Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Target trial emulation(Causal inference: What if, Chapter 22)

Shuntaro Sato
November 25, 2020

Target trial emulation(Causal inference: What if, Chapter 22)

Keywords: 因果推論, Target trial, Pragmatic trial, Time zero, ITT effect, Per protocol effect

Shuntaro Sato

November 25, 2020
Tweet

More Decks by Shuntaro Sato

Other Decks in Science

Transcript

  1. CAUSAL INFERENCE
    What If
    Part III
    2020年11⽉7⽇(⼟)
    Borealis@insearchofsth
    Chapter 22
    Target trial emulation

    View full-size slide

  2. III Causal inference from complex longitudinal data
    Section 22.1 The target trial (revisited)
    Section 22.2 Causal effects in randomized trials
    Section 22.3 Causal effects in observational analyses that emulate a target trial
    Section 22.4 Time zero
    Section 22.5 A unified analysis for causal inference
    Chapter Target trial emulation
    Technical Point 22.1
    Technical Point 22.2
    Technical Point 22.3
    Fine Point 22.1

    View full-size slide

  3. III Causal inference from complex longitudinal data
    Section 22.1 The target trial (revisited)
    Section 22.2 Causal effects in randomized trials
    Section 22.3 Causal effects in observational analyses that emulate a target trial
    Section 22.4 Time zero
    Section 22.5 A unified analysis for causal inference
    Chapter Target trial emulation
    Fine Point 22.1
    Technical Point 22.1 Technical Point 22.2 Technical Point 22.3

    View full-size slide

  4. p.37 (in Part I), p.277
    Observational
    Data
    Hypothetical
    Randomized Trial
    emulate
    Target trial
    “what randomized experiment are you trying to emulate?”
    Intro to Chapter 22
    これ以降、以下の⾊分けをしています
    n キーワードは⾚字
    n Observational Data or Study は⻘字
    n Randomized Trial or Experiment は⾚茶⾊

    View full-size slide

  5. Chapter 22 で勉強すること
    1. Sustained treatment strategy におけるTarget trialの概念
    2. Causal inferenceのフレームワーク
    (Randomized experiment だろうと Observational study だろうと)
    3. Causal effectの分類
    (Intention-to-treat effect, Per-protocol effect)
    p.277
    Target trial:
    A hypothetical randomized experiment to quantify a causal effect (p.37)
    Causal effect:
    A contrast between average counterfactual outcomes under different treatment values (p.37)

    View full-size slide

  6. 22.1 The target trial (revisited)
    Section 22.1 Section 22.2 Section 22.3 Section 22.4 Section 22.5

    View full-size slide

  7. p.277
    • Eligible participants: over 18 years old, no AIDS, no previous used of antiretroviral therapy
    • Randomly assigned to treatment strategy g or gʼ at the start
    • Start: time of assignment
    • End: death, loss to follow-up, 60M after baseline
    22.1 The target trial (revisited)
    Randomized trial to estimate the effect of antiretroviral therapy on the 5-year risk of death
    among HIV-positive individuals
    Pragmatic Trial
    Ex
    n The participants and their treating physicians are aware of the
    treatment they receive
    n Nobody receives a placebo (active treatments or no treatment)
    n Participants are monitored as frequently and intensely as regular
    patient outside of the study

    View full-size slide

  8. p.277
    22.1 The target trial (revisited)
    Pragmatic trial (本書内で定義の記載なし)
    A study (including randomized clinical trials) whose
    aim is to determine the effects of an intervention
    under the usual conditions in which it will be applied.
    A Dictionary of Epidemiology Six Edition

    View full-size slide

  9. 22.1 The target trial (revisited)
    “We resort to observational analyses of existing data because
    the randomized trial that would answer our causal question—
    the target trial—is not feasible, ethical, and timely. ”
    p.277
    Target trial
    Hernán MA, Robins JM. Am J Epidemiol 2016

