eTable 1. Model Fit Comparison Among LPA Models With Different Number of Latent Profiles by Reporting Groups
eTable 2. Model Fit Comparison Between Multi-Group LPA Models With and Without Restriction
eFigure. PROMIS Measures for Children and Proxies
Customize your JAMA Network experience by selecting one or more topics from the list below.
Weaver MS, Jacobs SS, Withycombe JS, et al. Profile Comparison of Patient-Reported and Proxy-Reported Symptoms in Pediatric Patients With Cancer Receiving Chemotherapy. JAMA Netw Open. 2022;5(3):e221855. doi:10.1001/jamanetworkopen.2022.1855
To what extent do family caregiver or proxy reports on symptom profiles align with self-reports of children with cancer?
In this cohort study of 436 children with cancer and 431 proxies, latent profile analysis found 3 predominant profiles: high symptom groups, moderate symptom groups, and low symptom groups. Although proxies rated individual symptoms and function as lower than the children’s self-reports, the model results show PROMIS symptom measure profiles do not significantly differ between children and proxies.
Grouping child and proxy reports according to profile classifications may help bring the reporting closer to representing the lived reality of the child.
The variability in individual symptom and adverse event reporting between pediatric patient-reports and proxy-reports is widely reported. However, the question of whether symptom profiles based on reports from children with cancer and their caregivers are similar or disparate have not yet been studied.
To compare proxy symptom reports with patient self-reports to assess alignment.
Design, Setting, and Participants
A multicenter cohort study was conducted from October 2016 to December 2018 from data collected at 9 pediatric cancer centers. Participants were a convenience sample of family caregivers or proxies of children aged 7 to 18 years who had received disease-directed oncology treatment in the form of chemotherapy for at least 1 month. Data were analyzed identifying clusters of individuals (ie, latent profiles) based on various responses (ie, indicators) in August 2021.
The children of proxy participants received upfront chemotherapy. Children and proxies completed Patient-Reported Outcomes Measurement Information System (PROMIS) surveys at 2 time points: within 72 hours preceding treatment initiation and following the course of chemotherapy.
Main Outcomes and Measures
The latent profile analysis methods were applied to caregiver-proxy reports of PROMIS Pediatric symptom and function measures (anxiety, depressive symptoms, pain interference, fatigue, psychological stress, and physical function–mobility). The instrument categorized respondents as high symptom suffering, medium symptom suffering, and low symptom suffering (hereafter, high, medium, and low symptom groups, respectively).
Of 580 approached proxies, 431 (368 [85.00%] were female) identified as legal guardians of children aged 7 to 18 years with a first cancer diagnosis (mean [SD] age, 13.03 [3.40] years; 235 [54.65%] were male). Proxy reports of children’s experiences based on the 5 proxy PROMIS measures comprised 3 distinct symptom profiles. The most common proxy assessments of children’s experiences were the moderate symptom groups (45.7% [197 of 431]) and the low symptom groups profiles (40.1% [173 of 431]). A high symptom groups profile emerged which represented 14.2% (61 of 431) of proxy assessments. The number of profiles and observed distribution of profile membership was similar between child and proxy reports. Proxy reports of individual symptoms generally recorded higher scores than child reports; however, no significant difference was observed between proxies and child profile model results for the PROMIS measures.
Conclusions and Relevance
Results of this cohort study suggest that, at the level of symptom severity profile, proxy caregiver reports may approximate the children’s reports and may serve as a guide to care when the child is not able to self-report.
The application of latent profile analysis (LPA) methods to self-reports of pediatric patients with cancer of their symptom burdens and treatment adverse events reveals distinct subgroups or profiles of children, as shown in analyses of both the Pediatric Patient-Reported Outcomes version of the Common Terminology Criteria for Adverse Events (Ped-PRO-CTCAE) and the National Institutes of Health Patient-Reported Outcomes Measurement Information System (PROMIS) Pediatric measures.1-3 Consistently identified pediatric profiles include high symptom groups, moderate symptom groups, and low symptom groups.1,3 Profiles of symptom experiences are clinically relevant to help identify and target support for children most affected by high symptom burden. While patient self-report is the clinical standard, at times in clinical practice the child may be too ill, too fatigued, unavailable or unwilling to self-communicate symptoms or function, resulting in the clinical team having to rely on family caregiver perspectives to guide care.
