|
|||||||||
|
|
|
|
|
|
|
|
|
|
|
CLINICAL OBSERVATIONS, INTERVENTIONS, AND THERAPEUTIC TRIALS
From the Departments of Leukemia and Biostatistics, The
University of Texas MD Anderson Cancer Center, Houston, TX.
We investigated treatment with gemtuzumab ozogamicin (GO) in 51 patients aged 65 years or older with newly diagnosed acute myeloid
leukemia (AML), refectory anemia (RA) with excess of blasts in
transformation, or RA with excess blasts. GO was given in doses of 9 mg/m2 of body-surface area on days 1 and 8 or,
therapeutically equivalently, on days 1 and 15, with or without
interleukin 11 (IL-11; 15 µg/kg per day on days 3 to 28), with
assignment to IL-11 treatment made randomly. Complete remission (CR)
rates were 2 of 26 (8%) for GO without IL-11 and 9 of 25 (36%) for GO
with IL-11. Regression analyses indicated that IL-11 was independently
predictive of CR but not survival. We compared GO with or without IL-11
with idarubicin plus cytosine arabinoside (IA), as previously
administered, in similar patients. The CR rate with IA was 15 of 31 (48%), and survival was superior with IA compared with GO with or
without IL-11 (P = .03). Besides accounting for possible
covariate effects on outcome, we also accounted for possible trial
effects (TEs) arising because IA and GO with or without IL-11 were not
arms of a randomized trial. Bayesian posterior probabilities that GO with or without IL-11 produced longer survival than IA, after accounting for covariates and TEs, were less than 0.01 in patients with
abnormal cytogenetic findings (AC) and less than 0.15 in patients with
normal cytogenetic findings (NC). Regarding CR, the analogous
probabilities were less than 0.02 for GO without IL-11 (all cytogenetic
groups), and for GO with IL-11, less than 0.25 for AC groups and about
0.50 for NC groups. TEs 2 to 5 times the magnitude of those previously
observed would be needed to conclude that survival with GO with or
without IL-11 is likely longer than with IA. Thus, there is little
evidence to suggest that GO with or without IL-11 should be used
instead of IA in older patients with newly diagnosed AML or
myelodysplastic syndrome.
(Blood. 2002;99:4343-4349) On the basis of evidence that it was less toxic
than, but as effective as, standard, cytosine arabinose-containing
regimens in patients 60 years of age or older with acute myeloid
leukemia (AML) in first relapse,1 gemtuzumab ozogamicin
(GO; Mylotarg; Wyeth-Ayerst Pharmaceuticals, Madison, NJ), a
combination of an anti-CD33 antibody and the cytotoxic agent
calicheamicin, has been approved for use in such patients. Because
there is doubt that the benefit-risk ratios associated with such
standard regimens are favorable even in newly diagnosed elderly
patients,2 a trial of GO in older patients with untreated
AML appeared reasonable.
In this study, we compared the complete remission (CR) and
survival rates with GO, adjusted for the effects of covariates such as
performance status and cytogenetic findings, with the rates we
previously observed in similarly aged ( GO trial
The initial 22 patients received 9 mg/m2 of body-surface
area GO on days 1 and 15, provided the marrow specimen obtained on day
14 was more than 10% cellular. Otherwise, the marrow was examined weekly until the response was known. If CR as defined by conventional criteria was not apparent, patients received other therapies. When only
5 of the 22 patients had CR, the remaining 29 patients received 9 mg/m2 GO on days 1 and 8. All patients were randomly
assigned to receive or not receive 15 µg/kg of body weight per day
interleukin-11 (IL-11) on days 3 to 28. Supportive care measures,
including use of a laminar-airflow room (protected environment [PE]),
were described previously.3 In remission, patients
received one course of GO, cyclosporine A, fludarabine, and cytosine
arabinoside (ara-C) alternating every 5 weeks for 10 months with one
course of IA. Doses were 6 mg/m2 GO on day 1; 6 mg/kg
cyclosporine A followed by 16 mg/kg daily on days 1 and 2 (continuous
infusion); 15 mg/m2 fludarabine twice daily on days 2 to 6;
and 0.5 g/m2 ara-C twice daily on days 2 to 6. IA treatment
consisted of 8 mg/m2 idarubicin daily on days 1 and 2 and
1.5 g/m2 ara-C daily on days 1 and 2 (continuous infusion).
