t-eff-adas-mmrm

id: t-eff-adas-mmrm
title: "ADAS-Cog (11) - Repeated Measures Analysis of Change from Baseline to Week 24"
regulatory_id: EFT08
type: table
dataset: adqsadas
analysis_set: efficacy
denominator: adsl
sources: phuse
group:
  - TRTP
total: false
analyses:
  - name: population
    method: continuous
    custom: population_n
    label: "Subjects in the analysis population"
    filter: "PARAMCD == 'ACTOT'"

  # Observed cases only. A repeated-measures model handles missing visits by
  # modelling the within-subject covariance, so feeding it carry-forward
  # records would count one measurement several times. Selecting this way gives
  # 539 records over 234 subjects, which is what the reference report's own
  # PROC MIXED output states it used.
  - name: mmrm
    method: continuous
    custom: mmrm_change
    label: "Mixed model for repeated measures of the change from baseline"
    variables: [CHG]
    filter: "PARAMCD == 'ACTOT' & ANL01FL == 'Y' & AVISITN %in% c(8, 16, 24) & !(DTYPE %in% 'LOCF')"
    response: CHG
    subject: USUBJID
    visit: AVISITN
    visit_levels: [8, 16, 24]
    treatment: TRTP
    treatment_levels: ["Placebo", "Xanomeline Low Dose", "Xanomeline High Dose"]
    factors: [SITEGR1]
    covariates: [BASE]
    # Treatment and the baseline score each interact with visit; site group does
    # not. This is the model the SAP names in section 10.1.1, and the reference
    # report's PROC MIXED output confirms it: 31 columns in X is exactly
    # 1 + 11 site groups + 3 visits + 3 treatments + 9 treatment-by-visit
    # + 1 baseline + 3 baseline-by-visit.
    visit_interactions: [BASE, TRTP]
    # The reference report's least-squares means are the treatment main effect,
    # averaged over all three post-baseline visits rather than conditioned on
    # Week 24. See the display's footnotes: the title says Week 24, the number
    # is a visit-averaged estimate, and this is where that is decided.
    lsmeans_over: all
    conf_level: 0.95
    contrasts:
      - { name: xan_vs_placebo, test: "Xanomeline Low Dose",
          reference: "Placebo", column: "Xanomeline Low Dose" }
      - { name: xan_vs_placebo, test: "Xanomeline High Dose",
          reference: "Placebo", column: "Xanomeline High Dose" }
      - { name: high_vs_low, test: "Xanomeline High Dose",
          reference: "Xanomeline Low Dose", column: "Xanomeline High Dose" }

