CHALLENGES POSED BY NON-RANDOM MISSING QUALITY OF LIFE DATA IN AN ADVANCED-STAGE COLORECTAL CANCER CLINICAL TRIAL

Effects of variations in agent, dose, and route of treatment administration on patient reported quality of life (QOL) were examined for 279 patients enrolled on a seven-arm randomized clinical trial (S8905) of 5-FU and its modulation for advanced colorectal cancer. Patients completed QOL questionnaires at randomization and weeks 6, 11, and 21 post-randomization with five QOL endpoints considered primary: three treatment-specific symptoms (stomatitis, diarrhea, and hand : foot sensitivity); physical functioning; and emotional functioning. Patient compli- ance with the QOL assessment schedule was good, supporting the feasibility of including QOL measures in cooperative group trials. However, death and deteriorating health produced substantial missing data. Cross- sectional analyses indicated that the seven therapeutic arms did not differ in their impact on QOL. Unfortunately, longitudinal analyses of the QOL data were inappropriate given non-random missing data. Graphical presentation of non-random missing data identified the seriousness of this problem and its effect on potential conclusions about QOL during treatment. This problem appears to be particularly challenging in the context of advanced-stage disease. Failure to recognize the presence of non-random missing data can lead to serious overestimates of patient QOL over time. Copyright © 2000 John Wiley & Sons, Ltd.


Clinical and quality of life issues in ad6anced colorectal cancer
Cancer of the colon and rectum is a common solid tumor, with approximately 140000 new cases diagnosed annually (Wingo et al., 1998).
Treatment for patients with advanced or recurrent colorectal cancer is usually palliative (Bengt et al., 1994). Systematic measures of patient quality of life (QOL) complement information about optimal therapeutic effects by addressing the degree of palliation achieved as a result of treatment.
The QOL of patients with advanced colorectal cancer treated with 5-FU has received little attention in clinical research. A six-arm trial reported by Poon et al. (1989) examined the effect of various biochemical modulations of 5-FU on the QOL of patients with advanced colorectal cancer. A regimen containing 5-FU and low-dose leucovorin (IV injection) showed improvements in three physician-rated measures of patient QOL (improvement in performance status, weight gain of 5% or more, improvement in symptoms). Hill et al. (1995) examined the effect on QOL of another agent, interferon alfa-2b (IFN), added to continuous infusion 5-FU for patients with advanced colorectal cancer. However, in this study, a comprehensive patient assessment of QOL was included, the European Organization for Research and Treatment of Cancer (EORTC) QLQ-C30 (Aaronson et al., 1993). Patients treated with protracted venous infusion 5-FU and IFN delivered subcutaneously did not differ from patients treated with protracted venous infusion 5-FU alone with respect to most clinical variables or domains of the EORTC QLQ-C30 (e.g. global QOL and psychosocial functioning). Certain physician-rated treatment-related toxicities were greater on the combined therapy arm (e.g. leukopenia and mucositis), but on the patientrated symptom scales, only appetite loss differentiated treatment arms with more appetite loss in the interferon modulation arm. Global QOL and emotional functioning did not differ by arm.

Describing missing QOL data in studies of ad6anced-stage disease
Recently, investigators have reported results of randomized trials in advanced-stage disease cancer sites in which missing data have been problematic for conducting longitudinal analyses . A conference was held in Switzerland in July, 1996 with two objectives: improving submission rates for QOL assessments in cancer clinical trials (Bernhard et al., 1998; and exploring appropriate options for conducting longitudinal analysis when substantial efforts to increase questionnaire submission rates still result in missing data . Hopwood et al. (1994) commented on this problem and suggested the use of graphic techniques, which have been extended in the analyses presented below. Donaldson has also recommended use of graphic techniques and described the bias associated with missing data (Bertsch and Donaldson, 1995); his suggestions guided methods to address this bias in several projects (Moinpour, 1994). It is necessary to inform readers about the extent of missing data in the analyses, as we have done, and note its presumed effect on conclusions about the effects of treatment on QOL. Staquet et al. (1996) noted the importance of describing missing data patterns in the report of QOL results; various analysis methods have been suggested (Staquet et al., 1998).
In this paper, we report the results of a QOL study conducted by the Southwest Oncology Group (SWOG) that failed to find QOL differences by treatment arm for patients with advanced colorectal cancer. The therapeutic trial, reported elsewhere (Leichman et al., 1995), will be briefly summarized. In particular, we will present an evaluation of the extent and nature of missing data found in this trial. As a result, we are able to highlight the challenges presented by non-random missing data for the analysis of change in QOL during treatment and the potential for serious overestimates of patient QOL over time when such missing data are not acknowledged. We also address the potential usefulness of baseline QOL data when they are prognostic for survival.

