Symptom Clusters in Patients With Advanced Cancer: A Reanalysis Comparing Different Statistical Methods

Vol. 44 No. 1 July 2012 Journal of Pain and Symptom Management 23 Original Article Symptom Clusters in Patients With Advanced Cancer: A Reanalysis Comparing Different Statistical Methods Emily Chen, BSc (C), Janet Nguyen, BSc (C), Luluel Khan, MD, Liying Zhang, PhD, Gemma Cramarossa, BHSc (C), May Tsao, MD, Cyril Danjoux, MD, Elizabeth Barnes, MD, Arjun Sahgal, MD, Lori Holden, MRTT, Florencia Jon, MRTT, and Edward Chow, MBBS, MSc, PhD, FRCPC Rapid Response Radiotherapy Program, Odette Cancer Centre, Sunnybrook Health Sciences Centre, University of Toronto, Toronto, Ontario, Canada Abstract Context. The clinical relevance of symptom cluster research remains questionable if inconsistencies, partially attributable to the varying statistical analyses used, exist. Objectives. To investigate whether symptom clusters identified were consistent using three different statistical methods and to observe the temporal pattern of clusters. A secondary objective was to compare symptom clustering in responders and nonresponders to radiotherapy over time. Methods. Reanalysis of an existing data set compiled from 1296 patients with advanced cancer was performed using hierarchical cluster analysis (HCA) and exploratory factor analysis (EFA) to extract symptom clusters at baseline, 1-, 2-, 4-, 8-, and 12-week follow-up time points. Findings were compared with results obtained using principal component analysis (PCA) in our previously published study. The original sample was further divided into two subgroups: responders and nonresponders. The symptom clusters present in each subgroup were examined using PCA, HCA, and EFA at the same time points as mentioned above. Results. The symptom cluster findings of HCA and PCA correlated more frequently with each other than either did with the results of EFA. Complete consensus in all three statistical methods was never reached at any assessment time point in the present study. Increasingly diverging patterns of symptom cluster development over time were observed in the responder vs. nonresponder subgroups. Symptom pairs comprising anxiety and depression or fatigue and drowsiness consistently presented in the same cluster despite the shifting of other symptoms in the cluster over time. Conclusion. The presence and composition of symptom clusters identified varied depending on which statistical analysis method was used. A key step in achieving consistency in symptom cluster research involves the utilization of Address correspondence to: Edward Chow MBBS, MSc, PhD, FRCPC, Department of Radiation Oncology, Odette Cancer Centre, Sunnybrook Health Sciences Ó 2012 U.S. Cancer Pain Relief Committee Published by Elsevier Inc. All rights reserved. Centre, 2075 Bayview Avenue, Toronto, ON M4N 3M5, Canada. E-mail: Edward.Chow@sunnybrook.ca Accepted for publication: July 17, 2011. 0885-3924/$ - see front matter doi:10.1016/j.jpainsymman.2011.07.011

24 Chen et al. Vol. 44 No. 1 July 2012 a common analytical method. J Pain Symptom Manage 2012;44:23e32. Ó 2012 U.S. Cancer Pain Relief Committee. Published by Elsevier Inc. All rights reserved. Key Words Symptom cluster, cancer metastases, statistical analysis Introduction Symptom experience among advanced cancer patients is often marked by a wide array of physical and psychological symptoms. Earlier research has demonstrated the tendency of two or more related symptoms to co-occur, forming symptom clusters. 1,2 Concurrent symptoms are burdensome for these patients and often impact their functional status, quality of life, and mood. 3 When multiple symptoms occur simultaneously, their interrelationship is likely complex and warrants multivariate analyses to describe it comprehensively. In recent years, the concept of symptom clusters has gained recognition as a valuable platform for symptom management in oncology. 4 However, the clinical relevance of symptom cluster research remains questionable if inconsistencies, partially attributable to the varying statistical analyses used, exist. A few previous studies have explored symptom clusters in cancer patients using more than one analytical method but were predominantly cross-sectional in design. Henoch et al. 5 examined symptom clusters in lung cancer patients using cluster analysis and factor analysis, and Pearson correlations. Maliski et al. 6 used these same three analytical methods in exploring symptom clusters in prostate cancer patients. Although the study by Gleason et al. 7 determined symptom clusters in patients with brain tumors longitudinally using four statistical methods, analysis was completed at only two time points. These studies observed symptom cluster findings that were comparable and somewhat consistent yet rarely in complete agreement when using different statistical methods. Previously, we explored symptom clustering in advanced cancer patients through applying principal component analysis (PCA) on data collected using the Edmonton Symptom Assessment System (ESAS). 8 Aside from PCA, two other analytical methods often used in symptom cluster research are hierarchical cluster analysis (HCA) and exploratory factor analysis (EFA). 9 These two statistical methods serve well to capture the nature of symptom clusters as an assembly of concurrent and related symptoms. Further symptom cluster analysis was conducted in the two subgroups: responders and nonresponders to radiation treatment. The primary objective of the present study was to establish if significantly different symptom clusters would result when analyzed with alternative statistical analysis approaches. A secondary objective was to compare symptom clusters in responders vs. nonresponders to radiation therapy over time. Patients and Methods The previously reported data set compiled from 1296 patients diagnosed with advanced cancer was used in the current analysis. 