`
`RESEARCH
`
`APPLICA TION NUMBER:
`
`2 1 -8 9 7
`
`STATISTICAL REVIEWg S)
`
`
`
`Office of Biostatistics
`
`US. Department of Health and Human Services
`Food and Drug Administration
`Center for Drug Evaluation and Research
`Office of Pharmacoepidemiology and Statistical Science
`
`STATISTICAL REVIEW AND EVALUATION
`
`CLINICAL STUDIES
`
`NDA/Serial Number:
`
`21-897
`
`Drug Name:
`Indication(s):
`
`I
`
`Medisorb Naltrexonc
`Treatment of alcohol dependence
`
`Applicant:
`
`Date(s):
`
`Alkermes, Incorporated
`
`Received 03/31/05; user fee (6 months) 09/30/05; extended date
`(major amendment) 12/3 0/05;
`
`Review Priority:
`
`Priority
`
`Biometrics Division:
`
`Division of Biometrics II
`
`Statistical Reviewer:
`
`Dionne L. Price, Ph.D.
`
`Concurring Reviewers: Thomas J. Permutt, Ph.D.
`
`Medical Division:
`
`Division of Anesthesia, Analgesia, and Rheumatology Products
`
`Clinical Team:
`
`Mwango Kashoki, M.D.
`
`Celia Winchell, M.D.
`
`Project Manager:
`
`Lisa Basham-Cruz
`
`Keywords: NDA review, clinical studies
`
`
`
`Table of Contents
`
`1
`
`EXECUTIVE SUMMARY .................................................................................................................................3
`
`1.1
`1.2
`1.3
`
`CONCLUSIONS AND RECOMMENDATIONS ....................................................................................................... 3
`BRIEF OVERVIEW OF CLINICAL STUDIES ........................................................................................................ 3
`STATISTICAL ISSUES AND FINDINGS ............................................................................................................... 4
`
`2.
`
`INTRODUCTION ............................................................................................................................................... 6
`
`2.1
`
`OVERVIEW ...................................................................................................................................................... 6
`
`DATA SOURCES .....................................................................‘......................................................................... 6
`2.2
`STATISTICAL EVALUATION ........................................................................................................................7
`
`3.
`
`3 .1
`3.2
`
`EVALUATION OF EFFICACY ...........................................................................‘................................................. 7
`EVALUATION OF SAFETY .............................................................................................................................. 20
`
`4.
`
`FINDINGS IN SPECIAL/SUBGROUP POPULATIONS ............................................................................ 21
`
`4.1
`4.2
`
`GENDER, RACE AND AGE ...................................-.......................................................................................... 2 1
`OTHER SPECIAL/SUBGROUP POPULATIONS .................................................................................................. 21
`
`5.
`
`SUMMARY AND CONCLUSIONS ................................................................................................................23
`
`5.1
`5.2
`
`STATISTICAL ISSUES AND COLLECTIVE EVIDENCE ....................................................................................... 23
`CONCLUSIONS AND RECOMMENDATIONS ............................................................................................'.........24
`
`5.2.1 Labeling ...................................................................................................................................................... 24
`
`APPENDICES ............................................................................................................................................................ 27
`
`DYNAMIC RANDOMIZATION ALGORITHM ..........................................................'....................................................... 28
`SUMMARY OF DEMOGRAPHIC AND BASELINE CHARACTERISTICS ............................................................................ 31
`STATISTICAL FORMULATION OF MODEL .................................................................................................................. 35
`RE-RANDOMIZATION ANALYSIS RESULTS .................................................................................................................37
`EXPLORATION OF PATIENT DISCONTINUATION ..................................................................................................... 3 8
`
`ANALYSES REPEATED USING ALTERNATE DEFINITION OF A HEAVY DRINKING DAY .............................................. 40
`
`
`
`1 EXECUTIVE SUMMARY '
`
`1.1 Conclusions and Recommendations
`
`Alkermes, Incorporated proposes Medisorb Naltrexone for the treatment of alcohol dependence.