Design and objecti6es
SWOG-8905 was a seven-arm Phase II/III randomized screening trial that investigated the role of biochemical modulators, route of administration differences, and dose intensity on survival of patients treated with 5-FU (Leichman et al., 1995) (see Figure 1). Treatment arms were designed to allow multiple, parallel comparisons. The single agent 5-FU bolus regimen (arm 1) was compared to 5-FU paired with a biochemical modulator: leucovorin (arm 2, low dose; arm 3, high dose); PALA (arm 7). In addition, route of administration and schedule differences were examined: bolus injection (arms 1-3); protracted or continuous infusion 5-FU for 28 days via a battery-powered pump (arms 4 and 5); infusion of 5 FU over a Psycho-Oncology 9: 340-354 (2000) 24-h/day for 28 days using a pump outside the hospital or an infusion bag in the hospital (arms 6 and 7). Single agent continuous infusion 5-FU was compared to continuous infusion plus leucovorin (arms 4 and 5) and high dose, intermittent infusion of 5-FU with the same regimen plus PALA (arms 6 and 7). The following eligibility criteria were employed for the therapeutic trial: recurrent or disseminated metastatic colorectal cancer; no previous chemotherapy for disseminated disease; at most, one prior immunotherapy regimen for metastases; at most, one prior adjuvant therapy and only if more than 1 year since discontinuation; prior surgery or radiotherapy only if more than 2 weeks since treatment; SWOG performance status of 0 -2; adequate hematologic, renal, hepatic, and cardiac function. Patients enrolled on the therapeutic trial who could complete a questionnaire in English were eligible for S9045, the companion QOL study. The planned sample for the QOL study was 40 patients per arm or 280 patients.
Endpoints and treatment comparisons. The overall objective of the QOL study was to examine the effect of different treatment combinations on patient-reported QOL: three treatment-specific symptoms (mouth pain, diarrhea, and hand/foot sensitivity) (Leichman et al., 1990); emotional functioning; and physical functioning. Table 1 describes the a priori hypotheses associated with each primary endpoint and assessment period. Although SWOG-9045 addressed other components of QOL , the focus of this paper is the set of QOL components designated primary endpoints. We also examined the ability of a baseline measure of general symptom status to predict survival.
QOL assessment schedule. The SWOG QOL questionnaire was administered at randomization and at weeks 6, 11, and 21 post-randomization. To minimize missing data due to deterioration of health status and death, the length of QOL follow-up was based on an estimated median survival of 24-46 weeks for patients treated with single-agent bolus 5-FU in previous studies (Nordic Gastrointestinal Tumor Adjuvant Therapy Group, 1989;Petrelli et al., 1989;Leichman et al., 1990). In this trial, our report of QOL is with respect to the past week because we wanted patients to be rating QOL at the same calendar times from randomization. Although we are missing acute toxicities that may occur just after treatment on the IV push and 24-h infusion arms, we are most interested in the overall effect of treatment at the stated time points. Two of the five primary QOL endpoints addressed the extension of treatment arm-related acute toxicities to broader areas of functioning (emotional and physical functioning).