8 Using the same data set compensates for possible inconsistencies that may result from factors such as different assessment tools or varied sample population characteristics. All English-speaking patients were asked to complete the ESAS questionnaire during their initial consultation in a palliative radiation therapy clinic. 8 Questionnaires were administered and all information was collected by a trained research assistant. ESAS questionnaires were administered as part of routine clinical monitoring at baseline and s 1, 2, 4, 8, and 12. The ESAS is an 11-point scale that evaluates the following nine symptoms on a scale from 0to10(0¼ absence of symptom and 10 ¼ worst possible symptom): pain, fatigue, nausea, depression, anxiety, drowsiness, appetite, sense of well-being, and dyspnea. 10 This tool has been successfully validated in the cancer population. 11,12 Patient demographics, cancer history, analgesic intake, and disease status also were recorded. All primary assessments were completed before radiation treatment. Ethics approval was obtained from Sunnybrook Health Sciences

Vol. 44 No. 1 July 2012 Symptom Clusters in Advanced Cancer 25 Center and verbal consent was provided by participating patients. Statistical Analysis Symptom clusters were derived using PCA, EFA, and HCA. Symptoms did not need to be present in a certain percentage of patients to be included in the analysis. PCA with varimax rotation was performed on the ESAS items to examine any interrelationships between symptoms at each followup time point. This analytical method groups variables (symptoms) together to form a component (cluster) by identifying which variables correlate with each other in a distinct pattern. 13 The highest loading factor score determines the assignment of symptoms to clusters. An eigenvalue greater than 1.0 and explaining almost 10% of the total variance was required to determine significant clusters. Cronbach s alpha was calculated to determine the internal consistency and reliability of the clusters. EFA is a familiar statistical approach for symptom cluster identification in oncology. EFA is the only method that functions on the assumption that symptoms in a cluster share a common underlying dimension, called a common or latent factor, which binds two or more symptoms together. 14,15 Symptoms associated with one latent factor covary more closely with each other compared with symptoms caused by a different latent factor. 16 EFA groups symptoms into a cluster based on covariance between symptoms. Factors with eigenvalues greater than 0.9 were retained, indicating that almost 10% of the variance in the symptom was shared with the latent factor after controlling for the correlation between factors. The maximum likelihood method with varimax orthogonal rotation was applied to approximately multivariate normal data to determine covariance between symptoms. Cronbach s alpha for each symptom cluster was calculated to assess internal consistency. PROC FACTOR procedure in Statistical Analysis Software (SAS version 9.2, SAS Institute, Cary, NC) was conducted for this analysis. HCA is another procedure that can be used to define a symptom cluster. Cluster analysis is an exploratory technique that is used to discover underlying groups of individuals who are similar in their symptom experience or symptom profile. 16 This method is focused on classification and trying to put similar entities together into a cluster and separate this cluster from other clusters. The PROC VAR- CLUS (SAS version 9.2, SAS Institute, Cary, NC) runs clusters on the basis of centroid components; R 2 values of each symptom with its own cluster and with its nearest cluster were calculated. The 1 R 2 ratio is the ratio of one minus the value in the Own Cluster column to one minus the value in the Next Closest column. Low ratios indicate well-separated clusters. The proportion of variation explained by clusters and the minimum proportion explained by a cluster were also reported. The PROC TREE (SAS version 9.2, SAS Institute, Cary, NC) procedure was applied to generate a graphical display of the clusters and demonstrate cluster hierarchy. Criteria for Responders and Nonresponders. To compare symptom cluster dynamics in responders and nonresponders, patients assessed at each follow-up were divided into these two categories based on their pain response to radiation. Response was calculated for each followup interval. Response definitions set by the International Bone Metastases Consensus 17 were adopted for this sample population. Responders were defined as individuals with complete response (CR) or partial response (PR). CR indicated a worst pain score of 0 at the site of radiation with no increase in analgesic intake. PR was described as 1) decrease in worst pain score of at least two at the treated site without analgesic increase or 2) analgesic reduction of at least 25% from baseline without pain increase. All other patients were considered nonresponders. PCA, HCA, and EFA were applied to both subgroups to determine symptom clusters in each. Results Symptom Clusters at Baseline We previously performed PCA with varimax rotation on the nine ESAS symptoms. Baseline patient characteristics are listed in Table 1.At baseline, three symptom clusters were identified. Cluster 1 consisted of lack of appetite, nausea, poor sense of well-being, and pain. Cluster 2 included fatigue, drowsiness, and dyspnea. Cluster 3 comprised anxiety and depression.

