Impact of non-participation bias due to psychiatric illness on mortality and cardiovascular event estimates: a Danish longitudinal population study

Introduction

Psychiatric disorders are associated with increased somatic morbidity and premature death, primarily due to preventable and treatable somatic diseases such as ischaemic heart disease, stroke, cancer, chronic lung diseases and infections.1–5 A reduced life expectancy of up to 10–20 years in severe mental disorders, for example, schizophrenia and other psychotic disorders, bipolar disorder and severe depression has been demonstrated.6–12 Thus, to accurately estimate the population’s health profile and disease burden, psychiatric patients must be represented in general population studies, as they may disproportionately contribute to disease burden and mortality in the population. Population studies usually require invited subjects to consent to participate, entailing a risk of selection bias and consequently distortion of study estimates. Well-known determinants of non-participation include sociodemographic factors such as young and old age, male sex, not being married, low socioeconomic status (SES) and foreign country of origin.13–17

Results from previous studies suggest an association between non-participation and psychiatric illness.15 18–23 In both a Finnish18 and Norwegian19 general health population survey and a Swedish population study on mental health,20 registry-based hospital-treated psychiatric illness was associated with non-participation. A population-based study in Norway15 estimated the prevalence of psychiatric illness in non-participants versus participants based on data from questionnaires and a random sample of general practice patient records and similarly found that psychiatric illness prevalence was increased among non-participants. Another Norwegian population study21 found that recipients of disability pension due to substance abuse, psychotic disorders or personality disorders were highly over-represented among non-participants. In studies of the Danish national breast cancer screening programme (Danish women aged 50–69 years)22 and the Danish national colorectal cancer screening programme (all Danish residents aged 50–74 years),23 people with mental disorder, defined as previous psychiatric diagnoses or use of psychoactive prescription medicine, were less likely to participate in the screening programmes compared with people without a history of mental disorders.

Based on the above evidence, we hypothesised that psychiatric diagnoses and psychotropic drug use would be more prevalent in the Danish General Suburban Population Study (GESUS) non-participants than participants, that psychiatric illness in the non-participant group would be of overall higher severity, and that the highest rates of cardiovascular events and all-cause mortality would be observed in the group of non-participating psychiatric patients.

No previous studies have to our knowledge investigated the potential impact of non-participation of psychiatric individuals on population study estimates of somatic morbidity and overall survival.

This present investigation is based on GESUS, a large general population health study cohort, in which we exploited that comprehensive registry-based health data were available for participants and non-participants alike. This enabled us to (1) characterise the distribution of all psychiatric diagnostic groups (cf. WHO International Classification of Diseases 10th revision (ICD-10)) and the use of psychiatric hospital services in GESUS non-participants compared with GESUS participants, (2) describe the use of psychotropic drugs in non-participants versus participants, (3) investigate the association between psychiatric morbidity and cardiovascular events and all-cause mortality in a 5-year follow-up and (4) test if participation status interacted with this association.

Methods

Results

A total of 49 807 Danish adults living in Naestved Municipality were defined as eligible for an invitation to participate in GESUS from 2010 to 2013. Due to the ongoing enrolment design, 100 were registered dead before receiving the invitation and hence excluded. Of the remaining 49 707 subjects, 21 203 (42.7%) consented to participate (figure 1). Table 1 lists the baseline sociodemographic characteristics of participants and non-participants. Compared with participants, non-participants were more often men (OR for non-participation 1.36 (1.31 to 1.41)), younger than 40 (OR for non-participation 2.95 (2.76 to 3.15)), originating outside the Northern countries (Denmark, Norway, Sweden, Finland, Greenland, Iceland, Faroe Islands) (OR for non-participation 1.12 (0.98 to 1.27)), residing in the rural part of Naestved Municipality (OR 1.20 (1.15 to 1.25)), not married (OR for non-participation among widowed 1.78 (1.65 to 1.92), unmarried 1.71 (1.62 to 1.81) and divorced 1.66 (1.57 to 1.75)) and had an overall lower SES indicated by highest attained education (OR for non-participation among individuals with basic education 1.81 (1.70 to 1.93)) and annual personal income (OR for non-participation among individuals with an income in the first quartile 1.77 (1.66 to 1.89)). Information on marital status was obtained after inclusion and was missing in 1818 subjects who died before the data extraction. A total of 997 subjects finished their education before 1981 and were therefore not registered in the Danish Population’s Education Register. Data on residence and annual personal income were missing by less than 0.1%.

Request permission

In a sensitivity analysis of the adjustment for SES (online supplemental table 1), adjustment for the highest attained education compared with adjustment for both the highest attained education and annual personal income overall resulted in the same estimates. Therefore, income was omitted and only education was used in the analyses to adjust for the confounding potential of SES.