IA trial
Statistical methods Design of trial of GO with or without IL-11.
We used the method of Thall and Sung,4 which extends the
method of Thall et al5 to randomized trials. Within each
of the 2 prognostic groups Analysis. Unadjusted survival probabilities were estimated by using the method of Kaplan and Meier.6 Unadjusted between-group comparisons of survival were done with the log-rank test.7 Logistic regression was used to assess the ability of patient characteristics or treatments to predict the probability of CR. The Cox proportional hazards regression model8 was used to assess the ability of patient characteristics or treatments to predict survival, with goodness of fit assessed by the Grambsch-Therneau test, Schoenfeld residual plots, martingale residual plots, and likelihood ratio statistics.9 All scatterplots were smoothed by using the lowess method of Cleveland,10 with predictive variables transformed as appropriate on the basis of these plots. Associations between pairs of binary variables were evaluated with the Fisher exact test. Computations were carried out in Splus11 by using standard Splus functions and the Splus survival analysis package of Therneau.12 The primary limitation of any inferences regarding a GO-IA treatment effect is that GO and IA were not arms of a randomized trial. Previously, we used the methods described in the preceding paragraph to account for prognosis and thus permit conclusions to be drawn about the efficacy of various treatments, with these treatments given in single-arm phase II trials.3,13-17 The underlying assumption was that, despite the sequential nature of these trials, TEs were inconsequential. We have recently begun to doubt this assumption.3 In particular, because GO and IA were studied in separate trials, the GO-IA treatment effect is confounded with the GO-IA TE and neither can be estimated individually from the available data. However, since the overall, confounded GO-IA TE plus treatment effect can be estimated, for a given assumed GO-IA TE, the treatment effect can be found by subtraction. Doing this for each of several reasonable TEs in turn yields corresponding reasonable treatment effects. In both the survival and the logistic regression models, we accounted for the unknown GO-IA (trials 1 and 2) TE by assuming in turn that it was the same as (1) that estimated from 2 separate trials (trials 3 and 4) of fludarabine, IA, and G-CSF (FAIG; enrolling 36 and 24 patients at least 65 years old, respectively13) and (2) that estimated from 2 separate trials (trials 5 and 6) of FAIG plus ATRA (17 and 44 patients at least 65 years old, respectively13). As detailed in the Appendix, we estimated the FAIG and FAIG plus ATRA TEs and, with these, the confounded GO-IA trial-treatment effect by using data from trials 1 to 6, along with the covariate parameters corresponding to performance status, PE, cytogenetic findings and, within the GO trial, the effect of IL-11. The posterior mean for the FAIG TE was ± 0.396 (SD, 0.318) and the posterior mean for the FAIG plus ATRA TE was ± 0.328 (SD, 0.291). The sign of each TE's posterior mean may be either plus or minus. For example, the mean trial 3 versus trial 4 effect here was 0.396. Equivalently, the trial 4 versus trial 3 effect was 0.396. Therefore, both cases (± 0.396)
must be considered. Although it is reasonable to assume that the GO-IA
TE was similar to the historical FAIG and FAIG plus ATRA TEs, it is
also possible that this was not the case. This motivated a sensitivity
analysis in which we varied the unknown GO-IA TE over a wide range of
possible values.