26 Chen et al. Vol. 44 No. 1 July 2012 Table 1 Baseline Patient Characteristics Characteristics n (%) Sex Male 677 (52) Female 619 (48) Median age at consultation, 69 (24e95) years (range) Median Karnofsky Performance 60% (10%e100%) Status (range) Primary cancer site Lung 415 (32) Breast 259 (20) Prostate 193 (15) Others 406 (31) Unknown 23 (2) Primary reason for referral Bone metastases 836 (64) Brain metastases 163 (13) Tumor mass 80 (6) Others 217 (17) Pertaining to EFA, Table 2 displays eigenvalues and proportion of variance for the nine ESAS symptoms. One of the three original symptom clusters from PCA was retained and a new symptom cluster emerged, both with eigenvalues >1.0 and proportions >10%. The two symptom clusters cumulatively accounted for 100% of the variance. Table 3 shows the two clusters and final communality from EFA. The new symptom cluster extracted included fatigue, drowsiness, poor sense of well-being, appetite, pain, nausea, and dyspnea, with a Cronbach s alpha of 0.78. The retained symptom cluster involving depression and anxiety had a Cronbach s alpha of 0.77. The internal consistencies of the two clusters were above 75%. Findings using HCA are displayed in Tables 4 and 5. Using this method, the centroid cluster algorithm split the nine ESAS symptoms Table 2 Eigenvalues and Proportions of Variance for Nine ESAS Symptoms at Baseline Component Eigenvalue Proportion Cumulative 1 6.37 0.78 0.78 2 1.78 0.22 1.00 3 0.50 0.06 1.06 4 0.04 0.004 1.07 5 0.02 0.002 1.07 6 0.05 0.006 1.06 7 0.05 0.007 1.06 8 0.20 0.02 1.03 9 0.26 0.03 1.00 ESAS ¼ Edmonton Symptom Assessment System. From eigenvalues and proportions of variance, two factors (clusters) were retained (eigenvalue > 1.0 and proportion > 10%) (bolded numbers), and the cumulative variance shown up to 100%. Table 3 Two Factors (Clusters) and Final Communality Determined Using EFA at Baseline Final Symptom Factor 1 Factor 2 Communality Fatigue 0.67 0.19 0.49 Drowsiness 0.67 0.14 0.47 Sense of well-being 0.61 0.35 0.50 Appetite 0.57 0.17 0.36 Pain 0.47 0.08 0.22 Nausea 0.45 0.20 0.24 Dyspnea 0.39 0.14 0.18 Depression 0.25 0.82 0.74 Anxiety 0.18 0.71 0.53 Percentage of variance 78 22 Cronbach s alpha 0.78 0.77 EFA ¼ exploratory factor analysis. Bolded values in each column indicate that the corresponding symptoms constituted the respective cluster determined using EFA as per the criteria of an eigenvalue > 1.0 and proportion > 10% (see Methods). into two clusters. Cluster 1 contained pain, fatigue, nausea, drowsiness, poor appetite, poor sense of well-being, and dyspnea. Cluster 2 included the symptoms depression and anxiety (Table 4). The first cluster only explained 43% of the total variation, whereas the second cluster explained 81% of the total variation. Cluster 1 was then further divided because of its smaller proportion of variance (43%), resulting in two groups: a cluster of fatigue, drowsiness, and dyspnea, and a cluster of pain, nausea, poor appetite, and poor sense of well-being. The tree diagram in Fig. 1 shows a visual representation of the three clusters identified using HCA. Table 4 Initial Two Symptom Clusters Identified Using Centroid Cluster Algorithm in HCA at Baseline Cluster and Symptoms Own Cluster R 2 Next Cluster 1 R 2 Own Cluster 1 R 2 Next Cluster Cluster 1 Pain 0.35 0.03 0.67 Fatigue 0.52 0.11 0.54 Nausea 0.36 0.07 0.68 Drowsiness 0.50 0.08 0.55 Poor appetite 0.47 0.07 0.57 Sense of 0.54 0.20 0.58 well-being Dyspnea 0.30 0.05 0.74 Cluster 2 Depression 0.81 0.19 0.23 Anxiety 0.81 0.12 0.21 HCA ¼ hierarchical cluster analysis. First cluster explained 43% and second cluster explained 81% of the total variation. Cluster 1 would be split because of the smallest proportion of variation (43%).