Survival data

In both non-psychiatric and psychiatric participants and non-participants, less than 5% of deaths in the follow-up period were due to unnatural causes (suicide, accidents, violence).

Figure 2A shows Kaplan-Meier-estimated survival curves in a 5-year follow-up after GESUS invitation, stratified by participation status and psychiatric morbidity. The 5-year survival probability was highest in non-psychiatric participants (97.5% (95% CI 97.2 to 97.7)), followed by psychiatric participants (95.5% (94.9 to 96.0)), non-psychiatric non-participants (92.6% (92.2 to 93.0)) and lowest in psychiatric non-participants (81.1% (80.4 to 81.9)) (log-rank test: p<0.0001). Cumulative incidence functions of MACE in the 5-year follow-up period are shown in figure 2B. Incidences of MACE also differed by participation status and psychiatric morbidity (Gray’s test: p<0.0001) with the highest registered incidence in psychiatric non-participants (12.2% (95% CI 11.5 to 12.8) and the lowest incidence in non-psychiatric participants (3.5% (3.2 to 3.8)) (psychiatric participants: 6.0% (5.4 to 6.7); non-psychiatric non-participants: 6.1 (5.8 to 6.5)).

Request permission

Figure 2

All-cause mortality and MACEs in the GESUS population (5-year follow-up). (A) Kaplan-Meier survival curve with 95% CIs for all-cause mortality and the cumulative incidence of all-cause mortality. (B) The cumulative incidence of MACEs with 95% CI. Data were obtained from the Danish Register of Causes of Death and the Danish National Patient Registry. GESUS, the Danish General Suburban Population Study; MACE, major adverse cardiovascular event.

Overall, GESUS non-participants had increased rates of both all-cause mortality (HR 2.53 (95% CI 2.32 to 2.76)) and MACE (HR 1.69 (95% CI 1.56 to 1.84)) relative to GESUS participants (ref.) (table 5). Likewise, individuals with psychiatric morbidity had higher rates of all-cause mortality (HR 1.52 (95% CI 1.42 to 1.62)) and MACE (HR 1.45 (1.35 to 1.55)) compared with individuals without psychiatric morbidity (ref.). Stratification by both participation and psychiatric morbidity status showed the same trend as presented in figure 2; compared with non-psychiatric participants (ref.), the rates of all-cause mortality and MACE were most increased in psychiatric non-participants (HR 3.37 (95% CI 3.01 to 3.78) and 2.30 (2.07 to 2.56), respectively), followed by non-psychiatric non-participants (HR 2.26 (2.02 to 2.54) and 1.65 (1.48 to 1.83), respectively) and psychiatric participants (HR 1.23 (1.05 to 1.44) and 1.39 (1.21 to 1.59), respectively). The association between psychiatric morbidity and (1) all-cause mortality and (2) MACE was altered depending on participation status, indicating an interaction between psychiatric morbidity and participation status (likelihood ratio test: p<0.0001 (for both outcomes)).

Table 5

View inline|View popup

Five-year all-cause mortality and MACEs according to GESUS participation and psychiatric morbidity

In a hypothetical cohort consisting of the complete GESUS population (N=49 707) minus the subpopulation of psychiatric non-participants (n=39 827), the estimated all-cause mortality and MACE incidence rates were 10.4 (95% CI 10.0 to 10.9) and 10.6 (10.2 to 11.1) per 1000 person-years, respectively. In comparison, the incidence rate estimates of all-cause mortality and MACE of the complete GESUS population (N=49 707)—including all subpopulations both participants and non-participations—were 16.3 (15.8 to 16.8) and 13.8 (13.3 to 14.3) per 1000 person-years (table 5, online supplemental table 2).

Discussion

Participation bias in many aspects of society’s activities is generally assumed, but scientific evidence and quantitative estimates are limited. This is to our knowledge the largest study to describe non-participation bias due to psychiatric illness in a general population survey and the first study to investigate the potential impact of psychiatric non-participation on population survey estimates of somatic health. In this study, we report a prevalence of severe psychiatric disorders more than twice as high in non-participants compared with participants in GESUS, as well as a more severe psychiatric illness course in non-participating compared with participating psychiatric patients. All psychiatric ICD-10 diagnoses—except eating disorders—were associated with non-participation, and psychiatric disorders particularly under-represented among study participants were psychotic disorders, substance abuse and dementia. The prevalence of individuals using psychotropic drugs was higher among non-participants, and antipsychotics, anti-dementia drugs and drugs used in addictive disorders were the most strongly associated with non-participation. Moreover, we report that even after adjustment for sociodemographic factors, non-participants with psychiatric morbidity had a twice as high rate of MACE and a more than three times higher rate of death than participants without psychiatric morbidity. In a hypothetical GESUS cohort including both participants and non-participants except the subpopulation of psychiatric non-participants, incidence rates of all-cause mortality and MACE were lower compared with incidence rates estimated from the complete GESUS population of all invited individuals. Thus, the GESUS population cohort—comprised of the GESUS participants only—underestimates MACE and the total mortality due to non-participation bias in general, and psychiatric non-participation in particular. In future studies, this bias can be minimised by various statistical methods, for example, stratification, matching, propensity score matching and inverse probability weighting.

