Emotional well-being does not predict survival in head and neck cancer patients

A radiation therapy oncology group study

Abstract

BACKGROUND.

The objective of the current study was to examine whether emotional well-being predicted survival in a large sample of patients with head and neck cancer who were participating in multicenter clinical trials.

METHODS.

Participants were enrolled in 2 Radiation Oncology Group (RTOG) clinical trials (RTOG 9003 and RTOG 9111) and completed a baseline measure of quality of life (the Functional Assessment of Cancer Therapy-General [FACT-G]), which included an Emotional Well-Being subscale. The outcome measure was overall survival. Main statistical analyses included overall survival rates, which were estimated by using the Kaplan-Meier method with univariate comparisons analyzed using the log-rank test. A multivariate Cox proportional hazards model was used to determine whether emotional well-being had prognostic impact on survival after accounting for tumor-related and sociodemographic variables. Additional exploratory analyses examined possible subgroup effects.

RESULTS.

No statistically significant univariate or multivariate effects were observed for emotional well-being, and there were no effects limited to subgroups. These results stand in sharp contrast to the prognostic value of a variety of demographic and clinical variables.

CONCLUSIONS.

The current results add to the weight of the evidence that emotional functioning is not an independent predictor of survival in cancer patients. The study had the advantage of a large number of deaths to be explained in a sample with the uniformity of treatment and quality of care that is required in clinical trials. Cancer 2007. © 2007 American Cancer Society.

Avariety of mechanisms have been posited to explain why emotional well-being should be related to disease progression and mortality among patients with cancer,1–3 and the notion that emotional state affects disease function and survival is core to a larger constellation of beliefs that are prevalent among both patients and professionals.4, 5 Related beliefs include that patients with cancer can influence the course and outcome of their cancer by making changes in their psychologic state, and, in particular, that psychotherapy can promote survival. Indeed, the assumption that emotional well-being affects cancer progression is basic to explanations of why psychotherapy should promote survival: In 2004, Spiegel noted that it is difficult to imagine that an intervention that does not benefit patients psychologically will extend survival.6 However, a recent comprehensive review concluded that no study reported an effect of psychotherapy on survival in which survival was an a priori primary outcome and in which the provision of psychotherapy was not confounded with improved medical care.7

Studies that are interpreted as indicating that better psychologic functioning predicts longer survival among cancer patients8–11 can be countered with other instances in which better psychologic functioning appears to predict shorter survival12–15 and with the larger number of studies that had null results.16–27 This literature is plagued by studies with samples that are heterogeneous with respect to staging and cancer site, unmeasured or poorly measured confounding biologic and treatment variables, and small numbers of deaths to adequately accommodate possible variability in patient and treatment characteristics. The ready rival explanation of some apparent demonstrations that emotional well-being predicts survival is that patients' self-reported emotional states reflect overall disease burden and symptom distress or their awareness of their medical condition and prognosis.

The belief that emotional well-being affects survival, nonetheless, has been remarkably resilient in the face of contrary data. Thus, Brown et al.28 noted a preponderance of studies with null results but still interpreted their own findings as establishing the importance of depressive symptoms in predicting cancer survival. However, Brown et al.28 were predicting only 80 deaths in a sample with 15 different tumor sites and 14 predictor variables, including 7 intercorrelated psychologic variables. Their regression equation was overfitted in terms of variables to patient ratio, and their overall analyses capitalized on chance in eliminating some medical and demographic predictors of mortality and some of the psychologic variables on the basis of preliminary inspections of the data (for a critique of these practices, see Babyak29).