All statistical regression analyses to assess the sensitivity of
possible GO-IA treatment effects to between-trial effects were based on
Bayesian models.18 Consequently, each effect was a random
quantity characterized by a probability distribution, rather than a
single value. Computations of posterior distributions were carried out
in BUGS 0.5.19 Dependence of patient survival on treatment
and prognostic covariates was assessed by using a Weibull survival-time
regression model. Uninformative priors having large variances were
assumed for all parameters. Variable selection was done in a step-down
fashion by computing the posterior probability of a positive effect,
Prob ( > 0), for each covariate, where is the covariate
effect, and dropping any covariate for which this probability was
between 0.10 and 0.90. Variables were dropped one at a time, with the
variable having Prob ( > 0) closest to 0.50 dropped at each
step, until all values of Prob ( > 0) were either more
than 0.90 or less than 0.10. Logistic regression was used to assess the
ability of patient characteristics or treatment to predict the
probability of CR, assuming uninformative priors, and variable
selection was carried out in the same manner as for the survival analysis.
Trial of GO with or without IL-11 The median age of the 51 patients was 71 years (range, 65-89). Seven (14%) were at least partly bedridden (Zubrod performance status 3 or 4). Thirty-seven (73%) had AML, 6 (12%) had RAEB-t, and 8 (16%) had RAEB. An abnormality in blood count documented for at least 1 month before presentation at MD Anderson (antecedent hematologic disorder [AHD]) was present in 35 (69%). Twenty patients (39%) had NC and 31 had AC. Fifteen (29%) had monosomies of chromosomes 5 or 7 (or both) or deletions of the long arms of these chromosomes (or both); these are collectively referred to here as "5/7." The remaining
31% of the patients had other abnormalities.
The CR rate was 11/51 (22%; exact 95% confidence interval [CI],
11%-35%) and was similar in patients given GO on days 1 and 15 and
patients given GO on days 1 and 8 (5 of 22 versus 6 of 29;
P = .86; 95% CI for the true difference in rates,
Given the similarity in results between the day 1 and 8 and day 1 and
15 schedules, we combined the data obtained using these schedules for
the randomized comparison of GO with IL and GO without IL-11. The CR
rate was considerably higher in patients given GO with IL-11 than in
patients given GO without IL-11 (Table 1; 9 of 25 versus 2 of 26; P = .02; 95% CI for the true
difference in rates, 0.07-0.50). The higher CR rate with GO with IL-11
was due to results in the NC group and the group with "other"
chromosome abnormalities (Table 1). The lower CR rate in the group
given GO without IL-11 did not result from an excess of near-CRs
with incomplete platelet recovery in this group, since no such events occurred in either arm. The 23 and 27 days from the start of
treatment required to reach a platelet count above
100 × 109/L in the 2 patients achieving CR in
the group treated with GO without IL-11 was well within the range of
the analogous times in the group given GO with IL-11 (median, 31 days;
range, 14-54 days). The higher CR rate with GO with IL-11 has not yet
translated into a significantly longer survival either in the entire
51-patient group (Figure 1A) or in the
patients with NC (Figure 1B). Any difference between the group given
IL-11 and that not given IL-11 became most apparent only once,
approximately 4 to 6 weeks after the start of therapy, ie, it does not
appear solely to reflect a difference in early death rates.