The childhood cancer literature based on subgroup analyses of validated patient-reported outcome measures for children with cancer shows substantial differences in the self-reports vs proxy reports.4 In a prior summary, family caregivers overestimated symptoms and underestimated mobility function compared with the children themselves.5 Family caregivers of children receiving cancer treatment rated children significantly worse in performance status over time, while the children’s self-reports of performance status rating remained stable.6 Systematic reviews reveal substantial heterogeneity in agreement and concordance between pediatric patient-reported outcomes compared with family caregiver reports of the same metrics for the child.7
The variability in single symptom and adverse event reporting between pediatric patient reports and proxy reports raises the question of whether there would also be similar or disparate symptom profiles according to the perspective of the 2 reporters. Whether and how family caregiver symptom profile reports align with the pediatric patient symptom profile reports have not yet been studied. The purpose of this study was to apply LPA methods to family caregiver (referred to as proxy) symptom reports to document the number and characteristics of symptom profiles and to compare these with previously described symptom profile findings from self-reports of pediatric patients with cancer. A secondary aim of this study was to assess whether any demographic or clinical factors, including baseline psychological stress, were significantly associated with patient membership in a particular latent profile symptom classification based on parent reports. If the profiles were found to be similar, this finding would help clinicians to identify children most at risk for adverse events and most at need for increased support even if the child was not able to self-respond.
This cohort study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline. In this convenience sample, eligible family caregivers were predominantly parents (proxies self-identified as the child’s legal guardian) of children aged 7 to 18 years who had received disease-directed oncology treatment in the form of chemotherapy for at least 1 month. Exclusion criteria included children receiving a phase 1 trial-based therapy or receiving active end-of-life care. The proxy had to be able to read and understand English and agree to participate. Race and ethnicity were identified through patient and proxy self-report. Proxies gave signed informed consent for themselves and their children, and children provided assent. Children and their proxies completed the PROMIS Pediatric or proxy report within 72 hours of the start of a scheduled treatment cycle and again 7 to 28 days later. Proxies were provided $10 gift cards at each time point, which totaled $20. Data collection occurred from October 2016 to October 2018.
The study received institutional review board approval at 9 pediatric oncology treatment setting recruitment sites in North America: Children’s Healthcare of Atlanta, Georgia; Children’s Hospital Los Angeles, California; UPMC (University of Pittsburgh Medical Center) Children’s Hospital of Pittsburgh, Pennsylvania; Children’s National Hospital, Washington, DC; Dana-Farber Cancer Institute, Boston Children’s Hospital, Boston, Massachusetts; Duke University, Durham, North Carolina; The Hospital for Sick Children, Toronto, Ontario, Canada; St Jude Children’s Research Hospital, Memphis, Tennessee; and the University of North Carolina at Chapel Hill.
The PROMIS measures, developed using item response theory (a psychometric paradigm that links item properties, individuals responding to these items, and the trait being measured), document pediatric patient and proxy reports of the child’s experience with illness and treatment. Previous findings3 indicate strong item and scale reliability, and predictive and construct validity of the report items in the 7 days since the data collection point. Items are scored from 0 (never) to 4 (almost always) for symptom interference and psychological measures, and from 0 (with no trouble) to 4 (not able to do) for the mobility measure.8,9
In the present study, the participating pediatric patients with cancer and their proxies completed the PROMIS Pediatric measures using the computerized adaptive testing approach. The pediatric and proxy scores of the PROMIS Pediatric symptom and function measures (anxiety, depressive symptoms, pain interference, fatigue, psychological stress, and physical function–mobility) were used for symptom descriptions and for profile analysis. Psychosocial stress was used as a covariate in the study. The instrument categorized respondents as high symptom suffering, medium symptom suffering, and low symptom suffering (hereafter, high, medium, and low symptom groups, respectively). Because this study was not a dyadic study, the profile focus does not investigate invariance at the individual child-proxy level.
