American Journal of Epidemiology Advance Access originally published online on March 22, 2006
American Journal of Epidemiology 2006 163(9):811-817; doi:10.1093/aje/kwj122
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Meta-Analysis |
A Bayesian Meta-analysis of Prophylactic Granulocyte Colony-Stimulating Factor and Granulocyte-Macrophage Colony-Stimulating Factor in Children with Cancer
1 Department of Pediatrics, University of Toronto and The Hospital for Sick Children, Toronto, Ontario, Canada
2 Department of Health Policy Management and Evaluation, University of Toronto, Toronto, Ontario, Canada
3 Department of Public Health Sciences, University of Toronto, Toronto, Ontario, Canada
4 The Institute for Work and Health, Toronto, Ontario, Canada
5 Division of Oncology, Children's Hospital of Philadelphia, Philadelphia, PA
Correspondence to Dr. Lillian Sung, Division of Hematology/Oncology, The Hospital for Sick Children, 555 University Avenue, Toronto, Ontario M5G 1X8, Canada (e-mail: Lillian.sung{at}sickkids.ca).
Received for publication May 16, 2005. Accepted for publication November 29, 2005.
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
The purpose of this analysis was to examine the efficacy of prophylactic hematopoietic colony-stimulating factors (CSFs) in pediatric cancer and to describe how a Bayesian meta-analysis can be conducted and then modified to incorporate information not readily included in a frequentist meta-analysis. Three Bayesian models were developed. The simplest model used the same data as a published frequentist meta-analysis. The second model included data that could not easily be incorporated into the frequentist meta-analysis, including data from different courses of chemotherapy and continuous outcomes that did not report variance estimates. The third model examined the effect of CSF type (granulocyte CSF vs. granulocyte-macrophage CSF). Compared with the frequentist model, the Bayesian model with the most data suggested a greater benefit of CSFs, with a 3.2-day reduction in duration of parenteral antibiotics (95% credible interval: –7.1, 0.7) in the expanded Bayesian model compared with a 0.8-day (95% confidence interval: –2.3, 0.7) reduction in the frequentist model. Bayesian meta-analysis also suggested that, compared with granulocyte-macrophage CSF, granulocyte CSF was associated with a 4.8-day decrease in the duration of parenteral antibiotics. Bayesian meta-analysis can readily include information not easily incorporated in a frequentist meta-analysis. Some treatment effect estimates were larger by a clinically important amount when additional data contributed to the pooled estimate.
Bayes theorem; granulocyte colony-stimulating factor; granulocyte-macrophage colony-stimulating factor; meta-analysis; neoplasms; pediatrics
Abbreviations: CSF, colony-stimulating factor; G-CSF, granulocyte colony-stimulating factor; GM-CSF, granulocyte-macrophage colony-stimulating factor
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Systematic reviews are a common methodology used to synthesize the results of multiple studies. Compared with individual studies, they usually provide improved power and more precise estimates of a treatment effect. These results are typically synthesized by using frequentist (or classical) approaches. However, because study designs may vary and data for some outcomes may be incomplete, it can be difficult to incorporate all available data into a frequentist meta-analysis.
We recently performed a frequentist meta-analysis of prophylactic granulocyte colony-stimulating factor (G-CSF) and granulocyte-macrophage colony-stimulating factor (GM- CSF) in children with cancer (1
). Chemotherapy commonly results in neutropenia (a low neutrophil count), which places patients with cancer at risk for potentially life-threatening infections. Both G-CSF and GM-CSF stimulate the regeneration of neutrophils in the bone marrow. These growth factors are given following completion of chemotherapy with the aim of shortening the duration of neutropenia. Although we knew that colony-stimulating factors (CSFs) could decrease the duration of neutropenia, it was unclear whether CSFs were associated with clinically meaningful outcomes. One such outcome is febrile neutropenia, which is fever that occurs during the period of neutropenia. Febrile neutropenia is important because children with this condition are typically admitted to the hospital and given intravenous (parenteral) antibiotics; some of these children will eventually be found to have a documented infection such as bacteremia (bacteria isolated in the bloodstream) and pneumonia. Consequently, other clinically important outcomes include duration of hospitalization, documented infections, days of intravenous antibiotic administration, and infection-related mortality.
