|Home | About | Journals | Submit | Contact Us | Français|
Although aggressive medical treatment protocols have led to 80% 5-year survival rates for most childhood cancers, many long-term survivors experience multiple troubling symptoms. Using data from 100 adult survivors of childhood cancers (ACC-survivors), we used latent variable mixture modeling to generate unique subgroups of survivors based on their experiences with a cluster of eight symptoms: lack of energy, worry, pain, difficulty sleeping, feeling irritable, feeling nervous, difficulty concentrating, and feeling sad (as measured by the Memorial Symptom Assessment Scale). We also examined factors that were likely to predict subgroup membership (chronic health conditions, health-promoting lifestyle, and demographic variables) and determined the extent to which satisfaction with quality of life (QoL) varied across the subgroups. The final mixture model included three subgroups of ACC-survivors: high symptoms (HS) (n=21), moderate symptoms (MS) (n=45), and low symptoms (LS) (n=34). ACC-survivors who reported at least one chronic health condition were six times as likely to be classified in the HS subgroup as compared to the LS subgroup. Mean health-promoting lifestyle scores were lowest in the HS subgroup and highest in the LS subgroup. Differences in QoL among the subgroups were statistically significant, thus validating that the subgroups were characterized uniquely for identifying those symptoms with highest life impact. To our knowledge, we are the first to identify distinct subgroups of ACC-survivors differentiated by symptom cluster experience profiles. The findings warrant additional research to confirm the subgroup-specific symptom cluster experience profiles in larger studies of ACC-survivors.
Prior to 1970, many children who were diagnosed with cancer had slim chances of being cured . Since then, 5-year survival rates for most childhood cancers have steadily increased to nearly 80% . Today, nearly one in 840 adults between the ages of 20 and 44 is a childhood cancer survivor . Among nearly 10,000 participants in the Childhood Cancer Survivor Study , 10%–23% of adult survivors of childhood cancers (ACC-survivors) reported moderate to extreme pain [5, 6]; 16%–40% significant fatigue [7–9]; 12%–16% problems sleeping [7, 8]; 5%–17% psychological distress (i.e., anxiety, depression, somatic distress) [5, 10–12]; and 14% difficulty concentrating . Most investigators, however, have studied all the symptoms separately, so whether or how the symptoms cluster in ACC-survivors is unknown.
Although emerging findings suggest that when symptoms cluster together, they act synergistically and magnify the impact of individual symptoms on QoL [13–19], studies of multiple symptoms and symptom clusters in ACC-survivors remain sparse. In ACC-survivors, Meeske and colleagues  found a clustering of pain, fatigue, sleep disturbance, depression, and neurocognitive symptoms. In these ACC-survivors, the combination of fatigue and depression was associated with more life impact (lower QoL) than either fatigue or depression alone. Of alarming note is that unrelieved symptoms in ACC-survivors have been associated with potentially fatal consequences, i.e., the combination of unrelieved pain and psychological distress was significantly associated with suicidal ideation or past suicidal attempt .
Although little is known about differences in what contributes to symptom clustering or the consequences of symptom clustering in ACC-survivors, these questions can be probed in illuminating ways by characterizing unique subgroups according to experiences with multiple symptoms. In three studies, other investigators [22–24] have demonstrated the feasibility of this approach by using cluster analysis with adult-onset oncology patients who were receiving cancer treatments. In all three studies, they found subgroups that differed according to symptom burden. In support of distinct subgroupings, differences in QoL among the subgroups were statistically significant, with the lowest QoL scores in the subgroup with the highest symptom burden. With further examination using subgroup approaches, ACC-survivors most in need of symptom relief can be identified and then in future studies, appropriate novel interventions can targeted way and tested according to explicitly defined subgroups.
We used latent variable mixture modeling to generate a model of subgroup-specific symptom cluster experience profiles in ACC-survivors. In this exploratory study, we had two aims: First, we identified subgroups of ACC-survivors using self reports of frequency, severity, and distress ratings for eight symptoms (lack of energy, worry, pain, difficulty sleeping, feeling irritable, feeling nervous, difficulty concentrating, and feeling sad) and factors that were likely to predict subgroup membership (chronic health conditions, health-promoting lifestyle, and demographic variables). Second, we determined the extent to which satisfaction with quality of life (QoL) varied across the subgroups.