We first examined the descriptive statistics of the individual PROMIS Pediatric measures in both the child self-reports and proxy reports, and then compared corresponding symptom and function scores. Then, single-group LPA models10,11 were applied to identify the number of latent symptom profiles in both children and proxy reports based on the PROMIS Pediatric measures and PROMIS proxy measures. Latent profile model and fit comparisons are presented in eTable 1 in the Supplement.
Depending on the number of symptom latent profiles statistically identified in the child and proxy reports, multiple-group LPA (MGLPA) models12 and Mplus version 8.6 (Muthén & Muthén) were used for further analysis. An observed grouping variable, ie, group 1 indicating children and group 2 indicating proxies, was added in the MGLPA model as a known class variable across which to test similarity of profile characteristics between children and proxies. Similarity of within-profile PROMIS score level and dispersion between children and proxies were then examined.
An unrestricted MGLPA model, in which within-profile means and variances of the PROMIS Pediatric measures were allowed to vary freely between children and proxies, was estimated; then a model with equality restriction on the means of the PROMIS measures between the 2 reporting groups was estimated and compared with the unrestricted model. A scaled χ2 difference test based on loglikelihood values and scaling correction factors obtained with the MLR estimator13 was used to test the similarity of means of the PROMIS Pediatric measures between the 2 reporting groups.
A similar approach was used to test the similarity of within-profile dispersion of the PROMIS Pediatric measures across children and proxies. An MGLPA model with equality restrictions on within-profile variances of the PROMIS measures across children and proxies was estimated and compared with the unrestricted model. The scaled χ2 test was used to test the similarity of variances of the PROMIS measures across populations.
Based on the results of the first and second similarity tests, an appropriate MGLPA model was retained for further analysis. Similarity of profile distribution across children and proxies was tested by regressing the symptom latent profile membership on the known class variable group in the retained MGLPA model.
In addition, the latent profile variable estimated from the retained model was saved as an observed categorical variable and was treated as the dependent variable in a multigroup multinomial logit model to test the association of sociodemographic and clinical factors with the symptom profile classification in each of the 2 reporting groups. If any of the associations were statistically significant, the similarity of such an association across the groups was tested using the Wald test of parameter constraint. Statistical significance was set at 2-tailed P < .05. All models were estimated using Mplus version 8.6 (Muthén & Muthén).
Of 580 child-caregiver proxies approached for participation, 88 (15.2%) declined and 10 (1.7%) withdrew before completing the survey. Of the remaining 482 families, we further restricted the proxy-based sample to those proxies who completed the survey on the same day as the child. A total of 436 children aged 7 to 18 years (mean [SD] age, 13.03 [3.40] years; 235 [54.65%] were male) and 431 proxies (368 [85.00%] were female) were included in the study. Sample demographic characteristics are shown in Table 1.
The mean scores of the pediatric and proxy PROMIS measures within 72 hours preceding treatment initiation are shown in Table 2. The mean scores are statistically significantly worse by proxy report for all 4 symptoms (anxiety, depressive symptoms, fatigue, and pain interference) and for mobility. The mean (SD) child anxiety score was 43.31 (9.98); the mean (SD) proxy anxiety score, 48.33 (11.41); the score difference, 5.02 (95% CI, 3.70-6.45). The mean (SD) child depressive symptoms score was 45.38 (10.59); the mean (SD) proxy depressive symptoms score, 48.08 (11.39); the score difference, 2.70 (95% CI, 1.23-4.04). The mean (SD) child fatigue score was 44.66 (11.91); the mean (SD) proxy fatigue score, 52.58 (12.80); the score difference, 7.91 (95% CI, 6.46-9.43). The mean (SD) child pain interference score was 43.28 (8.93); mean (SD) proxy pain interference score, 47.63 (9.89); the score difference, 4.35 (95% CI, 3.26-5.51). The mean (SD) child mobility score was 44.30 (9.95); the mean (SD) proxy mobility score, 40.74 (9.74); the score difference, −3.55 (95% CI, −4.60 to −2.56). Using the minimally important difference instrument score of 3 points as has been determined for these measures,14 3 of the symptoms are worse by proxy report (the exception being depressive symptoms).