Vol. 44 No. 1 July 2012 Symptom Clusters in Advanced Cancer 27 Table 5 Final Three Symptom Clusters at Baseline Determined Using HCA With Centroid Cluster Algorithm Cluster and Symptoms Own Cluster R 2 Next 1 R 2 Cluster Own Cluster 1 R 2 Next Cluster Cluster 1 Fatigue 0.67 0.23 0.42 Drowsiness 0.65 0.22 0.45 Dyspnea 0.49 0.10 0.58 Cluster 2 Depression 0.81 0.17 0.23 Anxiety 0.81 0.09 0.21 Cluster 3 Pain 0.42 0.12 0.66 Nausea 0.49 0.11 0.57 Poor appetite 0.58 0.16 0.49 Sense of well-being 0.59 0.24 0.53 HCA ¼ hierarchical cluster analysis. One cluster explained 39%, two clusters explained 51%, and three clusters explained 61% of the total variation. Potential one-cluster, two-cluster, and threecluster solutions explained 38.7%, 51.3%, and 61.1% of the total variance, respectively. The three centroid components (clusters) accounted for the highest total variability in the nine ESAS symptoms and were retained as the final solution (Table 6). Comparing Temporal Pattern of Symptom Clusters Between Different Analytical Methods Symptom clusters derived over time using PCA, EFA, and HCA are listed in Table 7. At the first week follow-up, the same three symptom clusters were identified using PCA and HCA and a similar two-cluster solution resulted from EFA. The first symptom cluster extracted unanimously using all three analytical methods Fatigue Drowsiness Shortness of Breath Pain Nausea Poor Appetite Sense of Well-being Depression Anxiety 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 Proportion of Variance Explained Fig. 1. PROC TREE procedure-generated dendrogram displaying three-cluster solution and cluster hierarchy. More similar symptoms were joined together earlier. included depression, anxiety, and a poor sense of well-being. The second symptom cluster identified through both PCA and HCA comprised fatigue, drowsiness, and dyspnea, and the third symptom cluster determined by PCA and HCA comprised pain, nausea, and poor appetite. Interestingly, the second EFA symptom cluster exactly encompassed the symptoms observed in the second and third clusters identified using the other two methods. Symptom clusters extracted at 2 varied depending on the type of statistical method used. All three methods extracted the cluster comprising depression, anxiety, and poor sense of well-being. With PCA, two other clusters, composed of fatigue, drowsiness, and dyspnea or pain, nausea and poor appetite, were identified. Only one other cluster was found using EFA and contained the collection of seven symptoms in the second and third PCA clusters. Similarly, HCA indicated a second cluster that included all these symptoms except dyspnea. At 4, the two symptom clusters identified using PCA, EFA, and HCA were almost in complete agreement. The cluster composed of anxiety, depression, and poor sense of wellbeing was consistent for all three methods. The other cluster determined using PCA and EFA contained pain, fatigue, drowsiness, nausea, poor appetite, and dyspnea. The second cluster defined through HCA included all these symptoms but pain. Again at 8, the two symptom clusters extracted were analogous among the three analytical methods used. All three methods agreed on the presence of a symptom cluster comprising depression, anxiety, pain, and poor sense of well-being. PCA and EFA both identified a second symptom cluster involving fatigue, drowsiness, nausea, poor appetite, and dyspnea. The similar finding from HCA included all these symptoms excluding dyspnea. The two symptom clusters identified using each of the three analytical methods at 12 resembled each other to a lesser degree. The first HCA cluster comprised fatigue, drowsiness, and pain, whereas the second contained depression, anxiety, nausea, poor appetite, and a poor sense of well-being. PCA findings were similar with the exception of an additional symptom of dyspnea in the second cluster. With EFA, the symptoms nausea, poor appetite, dyspnea, and poor well-being associated more