Over a four-month period in 2006, a convenience sample of 100 ACC-survivors from throughout the United States were recruited from a cohort of 117 who had completed a previous web-based survey on physical activity and who agreed to be contacted again for future studies. In the physical activity study, participants were recruited through Internet-based advertisements, flyers in specialty cancer survivor clinics, email messages to cancer camp alumni, and word of mouth. Inclusion criteria for the original study were: 1) self-reported diagnosis of cancer, 2) age less than 21 years at time of diagnosis, 3) at least two years beyond completion of cancer therapy, 4) between 18 and 39 years old at study entry, 5) English speaking, reading, and writing, and 6) access to the Internet. After receiving Institutional Review Board approval, we sent each potential participant an email message that included a link to the new survey website, instructions on how to complete the survey questions, and an identification number from the previous study. The survey website included an informed consent document that participants submitted before beginning the survey questions.
The multidimensional 32-item Memorial Symptom Assessment Scale (MSAS) was used to measure frequency, severity, and distress for symptoms experienced during the previous week. Participants used four-point categorical scales to rate frequency (1=rarely to 4=almost constantly) and severity (1=slight to 4=very severe) and a 5-point categorical scale to rate distress (1=not at all to 5=very much). The MSAS has had good psychometric properties in numerous studies with adult cancer patients [27, 28], adult patients with diagnoses other than cancer [29–31], and in children with cancer undergoing active treatments [32, 33].
The Health Promoting Lifestyle Profile II (HPLP-II)  is a 52-item instrument that contains items representing health responsibility, physical activity, nutrition, supportive interpersonal relations, spiritual growth, and functional stress management behaviors. Participants rated each behavior using a 4-point categorical scale (1=never; 4=routinely). Higher total scores indicate higher levels of a health promoting lifestyle (possible range 52–208). The HPLP-II has had good psychometric properties in other studies with ACC-survivors [35, 36]. Internal consistency reliability for our study was 0.95.
We included questions on age, education, income, marital status, race, ethnicity, cancer history, and types of cancer treatments (chemotherapy, radiation, and surgery). Using three open-ended questions, we asked respondents to list any: 1) permanent after-effects from treatment; 2) serious medical conditions, other than cancer, for which they were/are under a health care provider’s care; and 3) other medical conditions, disabilities, and chronic illnesses. Using the method described by Oeffinger and colleagues , we graded the responses from these three questions according to the Common Terminology Criteria for Adverse Events (version 3) : mild (grade 1), moderate (grade 2), severe (grade 3), or life-threatening or disabling (grade 4).
The Ferrans and Powers Quality of Life Index (QLI): Cancer III Version [39, 40] was used to measure QoL life in terms of life satisfaction. Quality of life was measured overall and in four major life domains: health and physical functioning, social and economic, psychological/ spiritual, and family. Scores range from 0 to 30 with higher scores indicating a better QoL. The QLI has good psychometric properties and has been used in approximately 200 published studies, 28 focusing on cancer survivors [39, 40]. For our study, the internal consistency reliability was 0.94 for the total instrument.
First, using Stata SE Version 9.2, we described the sample characteristics using means and standard deviations for continuous variables and frequencies and percents for categorical variables. Next, to construct the latent variable mixture models, we used items from the MSAS that corresponded to the most prevalent symptoms in our sample: lack of energy, worry, pain, difficulty sleeping, feeling irritable, feeling nervous, difficulty concentrating, and feeling sad. For each symptom, ACC-survivors rated frequency, severity, and distress. For analyses, the scales were collapsed into two categories: 1) low frequency (no symptom/rarely/occasionally) or high frequency (frequently/almost constantly; 2) low severity (no symptom/slight) or high severity (moderate/severe/very severe); and, 3) low distress (no symptom/not at all/a little bit/somewhat) or high distress (quite a bit/very much).