Among the tested LPA models with various numbers of latent symptom profiles, the 3-profile model fit the data best in both children and proxies. Model fit statistics and indices and model comparisons are presented in eTable 1 in the Supplement. This comparison demonstrates configural similarity of latent profiles using either the pediatric or proxy PROMIS measures. Although the latent symptom profiles in children and proxies appear to be similar (eFigure A and B in the Supplement), the specific characteristics of the profiles (eg, within-profile means and variances, profile distributions) may not be all similar across children and proxies.
The result of the scaled χ2 test shows that model fit between models with variant and invariant within-profile means of PROMIS measures differed significantly (scaled χ2 test, df = 9; P < .001) (Table 3). Information criterion indices (Akaike information criterion, bayesian information criterion, and adjusted bayesian information criterion) were also reported in the Supplement (eTable 2 in the Supplement) for model comparisons. The information criterion indices are much larger for the model with equality restrictions on within-profile PROMIS means (Akaike information criterion. 32 070.12; bayesian information criterion. 32 189.22; and adjusted bayesian information criterion, 32 109.82) than those for the model with no restriction (Akaike information criterion, 31 975.01; bayesian information criterion, 32 160.82; and adjusted bayesian information criterion, 32 026.95) (eTable 2 in the Supplement). This finding suggests that it may be better to allow outcome means in each profile to vary freely across the reporting groups (ie, children and proxies). However, the scaled χ2 test shows that models with and without equality restrictions on the within-profile variances of the PROMIS measures did not significantly differ (scaled χ2 test, df = 5; P = .08) (Table 3). Akaike information criterion, bayesian information criterion, and adjusted bayesian information criterion are very similar in the unrestricted model and the model with equality restriction on outcome variances (eTable 2 in the Supplement). This finding suggests that the variances of the PROMIS measures in specific symptom profiles are not significantly different between children and proxies. Based on the results of the similarity tests, the MGLPA model with variant within-profile means, but invariant within-profile variances across the reporting groups was considered the retained model.
The results of the retained model are presented in Table 4. The mean (SE) for each score is reported as higher by proxy than by child self-report (mild, moderate, and severe symptom levels in profiles 1, 2, and 3, respectively) in both groups. We, therefore, defined profile 1 as low symptom groups, profile 2 as moderate symptom groups, and profile 3 as high symptom groups. In cases from child self-reports, 45.6% of the cases (199 of 436) were classified in profile 1, 38.5% (168 of 436) in profile 2, and 15.8% (69 of 436) in profile 3. In cases from proxy reports, 40.1% (173 of 431) were classified in profile 1, 45.7% (197 of 431) were classified as profile 2, and 14.2% (61 of 431) were classified as profile 3 (Table 4). The quality of profile classification was good: the entropy statistic was 0.84, and the mean posterior probabilities of profile membership were 0.91 for profile 1, 0.95 for profile 2, and 0.91 for profile 3. The mean levels of anxiety, depressive symptoms, fatigue, and pain interference in all 3 profiles were consistently higher (worse), although the mobility mean score was lower (worse), for proxies compared with children. Both children and proxies in profile 3 reported the worst mean scores for depressive symptoms (mean [SE], children: 61.81 [1.29] and proxy: 64.74 [2.62]), anxiety (mean [SE], children: 59.50 [1.44] and proxy: 65.17 [3.30]), and pain interference (mean [SE], children: 57.27 [1.36] and proxy: 67.37 [1.53]) (Table 4).
The similarity of profile distributions was tested by regressing the latent categorical profile variable on the known class variable groups (1, children; 2, proxies) in the retained model. The odds of being classified in specific latent profiles were not significantly different between the 2 reporting groups. The odds ratio of being classified in profile 1 rather than profile 3 was 0.97 (95% CI, 0.52-1.78), and the odds ratio of being classified in profile 2 rather than profile 3 was 0.73 (95% CI, 0.29-1.34) (Table 3). Despite the overestimation of symptoms by proxies, the latent profile distributions are similar whether the PROMIS Pediatric or proxy measures are the basis of the symptom profiles.