28 Chen et al. Vol. 44 No. 1 July 2012 Number of Clusters Table 6 Total and Proportion of Variation at Baseline Explained by Varying Number of Clusters Using HCA Total Variation Explained by Clusters Proportion of Variation Explained by Clusters closely with fatigue, drowsiness, and pain. The remaining three symptoms of fatigue, drowsiness, and pain constituted the second cluster finding using EFA. Comparing Temporal Pattern of Symptom Clusters Within the Same Analytical Method When comparing symptom clusters over time using the same method, clusters remained fairly constant for PCA. For instance, well-being was the only symptom that shifted into a different cluster at s 1 and 2, with the latter two time points producing three identical clusters. From 4 onward, two symptom clusters instead of three emerged. Although there was more shifting among symptoms from 4 onward, certain symptoms consistently clustered together over the whole assessment period, including depression-anxiety, fatigue-drowsiness, and nausea-poor appetite. EFA did not produce many temporal changes in cluster composition, with identical clusters derived at s 1, 2, and 4, and at baseline and 12. Throughout the whole assessment period, the anxiety-depression and fatiguedrowsiness-nausea-dyspnea-poor appetite clusters remained stable. Pain was the only symptom that shifted between the two clusters. HCA did Minimum Proportion Explained by a Cluster MinimumR 2 for a Variable Maximum 1 R 2 Ratio for a Variable 1 3.48 0.39 0.39 0.27 2 4.61 0.51 0.43 0.30 0.74 3 5.50 0.61 0.52 0.42 0.66 HCA ¼ hierarchical cluster analysis. Bolded values indicate the proportion of variability in the nine symptoms explained by extracting one, two, or three clusters. A three-cluster finding is favorable as it accounts for the most total variation (61%). not produce any identical clusters at any assessment time points. It also was the only method that had one symptom excluded from all clusters from 2 onward. Although the composition of all symptom clusters derived by HCA changed over time, symptoms of anxiety-depression, fatigue-drowsiness, and nausea-poor appetite consistently remained together, whereas other symptoms shifted between clusters. Consistency of Symptom Cluster Findings Using PCA, EFA, and HCA Expectedly, variations in symptom cluster number and composition extracted using the three different statistical methods were repeatedly observed in this longitudinal study. However, also notable were core symptoms within clusters that consistently occurred in conjunction over time. Symptoms such as fatigue and drowsiness or anxiety and depression unfailingly presented together at each time point regardless of the analytical method used. Temporal Pattern of Symptom Clusters in Responders vs. Nonresponders Using PCA, symptom clusters observed in responders and nonresponders were fairly consistent, with shifting of no more than three Table 7 Symptom Clusters Identified at Baseline and s Using Three Statistical Methods Baseline One- Two- Four- Eight- 12- Follow- Up Symptom PCA EFA HCA PCA EFA HCA PCA EFA HCA PCA EFA HCA PCA EFA HCA PCA EFA HCA Depression D D D D D D D D D D D D D D D D D D Anxiety D D D D D D D D D D D D D D D D D D Fatigue Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Drowsiness Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Pain X Ο X X Ο X X Ο Ο Ο Ο d D D D Ο Ο Ο Nausea X Ο X X Ο X X Ο Ο Ο Ο Ο Ο Ο Ο D Ο D Poor appetite X Ο X X Ο X X Ο Ο Ο Ο Ο Ο Ο Ο D Ο D Dyspnea Ο Ο Ο Ο Ο Ο Ο Ο d Ο Ο Ο Ο Ο d D Ο d Poor well-being X Ο X D D D D D D D D D D D D D Ο D PCA ¼ principal component analysis; EFA ¼ exploratory factor analysis; HCA ¼ hierarchical cluster analysis. Symptoms with corresponding symbols indicate they were in the same cluster. Dash indicates the symptom was not present in any clusters.