Although the risks of Type II error with small samples is well known, there is also the less appreciated likelihood that apparent strong effects reported with small samples may be false-positive results, particularly when results were not anticipated or explained readily. Confirmatory publication bias is greater with small sample studies. Although null findings readily are dismissed or rejected for publication because of the low statistical power, positive findings appear remarkable and publishable because they were unexpected given the sample size. In samples that already are small in terms of numbers of patients, the number of relevant events (deaths) is even smaller. Moreover, the aberrant outcome of a very few patients can determine results, whether they are caused by selection bias or by chance. Statistical controls also are less effective in small samples: In 1990, Piantadosi noted that it is easy to envision a situation in which the marginal imbalances of prognostic factors are minimal but in which the joint distributions differ and are influential.30

The question of whether emotional well-being affects survival is an important one, but it is not going to be resolved with further accumulation of studies with small heterogeneous samples. Rather, what are needed are studies with ample numbers of deaths to be explained and better control of possible confounds. Large-scale, community-based clinical trials in which quality of life is assessed, including emotional well-being, present a particularly apt opportunity, in that patient populations are relatively homogeneous and well described, and there is some standardization of treatment. The current study took advantage of 2 head and neck radiation therapy trials to examine whether emotional well-being at study entry predicted survival in a combined sample of approximately 1100 patients and 650 deaths. Neither trial demonstrated a significant survival differences between the treatment arms. The number and advanced stage of the participating patients was such that we anticipated that we would have ample statistical power to detect even a small size effect for emotional well-being on survival.

MATERIALS AND METHODS

Patients who were participating in 2 Radiation Therapy Oncology Group (RTOG) study protocols (RTOG 9003 and RTOG 9111) were asked to complete the Functional Assessment of Cancer Treatment-General (FACT-G) at baseline and in follow-up. However, with RTOG 9003, the FACT-G was added after the protocol was opened to patient accrual. Therefore, the first 83 entries were excluded from this analysis, because they did not have the opportunity to participate in the quality-of-life (QOL) component.

Patients

RTOG 9003 was a randomized Phase III trial evaluating altered dose fractionation for patients with head and neck squamous cell carcinomas, specifically, 1) standard fractionated radiotherapy, 2) hyperfractioned radiotherapy; 3) accelerated hyperfractionated radiotherapy; and 4) accelerated fractionated radiotherapy with concomitant boost. The eligibility criteria were age ≥18 years; previously untreated stage III or IV but metastasis-negative (M0) squamous cell carcinoma of the oral cavity, oropharynx, supraglottic larynx, or stage II through IV carcinoma of the base of tongue or hypopharynx with staging according to the American Joint Committee on Cancer, 3rd edition, 1988; and a Karnofsky performance score (KPS) ≥60. Exclusion criteria were a prior (within 5 years) or synchronous malignancy other than nonmelanoma skin cancer. All the patients provided written informed consent in accordance with institutional guidelines. Further details were reported elsewhere.31 Between September 30, 1991 and August 1, 1997, 1113 patients were enrolled on this trial.

RTOG 9111 was a randomized Phase III trial comparing concurrent chemotherapy and radiotherapy for preservation of the larynx in locally advanced laryngeal cancer. The trial compared 1) induction chemotherapy and radiation therapy, versus 2) concomitant chemotherapy and radiation therapy, versus 3) radiation therapy alone. The eligibility criteria included age ≥18 years; biopsy-proven, previously untreated stage III or IV squamous cell carcinoma of the glottic or supraglottic larynx, the surgical treatment of which would require total laryngectomy; and a KPS ≥60. Patients who had a stage T1 primary tumor with large volume or who had large-volume stage T4 disease were not eligible. The disease had to be considered curable with surgery and postoperative radiotherapy. Finally, patients also had to have a white cell count ≥3500/cm³, a platelet count ≥100,000/cm³, a normal serum calcium level, and a creatinine clearance ≥50 mL per minute. All patients provided written informed consent in accordance with institutional guidelines. Further details have been reported elsewhere.32 Between August 17, 1992 and May 31, 2000, 547 patients were enrolled in this trial.

Measures

FACT-G

The FACT-G scale33 (version 29) is a 27-item core questionnaire that evaluates various domains of QOL, including, physical, functional, family-social, and emotional domains, and is used widely in North American cancer clinical trials. The 5-item Emotional Well-Being (EWB) subscale includes items such as “I feel sad” and “I am losing hope in my fight against my illness.” Items are summed to give scores for each domain and an overall QOL score. Higher scores on the EWB indicate better emotional well-being or less depressed mood.

