Prevalence and risk factors of depressive symptoms in a Canadian palliative home care population: a cross-sectional study

The interRAI Palliative Care (interRAI PC) assessments were completed between 2006 and 2012 by trained case managers using all sources of information available at the time of assessment. Assessors compared the information to standardized coding rules and item definitions, and followed a standard protocol outlined in the interRAI PC manual [33]. The interRAI PC is a copyrighted instrument (http://​www.​interrai.​org) designed for adults (18+) with palliative and end-of-life needs, regardless of setting (e.g., acute care, hospice, home care). The assessments were part of a pilot implementation of the interRAI PC in six of fourteen home care provider organizations in Ontario. All clients in the six sites who were classified as palliative were assessed with the interRAI PC. Patients were classified as palliative by the home care case manager if they were no longer responsive to curative treatment, considered to be dying, and the goal of care was to alleviate distressing symptoms in the last stage of their illness [34]. Informed consent was obtained from clients for participation in the pilot.

The study sample consisted of patients in the pilot who were 18+ living in Ontario, and not experiencing significant cognitive impairment (i.e., Cognitive Performance Score [CPS] < 4, see “Measures” below). Eliminating patients with significant cognitive impairment was done because neurodegeneration can interfere with patients’ emotional disposition, making it challenging for clinicians to diagnose depression [35].

All variables used in the analyses were derived from information in the interRAI PC, which is one of a number of interRAI instruments currently used as routine assessments in countries within North America, the Nordic region, Western Europe, the Czech Republic and Japan. InterRAI instruments feature a number of consistently-defined measures and scales. This facilitates their portability across service sectors and can foster continuous improvement, especially in palliative care services, which are increasingly being delivered by a range of multidisciplinary service providers.

The dependent variable in this study was the score on the interRAI Depression Rating Scale (DRS). The DRS is an assessor-rated score created by summing the responses to the following 7 items: made negative statements, persistent anger with self or others, expressions of seemingly unrealistic fears, repetitive health complaints, repetitive non-health complaints, sad/pained/worried facial expressions, and crying and tearfulness. The four item response categories are rescaled and the responses summed to generate a DRS score ranging from 0 to 14. A score of three or higher has been shown to be predictive of a clinically-confirmed depression diagnosis [36]. The DRS has been validated against the Hamilton Depression Rating Scale, the Cornell Scale for Depression in Dementia, and psychiatrists’ ratings [36]. Unidimensionality, reliability and validity of the DRS in a palliative care population were also recently demonstrated, with the evidence including: a Mökken Scale homogeneity coefficient of 0.34, ordinal reliability measures ranging from 0.78 to 0.86, higher correlations of the DRS with mood items compared to physical function items (indicating convergent and divergent validity), and the ability of the DRS to distinguish between subgroups of patients in the expected manner [37].

The interRAI PC contains a number of embedded health index scales. For example, the Cognitive Performance Scale (CPS) can range from 0 (cognitively intact) to 6 (very severe impairment), and it has been validated against the Mini-Mental State Examination [38]. Functional performance was measured using the Activities of Daily Living Self-performance Hierarchy Scale (ADL SHS), which categorizes 4 ADLs (locomotion, eating, personal hygiene, toilet use) on a scale ranging from 0 (independence) to 6 (total-dependence in late-loss ADLs) [39]. A score of 2 or more represents a conventionally-used cut-off indicating that limited assistance is required for at least one of the ADLs. Pain was assessed using a four-point pain scale ranging from 0 (no pain) to 3 (excruciating daily pain), with a cut-off of 2 used to identify patients experiencing daily pain. The pain scale has been validated against the Visual Analogue Scale [40] and is used as an outcome quality indicator in home care settings [9]. Health instability and illness severity was measured using the Changes in Health, End-Stage disease and Signs and Symptoms (CHESS) scale. Scores range from 0 to 5, with 5 representing the highest level of health instability. The CHESS score is a very strong predictor of mortality independent of age and a number of other factors [41], and successfully predicts adverse outcomes compared to other measures [42].

Other independent variables were identified from a review of the palliative care literature, and included in the model subject to availability within the interRAI PC and significance in terms of their association with the DRS. An alpha level of 0.25 was used for the χ ² significance tests to identify the initial list of independent variables to guard against the omission of potentially-important variables. Additional file 1: Table S1 provides the definitions for the independent variables. Some variables were composite measures that were either summative (sleep disorders, gastrointestinal disorders) or conditional on the presence of one or more situations (caregiver distress, life satisfaction). For life satisfaction, principal component factor analysis and the Cronbach’s alpha coefficient were used to confirm the validity of combining five (conceptually broad) items into one measure. Factor analysis confirmed a 1-factor solution, all factor loadings were high (0.67-0.81), and the Cronbach’s alpha was 0.77 suggesting acceptable internal consistency.