The results of multigroup multinomial logit model are shown in Table 5. None of the sociodemographic and clinical factors was associated with latent profile classification by child or proxy reports except age, with older age being more likely in the moderate symptom groups profile by child report than in the low symptom groups profile (odds ratio, 1.11; 95% CI, 1.02-1.21). Baseline psychological stress had a significant association with profile classification by both child and proxy reports, with higher scores more likely in the moderate and severe symptom profiles than in the mild symptom profiles.
Proxy reports of children’s experiences based on the proxy responses to the PROMIS Pediatric measures comprise 3 distinct symptom profiles. Mean symptom scores differed significantly among the 3 proxy-reported profiles. As expected, the largest differences occurred between the proxies’ high symptom groups and low symptom groups profiles.
Overall, proxy mean scores for all 5 PROMIS measures were worse than child-scores analyzed as a collective group. Recognizing that 3 points represent minimally important difference in the PROMIS scale for proxy data,14 anxiety, fatigue, pain, and mobility reached MID threshold and depressive symptoms were also closely aligned to this metric of meaningful changes in health status. Differences between proxy and child report increased from the mild symptom to the severe symptom profile although though size difference varied by symptom. Proxy-report PROMIS mean scores for this study’s collective convenience sample cohort are comparable to a representative sample of the US pediatric population (nononcologic cohort) in which the median proxy-report for anxiety was 43, depressive symptoms, 44, fatigue, 47, pain, 42, and mobility, 63.15
The number of symptom profiles was the same for both proxy-reported and child-reported data. Notably, the number of profiles was not determined a priori in this study but derived from statistical modeling. There existed 3 (mild, moderate, and high symptom groups) profiles in children undergoing cancer care with respect to their self-report using the same 5 PROMIS Pediatric measures.1,2 This study’s novel proxy profile finding parallels the model-identified profile group size resulting from children’s self-report.
At the individual level of symptom reporting, there are statistical and clinical differences consistently across symptom and function measures. Clinically, this study expands on prior reporting6 that proxies tend to report worse symptom pain and worse mobility function compared with children’s self-reports with potential underrecognition of mood values by proxies.5 However, at the profile level, the individual symptom differences are less pronounced. Although individual symptom reports remain available for the clinician, the profiles help make comparison with others. Notably, the child and proxy reports are most similar clinically in the low symptom groups profile at the level of the mean symptom or function scores. The exception is pain which is more than 2 times the difference between the 2 reporting groups than any of the other symptoms.
The lack of association of demographic characteristics (age, sex, race and ethnicity, and formal education of the proxy) with the profiles was a reminder that each patient lives their own personal cancer experience uniquely. The lack of clinical characteristic (diagnosis category, time since diagnosis, and hemoglobin level) association with profiles was unexpected because certain cancer types or treatment regimens are often thought to result in increased incidence of adverse events. Baseline psychological stress significantly increased the odds of being classified in the worse suffering profile by both child and proxy reports, indicating the clinical importance of this variable from both perspectives.
The findings of this study support the clinical concept that proxies offer an informative assessment of children’s cancer experiences.16 These data suggest that incorporating symptom pain profiles may allow for a more comprehensive summation of the child’s cancer experience with additional consideration of perceptions from those closest to the child. The actual factors associated with how closely proxy symptom reports align with child self-reports are complex, likely reflecting the proxy’s own wellness and stress, the child’s duration of illness, culture, and communication patterns within families, such as whether the child and proxy are trying to protect one another from awareness.5,17 Further assessment of ways to foster child and proxy concordance in symptom reporting is warranted. Historically, patient-reported outcomes literature in pediatric oncology has pointed out individual differences in symptom scores per respondent. Profiles provide opportunity to consider symptom reporting at a higher level of shared experience.
Study strengths include the multisite setting that yielded a large proxy sample. An additional strength is that some members of the study team had extensive prior application experience with LPA outcome data for child self-report large proxy sample.