In our pilot work for the current study, we examined the correlation between the FACT EWB and the Hopkins Symptom Checklist-25, a widely used screening instrument for depression,34 in an independent sample of 240 patients with breast cancer. We observed a correlation of .77, which is at the upper limit of what the reliabilities of these 2 instruments allow. In addition, as part of the current study, a depressed mood scale was derived from the 5-item FACT EWB scale by dropping 1 item: “I am proud of how I am coping with my illness.” The retained items were feeling sad, losing hope, feeling nervous, worrying about dying, and worrying that the condition would get worse. The coefficient α for the reduced scale was .70. However, in subsequent analyses, the correlation of the reduced scale was .94 with the full scale, and results were identical to those reported here for emotional well-being. Taken together, these findings provide some confidence in the comparability of our results with what would be obtained with self-report measures explicitly labeled as “depression”. However, we report only the results for emotional well-being.

Survival

Patients who were entered into the 2 phase III trials (RTOG 9003 and 9111) were followed routinely in those trials until death or until they refused further participation. Enrolling institutions followed patients according to the protocol at specified times and sent the case report (data collection) forms to RTOG. Reports of deaths were not verified independently by the RTOG. Patients were considered lost to follow-up if there was no report submitted within the last 3 years. Among the 1510 patients who were analyzed in this report, 63 patients (4.2%) were classified as lost to follow-up.

Statistical Methods

The outcome measure was overall survival (OS). All deaths were counted as failures for OS, and OS rates were estimated using the Kaplan-Meier method35 with univariate comparisons analyzed using the log-rank test.36 The multivariate Cox proportional hazards model was used to determine whether emotional well-being had prognostic impact on survival after accounting for tumor-related and sociodemographic variables.37 The initial analysis, as expected, showed a highly significant difference in survival between the protocols; so, the multivariate model was stratified by protocol. This model would account for the differences in patient eligibility and, thus, prognosis. It assumes that the hazards ratios (HR) for a variable in each stratum are similar.

There was no imputation for missing emotional well-being values, because it was being tested here as a key predictor of survival. However, imputation of missing values for other variables was performed with the Markov Chain Monte Carlo algorithm to minimize the potential bias from excluding patients from the analysis.38 (Missing values were encountered almost exclusively with the sociodemographic variables, such as precancer diagnosis income and marital status). Two SAS procedures (MI and MIANALYZE) were used to generate 10 imputed datasets and to combine the results from the analyses.39 The following variables were used in the imputations: protocol, sex, age, KPS, tumor (T) classification, lymph node (N) status, primary site, race, marital status, number of members in household, education level, precancer diagnosis income, type of employment, Canadian patient entry, survival time, and duration of locoregional disease control.

For each imputed dataset, a stepwise procedure was used to fit the Cox model. The greatest difference occurred with the selection of an income interval for the cut point. The $15,000 interval was chosen because it appeared in most of the derived models. The initial analysis was restricted to the eligible patients who had been included in the protocol treatment comparisons. In that analysis, the patients with and without baseline FACT-G were compared with respect to pretreatment characteristics, protocol treatment assigned, and survival. Then, the analysis was restricted to patients with baseline FACT-G to determine whether emotional well being had any independent prognostic impact on survival.

RESULTS

Table 1 presents the pretreatment demographic and clinical characteristics of the combined samples broken down by whether the patient provided a FACT-G. Not surprisingly, patients who did not provide a FACT-G tended to have poorer KPS and more missing values. Table 2 provides the numbers of deaths observed for both trials. There was no significant difference in survival between patients with and without baseline FACT-G data (HR, 0.954; 95% confidence interval [95% CI], 0.826–1.101; P = .58).