    View full-size slide

  10. p.277
    調べたい causal question を特定するためTarget trial のプロトコルを明確にする
    22.1 The target trial (revisited)
    n Eligibility criteria
    n Start and end of follow-up
    n Treatment strategies
    n Outcomes of interest
    n Causal contrast
    n Data analysis plan
    Key Components of Protocol

    View full-size slide

  11. p.277
    22.1 The target trial (revisited)
    Target trial
    Part I and II :Time-fixed treatment
    Part III : Sustained treatment strategy

    View full-size slide

  12. 22.1 The target trial (revisited)
    Population
    Intervention
    Comparator
    Outcome
    p.278
    PICO
    Richardson MA, et al. APC J Club 1995

    View full-size slide

  13. 22.2 Causal effects in randomized trials
    Section 22.1 Section 22.2 Section 22.3 Section 22.4 Section 22.5

    View full-size slide

  14. 22.2 Causal effects in randomized trials
    p.278‒
    Three types of causal effects
    1. The effect of assignment to the treatment strategy, regardless of
    treatment actually received (intention-to-treat effect)
    2. The effect of receiving the interventions as specified in the study
    protocol (per-protocol effect)
    3. The effect of receiving interventions other than the ones specified in
    the study protocol
    これまでの章で考えてきたこと
    Causal effect of treatment on an outcome Y measured at the end of follow-up
    本章で考えること
    Causal effect on survival (Technical Point 22.3)

    View full-size slide

  15. 参 考 (テキストの内容ではありません)
    <参考> Short Course on Causal Inference, Jun. 3-7, 2019 (Harvard T.H. Chan School of Public Health) Day5の講義資料
    Z A Y
    U
    Randomization to
    RTS arm vs non-malaria
    comparator arm
    Adherence Malaria status
    in 12 months of follow-up
    n アフリカの7カ国で⾏われたマラリアワクチン候補 RTS,S/AS01の第3相試験
    n ⽣後6〜12週の児 (N = 6,537) をRTS,S/AS01かnon-malaria comparator vaccineにランダムに割付
    n どちらの群もプロトコルに従って3回接種(予定通り接種できなかった児はnonadherentとする)
    A Phase 3 Trial of RTS,S/A01 Malaria Vaccine in African Infants
    n 重症マラリアに対する効果: Intention-to-treat population では 26.0% (95% CI, -7.4 to 48.6)
    Per-protocol populationでは 36.6% (95% CI, 4.6 to 57.7)
    RTS,S Clinical Trials Partnership, et al. N Engl J Med 2012
    (1 minus the risk ratio)

    View full-size slide

  16. RTS,S Clinical Trials Partnership, et al. N Engl J Med 2012 の Figure 1.を改変
    6537 Infants
    4538 Received dose 1 of RTS,S/AS01
    ITT population
    2179 Received dose 1 of control vaccine
    ITT population
    4235 Received dose 2
    4145 Received dose 3
    3995 Were included in the
    per-protocol population
    死亡、同意撤回、
    転居、追跡不能
    2179 Received dose 2
    2090 Received dose 3
    2008 Were included in the
    per-protocol population
    参 考 (テキストの内容ではありません)
    死亡、同意撤回、
    転居、追跡不能、
    医学的理由による接種
    計画からの逸脱
    死亡、同意撤回、
    転居、追跡不能
    死亡、同意撤回、
    転居、追跡不能、
    医学的理由による接種
    計画からの逸脱

    View full-size slide

  17. 22.2 Causal effects in randomized trials
    p.278
    Ak
    1: if the individual receives therapy at time k
    0: otherwise
    Ck
    1: if the individual remains uncensored at time k
    0: otherwise
    k 0, 1, 2...K with K =59 (5-year follow-up)
    Z
    Assigned to g1
    or g0
    1: g1
    “receive treatment Ak = 1 continuously during the follow-up
    unless a contraindication or toxicity arises”
    0: g0
    “receive treatment Ak = 0 continuously during the follow- up”
    Dk Indicator for death (1: yes, 0: no) by month K = 1, 2….K +1
    Key variables

