overall approximately 15% of data items) on health-care utilisation were missing, mainly representing the missing number and/or duration of contacts. The approach adopted to handle the missing data was imputation by the mean values of the respective variables at each time point. Owing to the large sample size and the fact that total outcomes are presented at a mean level, it was decided to use the mean instead of median imputation. In three observations where the number and/or duration of contacts were reported but the type of service was missing, the imputed value was assumed to be the most frequently used service across the study period.
Sample characteristics
Differences in sociodemographic variables between CHD patients with/without comorbid baseline depression were analysed using a two-sided t-test and chi-squared test. Depression status was measured by the HADS-D at baseline and at each follow-up point. Patients were classified into depressed and non-depressed, using a cut-off score of ≥ 8 for possible depression.114 At baseline, 13% of patients had depression.
Sociodemographic characteristics of the sample are shown in . Approximately two-thirds were male in both groups and the average age was 66 years and 71 years for the depressed and the non-depressed group, respectively. The majority of patients were married and retired, with only a small proportion (< 5%) cohabiting. In the non-depressed group, 53% received > 11 years of education while a similar proportion in the depressed group received < 10 years of education. Approximately half of either group were ex-smokers and consumed, on average, between 0 and 10 units of alcohol per week. The mean IMD score was 20 and 23, for the non-depressed and depressed group, respectively. The vast majority of the depressed patients reported chest pain, while most of the non-depressed patients did not have chest pain. Both groups had an average number of two additional comorbidities, with the most frequent ones being hypertension and diabetes. Approximately 65% of patients had depression when assessed by the CIS-R instrument. It is noticeable that approximately 80% of the depressed patients also had anxiety when measured by the HADS-A. Most of the differences were statistically significant at a 10% level (see ).
Sample characteristics by depression status (HADS-D) at baseline
Statistical analyses
Statistical analyses were carried out using Stata version 11. Regression analysis was used to estimate mean differences in service use and costs between groups. Baseline and follow-up costs were used as the dependent variables and the group identifier as the independent variable. To estimate whether or not the difference in costs between depressed and non-depressed patients was present at each time point, the HADS-D status was also assessed at consecutive follow-up times ().
Mean depression score (HADS-D) across time.
Sociodemographic and clinical baseline variables were tested for significance using simple ordinary least squares regression models. Each variable was regressed on the cumulative 3-year hospital care costs, community care costs, informal care costs, total health-care costs and total societal costs. Those variables with statistical significance at the 10% level (p < 0.10) in at least one of the above mentioned dependent variables were included in the final multivariate model to identify predictors of costs (). Certain other variables that were of interest, although not statistically significant in the first stage, were also included in the final models. Of the statistically significant variables, those with high numbers of missing values were excluded from the final models. The remaining variables were checked for correlation and those that were highly correlated (r ≥ 0.7) with each other were removed, one at a time, from the model. Two models were constructed using the cumulative health-care and societal costs as the dependent variables. The models included the following baseline covariates: sex, ethnicity, relationship status, age at referral, health literacy status, years in education, family history of depression or cardiac problems, smoking status, average units of alcohol per week, the body mass index score, number of CHD risk factors, depressive disorder assessed by the CIS-R, currently being treated for depression, anxiety status measured by the HADS, the IMD score, housing, working or financial problems, social contact and relationship problems, currently active comorbidities such as diabetes, arthritis, cancer, asthma, hypertension, COPD and chronic kidney disease. The models were also adjusted for informal care costs and total health-care costs at baseline. Approximately 100 observations had missing data on the health literacy variable. Thus, two additional models including all the above covariates except the health literacy variable were also run. To produce a more parsimonious model that aids interpretation, the backward elimination stepwise method was used. This method involved starting with all independent variables and then sequentially removing variables with the highest p-values until only those with p-values < 0.1 were included.
Significant covariates at 10% level (p < 0.1) for inclusion in the final model
It is commonly known that health-care data often violate the assumptions underlying ordinary least squares. Therefore, the models were run to test for the assumptions of normality and homoscedasticity of residuals. Both assumptions were violated and generalised linear models were used. The use of log transformation is common practice for dealing with skewed data, such as costs. However, log-scale results are of little interest in decision-making, while actual values (i.e. GBPs) are meaningful. Generalised linear models overcome the shortcomings of (log) ordinary least squares using the original scale of costs.182 To use generalised linear models, the distribution of cost data and a function that specifies the relationship between the mean and the linear specification of the covariates (the so-called link function) must be specified.182 The appropriate distribution of costs was identified by implementing the modified Park test.183 The gamma distribution, with a log-link function, was applied to the final model. The log-link function means that the value of the exponentiated coefficients is interpreted as proportional changes of the 3-year total health-care and societal costs (–).