Table 1. Pretreatment Characteristics of Patients With and Without a Baseline Functional Assessment of Cancer Therapy-General

Characteristic	No. of patients (%)
	Patients with baseline FACT-G	Patients without baseline FACT-G
FACT-G indicates Functional Assessment of Cancer Therapy-General; RTOG, Radiation Therapy Oncology Group; RX, protocol treatment prescription; STD RT, standard fractionated radiotherapy; HFX, hyperfractionated RT; AFX-S, standard accelerated hyperfractionated RT; AFX-CB, AFX with concomitant boost; CT, computed tomography; KPS, Karnofsky performance status.
Total	1093	417
RTOG 9003	689 (63)	303 (37)
Protocol RX
STD RT arm	172 (25)	76 (25)
HFX arm	168 (24)	74 (24)
AFX-S arm	179 (26)	76 (25)
AFX-CB arm	170 (25)	77 (25)
Primary site
Oral cavity	72 (10)	28 (9)
Oropharynx	413 (60)	187 (62)
Hypopharynx	99 (14)	32 (11)
Supraglottic larynx	105 (15)	56 (18)
RTOG 9111	404 (78)	114 (23)
Protocol treatment
Induction CT arm	134 (33)	39 (34)
Concurrent CT arm	139 (34)	33 (29)
RT alone arm	131 (33)	42 (37)
Primary site
Supraglottic larynx	270 (67)	87 (76)
Glottic larynx	134 (33)	27 (24)
KPS
90–100	749 (69)	244 (59)
60–80	344 (31)	173 (41)
Tumor classification
T1	41 (4)	15 (4)
T2	227 (21)	104 (25)
T3	580 (53)	201 (48)
T4	245 (22)	97 (23)
Lymph node classification
N0	349 (32)	130 (31)
N1	218 (20)	92 (22)
N2	426 (39)	168 (40)
Sex
Men	872 (80)	316 (78)
Women	221 (20)	101 (22)
Age, y
<60	543 (50)	176 (42)
60–74	480 (44)	196 (47)
≥75	70 (6)	45 (11)
Race
White	818 (75)	277 (66)
African American	208 (19)	58 (14)
Hispanic	46 (4)	41 (10)
Other	14 (1)	13 (3)
Unknown/not answered	7 (1)	28 (7)
Martial status
Married/other	573 (52)	186 (45)
Single/divorced	499 (46)	181 (43)
Unknown/not answered	21 (2)	50 (12)
Household income prior to illness, US$
<8000	222 (20)	75 (18)
8000–14,999	215 (20)	59 (14)
15,000–24,999	183 (17)	56 (13)
25,000–34,999	106 (10)	28 (7)
35,000–49,999	75 (7)	17 (4)
≥50,000	79 (7)	32 (6)
Unknown/not answered	213 (19)	157 (38)
Cigarette use at protocol entry
Yes	714 (65)	243 (58)
No	371 (34)	162 (39)
Unknown	8 (1)	12 (3)

Table 2. Deaths by Protocol and Baseline Functional Assessment of Cancer Therapy-General Cancer

Baseline FACT-G	Total no. of patients	No. of deaths (%)
FACT-G indicates Functional Assessment of Cancer Therapy-General; RTOG, Radiation Therapy Oncology Group.
RTOG 9003 protocol
Yes	689	488 (71)
No	303	210 (69)
RTOG 9111 protocol
Yes	404	158 (39)
No	114	54 (47)
Combined
Yes	1093	646 (59)

Table 3 presents results of the univariate analyses of emotional well-being with survival. No statistically significant association between well-being and survival was observed. Next, a base model that incorporated protocol, demographic variables, smoking status, staging, and KPS was built to account for as much variation as possible (Table 4). When emotional well-being was added to this model; it did not appear to have a significant impact on survival (HR, 0.1.016; 95% CI, 0.995–1.037 [P = .13]). The nonsignificant effect actually was in the opposite direction than was anticipated. These null results for emotional well-being stand in sharp contrast to the significant effects for demographic and clinical variables.