QOL assessment methods
When QOL is assessed in SWOG studies, the goal is to compare the effect of different treatment arms on the day-to-day functioning of patients receiving these treatments. For this reason, a comprehensive, patient-reported measure of QOL is included addressing physical, role, emotional, and social functioning as well as general symptom status and treatment-related side effects  (Moin pour et al., 1989). This approach is consistent with that adopted in other QOL research in cancer clinical trials (Aaronson et al., 1988;Gotay et al., 1992;Nayfield et al., 1992;Osoba, 1994). The QOL questionnaire was administered prior to the administration of treatment and QOL was rated with respect to the previous week; questionnaires were most frequently completed in the clinic. The components of the questionnaire are described below.
Treatment-specific symptoms. Stomatitis, diarrhea, and hand/foot sensitivity were designated as primary endpoints (see Table 1). Physician and nurse investigators who had treated patients with colorectal cancer developed all three measures. The QOL Study Coordinator determined the appropriateness of the symptom items as well as the entire questionnaire for this patient population by conducting a pilot test with ten patients.
The treatment-specific questions address symptoms usually included in the set of physician-rated toxicities; these items are presented individually and not summed for a total score. Treatmentrelated symptom items have five-or six-level response choices ranging from no problem with the symptom to severe problems. For analysis purposes, the three primary symptom endpoints were described as dichotomous variables: no mouth pain or some but not bothersome versus bothersome to unbearable mouth pain; none or occasional diarrhea versus frequent diarrhea (several times a week to several times a day); none or some hand/foot sensitivity versus bothersome to unbearable hand/foot sensitivity. Patients who had a colostomy were omitted from analyses examining the effect of treatment arm on occurrence of diarrhea.
General symptom status. The Symptom Distress Scale was used to assess patient symptom status (McCorkle and Young, 1978;McCorkle and Benoliel, 1983;Young and Longman, 1983;Mc-Corkle, 1987McCorkle et al., 1989. The scale's items address disease-and treatmentrelated symptoms commonly experienced by patients with cancer. McCorkle and colleagues have documented the psychometric properties of the Symptom Distress Scale (McCorkle and Young, 1978;McCorkle and Benoliel, 1983;Young and Longman, 1983;McCorkle, 1987;McCorkle et al., 1989;McCorkle et al., 1998). With permission of the questionnaire developers, two of the 13 items (cough and outlook) were omitted from the measure, leaving 11 items. Internal consistency reliability was 0.84 for the 11-item version of the Symptom Distress Scale administered to this patient sample (n= 287). Higher scores reflect worse symptom status.
Physical and emotional functioning. The Physical Functioning Scale and the Mental Health Index address two basic QOL domains from the General Health Survey Short Form 20 (SF-20) and the 36 (SF-36) developed as part of the Medical Outcomes Study (Stewart et al., 1988(Stewart et al., , 1989Ware and Sherbourne, 1992;McHorney et al., 1993McHorney et al., , 1994 and the Rand Health Insurance Experiment (Ware et al., 1980). The SWOG QOL studies were initiated during the transition between the SF-20 and the SF-36 in consultation with the developers of these questionnaires; this included a decision to change the time frame of the Short Form scales from 1 month to 1 week (there is now a 1-week time frame version of the SF-36) (John E. Ware Jr., 1988Jr., -1990. The SWOG QOL questionnaire also includes the SF-20 Role Functioning Scale, the SF-36 Social Functioning Scale, and the SF-20/SF-36 single item rating of general health. The ten-item Physical Functioning Scale and the five-item Mental Health Index come from the more recent and final version of these scales, the SF-36 Health Survey (Stewart et al., 1989;Ware and Sherbourne, 1992;McHorney et al., 1993). Psychometric properties of these scales have been established (Stewart et al., 1988(Stewart et al., , 1989Ware and Sherbourne, 1992;McHorney et al., 1993). Adequate internal consistency reliability was achieved for the patient sample reported in this paper (n= 287): 0.92 for Physical Functioning; 0.80 for the Mental Health Index. Higher scores reflect better functioning.

Statistical methods
Compliance and missing data. A number of quality control procedures followed in SWOG QOL studies have been described elsewhere . A 70% questionnaire submission rate was considered the lowest acceptable for data analysis. To evaluate compliance, the questionnaire submission rate at each time point was computed as the number of questionnaires submitted at that time point divided by the number of patients alive at that point. The decrease in sample size over time and the reasons for patient dropout were assessed. Average QOL scores at each time were compared for three groups of patients: those patients who completed all assessments, those who failed to complete all assessments due to death or deteriorating health, and those who failed to complete all assessments for other reasons.
Missing data within a QOL scale were handled as follows for Symptom Distress Scale, Physical Functioning Scale, and Mental Health Index total scores. If a patient failed to answer 20% or fewer items for a scale, his/her score for that scale was calculated with the mean of the remaining items (for that patient) substituted for missing items; if the scale had more than 20% of the items missing, the entire scale was considered missing.