Predictors of total health-care costs (cumulative for 3 years) (n = 624)
Predictors of total societal costs (cumulative for 3 years and excluding the health literacy status from the model) (n = 719)
Predictors of total societal costs (cumulative for 3 years) (n = 624)
Predictors of total health-care costs (cumulative for 3 years and excluding the health literacy status from the model) (n = 719)
Results
Health-care utilisation
and show the cumulative resource use by depression status. Depressed patients utilised more health-care resources than non-depressed patients during the 3-year study period (87.6 vs. 59 total health-care contacts; 95% CI 5.4 to 51.8). In particular, 41% of the depressed patients used the A&E department, compared with 33% of the non-depressed patients, but the average number of attendances was 0.3 lower in the depressed group than in the non-depressed group. However, the mean number of attendances was higher for the depressed patients in all periods except the last year. The differences were not statistically significant.
Cumulative resource use by type of service and by group for service users at 36 months
Grouped cumulative resource use for service users at 36 months
Patients with depression used inpatient care more and stayed in hospital significantly longer (7.8 days, 95% CI 0 to 15.6 days) than patients without depression. Approximately 5% of the patients were hospitalised for CHD reasons (1% of those had depressive symptoms). Nearly 80% of patients in both groups used outpatient care, with the depressed group using it more intensively (17.6 vs. 12 visits, 95% 0.7 to 10.9 visits). Patients in both groups made similar use of primary care, which included contacts with GPs, practice and district nurses, with the depressed group having a slightly higher frequency of contact. The mean duration of contacts with GPs and PNs was slightly, but not significantly, higher (1.4 minutes per patient) for the depressed patients than for the non-depressed patients. Patients with depression tended to spend approximately 5 minutes more with their GPs than non-depressed patients. Overall, secondary care, which included hospital care and care provided by other medical and care professionals (e.g. psychiatrists, psychologists, other community-based doctors, mental health nurses, social workers, housing workers and care attendants), was used more by the depressed group, with a significant difference of 27.9 contacts (95% CI 5.8 to 50 contracts) compared with the non-depressed group.
Resource use (by time point)
It was found that informal care was utilised more (54% vs. 30%) and more intensively (19.6 vs. 17.7 average hours per week per patient) by the depressed patients than by the non-depressed ones. Information on resource use by type of service and by group at each time point, separately, is given in –.
Resource use by type of service and by group for service users at baseline
Resource use by type of service and by group for service users at 36 months
Resource use by type of service and by group for service users at 6 months
Resource use by type of service and by group for service users at 12 months (by time point)
Resource use by type of service and by group for service users at 18 months
Resource use by type of service and by group for service users at 24 months
Resource use by type of service and by group for service users at 30 months
Costs
The cumulative costs for the 3-year study period are shown in and . Costs by type of service and by patient group at each time point for service users only (i.e. excluding those with zero costs) and for the whole sample are given in – and –, respectively. In almost all types of service, patients with depression incurred higher costs than patients without depression. The difference was statistically significant for costs of outpatient services and contacts with other medical professionals as well as for total costs. The aggregated 3-year costs for service users were significantly higher for the depressed group than for the non-depressed group, by approximately £4000 (see ). The highest costs were for inpatient care across the study period (). At baseline the non-depressed group incurred higher inpatient care costs per patient than the depressed group (–£1739, 95% CI –£3790 to £312). However, baseline costs were not included in the calculation of the aggregate costs because the estimates referred to the 6-month period prior to the study initiation. In all other follow-ups, there was a positive difference of approximately £3000 in inpatient care costs between patients with and without depression, except in the 36-month follow-up, in which case the difference remained positive but at a lower level (£1534, 95% CI –£1772 to £4840) (see –).
Cumulative costs per patient (£, in 2011–12 prices) by type of service and by group for service users at 36 months
Grouped cumulative costs per patient (£, in 2011–12 prices) by group for service users at 36 months
Costs per patient (£, in 2011–12 prices) by type of service and by group for service users at baseline
Costs per patient (£, in 2011–12 prices) by type of service and by group for service users at 36 months
Costs per patient (£, in 2011–12 prices) by type of service and by group for the whole sample at baseline
Costs per patient (£, in 2011–12 prices) by type of service and by group for the whole sample at 36 months
Hospital services costs per patient (£, in 2011–12 prices) by cost category and by group for service users over time.
Costs per patient (£, in 2011–12 prices) by type of service and by group for service users at 6 months
Costs per patient (£, in 2011–12 prices) by type of service and by group for service users at 12 months
Costs per patient (£, in 2011–12 prices) by type of service and by group for service users at 18 months
Costs per patient (£, in 2011–12 prices) by type of service and by group for service users at 24 months
Costs per patient (£, in 2011–12 prices) by type of service and by group for service users at 30 months
Costs per patient (£, in 2011–12 prices) by type of service and by group for the whole sample at 6 months
Costs per patient (£, in 2011–12 prices) by type of service and by group for the whole sample at 12 months
Costs per patient (£, in 2011–12 prices) by type of service and by group for the whole sample at 18 months
Costs per patient (£, in 2011–12 prices) by type of service and by group for the whole sample at 24 months
Costs per patient (£, in 2011–12 prices) by type of service and by group for the whole sample at 30 months
The vast majority of patients in both groups received outpatient care, and care from GPs and PNs. Outpatient care costs per patient were higher for patients with depression during the study period and significantly higher at 12, 24, 30 and 36 months, at around double the costs of the non-depressed group. However, the number of people who utilised the services was similar for both groups at each time period. Only a few patients used day hospital services at each time point; however, the costs incurred were high for both groups. This can be attributed to the high cost of some services, such as cataract operation. Average community care costs per patient were higher by £475 for patients with depression than for those without depression, and the difference was statistically significant (95% CI £66 to £883). Community costs and total health-care costs are shown in –. None of the differences in informal care costs during the study period was statistically significant. Societal costs are shown in –. Indirect costs included informal care, which was similar between groups, with slightly higher costs incurred by the depressed group (difference £43, 95% CI –£164 to £251), especially during the last three follow-up periods (see and ).