In our previous meta-analysis (1), which used a frequentist approach, we found that CSFs reduced the rate of febrile neutropenia (rate ratio = 0.80, 95 percent confidence interval: 0.67, 0.95; p = 0.01), decreased hospitalization length (weighted mean difference = –1.9, 95 percent confidence interval: –2.7, –1.1 days; p < 0.0001), and reduced the rate of documented infections (rate ratio = 0.78, 95 percent confidence interval: 0.62, 0.97; p = 0.02) when compared with placebo or no therapy. There were no differences in duration of parenteral antibiotic therapy (weighted mean difference = –0.8, 95 percent confidence interval: –2.3, 0.7 days; p = 0.3) or infection-related mortality (rate ratio = 1.02, 95 percent confidence interval: 0.34, 3.06; p = 0.97).
The analysis was complicated by two features of the randomized trials we included. First, some studies applied the assigned intervention after a single randomization to multiple cycles of chemotherapy and reported the data separately for different cycles (e.g., presenting the results after acute lymphoblastic leukemia induction and consolidation therapy separately). Because of nonindependence between results for subsequent cycles, we included data from only the first cycle in the frequentist meta-analysis. Second, some studies reported only an estimate of central tendency for continuous outcomes and did not report an estimate of variance. In the usual frequentist meta-analysis, these studies cannot be included. Both aspects resulted in a meta-analysis that omitted some of the available information for reasons that have no relevance to the question at hand.
Furthermore, two types of CSF products were used in these trials: G-CSF and GM-CSF. When indirect comparison of results stratified by treatment type was used, the classical analysis resulted in no differences in outcomes. However, a clinically meaningful analysis would directly compare the two types of CSF.
In this current meta-analysis, we attempted to overcome these limitations by using a Bayesian analytical approach. In this setting, a Bayesian analysis in essence determines how the results of a meta-analysis change the opinions held before the meta-analysis was conducted. This analysis requires an explicit mathematical expression of what is known about the effect of a treatment before the review (or meta-analysis) is conducted; this information is called the prior probability distribution. The data from the review are represented by the likelihood. The prior probability distribution is then updated by using the likelihood, and the resulting knowledge about the treatment effect is expressed as a posterior probability distribution. From this distribution, one can readily determine useful quantities such as the probability that two treatments differ by some clinically meaningful amount.
Our primary objective was to describe how a Bayesian meta-analysis can be conducted and then modified to incorporate information not readily included in a frequentist meta-analysis. Our secondary objective was to determine whether a Bayesian meta-analysis could shed further clinical insights not apparent from the frequentist meta-analysis.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Review methods
For this meta-analysis, 16 articles were included (2
–17). Studies were included if 1) the population consisted of children (age defined by the individual study) or data were extractable for those less than age 18 years in studies that included adults and children; 2) there was randomization between CSFs and placebo or no therapy; 3) CSFs were given after initiation of chemotherapy, prophylactically, before neutropenia or febrile neutropenia developed; and 4) identical chemotherapy immediately preceded CSF and placebo administration or no therapy. The approach to study selection, data extraction, and assessment of study quality, as well as the results of the frequentist analysis, have been described elsewhere (1). In total, the 16 studies included 1,183 children, 592 of whom were randomly assigned to CSFs and 591 to control arms. Table 1 describes the types of CSF and patient populations in included trials.
|
Statistical analysis
We limited our Bayesian analyses to outcomes reported in at least four studies. Therefore, this meta-analysis focused on six outcomes: three categorical (febrile neutropenia, documented infection, and infection-related mortality) and three continuous (days of neutropenia, hospitalization, and parenteral antibiotic therapy).
We developed three Bayesian models to examine the effect of CSFs. The first and simplest Bayesian model used the same data as the frequentist meta-analysis and served as a comparison. For studies that performed a single randomization and applied the allocated intervention following two cycles of chemotherapy and reported these results separately, the data from only the first cycle were included. In addition, studies that reported an estimate of central tendency for continuous outcomes but did not report an estimate of variance were excluded from the analysis of that outcome. The second Bayesian model incorporated data that had been excluded from the first model. The third Bayesian model examined the effect of CSF type (G-CSF vs. GM-CSF).
Development of the Bayesian models
Direct comparison with frequentist meta-analysis.
For the first Bayesian model, which mirrored the frequentist meta-analysis, we modeled categorical outcomes by assuming that the number of events in a study followed a Poisson distribution, with the natural logarithm of the underlying rate expressed as a regression equation. Continuous outcomes were analyzed by using linear regression. Both regression equations had random effects for the control group mean (intercept) and treatment effect (slope). The intercepts and slopes were assumed to come from a normal distribution with unknown mean and standard deviation, which themselves were given diffuse prior normal and uniform distributions.