Statistical comparisons
QOL treatment arm comparisons were made between combinations of treatment arms varying the route of administration or biochemical modulation at 6, 11, and 21 weeks after randomization (see Table 1). Since five primary endpoints were tested at each of three time points, the probability level for rejecting the null hypothesis at the traditional 5% (two-sided) alpha level was adjusted to 0.003. Power estimated prior to the activation of the QOL study was based on combinations of arms to address the hypotheses described in Table 1 and was deemed to be adequate for the five primary comparisons. The symptom outcome comparisons utilized onetailed tests, while the physical and emotional functioning comparisons involved two-tailed tests. Wilcoxon rank-sum tests were used to examine group differences at each of the three time points for the Physical Functioning Scale and the Mental Health Index. For the three symptom variables, chi-square tests were employed on binary data at each of the time points. To address the effect of actual sample size on ability to detect treatment arm differences, we present 95% confidence intervals (CI) for each proportion or mean estimate. The width of the CI represents the precision of the estimates.

Therapeutic trial results
The therapeutic trial was activated in August 1989 and closed in January 1993 with a total accrual of 620 patients; 599 patients were determined eligible for clinical evaluation.
No significant survival differences were detected when arms 2-7 were compared to standard single-agent bolus 5-FU (24). Patients treated with PALA (arm 7) showed the shortest survival duration. Although survival for the remaining arms did not differ statistically, it was better in the 5-FU infusion regimens. A followup trial with continuous low-dose 5-FU versus intermittent high-dose 5-FU is underway.

QOL accrual
The QOL study opened in October 1990 and closed in January 1993. Two hundred and eighty-seven patients were registered to the QOL companion study for the therapeutic trial. Eight of these patients were later deemed ineligible for the therapeutic trial and so were deleted from the QOL analyses. Of the remaining 279 patients, four did not complete a baseline QOL questionnaire, leaving 275 patients for the QOL analyses. Patient characteristics are summarized in Table 2.  a Other reasons at 6 weeks (n= 6): death or deteriorating health (n = 2); wife took home but patient did not complete (n=1); other (n= 3). Other reasons at 11 weeks (n = 7): deteriorating health (n= 1); off treatment or study (n = 3); out of the country (n = 1); patient refusal (n = 1); unknown (n= 1). Other reasons at 21 weeks (n = 11): death or deteriorating health (n =3); off study (n= 2); out of the country (n=1); patient refusal (n = 1); patient told QOL form not needed (n= 1); unknown (n=3).
with submitted questionnaires and if a questionnaire was not completed by the patient) for missing QOL questionnaires. Institution staff did not always use the categories for missing questionnaires as indicated by the examples of reasons in the 'other-specify' category. If 'other reasons' having to do with deteriorating health are added to the two specific categories of death and deteriorating health in Table 3(C), we have the following percentages for non-ignorable missing data at each time point: 6 weeks (11/49 =22%); 11 weeks (23/72 = 32%); 21 weeks (19/47 = 55%). The percentage of missing forms due to death or deteriorating health would be even higher if we added the 'off-treatment/off study' reasons to the deteriorating category since patients are often taken off treatment if they are not responding or are doing poorly. However, one still notes approximately 24%, 18%, and 10% of the questionnaires missing due to institution error at 6, 11, and 21 weeks, respectively. Figure 2 represents the missing data problem graphically; the 67 patients with intermit- Figure 2. QOL by type and length of follow-up. Six endpoints: mean symptom status total score (2A); % with bothersome stomatitis (2B); % with bothersome diarrhea (2C); % with bothersome hand/foot syndrome (2D); mean physical functioning score (2E); mean emotional functioning score (2). Patients with incomplete data are grouped by last questionnaire submitted: ---(complete QOL data at four time points); -(incomplete data because of death or deteriorating health); ---(incomplete data unrelated to health status).