`Based on an evaluation of the event rate of heavy drinking over 24 weeks, the applicant claims
`that Medisorb Naltrexone 380 mg reduces heavy drinking. My review of the statistical evidence
`suggests support for the claim. However, I believe that several additional factors warrant
`consideration when assessing Medisorb Naltrexone. First, protocol Violations were identified at
`two of the three sites inspected by the Division of Scientific Investigations. Alkermes’ failure to
`identify these violations prior to the submission of the NDA diminished my confidence in the
`overall conduct of the study and resulting data. Furthermore, analyses of the data including and
`excluding the sites with violations resulted in inconsistent findings further adding to my concern.
`Since support for Medisorb Naltrexone was derived from a single study, there was no replication
`of the findings to provide additional assurance. Lastly, multiple safety concerns, such as
`elevated transaminases and severe allergic reactions, were identified by the review team. While
`there is statistical evidence that the drug is active, the previously mentioned factors must be
`assessed collectively by the review team in order to evaluate the risks and benefits of Medisorb
`Naltrexone. In my opinion, this task is further complicated by the uncertainty surrounding the
`overall conduct of the study and resulting data.
`
`I 1.2 Brief Overview of Clinical Studies
`
`Oral naltrexone is approved for the management of alcohol dependence. Alkermes proposes an
`injectable depot formulation of naltrexone, namely Medisorb Naltrexone. The applicant asserts
`that Medisorb Naltrexone provides continued exposure for at least a month and may reduce the
`potential for hepatotoxicity associated with the oral formulation. The drug was introduced to the
`Division of Anesthesia, Analgesia, and Rheumatology Products via IND 61,138. The clinical
`development plan, endpoints, and statistical analyses were discussed during several meetings
`between the applicant and the division.
`’
`
`Prior to submission of the NDA, the applicant sought input from the division regarding the
`needed number of studies. At that time, the applicant proposed a single study to support the use
`of the drug. The division stated that two adequate and well-controlled studies were necessary
`unless the application was submitted under Section 505(b)(2) of the Food, Drug, and Cosmetic
`Act. On 31 March 2005, Alkermes submitted NDA 21-897 (pursuant to Section 505(b)(2)) in
`support of Medisorb Naltrexone. The application included a single, double-blind, placebo-
`controlled, multi-center study and relied on the agency’s previous findings of efficacy for oral
`naltrexone. In the study, patients were randomized to intramuscular injections of Medisorb
`Naltrexone 190 mg, Medisorb Naltrexone 380 mg, or placebo. Patients randomized to placebo
`received a matching volume of Medisorb microspheres (i.e. 2 mL or 4 mL) without naltrexone.
`
`
`
`Moreover, patients were allocated to treatment for balance On four baseline characteristics using
`a dynamic randomization scheme. Treatment was administered, along with biopsychosocial
`support therapy (using the BRENDA approach), during clinic visits occurring every four weeks
`for the duration of 24 weeks. Patients recorded their alcohol consumption using the timeline
`follow-back method (TLFB). The primary measure of efficacy was the event rate of heavy
`drinking over 24 weeks of treatment where a heavy drinking day was defined as a day on which
`a man consumed at least five drinks or a woman consumed at least four drinks. The applicant
`defined the event rate as the number of heavy drinking days divided by the number of days at
`risk for heavy drinking. Additionally, an alcoholic drink was defined as 13.6 grams of absolute
`ethanol. The applicant employed a stratified Andersen-Gill model for the primary analysis.
`
`1.3 Statistical Issues and Findings
`
`Since the event of interest (i.e. heavy drinking) could potentially occur on multiple days, the
`applicant employed an Andersen-Gill model to assess the overall effect of treatment. In general,
`the results produced by the model may be influenced by the non-proportionality of the hazard
`fimctions and/or by patient withdrawal that is treatment related. Thus prior to the submission of
`the NDA, the Division recommended that the applicant consider and propose methodology for
`use in the event that the proportional hazards (PH) assumption was seriously violated.
`Moreover, the applicant was urged to conduct a re-randomization test to validate the model
`inferences. The division additionally suggested the applicant justify and specify how missing
`data would be handled. To address the former recommendation regarding the PH assumption,
`the applicant used a stratified Andersen-Gill model. According to the applicant, “A stratified
`analysis adjusted for different baseline ‘hazards’ of the prespecified stratification factors. In this
`way, the treatment effect was not subject to the distortion that a covariate-by-time interaction
`would induce by inclusion of such a covariate in the model.” The applicant additionally
`proposed a nonparametric Wilcoxon test as an alternative method of analysis if the PH
`assumption was violated. Alkerrnes formally tested the assumption by inclusion of an interaction
`term in the model. To address potential missing data concerns, Alkennes assessed the
`randomness of the missing data via evaluations of the event rate of heavy drinking by the number .