Although treatment with GO without IL-11 and treatment with GO with
IL-11 were arms of a randomized trial, thus eliminating any possible
TEs in evaluating the effect of IL-11, the number of patients randomly
assigned was small enough that imbalances in important prognostic
covariates could have existed between the 2 arms. To address this
possibility, we used logistic regression and a Weibull survival-time
regression model to assess the relation between probability of CR or
hazard of death and treatment (GO without IL versus GO with IL-11),
age, performance status (Zubrod grade 0-2 versus Zubrod grade 3-4),
cytogenetic findings (normal versus
The distribution of CD33 positivity was similar in patients treated
with GO without IL-11 and those given GO with IL-11
(P = .73), with median values of 94% and 90% of blasts
expressing CD33 in the 2 groups respectively. Likewise, the
distribution of CD33 positivity was similar in patients having and not
having CR in the group given GO with IL-11 (P = .28, with
respective median values of 98% and 88%). The 2 patients achieving CR
in the group given GO without IL-11 had CD33 expression on 100% and 99% of blasts; the median value in the nonresponders was 93%. CR
rates according to morphologic category (AML versus MDS) and percentage
of CD33 positivity ( In terms of the trial's design, the posterior probabilities that, given the CR rates noted above, GO without IL-11 was worse than historical treatments were 0.999 and 0.96 in patients with AC and NC, respectively. The corresponding value for GO with IL-11 in the AC group was 0.96. The probability that GO with IL-11 would reduce the early mortality rate by 0.15 in the NC group was 0.21. These probabilities led to early closure of both arms of the trial in both the AC and NC groups. GO with or without IL-11 compared with IA: a Bayesian sensitivity analysis The patients given GO with or without IL-11 and those given IA were similar in age (Table 4). Poor performance status was most common in patients given GO without IL-11 and least common in those given GO with IL-11. The patients given IA were least likely to have an AHD or to be treated in PE rooms. Patients given IA had a higher CR rate (15 of 31 [48%]) than patients given GO with or without IL-11 (9 of 25 [36%] and 2 of 26 [8%], respectively). The difference in survival (P = .03; Figure 2) reflected deaths occurring beginning about 4 weeks after the start of treatment (Figure 2) and could not be accounted for by differences in survival among patients having CR (with survival dated from the date of CR) or among patients declared to have disease resistant to therapy (with survival dated from the date such resistance was recorded).
To begin the Bayesian sensitivity analyses, Table
5 shows the posterior means, SDs, and
resultant posterior probabilities that the covariates depicted in the
table, including use of GO with or without IL-11 rather than IA, were
associated with shorter survival independently of the other covariates
in the table. Specifically, after accounting for the prognostic
covariates in the table, the posterior probability that the combined
treatment effect and TE of GO with or without IL-11 was inferior to IA
was 0.995. Because the effect estimated here was the confounded GO-IA
trial-treatment effect, however, it is not immediately apparent how
much of this effect was due to treatment and how much was a TE.
Table 6 illustrates the method used to
estimate how much of this confounded GO effect might be attributed to
an effect of treatment. The table is concerned with survival in
patients with a normal karyotype. Starting with the posterior mean,
0.823, of the confounded GO effect shown in Table 5, we subtracted the posterior means of the TEs assumed from the FAIG and FAIG plus ATRA
trials. Because there were 2 TEs and, as explained earlier, the mean of
each could have either a plus or minus sign, there are 4 of these in
all, corresponding to the 4 rows in the table. These were a positive
and a negative effect based on the 2 FAIG trials (± 0.396) and a
positive and a negative effect based on the 2 FAIG plus ATRA trials
(± 0.328). Assuming in turn that each of these 4 possible TE
distributions was the GO-IA TE, the posterior distribution of the GO-IA
treatment effect was derived. From this, the 4 corresponding posterior
probabilities, Prob (
The method used to separate treatment effects and TEs for analysis of
CR was entirely analogous, with the only exception being that here we
also had to account for the IL-11 effect (Table 2). The covariates not
related to treatment were the same as those found predictive of
survival, although with quantitatively different posterior means and
SDs. After accounting for these covariates, the posterior probability
that treatment with GO without IL-11 was associated with a lower CR
rate than treatment with IA was 0.999, and the corresponding
probability for GO with IL-11 was 0.551. Table
8 illustrates the posterior probabilities
for GO without IL-11 and GO with IL-11 in the various cytogenetic
groups, assuming the same 4 types of TEs as in the survival analysis, although again with quantitatively posterior different means and SDs.
