Title
Which items on the Distress Thermometer Problem List are most distressing?
Authors
Dr Kerrie Ann Clover (1,2,); Dr Christopher Oldmeadow (3); Ms Louise Nelson (1);
Dr Kerry Rogers (1); Dr Alex J Mitchell (4); Prof Gregory Carter (1,2)
Affiliations
- Calvary Mater Newcastle, Psycho-Oncology Service, Newcastle
- University of Newcastle, Priority Research Centre for Translational Neuroscience and Mental Health, Newcastle
- Clinical Research Design, Information Technology and Statistical Support (CReDITSS) Hunter Medical Research Institute
- Department of Psycho-oncology, University of Leicester and Leicestershire Partnership Trust
Corresponding Author
Dr Kerrie Clover
P: +61 2 4014 4926
F: +61 2 4014 4933
Keywords
Neoplasms, psychology, distress, depression, anxiety
ABSTRACT
Purpose
The importance of distress identification and management in oncology has been established. We examined the relationship between distress and unmet bio-psychosocial needs, applying advanced statistical techniques, to identify which needs have the closest relationship to distress.
Methods
Oncology outpatients (n=1066) undergoing QUICATOUCH screening in an Australian cancer centre completed the Distress Thermometer (DT) and Problem List (PL). Principal components analysis (PCA), logistic regression and classification and regression tree (CART) analyses tested the relationship between DT score (at a cut-off point of 4) and PL items.
Results
16 items were reported by 5% of participants. PCA analysis identified four major components. Logistic regression analysis indicated three of these component scores and four individual items (20 items in total) demonstrated a significant independent relationship with distress. The best CART model contained only two PL items: “worry” and “depression”.
Conclusions
The DT and PL function as intended, quantifying negative emotional experience (distress) and identifying bio-psychosocial sources of distress. We offer two suggestions to minimise PL response time while targeting PL items most related to distress, thereby increasing clinical utility. To identify patients who might require specialised psychological services we suggest the DT followed by a short, case-finding instrument for patients over threshold on the DT. To identify other important sources of distress, we suggest using a modified PL of 14 key items, with a 15th item “any other problem” as a simple safety-net question. Shorter times for patient completion and clinician response to endorsed PL items will maximise acceptance and clinical utility.
INTRODUCTION
The importance of attending to the psychosocial concerns of oncology patients has been recognised [1]. Since psychosocial concerns often go undetected in clinical practice [2, 3], psychosocial screening and management has become a recommended standard for cancer care [4]. An important barrier to screening is “lack of time” in the clinical setting and the use of brief measures has been recommended to increase feasibility and acceptability [5].
One “ultra-short” measure is the Distress Thermometer (DT), a single item visual-analogue scale for distress [6-8]. Distress is conceptualised as a negative emotional experience arising from a range of bio-psychosocial stressors. A score of four or more on the Distress Thermometer has been suggested as warranting further assessment for possible referral to specialist services [6-8]. This cut off point has been widely validated against established measures of depression and anxiety [9]. In this manuscript we refer to patients with “distress” as those scoring four or more on the DT.
A companion “Problem List” (PL) for the DT was devised by a multi-disciplinary panel to guide the clinician in choosing the relevant clinical intervention(s) to manage identified unmet needs [6]. The five logically constructed domains of the PL (emotional; physical; spiritual; practical; and social or family) were intended to encompass potential sources of distress [6]. The number of PL items reported in individual articles varies due to revisions by the NCCN and others, but is generally in excess of 30 items. This significantly increases the time required for screening and lowers utility in clinical settings.
Additionally, there has been limited empirical examination of the concept that distress (as measured with the DT) is related to PL items. Statistical analysis has been limited to bivariate techniques in the majority of studies. The number of PL items related to distress in studies using chi-squared analysis has been reported as 22/33 [10]; 23/35 [11] and 32/33 [12]. Studies using correlational approaches have also reported significant relationships between distress and most PL items [13, 14].
Few published studies have used multivariate statistics to examine the relationship between distress and PL items [15-17] . It should be noted that these three studies used slightly modified versions of the PL, thus direct comparisons are somewhat limited. However, all studies reported that only some PL items were significantly independently related to distress, suggesting the possibility of modifying the list for clinical practice by retaining a reduced number of key items.