`of doses received, the Kaplan-Meier curves, and a pattern mixture model.
`
`According to the applicant, there was evidence of a severe violation of the proportional hazards
`assumption, both overall and for some strata. Additionally, the applicant stated that the re-
`randomization test based on the stratified Andersen-Gill model produced unstable results
`because of the small sizes of some of the strata. Based on the evaluation of drop-outs, the
`applicant concluded that study discontinuations were comparable across treatment groups and
`were therefore less likely to affect conclusions.
`I was not convinced that the violation of the
`proportional hazards could be ignored, nor was I convinced that the missing data occurred
`randomly. Thus, I focused significant attention on the nonparametric analysis. The
`nonparametric analysis conducted by the applicant essentially employed a last observation
`carried forward strategy for missing data. Since I had some concern regarding the possibility
`that patients withdrew for treatment-related reasons, I performed an additional analysis imputing
`heavy drinking days for all missing data days. My collective evaluation of the analyses and
`results suggested the existence of a treatment effect for the 380 mg dose of Medisorb Naltrexone.
`4
`
`
`
`The treatment effect was additionally explored via responder analyses. The applicant conducted
`a series of analyses exploring varying ‘categories’ of responders. Patients were classified into
`the following response categories: zero heavy drinking days per month, up to one heavy drinking
`day per month, up to two heavy drinking days per month, up to three heavy drinking days per
`month, and up to four heavy drinking days per month. The analyses provided some additional
`evidence of an effect. However, the analyses also raised questions regarding the clinical
`interpretation and meaningfulness of a reduction in the number of heavy drinking days among
`the population under study. These issues will be addressed in the medical review of Dr. Mwango
`Kashoki. To further explore the effects of the treatment, I conducted responder analyses on the
`subgroups of patients abstinent and non-abstinent at baseline. The response profile among the
`two subgroups suggested that a response to treatment was more likely to occur among patients
`abstinent at baseline.
`'
`
`An additional statistical concern was the appropriateness of pooling the placebo groups. The
`applicant contrasted the analysis based on the pooled placebo groups with the analysis
`considering separate placebo groups. Additionally, the applicant repeated the primary analysis
`exploring the treatment differences between the 4 mL placebo and 2 mL placebo groups. The
`results were consistent for pooled analyses and analyses with separate placebo groups.
`
`During the course of the review, the Division of Scientific Investigations identified various
`protocol violations affecting data collection at two sites. In response, the Division of Anesthesia,
`Analgesia, and Rheumatology Products subsequently requested that the applicant reanalyze the
`data excluding the sites. Since a stratified dynamic randomization scheme was used to allocate
`patients to treatment, I was uncertain about the validity of the model-based inferences when
`excluding data from the two sites. Thus, I also requested that the applicant use re-randomization
`tests to verify the results. Alkermes performed the requested analyses and concluded that the
`supplemental analyses confirmed the efficacy of Medisorb Naltrexone 380 mg. Alkermes
`maintained that the protocol violations did not affect the study blind. They additiOnally stated,
`“It is unlikely that the protocol deviations pertaining to the separation of roles — between the
`BRENDA therapist and the time line follow back collector — introduced bias into the study.” For
`these reasons, the applicant strongly believed that the data from the excluded sites should be
`included in the final analyses of the study. Upon thorough consideration by the review team, the
`Division was inclined to agree with the applicant’s assessment of the effect of the identified
`' violations. However, I did not agree with the applicant’s conclusions based on the analyses
`excluding the two sites.
`
`
`
`2.
`
`INTRODUCTION
`
`2.1 Overview
`
`Alkermes, Incorporated proposes Medisorb Naltrexone, an injectable depot formulation of
`naltrexone, for the treatment of alcohol dependence. According to the applicant, “Medisorb
`Naltrexone is a microsphere-based formulation composed of naltrexone incorporated into a
`biodegradable matrix of polyactide-co-glycolide.” Oral naltrexone is currently approved;
`however, the applicant asserts that the proposed formulation may reduce the potential for
`hepatotoxicity associated with oral naltrexone. The applicant also claims that the formulation
`provides continued exposure for at least one month.
