Background and methodology of MONITOR-GCSF, a pharmaco-epidemiological study of the multi-level determinants, predictors, and clinical outcomes of febrile neutropenia prophylaxis with biosimilar granulocyte-colony stimulating factor filgrastim

4.1. Study design 

MONITOR-GCSF is an international, prospective, observational, multi-level, pharmaco-epidemiological study in which chemotherapy-treated cancer patients are started on Zarzio® for the prophylaxis of febrile neutropenia per their prescribing physician's best clinical judgment. Potential centers and physician-investigators are identified by the local affiliate of the study sponsor (Sandoz Biopharmaceuticals, Holzkirchen, Germany) using a Study Briefing document summarizing the study (Fig. 2).

View Large Image Download to PowerPoint

Fig. 2. Study briefing document used for center and physician-investigator recruitment.

The multi-level design of this observational study is warranted for both clinical and statistical reasons. Clinically, it can be assumed that patients under the care of the same physician (or, for that matter, at the same center) are uniquely and exclusively exposed to that physician's knowledge, experience, expertise, and clinical practice patterns. At the center level, patient care may be influenced by clinical policies, procedures, and protocols. Statistically, this exposure to the same physician or center means that these patients are treated with a certain communality that may be different across investigators and centers participating in the study. Extending this across physician-investigators and centers, observations on the total sample of patients are not independent, thus violating a major assumption for statistical testing. In multi-level (or hierarchical linear) modeling, the effect of class (e.g., treating physician) on patient-level outcomes are estimated before the effect of between patient variability is determined; yielding an attribution of variance and identification of level-specific predictors of patient outcomes that separates between-class and within-patient variability.

Both CSF-naive patients and patients previously treated with other CSFs will be screened for eligibility. Those meeting the inclusion and exclusion criteria will be informed about the study and written informed consent will be obtained.

The enrollment period is 36 months. This long period was chosen to allow sufficient time for market entry of Zarzio® in the participating countries and the recruitment of sufficient study centers. Patients will be evaluated over six chemotherapy cycles. The total duration of the study protocol is approximately 42 months. As one aim of the study is to examine the influence of center and physician data, all participating centers and physician-investigators are asked to complete some questionnaires prior to enrolling their first patient.

Fig. 3 summarizes the main procedures of the MONITOR-GCSF study. Table 2 presents the study's timetable and the assessments to be performed at each time point.

View Large Image Download to PowerPoint

Fig. 3. Summary of study procedures.

Table 2. Timetable and assessment schedule of the MONITOR-GCSF study.

	Visit	1	2–5	6
	Chemotherapy cycle	1	2 through 5	6
	Inclusion/exclusion criteria	X
	Informed consent/patient permission	X
	Demographic data	X
	Height and weight	X
	Body surface area	X
	Medical history	X
	Cancer history	X
	History of repeated infections	X
	Prior cancer treatments (chemo- and non-chemotherapy)	X
	Cancer histology	X
	Chemotherapy-induced neutropenia (febrile and afebrile) history	X
	Prior treatments for chemotherapy-induced (febrile and afebrile) neutropenia	X
	Cancer stage	X
	Chemotherapy regimen (type, dose, frequency, NCI toxicity index)	X	X	X
	Changes in chemotherapy regimen (incl. dose reduction and dose delays)	X	X	X
	Radiotherapy regimen	X	X	X
	Changes in radiotherapy regimen (incl. delays, dose reductions, and cancellations)	X	X	X
	Surgery	X	X	X
	Changes in surgery (incl. delays, cancellations)	X	X	X
	Performance status (ECOG or Karnovsky)	X	X	X
	Recent or current infections	X	X	X
	Febrile neutropenia episode (start, end, duration, temperature, hospitalization, treatments)	X	X	X
	Antibiotic prophylaxis	X	X	X
	G-CSF prophylaxis	X	X	X
	Blood and urine cultures	X	X	X
	If febrile, reason for fever	X	X	X
	Target ANC for Zarzio® treatment	X	X	X
	Zarzio® dose and frequency	X	X	X
	Concomitant Antibiotics, antivirals, antifungals, corticosteroids, and analgesics	X	X	X
	Absolute neutrophil count (ANC), complete blood count (CBC), C-reactive protein, serum albumin (incl. assessment of anemia, leukocytosis, thrombocytopenia, …)	X	X	X
	Bone, joint, and muscle pain	X	X	X
	Headache/neurological symptoms	X	X	X
	Epistaxis, skin hemorrhage, gastro-intestinal, other bleeding	X	X	X
	Adherence assessment	X	X	X
	Physician's questionnaire (demographics, knowledge of neutropenia guidelines)	X
	Center characteristics, neutropenia protocols, adoption of neutropenia guidelines	X

4.2. Study populations 

The MONITOR-GCSF study is a pan-European study to which Austria, Belgium, France, Germany, Italy, Poland, Spain, and the United Kingdom have already committed (and assured the required sample size). Additional countries may join as Zarzio® market entry expands in the coming years.