Using Mplus Version 4.2 , through an iterative process to find the most parsimonious model , we fit a series of two- to four-subgroup latent variable mixture models to the data. The primary assumption with latent variable mixture models is that the data consist of different groups of individuals, but group membership is not observed . The groups are known as latent classes, which we refer to as subgroups. The modeling process included several steps that were incorporated within one statistical framework: 1) fitting preliminary models to determine the correct number of subgroups; 2) examining covariates that influenced the probability of subgroup membership; and 3) exploring the extent to which satisfaction with QoL varied across the subgroups.
In the first step, we fit preliminary models to determine the correct number of subgroups. In Figure 1, we show a diagram of the analytic model. Relationships between the categorical symptom items (see top box in Figure 1) and a categorical latent variable representing the subgroups (see middle circle in Figure 1) were explained by a multivariate logistic regression model. Model parameters included item parameters and subgroup probability parameters. Item parameters, which were specific to each subgroup, corresponded to the probability of an individual in that specific subgroup endorsing a specific category for a symptom item. Subgroup probability parameters specified the sample prevalence for each subgroup.
After fitting preliminary models, the second step involved including covariates that likely influenced the probability of subgroup membership. Using Stata, we explored between-subgroup differences (using one-way analysis of variance for continuous variables and Chi square for categorical variables) in each of the following covariates: health-promoting lifestyle, presence of at least one chronic health condition, and the characteristics of age, age at cancer diagnosis, number of years since completing cancer treatments, male gender, Hispanic ethnicity, married/partnered, had children, education, income, cancer type, and cancer treatment. After identifying between-subgroup differences in covariates that were significant at P < 0.05 (since this was a model-generating exploratory analysis, we did not adjust for multiple comparisons), we again used Mplus to fit additional models that included the covariates. Within the Mplus modeling framework, this second step was a multinomial logistic regression of the categorical latent subgroup variable on the covariates (see left box in Figure 1).
Finally in the third step, we fit models to explore the extent to which satisfaction with QoL varied with subgroup membership. Within the Mplus modeling framework, satisfaction with QoL was regressed on the categorical latent subgroup variable using linear regression (see right box in Figure 1).
During each step of the modeling process, model estimation was carried out by maximizing the fit of estimated cell frequency counts and means with the actual (observed) cell frequency counts and means found in the sample data. We used the Bayesian Information Criterion (BIC) to assist in identifying the correct number of subgroups. The BIC is based on the log likelihood of a fitted model and includes a penalty for the number of model parameters and sample size. Models with more parameters tend to have a better fit; whereas models with fewer parameters (more parsimonious models) usually have a poorer fit. The BIC optimizes the choice between a complex model with many parameters and good fit and a more parsimonious model with lesser fit. Successive models were compared, and the lowest value of BIC indicated the best fitting model. Importantly, models were evaluated for theoretical interpretability before deciding on the best fitting model.
Over two-thirds (71%) of the 100 respondents were female; 39% were married and 14% had children. At the time of survey completion, the respondents ranged in age from 18 to 37 years old (median 23); almost all (96%) were white; 9% were of Hispanic ethnicity; and the majority of the ACC-survivors (89%) reported completing at least some college. Most were survivors of leukemia (39%) and lymphoma (25%); 10% had survived malignant brain tumors, 11% malignant bone tumors, and the remaining 15% other types of childhood cancers. Age at cancer diagnosis ranged from less than 1 to 20 years old (median 11.5). Years since completion of cancer therapy ranged from 2 to 32 (median 10). Almost all respondents (95%) were treated with chemotherapy, and 51% received both radiation and chemotherapy. Over half (57%) of the sample reported at least one chronic health condition, with 45% reporting at least one Grade 1 chronic health condition; 12% at least one Grade 2; 5% at least one Grade 3; and 2% at least one Grade 4 chronic health condition. As shown in Table 1, the majority of self-reported chronic health conditions were cardiac (e.g., hypertension, cardiomyopathy), endocrine (e.g., hypothyroidism, diabetes), musculoskeletal (e.g., avascular necrosis, scoliosis) neurologic (e.g., stroke, seizures) and pulmonary/upper respiratory (e.g., diminished lung capacity, pulmonary fibrosis).
