The search for causal forces that provide the coordination between symptoms is an alternative to the current approach to the study of symptom clusters. We consider the correlations between symptoms to be artifacts that are created through the action of underlying causal effects. From this perspective, if the relevant causal forces change, the symptoms display new patterns of correlations. Studies using factor analysis presume that a stable common-cause underlies the symptoms in any given factor. We, on the other hand, remain receptive to the possibility that different and even changing causal structures underlie the connections between symptoms. By continuing to examining the causal networks underpinning a variety of symptoms, we hope to eventually target interventions that effectively address coordinated sets of symptoms. Symptoms reflecting a common cause would be most effectively addressed by proper management of that common cause, and symptoms linked in a causal chain would be most effectively addressed by proper management of the symptom heading the causal chain.
Early studies on symptom clusters [1, 2] were silent regarding the stability of symptom coordinations, but recently Chow and colleagues  used factor analysis to show that symptoms loading on various factors changed over time. They found that sometimes anxiety and depression loaded by themselves, and at other times anxiety and depression loaded with well-being, fatigue, or drowsiness, but they did not model the causal foundations of these shifts. We found that the effects coordinating some symptoms were stable over time, while others were not. Specifically, we found that the effects of anxiety on depression, of drowsiness on tiredness, and of appetite on well-being were stable, while the effects of pain on appetite, anxiety on well-being, and depression on well-being were not stable. Based on these findings, some may consider that we identified three clusters: anxiety/depression, drowsiness/tiredness, and appetite/well-being. These three pairs of symptoms do not constitute symptom clusters in our view. Rather, they are simply pairs of symptoms in which changes in the first symptom consistently lead to changes in the second symptom at two points in time. We do not know whether these relationships would hold over a longer period of time.
This study is the first to provide quantitative evidence of a stable causal relationship between appetite and well-being, but this effect is not surprising. A number of research groups have reported distress associated with loss of appetite [22–25]. In addition to the physiological benefits of the intake of nutrients, clinicians have noted the social importance of appetite. Appetite makes it possible for patients to share meal times with family and friends in a manner that feels "normal".
A number of authors have shown that anxiety and depression are common among palliative patients [26–28]. Wilson and colleagues reported that 24% of the participants in their sample of palliative care patients met the criteria for at least one anxiety or depressive disorder . The causes of anxiety and depression in this population range from medical complications to psychological and existential concerns , and anxiety and depression are often accompanied by the somatic complaints associated with advanced disease. These complexities make it tempting to fall back on common sense by viewing anxiety and depression as merely "somehow similar." A preferable approach is to tame the complexity by incorporating it into an appropriately structured model. Our clinical observation of patients moving into depression pointed to anxiety as a source of depression, irrespective of the multiplicity and diversity of additional sources of both depression and anxiety. Our model's specification mirrors what we saw at the bedside, regardless of the additional and even unknown factors that influence both anxiety and depression among palliative patients, anxiety led to depression. An insignificant effect estimate could have spoken against our understanding, but the estimate for the effect of anxiety on depression was significant at both one month and one week before death, suggesting effective management of anxiety would help reduce depression. Depression's subsequent but changing effect on well-being suggests that the degree to which anxiety management will carry over to patients' perceptions of their well-being depends on the proximity to death, but given the potential for at least some improvement in well-being, this seems worthy of future investigation.
Our modeling of drowsiness and tiredness provides an instance where we can informatively follow a reviewer's suggestion to clearly differentiate between common sense and our model's specification. We modeled drowsiness as a cause of tiredness. From a common sense perspective, one may think of drowsiness and tiredness as synonyms for sleepiness, and hence interpret our model as saying sleepiness causes sleepiness. Drowsiness as measured by the ESAS was conceptualized as medication or disease-induced neurological interference which could occur throughout the day and not only at the transition from wake to sleeping, Tiredness as measured by the ESAS was conceptualized as the lack ability to engage in desired activities, even when fully awake. Patients are regularly reminded of these important distinctions when we ask them to complete the ESAS. Thus the effect of drowsiness on tiredness is more akin to an effect of mind on body, than to the common-sense understanding of drowsiness and tiredness as sleepiness.