`
`Medisorb Naltrexone was introduced to the Division of Anesthesia, Analgesia, and
`Rheumatology Products (DAARP) via IND 61,138. During the development process, Alkerrnes
`submitted several study protocols for division comment. In addition, the product was discussed
`during a pre—lND meeting, a Type C industry meeting, a pre-NDA meeting, and a CMC meeting.
`Discussion topics included the clinical development plan, efficacy endpoints of interest, and the
`statistical analyses. In the pre—IND meeting, the Agency commented that a reduction in heavy
`drinking was a vague concept and recommended that a responder analysis with respect to
`absence of heavy drinking be conducted. The Agency fithher reiterated the recommendation to
`perform a responder analysis during the pre-NDA meeting. Additionally at the pre-IND
`meeting, the applicant was encouraged to explore analytical approaches appropriate for multiple
`failure times. The Agency also agreed that the study population consisting of currently abstinent
`alcoholics was suitable. During the.development process, the study population evolved to
`include non-abstinent alcoholics. Moreover, the event rate of heavy drinking over a period of
`time emerged as the primary outcome variable. Methodology appropriate for recurrent event
`data was proposed and utilized for the primary analysis. The statistical reviewer of the IND, Dr.
`Milton Fan, expressed several "concerns upon review of the draft statistical analysis plan. Dr.
`Fan’s concerns included the need to validate the model-based inference under the dynamic
`randomization algorithm, the handling of missing data, the appropriateness of pooling the
`placebo groups, and the validity of the proportional hazards assumption in the primary analysis.
`Currently, the applicant has submitted NDA 21—897 in supportof Medisorb Naltrexone for the
`treatment of alcohol dependence.
`
`2.2 Data Sources
`
`A single, randomized, placebo-controlled, multi-center, double-blind study was conducted to
`establish the efficacy of Medisorb Naltrexone. The data and final study reports for the
`completely electronic submission were archived in the Food and Drug Administration internal
`document room under the network path location \\Cdsesub1\evsprod\n02 il 897\0000.
`
`
`
`* 3.
`
`STATISTICAL EVALUATION
`
`3.1 Evaluation of Efficacy
`
`Study Design
`
`Eligible patients were randomized in a 2:2: 1 :1 ratio to receive intramuscular injections of
`Medisorb Naltrexone 190 mg, Medisorb Naltrexone 380 mg, placebo for Medisorb Naltrexone
`190 mg, or placebo for Medisorb Naltrexone 380 mg, respectively. Patients randomized to
`placebo received a matching volume of Medisorb microspheres (i.e. 2 mL or 4 mL) without
`naltrexone. Treatment was administered during clinic visits occurring at baseline and every 4
`weeks thereafter for a 24-week period. During clinic visits, participants also received
`biopsychosocial support therapy using the BRENDA approach. Alcohol consumption was
`recorded throughout the study using the timeline follow-back method (TLFB). In the NDA
`submission the applicant stated, “The BRENDA therapists did not collect the TLFB data
`reported in the study.”
`
`Patients were allocated to treatment for balance on four baseline characteristics using a dynamic
`randomization procedure. The characteristics were goal of abstinence, gender, abstinence prior
`to randomization, and investigative site or center. The former three characteristics had two
`levels while site had 24 levels. The dynamic randomization process was enacted via an
`interactive voice response system (IVRS). The randomization algorithm (biased coin, p=0.75) is
`provided in the appendix.
`‘
`
`The primary measure of efficacy was the event rate of heavy drinking over 24 weeks of
`treatment. This endpoint was defined as the number of heavy drinking days divided by the
`number of days at risk for heavy drinking. The applicant’s use of the event rate was motivated
`by the desire to evaluate the drinking events over a defined duration. In addition, a heavy
`drinking day was defined as a day on which a man consumed at least five drinks or a woman
`consumed at least four drinks. An alcoholic drink was defined as 13.6 grams of absolute ethanol.