Study limitations include lack of longitudinal study design and lack of analyses by protocol-specific treatment plans. Only certain PROMIS Pediatric measures were included; thus future profile testing that includes other symptoms might be important to examine. A limitation of the statistical analysis is the latent profile membership estimated from unconditional multigroup LPA model was used as an observed categorical variable for further analysis in the study. Given this situation, the measurement errors of latent profile membership were not taken into account in the multinomial logit model. These observations may warrant being replicated across different groups based on diagnostic stage, different disease sites, treatment types, and different levels of symptom severity to enhance generalizability. The study team did not specifically document the legal guardian’s relationship to the child (eg, parent, grandparent, or court-appointed guardian) and so, although most proxies were parents, the lack of documentation regarding relational role limited additional analyses on how familiarity or family structure may be associated with perspective. As the intention of this first assessment, to our knowledge, of caregiver perspectives was a comparison profile study, a future study may further engage in dyad-specific assessments to evaluate for symptom reporting congruence between individual child-proxy pairs.
In this cohort study, LPA models suggest that profile reports of children with cancer and their proxies clustered into 3 distinct symptom profiles based on responses to the self-report or proxy-report PROMIS Pediatric measures. Although proxies as a whole seemed to the view children’s symptom burden as worse than the children’s own reports, our model results show that within-profile variation of the PROMIS measures is not significantly different between children and proxies. Grouping child and proxy reports according to symptom profile classifications may bring the reporting closer to the lived reality for the child. Clinical use of the profiles may allow care teams to identify children with highest risk of sever adverse events and most in need of early intervention and a proactive supportive care approach. The recognition that this is a smaller cohort may help guide resource allocation toward additional supportive care interventions early in care for those experiencing the greatest symptom burden. This findings of this study suggest that clinicians may also use proxy reports to estimate profile membership to enhance the patient report when the child is unwilling or unable to self-report.
As the science of patient-reported and proxy-reported outcomes continues to evolve in the quality of the ask and reliability of the response, ideally care teams will query the children’s symptom burdens concurrent with obtaining the proxy perception of the child’s symptoms. This information may then be used to manage symptoms in a proactive way that may translate into improved care. Findings of this LPA of proxy perceptions of the symptom burden of pediatric patients with cancer suggest that although proxy report of the child’s symptom burden is not the standard,18 it offers beneficial inclusion.
Accepted for Publication: January 21, 2022.
Published: March 29, 2022. doi:10.1001/jamanetworkopen.2022.1855
Open Access: This is an open access article distributed under the terms of the CC-BY-NC-ND License. © 2022 Weaver MS et al. JAMA Network Open.
Corresponding Author: Meaghann S. Weaver, MD, MPH, PhD, Department of Pediatrics, University of Nebraska Medical Center, Omaha, NE 68198 (MeWeaver@childrensomaha.org).
Author Contributions: Drs Weaver and Hinds had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.
Concept and design: Weaver, Jacobs, Wang, Hinds.
Acquisition, analysis, or interpretation of data: All authors.
Drafting of the manuscript: Weaver, Withycombe, Wang, Hinds.
Critical revision of the manuscript for important intellectual content: Jacobs, Wang, Greenzang, Baker, Hinds.
Statistical analysis: Weaver, Wang, Hinds.
Obtained funding: Wang, Hinds.
Administrative, technical, or material support: Hinds.
Conflict of Interest Disclosures: Dr Withycombe reported receiving grants from National Institute of Arthritis and Musculoskeletal and Skin Diseases (U19AR069522; PIs: Reeve/Schanberg) and grants from National Cancer Institute (R01CA175759; PIs: Hinds/Reeve) during the conduct of the study. Dr Hinds reported receiving grants from NIH during the conduct of the study. No other disclosures were reported.
Funding/Support: Design and conduct of the study was supported by funding from the National Cancer Institute (R01CA175759) and the National Institute of Arthritis and Musculoskeletal and Skin Diseases (U19AR069522).
Role of the Funder/Sponsor: The funding organizations had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.
Disclaimer: Dr Weaver contributed to this paper in a private capacity. No official support or endorsement by the US Department of Veterans Affairs is intended, nor should be inferred.