The patient inclusion criteria are the following:

Not eligible will be patients meeting any of the following exclusion criteria:

While patients are the main study population, the multi-level design of the study also requires a minimum number of physician-investigators. At least 1000 patients (allowing for a 25% attrition rate) are to be recruited by a minimum of 75 physicians to achieve the statistical power needed to meet the study objectives. To enable the greatest possible diversity in physician-investigators so as to better reflect the heterogeneity of providers, physician criteria are limited to any person licensed to practice medicine in his/her country of origin and practicing, at least part-time, in oncology and/or hematology.

4.3. Data collection and management 

Being an observational study, all data will be recorded as available. There are no mandatory treatment regimens, assessments, and tests. Investigators enter data from source documents into the electronic CRF by the investigator (see Fig. 4 for an eCRF screen shot). Centralized monitoring will be performed by a contract research organization, which will also assure query management. To the centers’ knowledge, a random selection of 10% of patients will be identified for complete on-site monitoring.

View Large Image Download to PowerPoint

Fig. 4. Example of a webpage for electronic data collection.

4.4. Sample size calculations 

Because the diversity of this study's objectives and the associated diversity in statistical analyses, several sample size calculations were performed and subsequently integrated into the recommended sample sizes for, respectively, patients and practices. All calculations assume a desired power level of 0.80, alpha at 0.05 and two-tailed tests to detect a small effect size.

For Objectives 1, 2, 4, and 5: These are descriptive objectives on available data as generated by this study. The aim is mere description without inferences to populations at large. No sample size calculations based on either power or precision apply.

For Objective 3: The first research questions associated with this objective are descriptive; thus, no sample size calculations based on either power or precision apply. Using binary independent variables (congruent vs. non-congruent), a sample size of 800 observations (of which 66% respond and 34% do not respond to treatment) achieves 81% power to detect a minimal change in the probability of non-response from the baseline value of 0.340 to 0.442. This change corresponds to a minimal detectable odds ratio of 1.540. A two-sided log-rank test with an overall sample size of 800 subjects achieves 81% power to detect the difference in survival proportions between 0.76 and 0.67, the proportions surviving in two groups of the same size. This corresponds to a minimal detectable hazard ratio [HR] of 1.459. Using continuous independent variables (e.g. level of congruence) a sample size of 800 observations achieves 81% power to detect a minimal change in the probability of non-response from the value of 0.340 at the mean of X to 0.392 when X is increased to one standard deviation above the mean. This change corresponds to a minimal detectable odds ratio of 1.250. A minimal detectable log HR on a covariate with a standard deviation of 1.75 based on a sample of 800 observations achieves 82% power to detect minimal regression coefficients equal to ±0.125 (HR ≥1.133 or ≤0.882), in a multivariate Cox proportional hazards model with an expected model covariance of 0.10. ANC as a continuous variable will also be considered. We assume that up to 15 variables may be entered into the models and that small increments in R2 of 0.025 must be detectable. A sample size of 747 would be required to conduct a multiple regression procedure based on these assumptions.

For Objective 6: In the absence of formal sample size estimation procedures for multi-level modeling, we consider two approaches for sample size estimation for regression, one with a continuous (e.g., ANC-based outcomes; multiple regression) and one with a discrete criterion variable (e.g., ANC targets reached; logistic regression). In the case of the continuous criterion variable, we assume that up to 15 variables may be entered into the models and that small increments in R2 of .025 must be detectable. A sample size of 747 would be required to conduct a multiple regression procedure based on these assumptions. In the case of the discrete criterion variable, we conservatively assumed that analysis should be able to detect an odds ratio of 1.50, that the prevalence of the binary response with no exposure was random (0.50), and that the odds ratio for treatment with Zarzio® is 1 (as all patients would be treated with Zarzio®). Under these assumptions, a sample size of 774 patients would be required to achieve power of 0.80 at alpha=0.05.