Table 3. Univariate Cox Model: Emotional Well-Being

Emotional Well-Being*	HR	95% CI	P
HR indicates hazards ratio; 95% CI, 95% confidence interval; RTOG, Radiation Therapy Oncology Group. * The Emotion Well-Being (EWB) domain has 5 questions, and each is scored from 0 to 4; thus, total scores range between 0 and 20. In the current analysis, the EWB was considered as a continuous variable, with higher scores indicating better EWB. HR values >1.0 indicate that, with increasing EWB scores, survival decreased.
All patients (stratified)	1.018	(0.997–1.039)	.092
RTOG 9003	1.007	(0.984–1.031)
RTOG 9111	1.055	(1.010–1.106)

Table 4. Covariates Used in the Cox Multivariate Survival Analysis (n = 1093 Patients; 646 Deaths)*

Covariate and Subgroups	HR (95% CI)	P
HR indicates hazards ratio; 95% CI, 95% confidence interval; KPS, Karnofsky performance score.
Age, y
0: <60	1.297 (1.107–1.519)	.001
1: >60
KPS
0: 90–100	1.535 (1.297–1.816)	<.001
1: 60–80
Tumor classification
0: T1-T3	1.691 (1.411–2.027)	<.001
1: T4
Lymph node classification
0: N0-N1	1.880 (1.490–2.371)	<.001
1: N2-N3
Cigarette smoking at protocol entry
0: No	1.423 (1.191–1.699)	<.001
1: Yes
Income, $US
0: <15,000	1.300 (1.075–1.572)	.007
1: ≥15,000
Marital status
0: Married/live-in	1.269 (1.073–1.501)	.006
1: Single/divorced/separated/widowed
Emotional well being (continuous variable)	1.016 (0.995–1.037)	.13

Before simply conceding that the null hypothesis of no effect of emotional distress on survival could not be rejected, we undertook some exploratory analyses and were prepared to interpret any positive findings with caution. Part of our motivation for these analyses was that some sources have expressed considerable confidence in the claim that emotional well-being should predict survival in cancer patients; thus, they may be concerned if we appear to have overlooked a prognostic value for emotional well-being limited to some subgroup(s) of patients. With >600 deaths, the study provided an unusual opportunity to examine whether there may be differential effects of emotional distress across some key subgroups of patients. We addressed the question of subgroup effects by defining, a priori, a set of potentially interesting interactions and testing them in a Cox model in a pooled fashion (ie, an omnibus test for the entire set of interactions) to minimize Type I errors. Thus, these analyses examined the general question whether there may be subgroups for which the association of emotional well-being and survival may warrant further investigation. We selected 4 variables to be examined as main effects and then in interaction with a continuously scored baseline emotional well-being. The variables were 1) Protocol (9003 vs 9111); 2) sex (men vs women); 3) primary tumor site (oropharynx [reference group] vs oral cavity vs hypopharynx vs larynx); and 4) American Joint Commission on Cancer stage (II/III vs IV; stage II was combined with stage III, because there were only 16 patients with stage II disease). The interactions consisted of these 4 factors crossed with the EWB score. When all interactions were added simultaneously to the main effects in the Cox model, they did not improve the model significantly (P = .21); therefore, individual interaction effects were not examined further.

Finally, we rescaled the EWB score to rule out the possibility that our results were an artifact of our conventional scoring of the scale. This was done by dividing the score by 5, which is the range between the 75th and the 25th percentiles for the scores. This yields an HR that compares a patient in the middle of the upper half of the emotional distress distribution to a patient in the middle of the lower half. The results were the same as those achieved with the original scoring, except that the parameter estimate for the EWB score was 5 times greater.