`Secondary measures of efficacy included days to relapse of heavy drinking, days to relapse of
`any drinking, number of alcoholic drinks per day, percent of heavy drinking days, percent of
`days abstinent fi'om alcohol, and the event rate of drinking above the National Institute of
`Alcohol, Abuse, and Alcoholism derived “safe drinking” level (1 drink/day for women, 2
`drinks/day for men).
`
`A sample of size 600 was formulated using log-hazard ratio methods to detect a log event rate
`ratio of 0.50 to 0.55 with approximately 90% power. In the formulation of the sample size, the
`applicant assumed, “The proportion of subjects who will be ‘abstinent’ to heavy drinking will be
`0.775 at 24 weeks in 1 of the 2 Medisorb Naltrexone treatment groups as compared with 0.600 in
`the placebo group and 0.600 in the other Medisorb Naltrexone group.”
`
`
`
`Patient Disposition, Demographic and Baseline Characteristics
`
`Descriptive demographics and baseline characteristics were summarized for the intent-to-treat
`(ITT) population composed of all randomized patients who received at least one dose of
`treatment. The ages of patients ranged from 19 to 79 with a mean age of 45. In the study, 84%
`of patients were Caucasian, 8% were African-American, and 5% were Hispanic. Sixty-eight
`percent of the population was male, and the proportion of males to females was approximately 2
`to 1 across all treatment groups. Baseline characteristics included weight, height, type of
`treatment center (i.e. addiction and/or research), patients’ treatment goal, lead—in drinking (or
`abstinent at baseline), employment status, and smoking status. Ninety-two percent of
`participants consumed alcoholic beverages during the seven days prior to randomization. In
`addition, 43% of participants had a treatment goal of total abstinence. A detailed table outlining
`the composition of the study population with respect to demographic and baseline characteristics
`is presented in the appendix. Demographic and baseline characteristics were similar across the
`treatment groups.
`
`'
`
`Of the 627 randomized participants, 209 were randomized to placebo, 210 were randomized to
`190 mg, and 208 were randomized to 380 mg. Four-hundred and one participants received all
`six doses of the treatment. Table 1 was provided in the NDA submission and shows the reasons
`for incomplete treatment and incomplete data collection. During the review process, the
`applicant submitted data that further classified discontinuations. The reclassified
`discontinuations are presented by treatment in Table 2.
`
`Table 1: Reasons Patients Withdrew from Study Treatment and Withdrew from Data Collection
`(Source: Reproduced from Final Study Report ALK21-003, Table 14.1.2)
`Reason for Incomplete Data Collection
`
`
`
`
`
`Investigator
`Judgment
`
`Lost to
`Follow-up
`
`Withdrew
`Consent
`
`Other
`
`Reason for
`
`Incomplete
`Treatment
`
`Received 6
`doses of '
`treatment
`AB
`
`Investigator
`Judgment
`Lost to
`
`Follow-up
`Protocol
`Violation
`Withdrew
`Consent
`Other
`Total
`
`Complete
`Data
`Collection
`388
`
`8
`
`3
`
`40]
`
`30
`
`
`
`
`
`Table 2: Reasons Patients Withdrew from Study Treatment (Reclassified)
`(Source: Adapted from Table 1.1.1 submitted on 29 July 2005)
`190 mg
`380 mg
`Placebo
`(n=210
`n=208
`(n=209
`
`,
`
`Total
`n=627)
`
`Completed
`Adverse events
`
`Investigator Judgment
`Lack of efficacy
`Lost to follow-up
`Other
`Protocol violation
`
`Subject withdrew from
`consent
`
`137
`12
`
`2
`9
`31
`3
`2
`
`14
`
`'
`
`‘
`
`130
`27
`
`3
`9
`24
`0
`0
`
`15
`
`'
`
`.
`
`135
`13
`
`2
`16
`28
`2
`0
`
`13
`
`402
`52
`
`' 7
`34
`83
`5
`2
`
`42
`
`1 additionally used a bar graph to depict the percentage of patients within each treatment group
`that received 1, 2, 3, 4, 5, or 6 doses respectively. A distinguishable pattern of discontinuations
`was not apparent.