Copyright © 2000 John Wiley & Sons, Ltd.
Psycho-Oncology 9: 340-354 (2000) tent patterns of missing data are not included in Figure 2. Figure 2(A) portrays the effect of missing data on potential conclusions about symptom status during treatment. The following can only be descriptive in nature, as the informative dropout problem reflected in Figure 2 makes statistical comparisons inappropriate. Figure 2(A) indicates that patients who submitted all assessments had lower (better) Symptom Distress Scale scores at baseline and throughout the follow-up period. The average Symptom Distress Scale score for patients who dropped out of the QOL database due to death or deteriorating health status was worse at baseline than that of patients with complete assessments. This finding is consistent with the ability of the baseline Symptom Distress Scale scores to predict survival as presented below. Figure 2(A) further indicates that on average the Symptom Distress Scale score was worst (highest) at the last assessments taken before these patients dropped out. Average values for patients who had missing data for other reasons (e.g. institution error) looked more like those for patients who submitted complete QOL data. Conclusions about the effect of treatment on patient QOL over time are subject to bias by the tendency of patients with the poorest QOL to drop out. The effect of data missing because of death or deteriorating health was similar for physical and emotional functioning (Figure 2(E) and (F)) but did not appear to be as problematic for the three treatment-related symptom endpoints ( Figure  2(B), (C), and (D)).

Treatment-specific symptoms
Of the hypothesized differences for treatmentspecific symptoms, chi-square tests were significant only for hand/foot sensitivity at 11 (p= 0.003) and 21 (p B 0.001) weeks.
Stomatitis. Significant differences in stomatitis (percent with bothersome or worse symptoms) were not detected for patients with a regimen including leucovorin versus no leucovorin at any of the follow-up assessments, although differences were in the hypothesized direction (i.e. more stomatitis with leucovorin) (see Tables 1 and 4(A)). Figure 2(B) shows no evidence of bias due to incomplete follow-up.
Diarrhea. Significant differences in amount of diarrhea reported by patients receiving treatment by IV push were not detected at any of the follow-up assessments, although differences were again in the hypothesized direction (more diarrhea with IV push) (see Tables 1 and 4(B)). As with stomatitis, Figure 2(C) shows no evidence of bias due to incomplete follow-up.
Hand/Foot sensiti6ity. As hypothesized, patients whose treatment was administered via protracted continuous infusion reported significantly more hand and/or foot sensitivity at 21 weeks than did patients whose treatment was administered as a bolus dose (chi-square tests, pB0.001); the difference was also significant at 11 weeks (p= 0.003) (see Tables 1 and 4(C)). Figure 2(D) shows a slightly increasing percentage of patients reporting troublesome levels of hand/foot syndrome over time for those patients with complete QOL data, and shows little evidence of bias due to incomplete data.

Physical and emotional functioning
Hypothesized differences for physical and emotional functioning were not observed as tested by Wilcoxon rank sum tests. Figure 2(E) and (F) indicate poorer physical and emotional functioning for patients who drop out of the QOL study for health-related reasons relative to those with complete data. This bias prevented the conduct of longitudinal analyses for the emotional and physical functioning variables.
Physical functioning. Patients who received treatment via IV push or bolus dose did not report poorer physical functioning (lower SF-36 Physical Functioning Scale scores) than did patients receiving treatment via infusion (see Tables 1 and 5(A)). Figure 2(E) indicates poorer physical functioning for patients who drop out of the QOL study for health-related reasons relative to those with complete data.
Emotional functioning. Patients administered treatment via protracted continuous infusion reported similar emotional distress (Mental Health Index scores) to patients without a pump (  scores did not detect combined treatment arm differences at any of the time points, the difference was in the hypothesized direction at 21 weeks. Figure 2(F) suggests that patients with data missing due to death or deteriorating health show poorer emotional functioning.

O6erall symptom status
Although there were no specific hypotheses with respect to the total Symptom Distress Scale score, descriptive data for this questionnaire are presented in Table 6 broken down by one agent and two route of administration treatment arm combinations (see Table 1). In general, there were minimal differences. The data suggest better symptom status at the last assessment for patients receiving drug via continuous infusion versus 24-h infusion or IV push route of administration. However, this suggestion must be tempered, given the potential for missing data bias just demonstrated for symptom status data; in addition, a priori hypotheses regarding overall symptom status had not been identified.