Limitations and Strengths
This study is limited in that we had too few cases to be able to differentiate among types of cancer. Since type of cancer may influence symptom profile (e.g., breathlessness may be more common in lung cancer than in prostate cancer), there remains the possibility of finding even tighter symptom interconnections and simpler causal models for specific cancer types. The patient's type of cancer is not likely to change importantly in the last weeks of life, so even if we had included cancer diagnoses in our analysis, it would not have accounted for the changing patterns of effects observed in our models. In the future, we hope to investigate models of the symptoms displayed by different types of cancers via "stacked" or multi-group structural equation models.
The majority of our data were collected on an acute palliative unit or in hospice units in extended care centers. Thus the extent to which these findings generalize to other palliative care settings is unknown but could be addressed in future studies by collecting comparable data in home care and other hospice environments.
The two data collection points in this study were fairly close together. In a future study we plan to examine additional data collection points spanning a longer time period. Also, complete data were only available for only 82 of 140 potential participants and the symptoms that were assessed were limited to those measured by the ESAS. It is possible that results might differ with use of other assessment tools, or more cases. Each of the items included in the ESAS is far more complex and multifaceted than can be represented fully by a numerical rating scale. For example, recent developments in the assessment of advanced cancer pain suggest that in order to accurately assess pain, one must consider the mechanism of pain, the degree to which pain is a function of movement, related psychological distress, history of alcohol or drug addiction, and cognitive function [31, 32]. The theoretical needs of research with palliative patients, however, must be weighed against patient burden. The ideal data collection strategy for pain and other symptoms may require more effort than palliative patients can provide.
It is possible, but unlikely, that measurement of symptoms beyond those available in the ESAS would have altered the main findings of this study. New measures of variables that are causally down-stream from the variables in our model would by definition be incapable of influencing the variables in our model. New measures of causally up-stream variables that influence just the exogenous variables would have only more clearly specified the sources of the currently-free covariances between these variables. New variables influencing only specific endogenous variables would have replaced some of the modeled error variables and hence increased the proportion of explained variance in these variables but the current effect estimates would not change. Our primary concern was about whether symptoms that were common causes of two endogenous symptoms or one exogenous and one or more endogenous symptoms were missing because this would have rendered the model misspecified. While this is the most challenging kind of concern, the fact that both models fit with minimal model revision constitutes evidence that no such causal features are required. If such a symptom had been missed, the models would have failed to fit and would have provided covariance residuals diagnostically indicative of the location of the missed variable's effects but this did not happen, so this possibility seems unlikely in light of the available evidence.
Our above comments alluded to the detailed attention required in setting up an original structural equation model. This can be seen as a limitation of the method because it requires substantially greater effort and thoughtful engagement than is required by more exploratory methods. But encouraging thoughtful engagement with the relevant substantive variables can also be viewed as a great strength. The model examined in this study was constructed by several members of the study team over a period of several months, based on their clinical observation that as patients approached death, there were changing patterns in the patients' ESAS symptom profiles. We wondered about whether the patterns were real and if so, how the symptoms included in the ESAS were related to each other and to well-being. Structural equation modeling provided an opportunity to test ideas that came from our observations and discussions. From a treatment perspective, considerable advantage is gained by encouraging researchers to propose and test specific thoroughly-considered causal structures rather than stopping with the identification of correlated symptoms.
The decisions regarding which variables to model as exogenous or endogenous, and which specific effects to included or excluded, were not difficult in our case but we might have felt differently had our model failed. A strength of structural equation modeling is that it provides an opportunity to test ideas. The model may either fit the data, as it did in our case or not fit the data and be considered a "failed" model. Some researchers may be reluctant to expose their ideas to this kind of risk. The model test and diagnostics cannot detect all model specification problems  but the testing is strong enough that passed model tests (as in our models) provide some substantial reassurance unavailable with procedures like factor analysis and regression.
An additional methodological strength of our study design was that the data focus on fixed periods before death. This approach assisted in the identification of causal structures because the participants at each time point were more homogenous in their disease process than are typical palliative populations or participants recruited at the time of diagnosis or treatment.