`
`Figure 1: Number of Doses (per treatment group)
`
`
`
`
`70
`
`60
`
`50
`
`bO
`
`(A)O
`
`20
`
`10
`
`"InofPatlents
`
`
`
`
`
`
`
`
`
`
`
`
`llPlacebo
`D190 mg
`
`[1380 mg
`
`
`Number of Doses
`
`Statistical Methodologies
`
`The statistical methodologies utilized in the submission resulted from numerous correspondences
`' between the applicant and the agency. _As previously stated, concerns expressed by the statistical
`9
`
`
`
`reviewer of the IND included the handling of missing data, the appropriateness of pooling the
`placebo groups, and the validity of the proportional hazards assumption in the primary analysis.
`The applicant finalized the statistical analysis plan on 19 November 2003, prior to the unblinding
`of the study. As a result of additional feedback from the agency and statistical issues that arose
`after unblinding, subsequent analyses were performed by the applicant.
`
`Since the event of interest (i.e. heavy drinking) could occur on multiple days, the statistical
`methodology used by the applicant accounted for recurrent events across time. Specifically, an
`Andersen-Gill model was used to assess the overall effect of treatment. The model was stratified
`
`by gender, treatment goal of abstinence, and abstinence at baseline (i.e. no drinking seven days
`prior to the initial treatment administration). Indicator variables representing the treatment effect
`of the low dose relative to placebo and the high dose relative to placebo were included in the
`model. The applicant additionally repeated the analysis including a term for baseline percent of
`heavy drinking. The detailed statistical formulation of the model used in the primary analysis is
`provided in the appendix. Multiple comparisons, arising from testing each dose of treatment
`versus placebo, were accounted for via the method of Hochberg. To verify the validity of the
`model—based inference, statistical significance was evaluated via re-randomization tests. 'Re-
`randomization or permutation tests are advantageous in that few, if any, assumptions are required
`for their application. The following excerpt describes the general implementation of a re-
`randomization test:
`
`When you analyze an experiment or survey with a parametric test, you compare the observed
`value of the test statistic with the values in a table of its theoretical distribution. Analyzing the
`same experiment with a permutation test, you compare the observed value of the test statistic with
`the set of what-if values you obtain by rearranging and relabeling the data (excerptfrom
`Permutation Tests by Phillip Good).
`
`Through the use of the Andersen-Gill model, the applicant sought to provide evidence of a
`reduction in heavy drinking over time in patients receiving Medisorb Naltrexone. According to
`the applicant, “The method of analysis estimates the average event rate ratio over time taking
`into account patient discontinuation.” In general, the Andersen-Gill model is formulated by
`dividing the follow-up time for each patient into intervals defined by actual heavy drinking days.
`Thus, a patient only contributes data (and belongs to the risk set) for the days having a recorded
`measurement of the number of drinks consumed. The model assumes that multiple observations
`per patient are independent, that is, the numbers of events in non-overlapping intervals are
`independent (also see appendix). Furthermore, another assumption of the model is that of
`proportional hazards (i.e. the hazard or risk of experiencing a heavy drinking day is constant).
`
`To alleviate concern regarding the appropriateness of the assumption of independent
`observations, the applicant employed a robust variance estimator approach. Under the approach,
`the variance estimates were valid even if the dependence structure Was modeled incorrectly. The
`applicant proposed a stratified analysis over covariates for gender, prior drinking, and goal of
`abstinence to address concerns regarding the proportional hazards assumption. According to the
`applicant, “A stratified analysis adjusts for different baseline ‘hazards’ of the prespecified
`stratification factors.
`In this way, the treatment effect would not be subject to the distortion that
`
`10
`
`
`
`a covariate-by-time interaction would induce by inclusion of such a covariate in the model.”
`Moreover, the applicant supplemented the statistical analysis plan (after unblinding) to include _
`an alternate method of analysis if the assumption of proportional hazards was violated.
`I could
`not find a pre-specification of the alternate method in the final statistical analysis plan.
`However, the applicant stated that the approach was “prespecified” in a 21 June 2002 written
`response to agency comments. The correspondence stated, “. . .we prefer to provide an
`alternative to the Andersen-Gill model analysis if the proportional hazards assumption is not met,
`in the final analysis plan prior to unblinding. However, we will offer our most likely approach in
`brief here.” Following the proposal of a stratified analysis, the correspondence further stated,
`“Another approach is simply to collapse event rates over time for each patient such that the
`marginal event rate for each patient will be incorporated into an analysis of covariance or a non-
`parametric analysis of event rates (depending on the distribution of event rates over all
`subjects).”