In patients with
Thus far, we have assumed that the GO-IA TE was distributed identically
to one of those observed with either the 2 FAIG trials or the 2 FAIG
plus ATRA trials. However, it is also plausible that the magnitude of
the GO-IA TE differed from these previous TEs. Because of this
possibility, we performed sensitivity analyses in which the
distribution of the assumed GO-IA trial effect was varied over a wider
range of possibilities. Figure 3A-C shows the analyses for survival. On the horizontal axis in each figure, we
plotted hypothetical values for the mean GO-IA TE. The solid vertical
lines correspond to the mean TEs assumed earlier, on the basis of the
FAIG and FAIG plus ATRA trials (± 0.396 and ± 0.328, respectively).
The plotted curves are the posterior probabilities that GO with or
without IL-11 is harmful relative to IA (vertical axis) as a function
of the hypothesized mean GO-IA TE. For example, Figure 3A
(normal-karyotype group) shows that the mean TEs 0.396 and 0.328 produced the posterior probabilities of 0.832 and 0.868 noted in Table
7. It can be seen that for GO with or without IL-11 to be superior to
IA, corresponding to a posterior probability of less than 0.5, it would
be necessary to postulate a mean TE greater than 0.82, which is more
than twice the magnitude of that observed in the FAIG or FAIG plus ATRA
trials (Figure 3A, dotted lines). Furthermore, this TE must be in the
"correct" direction, ie, latent covariates must have led patients
given IA to have more, rather than less, favorable prognoses than
patients given GO with or without IL-11. The situation becomes even
more extreme in the analysis of
Standard 3-plus-7 regimens are not only usually unsuccessful in older patients with newly diagnosed AML but frequently produce considerable toxic effects, including those resulting in early death.2 Because of the recent approval of GO for use in elderly patients with relapsed AML, it would not be surprising if many physicians are tempted to use this agent in older patients with newly diagnosed AML. The principal point of this paper is that such use does not appear advisable. In particular, tables 7 and 8 indicate that, under reasonable assumptions, the probability is at least 83% that GO without IL-11 is inferior to our IA regimen with regard to both survival and CR, with the 83% pertaining only to patients with a normal karyotype and only if it is assumed that latent covariates (TEs) made such GO-treated patients have a worse prognosis than otherwise apparent. Although addition of IL-11 to GO improved outcome, this combination appeared to have an effect similar to that of IA only with respect to CR in patients with a normal karyotype. Even in these patients, its effect on survival was very likely unfavorable, and indeed this arm of the trial was stopped early in these patients because the early mortality rate was unlikely to be measurably improved. GO with or without IL-11 yielded particularly worse results compared with IA in patients with AC, especially patients with abnormalities of chromosome 5, 7, or both. Although only small numbers of patients with MDS were treated, results in RAEB and RAEB-t paralleled those in AML (Table 1 and Figure 1A legend). The sensitivity analysis indicated that TEs several times the magnitude of those observed in our previous trials and operating to favor GO would have to be assumed to make it plausible that treatment with IA and treatment with GO with or without IL-11 were equivalent (Figure 3A-C). Of course, it could be argued that there is no reliable method for estimating TEs such as those possibly arising between the trials of IA and of GO with or without IL-11. To postulate such a method is to imply that there is no need to randomize, a position with which we disagree. Indeed, in retrospect, we are not unsympathetic to the view that our randomization should have been between GO and IA rather than between GO with IL-11 and GO without IL-11. We note that follow-up is short and, in particular, relapses have occurred in only 2 of the 11 patients treated with GO with or without IL-11 who had CR. However, 45 of the 82 patients have died (Figure 2), and this prompted us to report these results. It is noteworthy that