Incorporation of additional data.
Data from multiple cycles.
For the second model, we used the following approach for studies that separately reported data following two cycles of chemotherapy. For categorical outcomes, we extended the first Poisson model to account for within-study correlation (clustering) by assuming that both time periods within a study had the same baseline rate, which was a study-specific random effect. For continuous outcomes measured at two time points, we assumed that the two correlated mean values followed a bivariate normal distribution. The correlation between observations in the two periods was not published for any of these studies, so, to fit our model, we had to assume a value. Our starting assumption was that this correlation was 0.5, but we examined the sensitivity of the results to values of 0.25 and 0.75.
Continuous data with missing variance.
We included continuous data that were missing an estimate of variance as follows. On the basis of the usual approach for sample variances, we assumed that each one has a distribution equal to the true study-specific variance times a chi-squared random variable divided by its degrees of freedom (18). We assumed that the true study-specific variances came from a "parent" lognormal distribution with an overall mean and precision. This between-study distribution of true variances was estimated from studies that reported variances and was then used to impute variances for studies that reported an estimate of central tendency but not variance. Web appendix 1 contains more details of this procedure, including the WinBUGS code used for the imputation. (This information from Drs. Tomlinson and Sung at the University of Toronto is described in the first of three supplementary appendixes; each is referred to as "Web appendix" in the text and is posted on the Journal's website (http://aje.oupjournals.org/).) It is important to note that this was a multiple imputation scheme because missing variances were replaced by multiple samples from the parent distribution, and the final results reflect the uncertainty in the imputation.
Incorporation of a treatment-type variable.
The third set of models examined differences between the effects of the two CSF types. If three or more studies reported a particular outcome for a given CSF type, the effect of that CSF type on that outcome was estimated with a random-effects model; otherwise, a fixed-effects model was assumed. In practice, the effect of G-CSF was always estimated by using random effects, and the effect of GM-CSF was estimated by using random or fixed effects, depending on the availability of data. In each case, direct comparisons were made within the same model between estimates of the effects of the two types of CSF. This Bayesian model also included the additional data excluded from the frequentist analysis.
Other statistical issues
We used published guidelines to report elements in our Bayesian analysis (19). For each of the analyses, we used diffuse priors, which allow the results to be driven by the data. Specifically, the prior distributions for the population mean treatment effects were normal, with mean = 0 and variance = 10,000, and the prior distribution for the standard deviations for the random-effect intercepts and slopes were flat relative to the scale of measurement. The lognormal prior distribution for true study variances had a normal (mean = 0, variance = 103) prior on the logarithm of the mean and a uniform (0.01, 100) prior on the standard deviation of the logarithm.
To assess how sensitive our inferences were to the particular prior distributions used, we also repeated the analysis by using skeptical priors (which express skepticism that CSFs are associated with a benefit) for the treatment effect parameter. For continuous outcomes, the prior distribution for the treatment effect was centered at 0, with variance such that there was a 95 percent chance that the treatment effect would be no more than a 2-day reduction in that outcome. For categorical outcomes, the prior distribution for the treatment effect was centered at a rate ratio of 1, with variance such that there was a 95 percent chance that the treatment effect would be no more than a 10 percent reduction in that outcome.
The Bayesian analysis was performed by using WinBUGS version 1.4, in which Markov chain Monte Carlo with Gibbs sampling is used to make inferences. After a burn-in of 5,000 updates, 50,000 iterations were performed and satisfactory convergence was observed by using triplicate runs (20). Examples of the programs used for the meta-analysis are available in Web appendix 1 and at the following website: http://fisher.utstat.toronto.edu/georget/GCSFMetaAnalysis.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Of the 16 randomized trials included, three contained information from multiple cycles of chemotherapy and reported these results separately (7
, 8, 12). Four reported at least one continuous outcome without providing an estimate of variance (7, 11, 13, 16). Table 2 demonstrates that the simplest Bayesian meta-analysis (the middle three columns) and the frequentist meta-analysis, which used the same data, yielded very similar results, albeit with slightly different interpretations. Specifically, there was a 22 percent decrease in the rate of febrile neutropenia associated with CSF administration, and there was a 99 percent probability that, compared with placebo or no therapy, CSFs were associated with less febrile neutropenia. As in the frequentist meta-analysis, the simple Bayesian analysis demonstrated that CSFs were associated with a 4-day reduction in the duration of neutropenia and a 2-day reduction in the duration of hospitalization but no difference in the duration of parenteral antibiotics or in infection-related mortality. In general, the credible intervals from the Bayesian analyses were wider than the corresponding confidence intervals from the frequentist analyses. Presumably, this finding results from the Bayesian models "averaging over" the posterior distribution of the unknown between-study variance parameter, whereas the frequentist models use a single estimate.
