The clustering of physical health conditions and associations with co-occurring mental health problems and problematic alcohol use: a cross-sectional study

Past year alcohol consumption was measured by the AUDIT [34] including two prior screening questions: “Do you ever drink alcohol nowadays?”, a “No” response triggered an additional question “Could I just check, does that mean you never have an alcoholic drink nowadays, or do you have an alcoholic drink very occasionally, perhaps for medicinal purposes or on special occasions like Christmas or New Year?”. Participants who responded “No” did not complete the AUDIT. An AUDIT score of 0 or a negative response to alcohol use screening questions indicated ‘non-drinkers’. An AUDIT score of 1–7 indicated “low risk”, while a score of 8 + was considered as hazardous drinking (“hazardous/harmful use or probable dependence”). A binary ‘hazardous drinking’ variable was created including two categories: AUDIT 0–7 = non/low risk drinking, AUDIT ≥ 8 = hazardous drinking.

CMDs included depression, anxiety, and social phobia and were measured by the Clinical Interview Schedule-Revised. Cases for mild, moderate and severe depression were grouped into an overall “depression” category to overcome the small cell sizes. Generalised anxiety disorder, obsessive compulsive disorder and panic disorder cases were clustered as “anxiety”, and social phobia, specific phobia and agoraphobia were classed as “phobia”. Post-traumatic stress disorder (PTSD) was assessed by the PTSD Checklist (PCL-C) corresponding to symptoms of DSM-IV PTSD in the past month [35].

Bipolar disorder and psychotic disorders are often grouped in the category of SMIs [36]. Bipolar disorder was screened by the 13-item Mood Disorder Questionnaire assessing manic/hypermanic symptoms and their co-occurrence over the lifetime causing moderate to severe problems (Cronbach’s α = 0.97) [37]. The five-item Psychosis Screening Questionnaire (PSQ) screened for symptoms of probable psychosis [38].

Other psychological problems included borderline personality disorder (BPD) and antisocial personality disorder (ASPD) assessed via the Structured Clinical Interview for DSM-IV Personality Disorders (SCID-II) over the lifetime [39]. Finally, attention-deficit hyperactivity disorder (ADHD) was measured by the six-item Adult Self-Report Scale [40].

Participants self-reported (yes/no) the occurrence of 21 health conditions (e.g. cancer, diabetes, stroke) since the age of 16 years [18]. This measure allowed the exploration of links between mental health, alcohol use and lifetime physical health conditions rather than only considering current physical conditions that occurred in the previous 12 months. In this study, 12 NCDs were considered: cancer, diabetes, epilepsy/seizures, stroke, heart attack/angina, high blood pressure (hypertension), bronchitis/emphysema, asthma, stomach ulcer/digestive problems, liver problems, bowel/colon problems, and arthritis. Selection of these health conditions was based on previous frameworks of multimorbidity [41, 42] and the categorisation was further amended following focus groups with general practitioners (GPs) who suggested that some conditions should be dropped. Therefore, ‘migraine or frequent headaches’, ‘cataracts/eyesight problems’, ‘bone/back/joint or muscle problems’, and ‘skin problems’ were not included in the analyses due to the lack of information regarding their chronicity.

Potential confounding variables included gender, age, ethnicity, and assessment of socioeconomic position through education, occupational grade, and housing tenancy [15]. Supplemental Table 1 (ST1) includes further details on the variables used.

Binary data were analysed by exploratory Latent Class Analysis (LCA) conducted in MPlus 8.5 to identify clusters of physical NCDs. Maximum conditional probability estimated the number of underlying homogeneous classes (clusters) of 12 physical NCDs. Decisions in relation to the best fitting latent class model was guided by statistical fit indices and conceptual and clinical considerations. The Akaike Information Criterion (AIC) [43], the Bayesian Information Criterion (BIC) [44], the sample-size-adjusted BIC (SSABIC) [45], the Lo–Mendel–Rubin Likelihood Ratio Test (LMR-LRT) [46] and entropy were used [47]. Lower values on the AIC, BIC and the SSABIC reflect a good-fitting latent class model. A non-significant LRT value (p > 0.05) suggests that the model with one less class is a better explanation of the data. The entropy statistics ranging between 0 and 1 measures the accuracy of classification of individuals into clusters according to their model-based posterior probabilities with higher values reflecting better classification of participants. The maximum likelihood estimation method was employed when analysing ordinal observed variables.

Multinomial logistic regressions (MLR) in MPlus 8.5 were conducted with a designated reference class to determine how the mental health and alcohol use categories were associated with class membership. Results report unadjusted multinomial OR (MOR) and adjusted MORs (AMOR) with 95% CIs controlling for age, gender, ethnicity, education, and occupational grade.

Previous research has demonstrated that individuals with mental health conditions have a higher risk of developing multimorbidity [18]. Therefore, the MLR analysis was repeated using the MHPs alone group as the reference group to explore whether the group with co-occurring MHPs and HD had a further increased risk. Additionally, MLR in MPlus investigated the association between type of mental health problem (i.e. CMD or SMI) with class membership, with ‘No CMD’ and ‘No SMI’ as reference categories.

To identify latent classes, seven latent class models were estimated (N = 7543). The goodness-of-fit statistics to identify the best fitting model are shown in Table ST3. There was a decrease in AIC throughout the six models. The BIC and SSABIC decreased to the four-class model and increased for the further models. The highest entropy statistics was found in the four-class model. Based on the goodness-of-fit indices, data could be well explained by a latent class model with three to six classes. The non-significant LMR-LRT statistics (P > 0.05) in the five-class model suggested that this model is not favourable to the four-class model. Additionally, while entropy with values closer to 1 (> 0.80) demonstrate clear delineation of classes [48], poor entropy is more difficult to specify as the quality of classification has different impact in different settings and even poor entropy can clearly differentiate some of the classes [49]. Although the entropy was slightly lower in the five-class model compared to the four-class model and the non-significant LRT value (p > 0.05) suggests that the four-class model is statistically a better explanation of the data, we chose the five-class model as it provided a clinically more informative and interpretable explanation of the clustering of physical health conditions following discussion with GPs.

Conditional probabilities are presented in Table 2 and a profile plot for the 5-class model is shown in Fig. 1. The largest class accounted for 76.6% of participants. It was labelled the ‘Physically Healthy’ class as the probabilities for reporting 10 out of 12 physical health conditions were ≤ 5%. There was a 7% likelihood of hypertension and 9% likelihood of asthma. The second largest class, called ‘Hypertension & Arthritis’, accounted for 14.3% of participants. In this class, there was a 65% probability for hypertension and an increasing probability of having arthritis (39%) and diabetes (20%). The third class named ‘Digestive & Bowel Problems’ comprised 3.2% of participants, in which the probability of reporting stomach ulcer/digestive problems was 55% and bowel problems was 48%, with an increasing probability of arthritis (26%). The fourth class named ‘Emerging Multimorbidity’ included 3.1% of individuals with an increasing probability of bronchitis/emphysema (35%), asthma (32%), and arthritis (25%). The fifth and final class accounted for 2.8% of individuals with ‘Complex Multimorbidity’ with a probability of 72% for having arthritis, 65% for having hypertension, 47% for bronchitis/emphysema, and 43% for stomach ulcer/digestive problems. This class was also characterised by an increasing probability of other conditions: bowel problems (37%), asthma (34%), diabetes (28%), and cancer (25%).

When conducting the MLR analysis with the MHPs alone group as the reference group, and after adjustment, those with co-occurring MHPs and HD had 1.7-fold increased odds of being assigned to ‘Digestive & Bowel Problems’ class compared to those with MHPs alone (Table 4).