our CR rate with GO in elderly patients with newly diagnosed disease is lower than the CR rate reported in elderly patients with relapsed AML.1 We speculate that this may reflect the need for patients with relapse to have had a remission lasting at least 3 months. As a result, they may have had better prognoses than our newly diagnosed patients. Certainly, however, the discrepancy suggests that our results may not be generalizable, as does (with respect to the comparison of GO with and without IL-11) the observation (Table 3), which contrasts with previous observations by us and others, that an AHD is associated with longer survival in the data set for GO with or without IL-11. Finally, we emphasize that assignment to the day 1 and 8 and day 1 and 15 schedules for GO was not random, hindering scientific comparisons of these schedules. We have no explanation for the apparently poor results with GO. Although there was no relation between CD33 expression and response, our measurements of CD33, although consistent with those available to most physicians, were relatively crude. Thus, it is conceivable that responders and nonresponders to GO differed in the amount of CD33 expressed per cell, a difference that we would have been unable to detect. Nor did we examine the relation between multidrug resistance 1 (MDR1) status and response, although it has been reported that MDR1 expression unfavorably affects the response rate to GO.21 Lastly, we do not know why addition of IL-11 to GO improved the CR rate. Because this effect appeared unrelated to an effect on platelet recovery, it is possible that IL-11 has an antileukemic effect, a possibility that might be investigated in trials involving other drugs. Despite the uncertainties noted above, we conclude that there is insufficient evidence to warrant use of GO with or without IL-11 in patients 65 years of age or older with newly diagnosed AML, RAEB-t, or RAEB.
We thank Angela Culler for expert secretarial assistance.
Submitted October 22, 2001; accepted February 4, 2002.
Supported in part by a grant from Wyeth-Ayerst Pharmaceuticals.
The publication costs of this article were defrayed in part by page charge payment. Therefore, and solely to indicate this fact, this article is hereby marked "advertisement" in accordance with 18 U.S.C. section 1734.
Reprints: Elihu H. Estey, Department of Leukemia, Box 428, The University of Texas MD Anderson Cancer Center, 1515 Holcombe Boulevard, Houston, TX 77030; e-mail: ehestey{at}mdanderson.org.
1.
Sievers EL, Larson RA, Stadtmauer EA, et al.
Efficacy and safety of gemtuzumab ozogamicin in patients with CD33-positive acute myeloid leukemia in first relapse.
J Clin Oncol.
2001;19:3244-3254
2.
Estey E.
How I treat older patients with AML.
Blood.
2000;96:1670-1673
3.
Estey EH, Thall PF, Cortes J, et al.
Comparison of idarubicin + ara-C-, fludarabine + ara-C-, and topotecan + ara-C-based regimens in treatment of newly diagnosed acute myeloid leukemia, refractory anemia with excess blasts in transformation, or refractory anemia with excess blasts.
Blood.
2001;98:3575-3583 4. Thall PF, Sung H-G. Some extensions and applications of a Bayesian strategy for monitoring multiple outcomes in clinical trials. Stat Med. 1998;17:1563-1580[CrossRef][Medline] [Order article via Infotrieve]. 5. Thall PF, Simon R, Estey EH. New statistical strategy for monitoring safety and efficacy in single-arm clinical trials. J Clin Oncol. 1996;14:296-303[Abstract]. 6. Kaplan EL, Meier P. Nonparametric estimator from incomplete observations. J Am Stat Assoc. 1958;53:457-481[CrossRef]. 7. Mantel N. Evaluation of survival data and two new rank order statistics arising in its consideration. Cancer Chemother Rep. 1966;60:163-170. 8. Cox DR. Regression models and life tables (with discussion). J R Stat Soc B. 1972;34:187-220. 9. Therneau TM, Grambsch P. Modeling Survival Data. New York, NY: Springer; 2000. 10. Cleveland WS. Robust locally weighted regression and smoothing scatterplots. J Am Stat Assoc. 1979;74:829-836[CrossRef]. 11. Venables WN, Ripley BD. Modern Applied Statistics with Splus. 3rd ed. New York, NY: Springer; 1999. 12. Therneau TM. A Package for Survival Analysis in S. Rochester, MN: Mayo Clinic Foundation; 1994.
13.