The increased somatic morbidity and mortality in individuals with psychiatric illness are well-established,6–12 but the demonstration of a twofold incidence of MACE and a 14% reduced survival probability (predominantly due to death of natural causes) in non-participants with psychiatric illness compared with participants with psychiatric illness has not been described before. Our findings indicate that not only do psychiatric patients participate less frequently in population studies, but the group of participating psychiatric patients is also not representative of the overall psychiatric population in terms of either psychiatric diagnostic group, psychiatric illness severity, somatic comorbidity or all-cause mortality. One could hypothesise that participation in health studies—independently of the presence of psychiatric morbidity—is associated with better health behaviour, overall resulting in both more well-treated psychiatric disorders and somatic diseases. Our results also imply that population surveys aiming at determining psychiatric and somatic disease prevalence may underestimate the true burden, and survey subgroup analyses of specific psychiatric diagnostic groups—especially those with low participation rate (psychotic disorders, substance abuse, dementia)—constitute a great challenge as they may be inaccurate due to group heterogenicity and non-representative responders.

Our finding of an association between psychiatric disorders and non-participation in population studies was consistent with that of previous studies.15 18–23 Across all studies, a previous hospital-registered diagnosis of schizophrenia or substance abuse was most strongly associated with non-participation (no other studies included diagnoses of dementia).18 19 22 However, the association between mood disorders and non-participation was inconsistent, perhaps due to different study purposes or demands.18 20–22 The prevalence of individuals with hospital-treated psychiatric disorders was markedly higher in GESUS (8.6% in non-participants, 3.9% in participants) compared with other studies (4.2%–5.1% in non-participants, 2.0%–2.5% in participants),18–20 which may be explained by a biased difference due to inclusion of specific age groups or specific psychiatric diagnoses or a true difference related to sociodemographic variations. After including psychotropic drug use in the definition of psychiatric morbidity, the prevalence of psychiatric morbidity in GESUS increased to 31.1%, like the findings of Jensen et al.22

Limitations of our study include the lack of registry data on psychiatric diagnoses in primary care. Although we tried to accommodate by including psychotropic drug use as a proxy measure of psychiatric illness, this still leaves mentally ill individuals without pharmacological or specialty psychiatric treatment unidentified. We only included the primary diagnosis in the definition and description of psychiatric illness, and we cannot rule out that the presence of psychiatric comorbidities (particularly substance abuse) would be associated with even higher rates of excess morbidity and mortality. Another challenge is the registry-based qualitative description of psychiatric illness. However, in line with previous studies,22 33 34 we included the cumulative use of psychiatric hospital services and time spent in inpatient care as proxy measures of illness severity and chronicity, although these are not necessarily correlated. We adjusted estimates for sociodemographic factors well-known to be determinants of non-participation,13–17 but due to the close relation between psychiatric morbidity and sociodemographic factors,35 this might have diluted the true associations, just as we cannot exclude residual confounding. It is also likely that the GESUS participation rate and pattern were affected by the study purpose and demands, which were more extensive than demands of most previous population studies evaluating mental health bias,15 18–23 perhaps leading to a lower participation rate (43%). The possible effect of this aspect on the differential participation of individuals with psychiatric illnesses is difficult to evaluate and thus unknown. Although a wide age range was included in GESUS, only 25% of individuals aged 20–30 years were invited, possibly leading to a slight underestimation of the early-onset psychiatric disorders. It should also be considered that GESUS was conducted in a suburban and rural population, potentially limiting the extrapolation to populations of other sociodemographic compositions.

To conclude, this is to our knowledge the first study to investigate the impact of psychiatric non-participation on population survey estimates of somatic health. By unique access to data on both study participants and non-participants enabled by cross-linkage to Danish nationwide registries, we demonstrate considerable non-participation bias in a large general population study by under-representation of psychiatric patients leading to distorted estimates of both psychiatric and somatic morbidity prevalence and severity as well as all-cause mortality. To obtain accurate estimates from future population studies to use for strategies in public health efforts—particularly those aimed at psychiatric conditions—initiatives are necessary. We suggest considering the following: optimisation of recruitment strategies specifically targeting individuals with psychiatric illness and use of statistical sampling weights36 to adjust for differences between sample and background population, particularly focusing on the representation of psychiatric illnesses.