Since this study is also focused on the center-level determinants associated with clinical and safety outcomes, we estimated the number of centers that must be recruited to allow inferences at that level. Specific to sample size estimations, we assumed that center-level variables would be secondary to patient-level variables as determinants/predictors of ANC and safety outcomes, and therefore that only robust determinants should be retained in modeling efforts. To this effect, we assumed that up to 5 variables may be entered into the models and that increments in R2 of 0.15 would be meaningful and must be detectable. In the case of the continuous criterion variable, a minimum of 75 centers would be required to conduct a multiple regression procedure. In the case of the discrete criterion variable, we conservatively assumed that analysis should be able to detect an odds ratio of 3.5 for a binary and 2.5 for a continuous independent variable. These ratios are higher than in the patient calculations, however we argue that only solid center-level differences are of interest, not minimal variations between centers. We assumed that the prevalence of the binary response with no exposure was random (0.50). Under these assumptions, a minimal sample size of 75 centers would be required to achieve power of 0.80 at alpha=0.05. Using established techniques [26], [27], a two-sided log rank test with an overall sample size of 800 subjects achieves 81% power to detect the difference in survival proportions between 0.76 and 0.67, the proportions surviving in two groups of the same size. This corresponds to a HR of 1.459. A minimal detectable log HR on a covariate with a standard deviation of 1.75 based on a sample of 800 observations achieves 82% power to detect minimal regression coefficients equal to ±0.125 (HR ≥1.133 or ≤0.882), in a multivariate Cox proportional hazards model with an expected model covariance of 0.10.

For Objective 7: There are no formal sample size estimation procedures for clustering and related data mining techniques, except the expectation that sample sizes be “as large as possible.” Assuming that the data model for this study includes X physician level data, and Y patient data, the total would be V=X+Y variables, if all variables were included in the analysis. As the latter is unlikely, we can set V<X+Y. Further, it will be important to test whether profiles/cohorts identified by the above techniques are statistically different from each other on key parameters. Sample size estimation procedures for comparing two or more groups were applied to determine minimum sample sizes for comparing two, three, four, and five cluster solutions. For univariate comparisons of two clusters, a sample size of 394 patients per cluster (total N=788) would be required to detect a small standardized effect size of f=0.20, with a sample standard deviation of 2.0 following the non-central t distribution. For univariate comparisons among three clusters, a sample size of 257 patients per cluster (total N=741) would be required to detect a small standardized effect size of f=0.114, with a sample standard deviation of 1.75 following the non-central F distribution. For univariate comparisons among four clusters, a sample size of 155 patients per cluster (total N=620) would be required to detect a small standardized effect size of f=0.133, with a sample standard deviation of 1.50 following the non-central F distribution. For univariate comparisons among five clusters, a sample size of 136 patients per cluster (total N=680) would be required to detect a small standardized effect size of f=0.133, with a sample standard deviation of 1.50 following the non-central F distribution.

For Objectives 8, 9, and 10: A logistic regression of a binary response variable (target ANC achieved) on a binary independent variable (X) with a sample size of 800 observations (of which 66% achieve target ANC and 34% are non-responders) achieves 81% power to detect a minimal change in the probability of non-response from the baseline value of 0.340 to 0.442. This change corresponds to a minimal detectable odds ratio of 1.540. A logistic regression of a binary response variable (target ANC achieved) on a continuous, normally distributed variable (X) with a sample size of 800 observations achieves 81% power to detect a minimal change in the probability of non-response from the value of 0.340 at the mean of X to 0.392 when X is increased to one standard deviation above the mean. This change corresponds to a minimal detectable odds ratio of 1.250. A two-sided log-rank test with an overall sample size of 800 subjects achieves 81% power to detect the difference in survival proportions between 0.76 and 0.67, the proportions surviving in two groups of the same size. This corresponds to a minimal detectable HR of 1.459. A minimal detectable log HR on a covariate with a standard deviation of 1.75 based on a sample of 800 observations achieves 82% power to detect minimal regression coefficients equal to ±0.125 (HR ≥1.133 or ≤0.882), in a multivariate Cox proportional hazards model with an expected model covariance of 0.10.

Summary and recommendation: Required patient sample size estimates range from 439 (for two-sample t-tests) to 800 (for time-to-event and non-responder analyses). Taking the higher estimates, we propose to add a 25% margin for patient attrition, thus raising the required sample size to a minimum of 1000 patients to be recruited from at least 75 centers. Each center will be asked to enroll up to 15 patients.

4.5. Statistical analysis 

Objectives 1 and 2: Descriptive statistics of frequency, central tendency, and dispersion will be used under consideration of applicable levels of measurement. Secondary correlational/associative analyses may be performed if descriptive analyses suggest the presence of associations among variables. In that case, associative analyses will be performed under consideration of the levels of measurement of the variables involved. Differences in treatment by patient subgroups will be evaluated using Fisher's exact, χ2, Student's t, analysis of variance, Mann–Whitney U, or Kruskal–Wallis tests where appropriate.