A multivariate logistic regression model was created using the DRS as the dichotomous dependent variable (DRS ≥ 3 versus <3) and the interRAI functional indicators and other significant items (p ≤ 0.25) as independent variables. The independent variables considered for inclusion in the regression model are listed in the first column of Table 1. All variables were categorical with the categories determined by examining logit plots and univariate statistics measuring distributional properties. The derivation and meaning of response categories for the independent variables is provided in the (Additional file 1: Table S1, second and third columns).

Multicollinearity was assessed using polychoric correlations because these are preferred as a measure of association for ordinal data [43, 44]. A cutoff of 0.40 was used to identify variable pairs where elimination of one item was required to avoid mulit-collinearity problems. The following variable pairs exceeded the cutoff:

Prognosis and the CHESS score are both mortality measures. CHESS was retained in the model instead of prognosis because it has been validated and displays less assessor variability in use and judgement [41]. Dyspnea, appetite problems and the ADL SHS score were eliminated because each of these directly relates to items used in calculating the CHESS score [41]. The CPS score was retained in the model instead of communication problems because the validity and reliability of the former is well established [38], and many studies report an association between cognitive impairment and depression [9, 45, 46]. Living alone was retained in the model instead of marital status because of the strong bi-variate association it shows with the DRS (Table 1), and because it better captures the extent of social isolation, which has been long recognized as a risk factor for depression [47, 48].

The impact of missing data was also considered. Although the extent of missing data for most items is below 5% (Table 1), restricting the analysis to clients having a response for all variables included in the model (a complete case analysis) reduces the sample from 5,144 to 3,734, a reduction of 27%. While a complete case approach is often used in health research [49], other methods are receiving increased attention with the choice of method depending on the pattern of missing data and the mechanisms causing it [50]. We believe that our missing data reflect a random process rather than systematic bias. However, we cannot be certain which of the three randomness patterns described in the literature applies to our data: missing completely at random (MCAR), missing at random (MAR), or missing not at random (MNAR) [50]. There is no universal method of handling MNAR, but the pattern is rare [51]. MCAR has been found to be insensitive to the method of handling missing data [50], but this pattern is also rare [51, 52]. MAR characterizes most missing data [51, 52], and multiple imputation is recommended for this pattern because it leads to unbiased results with correct standard errors [50, 52]. We performed multiple imputation, and provide these results alongside those for the complete case analysis, as has been previously recommended [53]. Logistic regressions were performed using backward selection and a significance level of 0.05 for retaining model variables. Risk factors were considered significant if they were selected in at least 50% of the logistic regressions (i.e., 50 regressions were run, one for each of the 50 imputations conducted, with factors considered significant if they were retained in at least 25 runs). Multiple imputation used continuous-based imputation with rounding [54] and the results for 50 imputations were pooled using normalizing transformations [55]. Imputation included the outcome variable (DRS) [56] and all independent variables left after addressing the multi-collinearity concerns discussed above.

A special sub-analysis was undertaken to explore whether prognostic awareness was a potential risk factor for depression. This analysis employed a smaller sample consisting of patients with an estimated prognosis of less than six months, because prognostic awareness in the interRAI PC pertains only to this subgroup. We also tested whether the effect of prognostic awareness on the DRS was mediated by life satisfaction, using the methodology recommended by Frazier et al. [57]. Mediation was examined because of the complex way that prognostic awareness appears to interact with acceptance in shaping depression [27]. In particular, we were interested in whether life satisfaction (which includes acceptance and optimism) mediated the relationship between prognostic awareness and the DRS, in light of the research linking “peaceful awareness” with more positive end-of-life outcomes [3, 27].

It is recommended that the goodness-of-fit of prognostic models be assessed using measures of both discrimination and calibration [58, 59]. The c statistic, which corresponds to the area under the receiver operating characteristic (ROC) curve, was chosen as the measure of discrimination, and the Hosmer and Lemeshow statistic was chosen to assess calibration. A c statistic of 0.70 or greater indicates good sensitivity and specificity, and a Hosmer and Lemeshow statistic that is small with a large p value indicates acceptable model fit.

SAS Version 9.2 was used for all of the statistical analyses (http://​www.​sas.​com). The study was approved by the Office of Research Ethics at the University of Waterloo. This institution is responsible for managing and controlling access to the data used in this study.