Community services costs per patient (£, in 2011–12 prices) by cost category and by group for service users over time.
Informal care costs per patient (£, in 2011–12 prices) by group for service users over time.
Cumulative informal care costs per patient (£, in 2011–12 prices) by group for service users at 36 months.
Mean total societal costs per patient (£, in 2011–12 prices) across time.
Total societal costs per patient (£, in 2011–12 prices) by group for service users over time.
Total health-care costs per patient (£, in 2011–12 prices) by group for service users over time.
Cumulative health-care costs per patient (£, in 2011–12 prices) by group for service users at 36 months.
Cumulative societal costs per patient (£, in 2011–12 prices) by group for service users at 3 years.
Grouped cumulative costs (£, in 2011–12 prices) by group and by cost category for service users at 36 months.
Predictors of costs
– show the output of the four generalised linear models, having accounted for the impact of depression status on costs by using the depressive disorder assessed by the CIS-R as an input variable. The results reveal that health literacy had a significant negative impact on total costs from the health-care perspective. It would decrease costs by approximately 40%. From both the health-care and societal perspective, ethnicity was significantly associated with total costs. More specifically, being Asian, black or an ethnicity other than white is predicted to reduce costs by approximately 45%. Similarly, higher costs were associated with the prevalence of depressive disorder (measure by the CIS-R), with housing and relationship problems, and with currently suffering from cancer.
When the health literacy variable was excluded from the first two models, health-care and societal costs were shown to be significantly and positively associated with having an additional comorbidity other than cancer, and this included chronic kidney disease. Another new factor that was present in these models was reported chest pain, based on responses to the Rose angina questionnaire for angina. The association this had with costs was positive and the magnitude was similar (+23%) from both perspectives. It is notable that smoking status is not now a significant predictor of total costs. Total health-care costs at baseline are predicted not to have any impact on the cumulative health-care or societal costs in any of the four models.
Discussion
The aim of the economic analysis was to estimate the cumulative health-care utilisation and costs of patients with cardiovascular heart disease with and without baseline and subsequent depressive symptoms. The cumulative cost per patient incurred by depressed patients was approximately double the cost incurred by non-depressed patients during the 3-year study period. Overall, inpatient care services dominated costs at baseline and follow-up, while the majority of patients in both groups received outpatient care and care from GPs and PNs.
Across the study period, total costs were increasing, except in the second half of the second year, when they declined dramatically (see ). Differences in costs between depressed and non-depressed patients were statistically significant at almost all time points during the study period. Statistically significant predictors of higher societal costs were depressive disorder (CIS-R), white ethnicity, housing problems, relationship problems, having reported cancer as an additional comorbidity and baseline health-care costs. The prevalence of depressive disorder measured by the CIS-R instrument predicted increases in total costs by approximately 50% in all four multivariate models. Inadequate health literacy was associated with reduced health-care cost and this may reflect poor access to care for this group.
There are several reasons why costs might have been under- or overestimated in this economic analysis. With regard to direct costs, medication costs were not taken into account because such data (self-reported) were available at baseline only. However, some patients were being treated for conspicuous depression and other comorbidities during the study. Although this is a limitation, antidepressant treatment is generally inexpensive, given that the most frequently used drugs are ‘off patent’. Another limitation is that indirect costs included only informal care. The high level of missing data in variables related to employment status and sickness absence made it impossible to calculate productivity losses, which is a very important component of indirect costs. Finally, the subgroup analysis was based on depression status as assessed by the HADS score.
Overall, the results could be considered as robust, as data on the important variables for this economic analysis (e.g. service use) were mostly complete, with relatively few having extensive missing data. Further analyses could consider the patterns of depression, examining the varying cost among the different patterns of depression and anxiety. Other future work could look at the economic impact of comorbid depression as a disorder rather than as a symptom-based condition. Similarly, the examination of the relationship between anxiety and costs would be of interest, considering that the majority of patients in the depressed group of this analysis also presented symptomatic anxiety (HADS-A). These findings could drive further research to investigating the excess cost of depression (and/or anxiety) in this patient group by conducting modelling studies on diagnostic and therapeutic approaches that might help to improve health-care outcomes and reduce total costs.