Vol. 44 No. 1 July 2012 Symptom Clusters in Advanced Cancer 29 symptoms from baseline to the 2 follow-up. Cluster compositions for each subgroup began to significantly diverge from 4 onwards. At 4, although clusters in the nonresponder population remained comparable to those identified at earlier time points, notably different clusters were extracted in the responder subgroup. At s 8 and 12 after treatment, symptom clusters in both subgroups were significantly different from those observed at baseline. PCA findings are detailed in Appendix 1. EFA results illustrated the most striking contrast between symptom clusters in responders and nonresponders. As detailed in Appendix 2, two symptom clusters were present at baseline in both subgroups. Minor changes in cluster constituents were observed from baseline to 4 for both populations, with slightly more symptoms shifting in responder symptom clusters. However, remarkably, in s 8 and 12, whereas clusters in the nonresponder group continued to remain quite consistent, no symptom clusters were observed in the responder subgroup. Symptom clusters identified using HCA in responders and nonresponders presented in a similar manner. With the exception of symptom clusters in responders at 1, clusters in both subgroups were fairly stable from baseline to the two-week follow-up. Cluster constituents for responders and nonresponders began shifting significantly in different manners from 4 onwards. The clusters results are summarized in Appendix 3. Despite the varying statistical analyses used, a common trend in the temporal pattern of symptom clusters emerged in responders and nonresponders. At baseline, cluster compositions were nearly identical between the two subgroups for each method. From baseline to at least 2, relatively minimal changes were observed in symptom cluster composition for both responders and nonresponders. However, by the eight-week follow-up, notable symptom cluster disparities were observed between the two subgroups. The findings of both EFA and HCA demonstrated symptom clusters in the responders subgroup deviated less over time from the clusters determined at baseline. At each time point, the majority of the patient sample consisted of nonresponders; therefore, the temporal pattern of symptom clusters in the general sample population largely resembled that of the nonresponder sample. Discussion To our knowledge, this is the first longitudinal study to compare symptom clusters identified by three different statistical methods (PCA, EFA, and HCA) in advanced cancer patients. The large sample size of the original study allowed for further analysis through dividing the study patients into two subgroups: responders and nonresponders to radiation treatment. Uniquely, symptom clusters were extracted through the same process for each subgroup and compared at baseline and each follow-up time point. The present study reanalyzed the same data set from our previous study to eliminate extraneous factors that may affect symptom cluster results, including varying sample population and assessment tools. All three statistical methods (PCA, EFA, and HCA) used in this study are exploratory and descriptive. The present study revealed stronger correlations between the symptom cluster finding of PCA and HCA. The only significant disparity in PCA and HCA cluster results was observed at the 2 follow-up. Although the findings of PCA more frequently supported that of HCA and vice versa, precautions should be taken in simply disregarding the outcomes of EFA. Discrepancies between EFA and PCA or HCA generally involved disagreements in the partitioning of a similar group of symptoms. EFA had a tendency to group more symptoms together into a larger cluster, whereas PCA and HCA divided the same symptoms into two separate clusters. Despite correlations in symptom clusters extracted among the three methods, the lack of absolute consensus indicates that type of statistical analysis used directly influences the number and composition of derived symptom clusters. HCA and EFA are both conceptually suitable methods for clustering symptoms based on cooccurrence and relatedness. 13 PCA is a less favorable method because it does not recognize the occurrence of errors in the symptom clustering and may lead to overextraction. 18 EFA is the only method of the three examined in this study with the theoretical basis to propose an underlying factor that may account for the clustering of symptoms. However, cluster analysis could be useful clinically to identify subgroups of individuals who have distinctive profiles of symptoms, allowing clinicians to target specific interventions for each subgroup. In