`
`Since the agency did not fully concur with the proposal to pool placebo groups, the applicant
`provided a justification. The applicant stated that the low dose injection required a lower volume
`of microspheres than the high dose injection. Thus, the microsphere volume for placebo
`injections was matched to active injections for the sole purpose of maintaining the blind.
`Furthermore the applicant stated, “The undisputed assumption of the study design in the original
`prbtocol is that drinking outcomes are independent of whether subjects receive a low volume or
`high volume placebo injection.” The applicant therefore concluded that pooling of the placebo
`groups was appropriate. The applicant also contrasted the analysis based on the combined
`placebo groups with the analysis considering separate placebo groups to alleviate the concern
`regarding the pooling. Additionally, the applicant repeated the primary analysis exploring the
`treatment difference between the 4 mL placebo and the 2 mL placebo groups.
`
`Event rates obtained via the Andersen-Gill model were based on available data only. The
`applicant assumed that uncaptured or missing data occurred randomly and provided no additional
`insight into the effect of the treatment. To assess the assumption that missing data occurred
`randomly, the applicant examined the comparability of the treatment groups for subject
`discontinuation and outcomes via several techniques. The applicant examined the event rate of
`heavy drinking by the number of doses received, the Kaplan-Meier curves, and a pattern mixture
`model. In general, a pattern mixture model is a statistical tool designed to model the available or
`observed data and the missing data mechanism. Using the pattern mixture model approach
`employed by the applicant, the data was initially stratified by the number of doses (i.e. the
`missing data pattern). Estimates of the high and low dose treatment effects were then obtained
`within each stratum. The estimates were. subsequently weighted (by 1/variance), and pooled
`estimates and variances were obtained to formulate conclusions. In the construction of a general
`pattern mixture model, strata are selected by combining groups with similar missing data
`patterns. Moreover, an assumption of the approach is that uncaptured data within each stratum is
`missing randomly.
`
`The applicant conducted a responder analysis whereby patients were classified into the following
`categories: zero heavy drinking days per month, up to one heavy drinking day per month, up to
`two heavy drinking days per month, up to three heavy drinking days per month, and up to four 1
`
`1]
`
`
`
`heavy drinking days per month. Heavy drinking days per month were computed via the formula,
`Heavy Drinking Days per month = (Percent Heavy Drinking Days *3 0. 4)/1 00.
`Differences between the proportions for patients on active treatment versus placebo were
`compared via chi-square tests.
`
`Results and Conclusions
`
`The results of the applicant’s primary analyses are shown in Table 3. The applicant concluded
`that the 380 mg dose of Medisorb Naltrexone significantly reduced the event rate of heavy
`drinking as compared to placebo. Specifically, patients receiving 380 mg of Medisorb
`Naltrexone experienced a 25% reduction, as indicated by the hazard ratio of 0.75, in the event
`rate of heavy drinking compared to the placebo group. An equivalent conclusion was attained
`when the analysis was adjusted for the percent of heavy drinking at baseline.
`
`Table 3: Event rate of Heavy Drinking”: Test for Treatment Effect in ALK21-003:
`Andersen-Gill (Robust Variance) Stratified Analysis
`(Source: Adapted from Final Study Report ALK21-003, Table 8)
`Estimate
`Hazard ratio(95% CI)
`Unadjusted p-
`Adjusted p-value
`value
`
`.
`
`-0.19
`
`0.83 (068,102) V
`190 mg vs.
`'
`placebo
`0.02
`'
`0.01
`0.75 (060,094)
`-0.29
`380 mg vs.
`placebo
`For each variable (190mg or 380 mg) in the analysis, parameter estimates are obtained for each stratum and pooled
`by weighting each stratum by 1/var (as described by Wei and Johnson, Biometrika, 1985). The hazard ratios are
`obtained by exponentiating the parameter estimates.
`THochberg method was used to adjust p-value of 190 vs. placebo and 380 mg vs. placebo.
`
`0.07
`
`0.07 '
`
`The planned stratified analysis across gender, treatment goal of abstinence, and lead-in drinking
`(or abstinent at baseline) resulted in eight possible strata. The stratum formed by females, no
`lead-in drinking, and a