Missing data problems in ad6anced-stage disease
This study showed that good quality control procedures can result in good compliance with QOL assessment schedules in cooperative group trials. Submission rates for live patients ranged from 99% at baseline to 78% at 21 weeks. These submission rates were achieved in the cooperative group context where many institutions contribute patients to trials, complicating quality control procedures. In general, the QOL questionnaires were administered on schedule, in the face of  demanding data management requirements associated with clinical trials research. Patients also demonstrated their willingness to complete questionnaires. However, missing data due to death and deteriorating health can make it impossible to conduct of longitudinal analyses due to reductions in sample size and non-ignorable missing data biases. With advanced-stage disease, even the best quality control procedures cannot avoid missing data that are associated with deteriorating patient health and patient death. Our experience in this trial suggests that it is risky to draw firm conclusions about the effect of treatment on patient functioning over time, given the substantial number of missing follow-up assessments. Patients who died or who had deteriorating health status were increasingly not available for analyses over the 21-week QOL assessment period. Using only QOL data for patients who survive and are well enough to complete all questionnaires clearly leads to an overestimate of QOL status and is only informative for patients who survive. When QOL data are missing related to an aspect of the QOL endpoint (i.e. the health status of the patient), the resulting bias is problematic: it compromises our ability to describe patterns of QOL change, to draw conclusions about the maintenance of QOL during the course of treatment, and to document the presence or absence of palliation.

QOL and patient benefit
In this study, we were unable to determine the extent to which the treatment arms differed in their ability to achieve palliation or confer patient benefit (Moinpour et al., 1999). Substantial problems with missing data due to death and deteriorating health complicated our ability to evaluate patient benefit by treatment arm. In addition, confidence intervals indicated a wide range of potential values for the reported measures. For example, the confidence interval for the continuous infusion arms at 21 weeks indicates that the proportion of patients reporting bothersome or worse hand/foot sensitivity could be as high as 44%. Hill et al. (1995), using the comprehensive EORTC QC-30 measure, did not find QOL differences by treatment arm in the advanced-stage disease setting. The Hill et al. trial compared protracted infusion of 5-FU alone or with a biochemical modulator (IFN). Although the biochemical modulator in the Hill study differed from that used in this study, the continuous infusion 5-FU alone arm was the same. It may be that  patients with advanced-stage disease, even those with good performance status, who respond modestly to treatment, are not likely to show significant improvements in broader aspects of day to day functioning such as physical and emotional functioning. That is, the modest clinical outcomes may not translate to significant improvement in ability to carry out normal activities and in emotional outlook. In addition, modest clinical improvements may also be associated with substantial levels of toxicity. Saltz et al. (1999) reported confirmed response rates of 33% for a combined regimen of 5-FU, leucovorin, and CPT-11 compared to confirmed response rates of B20% for 5-FU/leucovorin and for CPT-11 alone. Although QOL was not measured in this study, the authors reported non-trivial problems with grade 3 or 4 diarrhea in the arms containing CPT-11 and grade 4 neutropenia in the arms with 5-FU and leucovorin. It may be that in the ad-vanced-stage disease setting, stability in the patient's report of both symptom status and general QOL dimensions is the best we can expect. However, QOL data have the potential to help us identify patient benefit (or harm) attributable to treatment regimens when response rates and survival are minimally affected. Attempts to compare the toxicity profile reported for the therapeutic trial (Leichman et al., 1995) with that available for the companion QOL study are problematic given that our analyses were based on combinations of arms rather than individual arms as reported in the therapeutic paper. For example, we combined all arms with the modulator, leucovorin, and compared symptom reports to those provided by patients not treated with leucovorin. The direction of the effect was as hypothesized but not significantly different from zero. The report of the therapeutic trial showed significantly more severe stomatitis in Copyright © 2000John Wiley & Sons, Ltd. Psycho-Oncology 9: 340-354 (2000 arms 2 and 5, two of the three arms in our combined comparison. It also reported the highest proportion of severe diarrhea in arm 3 (IV bolus administration of 5-FU plus high dose leucovorin). We reported more diarrhea (not significantly so) in arms 1 -3 versus arms 4 -7. Arms 1 and 2 differed from arm 3 in that arms 1 and 2 administered 5-FU over 5 days versus a 1-day administration in arm 3; the single day administration would be expected to result in more diarrhea.
The Meta-analysis Group in Cancer evaluated administration of 5-FU by bolus versus continuous infusion (Meta-analysis Group in Cancer, 1998); arm 7 and the two arms with high dose leucovorin included in the Southwest Oncology Group trial were omitted. Although the authors of the meta-analysis reported, as we did, significantly more hand/foot sensitivity for continuous infusion compared to other treatment administration routes, its conclusion was that continuous infusion administration of 5-FU was superior to bolus administration for a number of clinical outcomes. A follow-up SWOG trial (SWOG-9420) used continuous infusion of two doses of 5-FU (no QOL data were collected). In a post hoc examination of QOL data for SWOG 8905/9045, we did not observe superiority in QOL when 5-FU alone was administered by continuous infusion compared to the arms where 5-FU was modulated by leucovorin (bolus or infusion).
Patients with complete data had relatively stable QOL (Symptom Distress Scale, Physical Functioning, and Mental Health Index scores) over the QOL study period, unlike those who dropped out of the study as they deteriorated or died. However, information based only on patients providing a complete schedule of assessments is not useful for new patients at the initiation of treatment. As noted above, the inappropriate use of complete case analysis approaches would suggest a palliative effect of treatment. Patients who are not doing well at the last time they contributed data are not represented in such analyses. Their reported QOL would certainly temper conclusions regarding palliation. From a common sense perspective, we would conclude that palliation did not occur for the large number of patients whose condition deteriorated while on treatment. The physician-rated improvements reported by Poon et al. (1989) for patients receiving 5-FU and lowdose leucovorin documented palliation but this research involved a less comprehensive measure of QOL and one that did not include patient selfadministered questionnaires.
Finally, we must remember that the opposite of palliation can occur in treatment trials. It is possible that treatment can worsen patient functioning; for example severe treatment side-effects can negatively affect patient emotional or physical functioning. An example from advanced-stage disease in prostate cancer documented negative impacts on emotional functioning at months 1, 3, and 6 months post-randomization for patients receiving total androgen ablation (orchiectomy plus flutamide) versus treatment with orchiectomy alone (plus placebo) . Hill et al. (1995) noted that transient symptom problems during the trial did not affect broader areas of patient functioning, providing some evidence that the treatments did achieve palliation (i.e. overall patient QOL and functioning were maintained during the treatment period). Symptom Distress Scale scores did not go above 18 for any of the combination of arms over the 21-week assessment period in this trial. Symptom Distress Scale total scores above 24-26 points have been found by the developer of the questionnaire to reflect moderate symptom problems (McCorkle et al., 1998). This would suggest that, although we did not observe differences in symptom palliation by treatment arm, most of the patients treated with these approaches appeared to be maintaining a fairly good level of symptom status.
We believe that until there is a more complete QOL database for patients with advanced-stage disease, it is important to continue to include comprehensive measures of QOL in clinical trials addressing QOL issues. Without a comprehensive measure, we cannot properly evaluate the extent to which palliation objectives are achieved or the full extent of negative effects of treatment such as those found in the advanced-stage prostate cancer trial described above .