VanHoose et al [15] conducted a logistic regression with categorised distress score (0-3 vs 4-10) as the dependent variable and all of the original PL items as predictor variables along with age and marital status. They found five variables independently contributed to distress: worry; getting around; financial; sleep and nervousness. Buchman et al [16] conducted two logistic regression analyses, one using domain scores and one using items significantly associated with categorised distress score. Four independent predictors of distress were identified: emotional concerns; family concerns; physical concerns and depression history (a new item). A third study [17] modified the PL so that each item was rated on a 0-10 scale and analysed as a continuous variable. Four items were found to independently contribute to DT score: nervousness, pain, emotional control (a new item) and physical fitness (a new item).
Differences between studies in statistical methods and PL items limit study comparisons. However, emotional aspects consistently demonstrated a strong relationship with increased distress. Items on the physical domain also appeared to be frequently related to increased distress while practical and family problems had a less consistent relationship. Spiritual concerns were the least often observed to have a statistically significant relationship with distress as measured on the DT.
AIMS
Given that statistical approaches to determining the relationship between distress and problem list items have been limited to bivariate or regression analyses, we were interested to employ other statistical techniques, specifically Principal Components Analysis (PCA) with logistic regression and Classification and Regression Tree (CART) analysis, to examine the relationship between PL items and distress.
Our secondary aim was to develop a shorter PL, by identifying items with a significant independent relationship to distress. We hope that a brief PL might be useful in busy clinical settings.
Methods
Setting
The Calvary Mater Newcastle (CMN) is a major regional cancer treatment centre in Australia. Oncology and haematology outpatients completed the DT, PL and other measures on a touchscreen computer (QUICATOUCH) prior to their oncology appointment. Details of the QUICATOUCH assessment and scoring algorithms have been described previously [18-20].
Study Design
The study is a cross-sectional survey of patients undergoing their first occasion of QUICATOUCH screening in the first three months of the program (Oct-Dec 2007) when all patients completed the problem list regardless of DT score. After this time we changed our procedures and only asked the PL of people scoring over threshold (four or more) on the DT.
Ethics
The Hunter New England Research Ethics Committee formally approved the use of these data as a quality improvement project.
Participants
Eligibility criteria for QUICATOUCH assessment included age over 18 years, sufficient English language skills and being well enough to participate. Patients are eligible for assessment at any clinic visit after their initial appointment. Resource constraints prevented us from tailoring screening to stage of illness or time since previous screening.
Measures
DT and PL
Patients indicated the amount of distress experienced “in the past week including today” on the Distress Thermometer (DT), a visual-analogue scale from zero (no distress) to 10 (extreme distress). They then answered (yes/no) the PL items. We used the 2005 version of the PL with 35 items [6] (Appendix 1).
Demographic and Clinical Variables
Age, gender and whether the patient was currently undergoing treatment with either radiation therapy or chemotherapy were routinely collected.
Statistical Analysis
CART analyses were performed using RPART library in R (version 3.1.1). All other statistical analyses were programmed using SAS v 9.4 (SAS Institute, Cary, North Carolina, USA).
Prevalence
Percentages were used to rank the items in terms of frequency. Items endorsed by 3-5% of patients were deemed ‘low prevalence’ items and items indicated by 0-2% of patients were deemed ‘very low prevalence’ items.
Principal Component Analysis
Principal Component Analysis (PCA) is a data reduction technique that identifies patterns of correlation among a set of variables. Unlike factor analysis, PCA is directly applicable to binary data (like the PL) and does not assume that item responses are causal outcomes from a latent trait. We used eigenvalues (>1), the scree plot, and the amount of variance explained to inform component extraction. A minimum component loading of at least 0.4 was used to determine significant items. A component on which three or more items loaded was termed a major component.
Logistic Regression Analysis
The results of the PCA analysis were used to select potential predictor variables for a logistic regression model to determine relationship with DT score ≥4 as the dependent variable. Item sum scores were entered for four major components as continuous independent variables. Items which did not load on one of the major components were entered as single item dichotomous independent variables. Age (years – continuous variable) and gender (dichotomous) were also included as independent variables as they have previously been shown to demonstrate a relationship with distress (15-16).