DISCUSSION

The current results provided no support for the hypothesis that negative emotional well-being predicts poorer survival among this large sample of patients with head and neck cancer. No effects were observed in either univariate or multivariate analyses or in exploratory analyses that examined interactions between emotional well-being and study protocol, sex, primary cancer site, or staging. Thus, this psychologic variable neither affected progression or death directly nor functioned as a lurking variable,40 and it only emerged when other prognostic factors were controlled. In contrast, a number of demographic and clinical variables significantly predicted survival. These results are entirely consistent with another study of 57 deaths among 208 patients with head and neck cancer patients17 that included both a measure of emotional well-being from a QOL instrument and the Center for Epidemiologic Studies-Depression scale.41 The current study had the advantage of a pair of large, well-described samples with the standardization of treatment available in a controlled clinical trial. The number of deaths to be explained, 646, was larger than the total sample size of most previous studies. It should be noted that the number of events, and not the sample size, determined the ability to detect statistical differences and the ability to avoid spurious results from overfitting the regression models.42

Some limitations of our study should be noted. First, patients had to be judged mentally reliable to follow instructions and to keep appointments, and they had to provide informed consent to participation in a clinical trial. The resulting sample may not be truly representative of the larger head and neck cancer population with respect to emotional well-being. Second, it is both a strength and a limitation that we confined ourselves to a pair of homogeneous samples of head and neck cancer patients who satisfied the protocol and agreed to participate. The strength is that we can have greater confidence that an effect was not obscured by unmeasured clinical differences. However, our results cannot necessarily be extrapolated to other tumor sites. It is possible that breast cancer and prostate cancer would yield different results, which are more plausible candidates for an effect of emotional functioning because of the greater role that can be accorded to endocrine factors. However, recent large-scale studies of patients with metastatic and nonmetastic breast cancer yielded similar results with a similar measure.42–44 Third, our assessment of emotional well-being was limited to what can be derived from the FACT-G, namely, the EWB scale. An important advantage of this measure is that the FACT-G is a widely used measure of QOL in large-scale clinical trials for a full range of cancers; thus, our study may be replicated using other large data sets with possible direct comparison of scores on this measure. This advantage cannot be claimed for most other measures. Moreover, a depression scale that was derived from this measure of emotional well-being by excluding 1 potentially irrelevant item also was examined, but it was highly correlated with the existing EWB scale and produced identical results. Finally, in our pilot work, we discovered that our measure of emotional well-being was correlated as highly with an instrument that is widely used for screening cancer and other medical patients as the reliabilities of the 2 measures would allow. It should be noted that there has been no consensus in the existing literature regarding which aspect of emotional functioning or negative affect should be most crucial in predicting survival and with what measure it should be assessed, and there is no consistency regarding which measures have yielded positive findings. Indeed, positive findings often have emerged in studies in which there were multiple measures administered, and the claim of an effect often has depended on emphasizing a particular measure when similar measures failed to yield an effect. Finally, in the range of negative affect, which tends to characterize cancer patients, there is a tendency for measures of negative affect to converge on a single factor of distress with little specificity apparent.45 Thus, there is no good a priori reason to expect 1 measure of affect to work better than others in predicting survival. Claims for the superiority of particular measures that have been made in the past have been inconsistent and typically have emerged post hoc and in the service of a seeming confirmatory bias for a specific set of findings.

In summary, emotional well-being was not a predictor of survival among patients with head and neck cancer who were participating in a pair of clinical trials. The current retrospective study was unusual in the large number of deaths being explained, but the results add to the weight of evidence provided by smaller studies with less appropriate analyses of data. Despite such negative evidence, positive assessments of the evidence for an association between negative emotional well-being and reduced survival persist. Thus, it was claimed recently that evidence is growing of a correlation between depression and cancer incidence and progression.6 It is not clear what it would take to move the field beyond an appraisal of the literature even as simply “mixed” or “contradictory.” With a literature dominated by studies with small, heterogeneous samples and inadequately assessed and controlled clinical variables, box scores and meta-analyses become inappropriate. Kraemer et al.46 have cautioned against any optimism that similarly flawed studies, particularly those with small sample sizes, can be combined into a meta-analysis as an informative contribution to the literature. They noted that the twin threats of low statistical power and spurious results from multivariate analyses should cause the smaller studies from being considered in the same context as larger, better controlled studies.