CONCLUSIONS
The primary purpose of the QOL companion study was to monitor the effects on patient QOL of agent and route of administration variations in an advanced colorectal cancer treatment trial. Although we comprehensively measured the effect of treatment on QOL, we could not complete the planned analyses due to the presence of non-

Copyright © 2000 John Wiley & Sons, Ltd.
Psycho-Oncology 9: 340-354 (2000) ignorable missing data. Therefore, the primary purpose of this paper became the interpretation of QOL results in the context of missing longitudinal data. Quality control procedures were in place to ensure adequate submission rates for patients who were alive and not experiencing deteriorating health status. However, institution error accounted for 10-24% of missing questionnaires over the follow-up period, indicating that quality control procedures can always be improved. Although we observed adequate submission rates for living patients, these rates did not make up for the decreasing sample size over time and the bias associated with attrition due to death. For example, by the end of the QOL assessment schedule (21 weeks), 46% of the reasons for missing questionnaires were associated with death and deteriorating health. Therefore, longitudinal analyses of the QOL data could not be justified. The challenges associated with QOL assessment in the advanced-stage disease setting have been noted (Bernhard et al., 1998). For this reason, many investigators 'front-load' QOL assessments in order to capture patient functioning before deteriorating health prevents submission of questionnaires. The disadvantage to this approach is the failure to capture longer-term positive effects of treatment on patient QOL as well as the potential for later negative effects. However, in advancedstage disease, we are trying to document support for new therapies where we do not expect to see substantial gains in survival. The ability to differentiate treatment arms with respect to symptom status and general areas of functioning can be important, although some justification of the clinical meaningfulness of differences is required (Lydick and Epstein, 1993;Osoba et al., 1998;Sloan et al., 1998). In this context, a focus on QOL assessment during treatment is probably defensible and may result in the identification of outcomes other than survival that demonstrate patient benefit.