id: t-eff-adas-mmrm
title: "ADAS-Cog (11) - Repeated Measures Analysis of Change from Baseline to Week 24"
study: CDISCPILOT01
population_label: "Efficacy Analysis Set"
cutoff: "2015-02-17"
source: "Source: adqsadas (ADAS-Cog (11) total score, PARAMCD ACTOT) and adsl, from the CDISC pilot's own ADaM package (phuse-org/phuse-scripts, MIT). Supportive analysis as specified in the CDISCPILOT01 statistical analysis plan, section 10.1.1. Reference report display: Table 14-3.11. Data cut-off: 2015-02-17."
footnotes:
  - "[1] Mixed model for repeated measures of the change from baseline, fitted by restricted maximum likelihood with an unstructured within-subject covariance matrix. Fixed effects: treatment, site group, visit, treatment by visit, the baseline score, and the baseline score by visit."
  - "[2] Pairwise comparison of least-squares means; p-values are not adjusted for multiple comparisons."
  - "Observed cases only: 539 assessments from 234 subjects at Weeks 8, 16 and 24. The model accounts for a missing visit through the covariance structure, so no value is carried forward."
  - "WHAT THE LEAST-SQUARES MEANS ESTIMATE: they are the treatment main effect — the model's prediction averaged over all three post-baseline visits and over the eleven site groups equally, with the baseline score held at its mean. They are not the Week-24 visit conditioned on, which the title might suggest and which is a different and less precise quantity (for placebo, 2.3 with a standard error of 0.69 rather than 1.6 with 0.49). This display reproduces what the reference report computed; the distinction is stated here because the number cannot state it itself."
  - "MEASURED, NOT ASSUMED: the fit reproduces the reference report's own PROC MIXED output to six significant figures on all six unstructured covariance parameters, and its REML criterion of 3087.84303515 to ten. It uses the same 539 observations from the same 234 subjects. Two programs can agree on a rounded least-squares mean by accident; they cannot agree on that by accident, so the model is identified far more sharply than the rounded cells above can show."
  - "DIFFERS FROM THE REFERENCE, DELIBERATELY: standard errors and degrees of freedom here are the model-based ones, on 517 residual degrees of freedom. The reference used Kenward-Roger degrees of freedom with Prasad-Rao-Jeske-Kackar-Harville standard errors, which inflate both by roughly one per cent to allow for the covariance parameters having been estimated rather than known. That adjustment is not implemented here. Of the twelve cells above, seven are identical to the reference and five differ at the last digit shown: the three p-values, each 0.001 lower (0.954, 0.555 and 0.605 against 0.955, 0.556 and 0.606); the high-dose least-squares mean standard error (0.55 against 0.56); and one confidence-interval bound (-1.8 against -1.9). Every point estimate agrees. The difference is measured and recorded in quality/data/efficacy-agreement.json."
  - "Treatment columns are planned treatment (TRTP), which is how the CDISCPILOT01 reference report presents them."
  - "CDISCPILOT01 states no data cut-off date. The date shown is the latest ADAS-Cog assessment date present in the vendored ADaM package."
columns:
  order: ["Placebo", "Xanomeline Low Dose", "Xanomeline High Dose"]
format:
  est_se: "{estimate} ({se})"
  ci: "({lcl};{ucl})"
  p_value: "{pvalue}"
  digits:
    estimate: 1
    se: 2
    lcl: 1
    ucl: 1
    pvalue: 3
rows:
  - { analysis: mmrm, variable: CHG, level: lsmean, pattern: est_se,
      label: "LS means (SE) [1]" }
  - { label: "Comparison of least-squares means", section: true }
  - { analysis: mmrm, variable: CHG, level: xan_vs_placebo, pattern: p_value,
      label: "p-value (Xanomeline - Placebo) [1][2]", indent: 1 }
  - { analysis: mmrm, variable: CHG, level: xan_vs_placebo, pattern: est_se,
      label: "Difference of LS means (SE)", indent: 2 }
  - { analysis: mmrm, variable: CHG, level: xan_vs_placebo, pattern: ci,
      label: "95% CI", indent: 2 }
  - { analysis: mmrm, variable: CHG, level: high_vs_low, pattern: p_value,
      label: "p-value (Xanomeline High - Xanomeline Low) [1][2]", indent: 1 }
  - { analysis: mmrm, variable: CHG, level: high_vs_low, pattern: est_se,
      label: "Difference of LS means (SE)", indent: 2 }
  - { analysis: mmrm, variable: CHG, level: high_vs_low, pattern: ci,
      label: "95% CI", indent: 2 }
variants:
  post_text: {}

# Display-specific statistics (contract section 2).
#
# Thin wrappers, as the other ADAS-Cog displays are: the statistics live in
# {opencsr} so they are tested once and shared, and every model term — the
# factors, the covariate, which of them interact with visit, and which visits
# the least-squares means average over — is declared in this display's
# analysis.yaml rather than written here.

mmrm_change <- function(data, spec, denominator = NULL) {
  opencsr::ard_mmrm(data, spec, denominator)
}

population_n <- function(data, spec, denominator = NULL) {
  opencsr::ard_population_n(data, spec, denominator)
}