30 Chen et al. Vol. 44 No. 1 July 2012 determining the optimal method for extracting useful symptom clusters, researchers should focus on recognizing statistical results that are most clinically meaningful. To be clinically significant, these symptom cluster results should occur commonly in patients and provide practical insights in symptom management. 9 Symptom cluster stability over time is also valuable for the development of effective symptom management interventions. However, previous studies have demonstrated symptom experience is dynamic and, as a result, will likely trigger different symptoms to cluster. 19,20 In examining the longitudinal patterns of symptom clusters in the total patient population,itwasevidentthatthecompositionof many clusters derived at baseline, regardless of statistical method, did change at followups.thismaymeanthatlatentfactorsdo,in fact, change over time and in response to treatment. Within the responder and nonresponder subgroups, variations in symptom cluster composition emerged over time. There were also evident discrepancies in symptom cluster temporal patterns between the two subgroups. Symptoms in a cluster are interrelated; therefore, the resolution, improvement, or exacerbation of one symptom over time may explain the transforming symptom cluster results over time in responders vs. nonresponders. In the present study, radiotherapy treatment of pain can affect other symptoms in tandem, altering the composition of symptom clusters longitudinally. The decrease in analgesic intake of some responders also may play a role in the different pattern of symptom cluster development in the two subgroups, as pain medications often produce side effects such as fatigue, drowsiness, and nausea. Other possible factors affecting symptom clusters over time include long-term effects of treatment, response adjustments to persistent symptoms, and disease progression. Although changes in symptom cluster compositions were noted over time, an alternative explanation may be that at least some of the clustering noted may be attributed to a Type II error. Therefore, although certain symptom may be statistically deemed within a symptom cluster, it may not be clinically real. Kirkova and Walsh 21 addressed the dilemma of reaching a balance between the need to establish stability in practical symptom cluster findings and the inevitable modification of clusters over time. They suggested a minimum of 75% of symptoms in a cluster, including the most prevalent symptom, are required for the symptom cluster to be considered present and reliable longitudinally. Because of the low number of symptoms assessed in the present study, rarely did any clusters identified at baseline retain 75% of its symptoms throughout all the follow-up time points. Additional future longitudinal studies are required to investigate the presentation of symptom clusters over time. The major limitation of our present study was the use of the nine-item ESAS as the assessment tool in collecting symptom data. It is likely that the heterogeneous sample of advanced cancer patients studied experienced a wide range of symptoms that was not comprehensively captured using the ESAS alone. A representative range of symptoms is crucial for the context and validity of symptom cluster research, as the omission of symptoms may result in underidentification of clusters. The findings of our study should be interpreted with consideration. Our findings draw attention to one barrier in cancer symptom cluster research associated with the inconsistency in analytical methods used among previously published studies. This gap prevents justified comparisons among study results. Additional areas that may help advance symptom cluster research include examining relationships between symptoms in a cluster to determine if they are related through a shared common variance, etiology, and/or whether they produce synergistic effects on patient outcomes. Determination of the optimal approach to evaluate the underlying molecular mechanisms for symptom clusters will inform whether certain symptoms in a cluster should be targeted to alleviate the other symptoms (i.e., pain causes sleep disturbance resulting in fatigue), or whether it is best to treat all symptoms simultaneously. Further considerations should include comparisons of methodological approaches in identifying symptom clusters, including the use of symptom severity scores vs. symptom frequency scores to create symptom clusters, or using generic- vs. disease-specific assessment tools. 22 In a collective effort to assemble consistent and comparable results when determining symptom clusters from a pool of symptoms, clinicians and statisticians should come to terms with one optimal statistical method that should

Vol. 44 No. 1 July 2012 Symptom Clusters in Advanced Cancer 31 be used. Implementing a constant analytical method across all relevant future symptom cluster research is an imperative step in producing comparable clinically significant findings that may prove useful for bettering symptom diagnosis, treatment, and management among cancer patients. Disclosures and Acknowledgments This study was supported by the Michael and Karyn Goldstein Cancer Research Fund. The authors have no conflicts of interest to disclose. The authors thank Stacy Yuen and Kristina Facchini for their administrative assistance. References 1. Dodd MJ, Miaskowski C, Paul SM. Symptom clusters and their effect on the functional status of patients with cancer. Oncol Nurs Forum 2001;28: 465e470. 2. Kim HJ, McGuire DB, Tulman L, Barsevick AM. Symptom clusters: concept analysis and clinical implications for cancer nursing. Cancer Nurs 2005; 28:270e282. quiz 283e284. 