ACC-survivors reported a range of 1 to 26 symptoms (median 8). The most prevalent symptoms were lack of energy (78%) and worry (77%). At least 50% of the sample reported one or more of the following symptoms: lack of energy, worry, pain, difficulty sleeping, irritability, nervousness, difficulty concentrating, and feeling sad. The mean health promoting lifestyle (HPLP-II) score was 135.6 (standard deviation [SD] 24.1), with total scores ranging from 88 to 201. Total scores on satisfaction with QoL, (QLI) ranged from 10.9 to 30 (mean 23.3; SD 4.8).
In models without covariates, significant improvements in model fit were observed for up to three subgroups (two-subgroup BIC= 2787.1; three-subgroup BIC= 2772.2; four-subgroup BIC= 2799.1). Between subgroup differences were significant for two covariates: presence of at least one chronic health condition (Χ2 [2, n =100] = 7.26, P = 0.027) and mean health-promoting lifestyle scores (F(2, 97) = 12.41, P < 0.0001). Both of these covariates contributed to the classification and decreased the three-subgroup BIC further to 2712.2. For the final three-subgroup model, average posterior probabilities (how well each participant fit his or her assigned subgroup) ranged from 97% to 99%, indicating a high degree of confidence in subgroup assignment and model fit.
Figure 2 indicates symptom characteristic probability plots for the three-subgroup model. The x-axis includes the three characteristics (high frequency, high severity, and high distress) for each of the eight symptoms. The y-axis represents the probability of endorsing each symptom characteristic, given membership in a particular subgroup. The three subgroups are represented by red, green, and blue lines. The top profile (red line) is the “high-symptom” (HS) subgroup. When compared to the other two subgroups, the 21 respondents in this subgroup had the highest probabilities of high frequency, severity, and distress ratings for all symptoms. In contrast to the HS subgroup, the “low-symptom” (LS) subgroup (represented by the bottom profile [blue line] in Figure 2) included 34 respondents who had the lowest probabilities of high frequency, severity, and distress ratings. The 45 “moderate-symptom” (MS) subgroup members (represented by the middle profile [green line] in Figure 2) were similar to LS subgroup members in relation to symptom distress ratings; both subgroups had nearly zero probabilities of rating high distress for each of the eight symptom items. Moderate-symptom subgroup members, however, had higher symptom severity conditional probabilities than LS subgroup members for all eight symptoms.
In Table 2, along with the estimated conditional probabilities for each symptom (also graphically depicted in Figure 2), we include the sub-group sample frequencies and means for the covariates and the estimated conditional QLI means. ACC-survivors in the HS subgroup had the lowest mean QLI scores, and differences among the three subgroups in QLI means scores were statistically significant (P<0.001). ACC-survivors who reported at least one chronic health condition were six times as likely to be classified in the HS subgroup instead of the LS subgroup (OR 6.1, P = 0.01); however, there were no significant differences in reported chronic health conditions between the MS and LS groups (OR 1.6, P = 0.37). Mean HPLP-II scores were lowest in the HS subgroup and highest in the LS subgroup. Compared to the LS subgroup, for every one standard deviation increase in HPLP-II scores, the odds of being classified in the HS subgroup decreased by 75% (OR 0.25, P = 0.002) and the odds of being classified in the MS subgroup decreased by 54% (OR 0.46, P = 0.032). There were no differences among subgroups in other demographic or health characteristics (see Table 3).