Estey EH, Thall PF, Pierce S, et al.
Randomized phase II study of fludarabine + cytosine arabinoside + idarubicin ± all-trans retinoic acid ± granulocyte colony-stimulating factor in poor prognosis newly diagnosed acute myeloid leukemia and myelodysplastic syndrome.
Blood.
1999;93:2478-2484
14.
Estey EH, Thall PF, Pierce S, et al.
Treatment of newly diagnosed acute promyelocytic leukemia without cytarabine.
J Clin Oncol.
1997;15:483-490 15. Estey EH, Thall PF, Andreeff M, et al. Use of granulocyte colony-stimulating factor before, during, and after fludarabine plus cytarabine induction therapy of newly diagnosed acute myelogenous leukemia or myelodysplastic syndromes: comparison with fludarabine plus cytarabine without granulocyte colony-stimulating factor. J Clin Oncol. 1994;12:671-678[Abstract].
16.
Estey EH, Thall PF, Kantarjian H, et al.
Treatment of newly diagnosed acute myelogenous leukemia with granulocyte-macrophage colony-stimulating factor (GM-CSF) before and during continuous-infusion high-dose ara-C + daunorubicin: comparison to patients treated without GM-CSF.
Blood.
1992;79:2246-2255
17.
Estey EH, Thall PF, Beran M, Kantarjian H, Pierce S, Keating M.
Effect of diagnosis (refractory anemia with excess blasts, refractory anemia with excess blasts in transformation, or acute myeloid leukemia [AML]) on outcome of AML-type chemotherapy.
Blood.
1997;90:2969-2977 18. Gelman A, Carlin JB, Stern HS, Rubin DB. Bayesian Data Analysis. New York, NY: Chapman Hall; 1995. 19. Spiegelhalter D, Thomas A, Best N, Gilks W. Bayesian Inference Using Gibbs Sampling Manual (BUGS 0.5). Cambridge United Kingdom: Medical Research Council Biostatistics Unit; 1995. 20. Giles FJ, Kantarjian HM, Kornblau SM, et al. Mylotarg (gemtuzumab ozogamicin) therapy is associated with hepatic venoocclusive disease in patients who have not received stem cell transplantation. Cancer. 2001;92:406-413[CrossRef][Medline] [Order article via Infotrieve].
21.
Linenberger ML, Hong T, Flowers D, et al.
Multidrug-resistance phenotype and clinical responses to gemtuzumab ozogamicin.
Blood.
2001;98:988-994
|
Top
Abstract
Introduction
65 years) newly diagnosed
patients given idarubicin plus cytosine arabinoside (IA). Because
IA and GO were not administered in arms of a randomized trial, any
difference between them may reflect not only treatment or a more
favorable configuration of known prognostic covariates in 1 of the 2 trials but also differences between the trials with respect to unknown
("latent") variables. We here refer to the difference between
average patient outcomes due to such latent variables in 2 trials as a
"trial effect" (TE). By definition, the TE accounts for the
difference in average covariate-adjusted outcome observed in 2 trials
of the same treatment. Comparing the outcome rates for 2 treatments
studied in separate trials, such as IA and GO, by adjusting for known
patient covariates by means of statistical regression, yields an
estimated effect that is the sum of the actual treatment effect and the
TE. The treatment effects and TEs are confounded, and neither can be
estimated individually. To deal with this problem, we estimated the
magnitude of the TEs observed in pairs of previous trials at MD
Anderson of the same treatment in patients with newly diagnosed AML or
myelodysplastic syndrome (MDS). We used this historical information to
perform a sensitivity analysis that provides a plausible estimate of
the actual covariate-adjusted GO-versus-IA (GO-IA) treatment effect. Although not a substitute for results from a randomized trial, the
findings of our analyses were sufficiently striking to warrant reporting.
namely, fludarabine with ara-C with or without idarubicin (FA)
or topotecan with ara-C (TA)