Objective 3: Descriptive statistics of frequency, central tendency, and dispersion will be used under consideration of applicable levels of measurement. Secondary correlational/associative analyses may be performed if descriptive analyses suggest the presence of associations among variables. In that case, associative analyses will be performed under consideration of the levels of measurement of the variables involved. Hierarchical modeling will be used to determine if treatment congruence is associated with differences in clinical and safety outcomes.

Objective 4: In addition to appropriate descriptive statistics as above, an index of intra-individual variation in ANC (“Intra-ANC”) levels will be calculated following Nesselroade and Salthouse's [28] proposal to calculate the standard deviation of a subject's multiple ANC measurements. While one could also use the variance, the standard deviation approach yields a metric scaled on these original data and respective means. One-way (within-subjects) repeated measures analysis of variance and factorial (within-subjects and between-subjects) using the Geisser-Greenhouse corrected F-test will be used to examine this objective. In addition, though in an exploratory fashion and data permitting, we will attempt to perform time series analysis under various scenarios of autoregressive lag to explore the presence of trend effects in the 6-cycle data.

Objective 5: Proportions, median follow-up duration, median event-free survival times, and events per patient year will be used to describe the safety endpoints of hospitalization and all-cause mortality over the entire study period (0–6 cycles). The distribution of events and event-free survival will be described for the entire sample using Kaplan–Meier estimator modeling. Mantel–Cox log-rank tests (weighing all time points equally) or generalized Wilcoxon–Breslow tests (weighing time points relative to the number of cases at risk at each time point) will be used to evaluate bivariate differences in survival distribution between identified subgroups. Cox proportional hazards modeling will be used to analyze the multivariate effects of identified risk factors and other determinants on survival under the proportional hazards assumption. Adjusted HRs and 95% confidence intervals will be calculated to test the direction and strength of the influence of individual factors on event-risk and event-free survival. Omnibus tests of model coefficients, −2log likelihood and Nagelkerke pseudo R2 also will be calculated to determine variable block and overall model fit and significance.

Objective 6: We will apply hierarchical modeling to test the relationship of patient- and physician/center-level variables and treatment response. This is a two-level model and consists of two sub-models, one for each level. In this study, patients are nested within centers and physicians, and two-level models will be used for patient/center and patient/physician. We will apply this method in multivariate linear regression and logistic regression models of ANC outcomes in which slope coefficients or odds ratios and 95% confidence intervals will be calculated. Adjusted R2 or Nagelkerke pseudo R2 will be calculated where appropriate. The intraclass correlation coefficient (range 0.00–1.00) will be used to quantify the variability in patient outcome attributable to within-center or within-physician variability before any patient-level determinants are considered. We will also apply the above method in Cox proportional hazards modeling (calculating adjusted hazards ratios and 95% confidence intervals) of safety outcomes. In addition, in an exploratory fashion and data permitting, time-dependent covariates will be assessed in multivariate Cox proportional hazards models, and we will evaluate the multi-level determinants of the count of overnight hospitalizations using hierarchical Poisson regression modeling, calculating incident rate ratios and 95% confidence intervals.

Objective 7: To identify cohorts (“clusters”) we will apply sequentially a series of increasingly more complex procedures: aggregation techniques (“classification and clustering methods”), differentiation techniques (“tree-based models”), and associative and pattern recognition techniques (e.g., “neural networks”). Aggregation procedures will be done on the sample as a whole. However, if we need to migrate to differentiation and associative/pattern recognition procedures, we will randomly divide the sample into a training set (for model development) and a testing set (for model validation). Once (two or more) cohorts have been identified, we will use both univariate and multivariate statistical significance testing procedures to measure differences between the two or more groups on key differentiating variables. In the event of 2 cohorts, the univariate test will be the Student's t-test for independent samples and the multivariate test will be the Hotelling's T2. In the event of 3 or more cohorts, the univariate test will be analysis of variance and the multivariate will be multivariate analysis of variance. As some key differentiating variables may be at ordinal or nominal levels of measurement, appropriate adjustments will be made in the choice of statistical models – recognizing that multivariate comparisons may not be possible.

Objectives 8, 9, and 10: Multivariate logistic regression, Kaplan–Meier estimation, and Cox proportional hazards modeling will be used to address these objectives.

4.6. Interim analyses 

In addition to the end-of-study analyses, interim analyses are planned after patients enrolled in the first 3 months of the enrollment period have completed 6 cycles of follow-up (approx. month 9); and 12 (approx. month 21) and 24 (approx. month 33) months later, using a Pocock-adjusted level of statistical significance [29], [30].