|
The three columns on the right side of table 2 demonstrate the impact of including more data in the Bayesian meta-analysis. When studies with two time periods and studies with missing variance estimates were included, the effect of CSFs on reducing febrile neutropenia, duration of neutropenia, and documented infections was similar compared with the frequentist meta-analysis. There was a 99 percent probability that CSFs reduced the rate of febrile neutropenia and a 94 percent probability that CSFs reduced the rate of documented infections. In contrast to the simpler Bayesian model, the model including additional data demonstrated a greater reduction in duration of hospitalization (3.4 days for the expanded model compared with 1.9 days for the simpler model and frequentist analysis) and duration of parenteral antibiotic therapy (3.2 days for the expanded model compared with 0.8 days for the simpler model and frequentist analysis). There was a 95 percent probability that CSFs were associated with a reduction in duration of parenteral antibiotic administration.
Table 3 illustrates the difference in effect between G-CSF and GM-CSF. Compared with GM-CSF, G-CSF appeared to be associated with a greater decrease in the rate of febrile neutropenia, duration of neutropenia, rate of documented infection, and days of parenteral antibiotics. There was a 90 percent probability that G-CSF was more effective than GM-CSF in reducing the rate of febrile neutropenia. The use of G-CSF was associated with a 1.6-day greater decrease in duration of hospitalization compared with GM-CSF, and there was a 68 percent probability that G-CSF was better than GM-CSF with respect to this outcome. The use of G-CSF also was associated with a 4.8-day greater decrease in duration of parenteral antibiotic therapy compared with GM-CSF, and there was a 98 percent probability that G-CSF was better than GM-CSF with respect to this outcome.
|
The results of the sensitivity analyses using correlation coefficients of 0.25 and 0.75 for continuous outcomes reported for two periods were very similar to that in which a correlation coefficient of 0.5 was used (Web appendix 2).
When skeptical priors were used, the evidence for a CSF effect was still convincing. The probabilities that CSF was better than placebo or no therapy were 80 percent for the rate of febrile neutropenia, 99.8 percent for the duration of neutropenia, 91 percent for the duration of hospitalization, 70 percent for the documented infection rate, and 80 percent for the duration of parenteral antibiotic (Web appendix 3).
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
We demonstrated how Bayesian meta-analysis can readily overcome some of the methodological difficulties encountered during a frequentist meta-analysis that arose as a result of heterogeneity in the design of studies and reporting of data. This heterogeneity is a common problem when using secondary data sources. In this particular case, we found that inclusion of the additional information altered some of the conclusions from the more limited analyses.
The results of the frequentist analyses and simplest Bayesian models were virtually identical, which is to be expected because the Bayesian analyses used noninformative priors. Nonetheless, the simpler Bayesian analysis did have an advantage; the Bayesian paradigm makes it possible to express the probability that CSF is better than placebo or no therapy. Although it did not change the results, the Bayesian Poisson approach has a feature that is appealing in contrast to the corresponding frequentist approach. In our frequentist model, computation of variances to use as study weights required adding a small and arbitrary constant to studies with no observed events. Our Bayesian approach models event counts directly by using the Poisson distribution, so zero event counts in studies pose no problem. Finally, all of the Bayesian models, simple and otherwise, share the property that their results are exact (given the model). They do not, unlike the frequentist models for the same data, rely on the asymptotic normality of estimates for inference.
In contrast to the simpler model, several results were different in the expanded Bayesian model that made use of more information. These differences included a greater reduction in the duration of hospitalization and a reduction in parenteral antibiotic administration associated with CSF administration in the expanded Bayesian model. Upon inspection of the data (Web appendix 1), this difference can be attributed at least in part to observations with large differences in the treatment and control groups, in which an estimate of variance was omitted from the publication and could not be retrieved from the authors. Compared with findings from our previous publication (1), these findings suggest that CSFs are associated with greater benefits. However, it also is possible that publications that did not report variances were of poorer quality. If this hypothesis is true, then this current analysis may overestimate the treatment effect.