3. Kwekkeboom KL, Cherwin CH, Lee JW, Wanta B. Mind-body treatments for the pain-fatigue-sleep disturbance symptom cluster in persons with cancer. J Pain Symptom Manage 2010;39:126e138. 4. Paice JA. Assessment of symptom clusters in people with cancer. J Natl Cancer Inst Monogr 2004;32:98e102. 5. Henoch I, Ploner A, Tishelman C. Increasing stringency in symptom cluster research: a methodological exploration of symptom clusters in patients with inoperable lung cancer. Oncol Nurs Forum 2009;36:E282eE292. 6. Maliski SL, Kwan L, Elashoff D, Litwin MS. Symptom clusters related to treatment for prostate cancer. Oncol Nurs Forum 2008;35:786e793. 7. Gleason JF Jr, Case D, Rapp SR, et al. Symptom clusters in patients with newly-diagnosed brain tumors. J Support Oncol 2007;5:427e433. 436. 8. Fan G, Hadi S, Chow E. Symptom clusters in patients with advanced-stage cancer referred for palliative radiation therapy in an outpatient setting. Support Cancer Ther 2007;4:157e162. 9. Xiao C. The state of science in the study of cancer symptom clusters. Eur J Oncol Nurs 2010;14: 417e434. 10. Bruera E, Kuehn N, Miller MJ, Selmser P, Macmillan K. The Edmonton Symptom Assessment System (ESAS): a simple method for the assessment of palliative care patients. J Palliat Care 1991;7:6e9. 11. Chang VT, Hwang SS, Feuerman M. Validation of the Edmonton Symptom Assessment Scale. Cancer 2000;88:2164e2171. 12. Moro C, Brunelli C, Miccinesi G, et al. Edmonton Symptom Assessment Scale: Italian validation in two palliative care settings. Support Care Cancer 2006;14:30e37. 13. Skerman HM, Yates PM, Battistutta D. Multivariate methods to identify cancer-related symptom clusters. Res Nurs Health 2009;32:345e360. 14. Kim HJ, Abraham IL. Statistical approaches to modeling symptom clusters in cancer patients. Cancer Nurs 2008;31:E1eE10. 15. Kim J, Muller CW. Introduction to factor analysis: What it is and how to do it. Beverly Hills, CA: Sage Publications, Inc, 1978. 16. Barsevick AM, Whitmer K, Nail LM, Beck SL, Dudley WN. Symptom cluster research: conceptual, design, measurement, and analysis issues. J Pain Symptom Manage 2006;31:85e95. 17. Chow E, Wu JS, Hoskin P, et al. International consensus on palliative radiotherapy endpoints for future clinical trials in bone metastases. Radiother Oncol 2002;64:275e280. 18. Velicer WF, Jackson DN. Component analysis versus common factor analysis: some issues in selecting an appropriate procedure. Multivariate Behav Res 1990;25:1e28. 19. Donovan KA, Jacobsen PB, Andrykowski MA, et al. Course of fatigue in women receiving chemotherapy and/or radiotherapy for early stage breast cancer. J Pain Symptom Manage 2004;28:373e380. 20. Glaus A, Boehme C, Thurlimann B, et al. Fatigue and menopausal symptoms in women with breast cancer undergoing hormonal cancer treatment. Ann Oncol 2006;17:801e806. 21. Kirkova J, Walsh D. Cancer symptom clustersda dynamic construct. Support Care Cancer 2007;15: 1011e1013. 22. Miaskowski C, Aouizerat BE, Dodd M, et al. Conceptual issues in symptom clusters research and their implications for quality-of-life assessment in patients with cancer. J Natl Cancer Inst Monogr 2007;37:39e46.

32 Chen et al. Vol. 44 No. 1 July 2012 Appendix 1 Symptom Clusters in Responder vs. Nonresponder Subgroups Using PCA Baseline One- Two- Four- Eight- 12- Method Symptom NR R NR R NR R NR R NR R NR R PCA Depression D D D D D D D D D D D D Anxiety D D D D D D D D D D D D Fatigue Ο Ο Ο Ο Ο Ο Ο Ο Ο D Ο D Drowsiness Ο Ο Ο Ο Ο Ο Ο Ο Ο D Ο D Pain X X X X X X d X D d Ο Ο Nausea X X X X X X Ο X Ο Ο X D Poor appetite X X X Ο X Ο Ο Ο D Ο X D Dyspnea Ο Ο Ο Ο Ο Ο D Ο d Ο X Ο Poor well-being X X D D D Ο D D D D D D PCA ¼ principal component analysis; NR ¼ nonresponder; R ¼ responder. Symptoms with corresponding symbols indicate they were in the same cluster. Dash indicates the symptom was not present in any clusters. Appendix 2 Symptom Clusters in Responder vs. Nonresponder Subgroups Using EFA Baseline One- Two- Four- Eight- 12- Method Symptom NR R NR R NR R NR R NR R NR R EFA Depression D D D D D D D D D d D d Anxiety D D D D D D D D D d D d Fatigue Ο Ο Ο Ο Ο Ο Ο Ο Ο d Ο d Drowsiness Ο Ο Ο Ο Ο Ο Ο Ο Ο d Ο d Pain Ο Ο Ο Ο Ο D Ο D Ο d Ο d Nausea Ο Ο Ο D Ο Ο Ο Ο Ο d Ο d Poor appetite Ο Ο Ο Ο Ο Ο Ο D Ο d Ο d Dyspnea Ο Ο Ο Ο Ο Ο Ο Ο D d Ο d Poor well-being Ο Ο D D D Ο D D D d Ο d EFA ¼ exploratory factor analysis; NR ¼ nonresponder; R ¼ responder. Symptoms with corresponding symbols indicate they were in the same cluster. Dash indicates the symptom was not present in any clusters. Appendix 3 Symptom Clusters in Responder vs. Nonresponder Subgroups Using HCASQ Baseline One- Two- Four- Eight- 12- Method Symptom NR R NR R NR R NR R NR R NR R HCA Depression D D D D D D D D D D D D Anxiety D D D D D D D D D D D D Fatigue Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο D Drowsiness Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο Ο D Pain X X X d X D d X X D Ο Ο Nausea X X X D X X Ο X Ο X X d Poor appetite X X X Ο X X Ο Ο X Ο D D Dyspnea Ο Ο Ο Ο Ο Ο D D D X X Ο Poor well-being D X D D D X D D D Ο Ο D HCA ¼ hierarchical cluster analysis; NR ¼ nonresponder; R ¼ responder. Symptoms with corresponding symbols indicate they were in the same cluster. Dash indicates the symptom was not present in any clusters.