Classification and Regression Tree (CART) analysis
CART analysis develops a “decision tree” structure for data. The CART algorithm selects the items that provide the optimal partition of the data in respect to discrimination of the outcome variable (measured by the Gini index). This process of splitting (subgrouping) is repeated until a minimum group size is reached (here we used 1). To minimize over-fitting the tree is then pruned using a cost-complexity parameter that minimizes the cross-validated error. The cost-complexity parameter reflects the trade-off between the tree complexity and how well the tree fits the data. The best model has the lowest cross-validated error. Unlike regression analysis, CART analysis is not affected by strong correlations between variables. In the first CART analysis all of the individual PL items were included as potential predictors (independent variables); with distress (at a cut-off point of ≥ 4on the DT) as the outcome variable. A second CART analysis was conducted which included age (years), gender, cancer type (haematological, breast, bowel, prostate, lung and “other”) and current treatment type (chemotherapy, radiotherapy, both, none) with all of the individual PL items as potential independent variables.
Results
Data were obtained from 1066 participants (female n=555, 52%) with various types of cancer including haematological (n=194, 18%), breast (n=191, 18%); bowel (n=118, 11%); prostate (n=80, 8%); lung (n=41, 4%) and other (n=440, 41%). Ages ranged from 18 to 90 years (mean age = 62 years, standard deviation = 13 years). Most (n=839, 78%) were not currently on treatment, with chemotherapy-only being received by n=187 (18%); radiotherapy-only by n=12 (1%) and combined chemo and radiotherapy by n=30 (3%). Distress over threshold (>=4) was reported by n=292 (27%) of the sample (males: n=113/511 (22%); females n=179/555 (32%)).
Prevalence
The overall prevalence of endorsement of PL items is shown in Figure 1. There were eleven low prevalence items: problems with children; indigestion; mouth sores; partner; dry nose; changes in urination; insurance or finances; appearance; fever; bathing or dressing; transport; and five very low prevalence items: problems with work or school, sexual, housing, childcare, spiritual.
INSERT FIGURE 1 AROUND HERE
Principal Component Analysis
There were 11 components with eigenvalues exceeding 1, which cumulatively explained 53% of the variance in the data. There were four major components (three or more items) and seven other components which consisted of only 1 or 2 items. Six items loaded highly (rotated factor loadings) on the first component, which appeared to represent an “emotional” problem group: worry (.68), fears (.68), nervousness (.67), sadness (0.66), depression (.62) and problems sleeping (.46). The second component appeared to represent a “physical functioning” problem group and consisted of seven items: getting around (.70), pain (.64), transport (.44), breathing (.43), fatigue (.41), memory (.41) and bathing (.40). The third component was a “gastro-intestinal symptoms” problem group and consisted of nausea (.60), diarrhoea (.59) and eating (.58). The fourth component consisted of three items which might be a “support” problem group: partner (.66), financial (.61) and spiritual (.57). The scree plots indicated there was no clear cut-point where increasing the number of components had little effect on the explained proportion of variance, indicating that many of the items were largely independent of each other.
The mean component score on Component 1 for those over threshold for distress was 2.32 (SD=1.66) and for those under threshold was 0.32 (SD=0.75). The respective figures for the other components were: Component 2: 1.82 (1.54) and 0.54 (0.99); Component 3: 0.54 (0.81) and 0.14 (0.45) and Component 4: 0.24 (.53) and 0.04 (0.24).
Multiple Logistic Regression Analysis
Component 1 (emotional), Component 2 (physical function), Component 4 (support) scores and four individual PL items (problems with appearance, children, loss of interest and skin) were significantly independently related to distress (Table 1, Area Under the Curve = 92%; -2 log likelihood = 708; Akaike Information Criterion = 758). Component three (gastro-intestinal) was not significantly associated with distress after adjusting for other variables. Other items which did not show a significant independent relationship with distress were: childcare, housing, work/school, changes in urination, constipation, fever, indigestion, mouth sores, nose dry/congested, sexual, feeling swollen and tingling.
INSERT TABLE 1 AROUND HERE
CART analysis
The best model in the first CART analysis (PL items only) needed just two items: worry and depression to predict distress categorised at a cut-off point of ≥ 4 on the DT (Figure 2). Twenty four percent of patients experienced worry, of which 73% (185/251) had a distress score over threshold. The majority (81%, 30/37) of those that did not endorse worry but endorsed depression also had distress over threshold. The majority (90%, 701/778) of the remaining patients (no endorsement of worry or depression) did not experience distress over threshold. In combination the two PL items (worry or depression) had a sensitivity of 73%, specificity of 93% for distress above the cut-off point of four (with positive predictive value of 79% and negative predictive value of 90%). The best model in the second CART analysis (which added demographic and clinical variables as potential predictors) was identical to the first model, with just two items: worry and depression, using the same rules for accurate classification.