The third Bayesian model that included a variable for treatment effect suggested that G-CSF is associated with greater benefit than GM-CSF, in particular with respect to the rate of febrile neutropenia and duration of parenteral antibiotic therapy. Since GM-CSF administration itself can be associated with fever (in contrast to G-CSF) (21), this property may explain these differences in treatment effect associated with a different CSF type.
In theory, these analyses all could have been performed by using frequentist methods, although implementation of complicated meta-regression, particularly using random effects and imputation of missing variances, is not straightforward. It does require some statistical expertise to specify the appropriate models in WinBUGS. However, the software does not put unnecessary limitations on the structure of these models, so a model can be built and fitted that is appropriate to the data analysis at hand. We are not aware of an off-the-shelf statistical software package that would allow an analyst to specify and fit a frequentist version of model 2 in Web appendix 1 of this paper, for example. So, although there is nothing inherently Bayesian about imputation of missing variances and inclusion of both univariate and bivariate outcomes in a single meta-analysis, the availability of the WinBUGS software means that a Bayesian analysis may be the best current route to implementing solutions to these problems. We also would argue that the Bayesian hierarchical model makes it easier to conceptualize the solutions. Furthermore, a Bayesian approach enabled us to compute the probability that CSFs were better than placebo or no therapy and the probability that G-CSF was more effective than GM-CSF. These comparisons are clinically meaningful, and the ability to make such probabilistic inferential statements under a Bayesian paradigm could have important clinical implications.
A useful way to think about the posterior probabilities from a Bayesian model is that they are a function of the prior distributions used (22). Different priors will necessarily lead to different posterior probabilities, even if the differences are minor. We have shown that the strong, positive findings for the rate of febrile neutropenia, for duration of neutropenia, and for duration of hospitalization remained relatively strong even with skeptical prior distributions for the treatment effect that put most of their weight on values of little clinical importance.
Previous research compared the findings of Bayesian and frequentist analyses, in which eight meta-analyses were reanalyzed by using a Bayesian approach after an initial frequentist analysis (23). The results of four meta-analyses were similar and four were discrepant, with the Bayesian analyses concluding that the intervention was efficacious in contrast to the frequentist analysis. Our research is somewhat different from these other meta-analyses, since ours demonstrated that the Bayesian analysis of the same data was similar to the frequentist analysis; only upon inclusion of previously excluded data did differences emerge between the Bayesian and frequentist approaches.
In summary, Bayesian meta-analysis can readily overcome some of the methodological difficulties encountered when synthesizing the results of multiple trials. We found that some treatment effect estimates were larger by a clinically important amount when additional data contributed to the pooled estimate.
|
ACKNOWLEDGMENTS |
|---|
The authors thank Elizabeth Uleryk for her expert assistance with the literature searches.
Dr. Sung is currently developing a study with Amgen (Thousand Oaks, California), not related to a colony-stimulating factor. There are no financial associations to disclose.
|
References |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
- Sung L, Nathan PC, Lange B, et al. Prophylactic granulocyte colony-stimulating factor and granulocyte-macrophage colony-stimulating factor decrease febrile neutropenia after chemotherapy in children with cancer: a meta-analysis of randomized controlled trials. J Clin Oncol 2004;22:3350–6.
[Abstract/Free Full Text] - Burdach SE, Muschenich M, Josephs W, et al. Granulocyte-macrophage-colony stimulating factor for prevention of neutropenia and infections in children and adolescents with solid tumors. Results of a prospective randomized study. Cancer 1995;76:510–16.[CrossRef][Medline]
- Calderwood S, Romeyer F, Blanchette V, et al. Concurrent RhGM-CSF does not offset myelosuppression from intensive chemotherapy: randomized placebo-controlled study in childhood acute lymphoblastic leukemia. Am J Hematol 1994;47:27–32.[Medline]
- Channa J, Hashmi KZ. Role of recombinant granulocyte-macrophage colony–stimulating factors in reducing the duration of neutropenia. J Coll Physicians Surg Pak 2002;12:538–41.