Sample characteristics are provided in Table 1. Two-thirds of the sample was over the age of 65, with an average age of 70.0 years and a range of 19.6-107.2 years. The study sample was evenly distributed by gender (49.1% male), and over half (61%) were currently married. The prevalence of symptoms of depression was 9.4% in the full sample (patients with DRS ≥3, footnote h of Table 1). Eighty-three percent of the sample had a CHESS score of 2 or more, with this high percentage reflecting the severity of illness characterizing this palliative population. Sixteen percent of the sample had a CPS score of 2 or more, which is approximately equal to a score of 19.2 on the Mini-mental State Examination [38]. The assessor-reported item on prognostic awareness, which applies to patients with an estimated prognosis of less than six months, indicated that just over half of these patients were aware of their prognosis. Thirty-seven percent of the sample experienced at least a moderate level of functional impairment, almost two-thirds had daily pain, and 42% had at least 3 co-morbidities, with cancer being the main diagnosis (86.5%). One quarter of the patients reported that their caregivers were distressed, although the vast majority (95%) reported having a supportive family.

Table 1 provides the bivariate associations between the independent variables and the DRS (ORs and 95% CI, χ ² p values). The DRS was significantly (p ≤ 0.05) associated with the majority of items. The proportion of patients with DRS scores ≥ 3 was higher for patients who: were younger, female, had with more co-morbidity, pain, cognitive impairment, health instability (CHESS), needed more assistance with ADLs, sleep disorders, appetite problems, gastrointestinal problems, communication problems, dyspnea, and caregiver distress. On the other hand, the proportion of patients with DRS scores ≥3 was lower for patients with an estimated prognosis of more than 6 months, high life satisfaction and more family support. Among those with an estimated prognosis of less than six months, prognostic awareness was associated with lower DRS scores. There is little evidence of the DRS varying by marital status, site, or living arrangement (alone versus with others). There were also no significant differences in the DRS across diagnostic categories, consistent with a recent study by Steinhauser et al. [60].

Table 2 provides the logistic regression results for the complete case and multiple imputation models. The results from the two methods were similar for the risk factors selected into the models and the relative strength of their association with the DRS. Both methods found that life satisfaction had the strongest association with the DRS, and both selected the following six risk factors: life satisfaction, CHESS, sleep disorders, pain, caregiver distress and gender. The multiple imputation analysis identified two additional risk factors associated with the DRS - cognitive impairment and gastrointestinal symptoms. The c statistic was above 0.70 for the complete case and all multiple imputation models, indicating good sensitivity and specificity. The Hosmer and Lemeshow statistic for the complete case model was relatively small with a large p value (6.97; p = 0.54), indicating model acceptability. For the multiple imputation models, this statistic showed much more variation compared to the c statistic, as expected [59], but in all cases the statistic indicated acceptable model fit.

We note that prognosis is closely related to CHESS and performs a similar function in the model – i.e., if prognosis is included instead of CHESS, it is a significant risk factor along with the seven other variables selected into the CHESS model (see Table 2). The prognosis model also identifies one additional risk factor, number of co-morbidities, although the OR is significant for the highest category only (5+). These results reflect the conceptual overlap between CHESS and prognosis and the strong relationship both show with the DRS (see Figures 1 and 2).

We also examined whether prognostic awareness was a risk factor associated with depressive symptoms. We used multiple imputation, and ran the logistic regression for patients with an estimated prognosis of less than six months (n = 2547). Table 3 (last column) provides the regression results for the model including CHESS and prognostic awareness, and shows that the results were similar to the model including CHESS only (first column). An important finding was that prognostic awareness was not a risk factor associated with depressive symptoms. We also found that life satisfaction met Frazier et al.’s [57] four mediation conditions, specifically: significant relationships existed between prognostic awareness and both the DRS (pooled p = 0.03) and life satisfaction (pooled p < 0.0001), a significant relationship existed between life satisfaction and the DRS (pooled p < 0.0001), and the relationship between prognostic awareness and the DRS was significantly reduced once life satisfaction entered the model (pooled z statistic for mediation effect = 6.82, which exceeds 1.96, the p = 0.05 significance level). Therefore, prognostic awareness may not be a significant risk factor because life satisfaction mediates its effect on depressive symptoms. However, prognostic awareness may have an important indirect effect on depressive symptoms, given the nature and significance of its relationship with life satisfaction. For example, our data indicate that life satisfaction was significantly stronger among those that were aware of their prognosis (χ ² = 98.11, p < 0.0001).