To our knowledge, we are the first to identify distinct subgroups of ACC-survivors differentiated by unique symptom cluster experience profiles. Despite being off cancer treatments for an average of 11 years prior to completing the survey, symptom frequency, severity, and distress ratings in the HS subgroup closely mirrored symptom characteristics in another study using the MSAS , in which adult oncology patients were much older (average age 56) and sicker (56% had metastatic cancers and 56% were inpatients) than the ACC-survivors in our study. In our study, it might be speculated that physiological disease mechanisms or underlying distress disorders  (e.g., depression, anxiety, and post-traumatic stress disorder)  perpetuate the high symptom burden in the HS subgroup. Frequent and severe pain also characterized the HS group. Additional information about the pain and past treatment history are necessary to determine if the ACC-survivors in our study are experiencing persistent neuropathic pain from chemotherapeutic drugs such as vincristine and cisplatin [47, 48]. Although more studies are needed, pain and other symptoms (e.g., fatigue, difficulty concentrating, anxiety, and depression) may also share common underlying biological mechanisms as part of a cytokine-immunologic model proposed by Cleeland and others [49–51], in which co-occurring cancer symptoms may be mediated by pro-inflammatory cytokines (e.g., interleukin-1 [IL-1], IL-6, tumor necrosis factor-alpha) acting on the peripheral and central nervous systems.
Differences among the three subgroups in QoL scores (at least four points) can be deemed clinically meaningful. In studies assessing QoL change over time, a difference of 2–3 points in the total QLI score was clinically significant and associated with significant improvement in self-image, physical symptoms, and general health [39, 52–54]. Not surprisingly, ACC-survivors in the HS subgroup had the lowest mean QLI scores. From approximately 200 studies using the QLI, only one sample of persons with chronic fatigue syndrome had lower mean QLI scores (12.6) than our HS subgroup ACC-survivors (17.0). When compared with a sample of healthy young adults (mean QLI score 23.1) , ACC-survivors in our MS subgroup had an equivalent mean QLI score (23.1) and ACC-survivors in our LS group had a substantially higher mean QLI score (27.5). As other investigators have postulated, after experiencing the stress of cancer diagnoses and treatments, ACC-survivors may be satisfied with a lower health status than their healthy young adult counterparts (QoL response shift) [57–59] or they may have embraced their cancer experiences, finding new meaning in life and actually moving to a higher level of life satisfaction (post-traumatic growth) .
Since reporting a health-promoting lifestyle was associated with an increased likelihood of being classified in the MS or LS subgroup rather than the HS subgroup, perhaps health behaviors contributed to differences in symptom cluster profiles across the subgroups. Further research is warranted to test this hypothesis, but investigators in studies with survivors of adult-onset cancers have demonstrated that physical activity interventions reduced fatigue, increased happiness , and decreased depression and anxiety symptoms . Likewise, in other studies with survivors of childhood cancers, uncertainty  and worries  were associated with unhealthy lifestyles.
This exploratory, model-generating study had several limitations. Generalization of study findings is limited by the relatively small, predominately white, female, well-educated convenience sample of ACC-survivors who completed the internet survey. Rather than self-report, medical record abstraction for cancer history, cancer treatments, and chronic health conditions and objective health behavior measures such as pedometers or actimeters for physical activity could improve accuracy and minimize response bias inherent in self-report measures. The cross-sectional design limits interpretation of results regarding the reciprocal relationships among health-promoting behaviors, QoL and symptom cluster experiences. While health-promoting behaviors could have a beneficial effect on symptom reporting, experiencing worse symptoms could retard engaging in health-promoting behaviors. Likewise while symptoms might dampen the reporting of satisfaction with QoL, a lower overall QoL could make reporting of symptoms more negative.
Despite these limitations, the model of subgroup-specific symptom cluster experience profiles that we generated can now be tested in a larger stratified heterogeneous sample of childhood cancer survivors to explore further whether demographic, disease/treatment-related, and/or common underlying biologic mechanisms differentiate the subgroups. With further testing, the subgroup model can be a useful way to classify ACC-survivors and identify subgroups for testing targeted interventions aimed at improving QoL. For example, interventions might range from web-based programs for maintaining high level wellness and early detection of symptom cluster patterns in LS subgroups to intense symptom management interventions using behavioral and pharmacological interventions in HS subgroups with underlying chronic conditions.
This project was supported by: 1) a University of Illinois College of Nursing Dean’s Fund Grant, 2) Grant # P30 NR009014 Center for Reducing Risks in Vulnerable Populations (CRRVP) from the National Institute of Nursing Research, and 3) an Oncology Nursing Society Foundation Small Research Grant. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institute of Nursing Research or the National Institutes of Health.
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.