- Clarke V, Dunstan FD, Webb DK. Granulocyte colony-stimulating factor ameliorates toxicity of intensification chemotherapy for acute lymphoblastic leukemia. Med Pediatr Oncol 1999;32:331–5.[CrossRef][Medline]
- Dibenedetto SP, Ragusa R, Ippolito AM, et al. Assessment of the value of treatment with granulocyte colony-stimulating factor in children with acute lymphoblastic leukemia: a randomized clinical trial. Eur J Haematol 1995;55:93–6.[Medline]
- Heath JA, Steinherz PG, Altman A, et al. Human granulocyte colony-stimulating factor in children with high-risk acute lymphoblastic leukemia: a Children's Cancer Group Study. J Clin Oncol 2003;21:1612–17.
[Abstract/Free Full Text] - Laver J, Amylon M, Desai S, et al. Randomized trial of r-metHu granulocyte colony-stimulating factor in an intensive treatment for T-cell leukemia and advanced-stage lymphoblastic lymphoma of childhood: a Pediatric Oncology Group pilot study. J Clin Oncol 1998;16:522–6.[Abstract]
- Little MA, Morland B, Chisholm J, et al. A randomised study of prophylactic G-CSF following MRC UKALL XI intensification regimen in childhood ALL and T-NHL. Med Pediatr Oncol 2002;38:98–103.[CrossRef][Medline]
- Michel G, Landman-Parker J, Auclerc MF, et al. Use of recombinant human granulocyte colony-stimulating factor to increase chemotherapy dose-intensity: a randomized trial in very high-risk childhood acute lymphoblastic leukemia. J Clin Oncol 2000;18:1517–24.
[Abstract/Free Full Text] - Michon JM, Hartmann O, Bouffet E, et al. An open-label, multicentre, randomised phase 2 study of recombinant human granulocyte colony-stimulating factor (filgrastim) as an adjunct to combination chemotherapy in paediatric patients with metastatic neuroblastoma. Eur J Cancer 1998;34:1063–9.[Medline]
- Patte C, Laplanche A, Bertozzi AI, et al. Granulocyte colony-stimulating factor in induction treatment of children with non-Hodgkin's lymphoma: a randomized study of the French Society of Pediatric Oncology. J Clin Oncol 2002;20:441–8.
[Abstract/Free Full Text] - Pui CH, Boyett JM, Hughes WT, et al. Human granulocyte colony-stimulating factor after induction chemotherapy in children with acute lymphoblastic leukemia. N Engl J Med 1997;336:1781–7.
[Abstract/Free Full Text] - Riikonen P, Rahiala J, Salonvaara M, et al. Prophylactic administration of granulocyte colony-stimulating factor (filgrastim) after conventional chemotherapy in children with cancer. Stem Cells 1995;13:289–94.[Abstract]
- van Pelt LJ, de Craen AJ, Langeveld NE, et al. Granulocyte-macrophage colony-stimulating factor (GM-CSF) ameliorates chemotherapy-induced neutropenia in children with solid tumors. Pediatr Hematol Oncol 1997;14:539–45.[Medline]
- Welte K, Reiter A, Mempel K, et al. A randomized phase-III study of the efficacy of granulocyte colony-stimulating factor in children with high-risk acute lymphoblastic leukemia. Berlin-Frankfurt-Munster Study Group. Blood 1996;87:3143–50.
[Abstract/Free Full Text] - Wexler LH, Weaver-McClure L, Steinberg SM, et al. Randomized trial of recombinant human granulocyte-macrophage colony-stimulating factor in pediatric patients receiving intensive myelosuppressive chemotherapy. J Clin Oncol 1996;14:901–10.
[Abstract/Free Full Text] - Rosner B. Fundamentals of biostatistics. Pacific Grove, CA: Duxbury Press, 2000.
- Sung L, Hayden J, Greenberg ML, et al. Seven items were identified for inclusion when reporting a Bayesian analysis of a clinical study. J Clin Epidemiol 2005;58:261–8.[Medline]
- Brooks SP, Roberts GO. Convergence assessment techniques for Markov chain Monte Carlo. Stat Comput 1998;8:319–35.[CrossRef]
- Lehrnbecher T, Welte K. Haematopoietic growth factors in children with neutropenia. Br J Haematol 2002;116:28–56.[CrossRef][Medline]
- O'Rourke K. Two cheers for Bayes. Control Clin Trials 1996;17:350–2.[Medline]
- Bloom BS, de Pouvourville N, Libert S. Classic or Bayesian research design and analysis. Does it make a difference? Int J Technol Assess Health Care 2002;18:120–6.[Medline]
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||