    View full-size slide

  18. overbar について
    p.235 (in Chapter 19)
    time 0から time k までの治療歴
    “We use an overbar to denote treatment history, ...”
    ̅
    "k
    = "0
    , "1
    , … ")
    ̅
    " = "0
    = 1, "1
    = 1, … "59
    = 1 = 1, 1, … 1 ̅
    " = ,
    1
    ̅
    " = 0, 0, … 0 = ,
    0
    or
    フォローアップ中継続して治療を受けた場合の治療歴
    フォローアップ中ずっと治療を受けなかった場合の治療歴
    治療歴全体 K を⽰す場合は ̅
    "-ではなく以下のように表記
    ̅
    "

    View full-size slide

  19. 22.2 Causal effects in randomized trials
    p.279
    Intention-to-treat effect
    The effect of assignment to the treatment strategy
    regardless of treatment actually received
    Pr D
    $
    %&',)
    *
    $
    &)
    + = 1 −Pr D
    $
    %&+, ̅
    *
    $
    &)
    + = 1
    time k における intention-to-treat effect は
    the counterfactual risk of death
    under assignment to strategy g1
    (z = 1) vs g0
    (z = 0)
    if nobody had been lost to follow-up through time k ()
    01
    = )
    2)
    3つの causal effect
    〜その1〜

    View full-size slide

  20. 22.2 Causal effects in randomized trials
    p.279
    n be assigned to strategy g1
    at baseline and remain
    under study until the end of follow-up
    n be assigned to strategy g0
    at baseline and remain
    under study until the end of follow-up
    Intention-to-treat effect

    View full-size slide

  21. 22.2 Causal effects in randomized trials
    p.279-280
    治療へのassignmentと治療のinitiationが同時に発⽣する場合
    Intention-to-treat effect は以下の⼆つの意味を持つ:
    n The effect of assignment
    n The effect of initiation
    Pr D
    $
    %
    &
    '(,*
    +
    $
    '*
    , = 1 −Pr D
    $
    %
    &
    ',, ̅
    +
    $
    '*
    , = 1
    Ex
    ・g1
    に割り付けられた全ての⼈が time 0 の時点で治療を受ける
    (その後、継続治療を受けようとそうでなかろうと)
    ・g0
    に割り付けられた全ての⼈が time 0 の時点で治療を受けない
    (その後、治療を受けることになろうとも)

    View full-size slide

  22. 22.2 Causal effects in randomized trials
    p.280
    Intention-to-treat effect is agnostic about
    any treatment decisions made after baseline, including:
    n Discontinuation or initiation of the treatments of interest
    n Use of non-approved concomitant treatments
    n Any other deviations from protocol
    ある2つの研究について、
    どちらにもバイアスがなく同じプロトコールであったとしても
    異なった状況で⾏われれば得られる intention-to-treat effect は違ってくる可能性がある

    View full-size slide

  23. 22.2 Causal effects in randomized trials
    p.280
    Per-protocol effect
    The effect of receiving the interventions as specified
    in the study protocol
    time k における per-protocol effect effect は…
    the counterfactual risk of death
    under full adherence to strategy g1
    vs g0
    if nobody had been lost to follow-up through time k (4
    56
    = 8
    9)
    3つの causal effect
    〜その 2〜
    The comparison of dynamic strategies rather than static strategies (Fine Point 19.2)
    Pr D
    <
    =
    >
    ,8
    @
    <
    A8
    B = 1 −Pr D
    <
    =
    B
    , ̅
    @
    <
    A8
    B = 1

    View full-size slide

  24. 22.2 Causal effects in randomized trials
    p.280
    n receive treatment strategy g1
    continuously between
    baseline k = 0 and end of follow-up
    n receive treatment to strategy g0
    continuously between
    baseline k = 0 and end of follow-up
    Per-protocol effect

    View full-size slide

  25. Fine Point 19.2
    Per-protocol effects to compare treatment strategies
    p.239 (in Chapter 19)
    n Valid estimation of the per-protocol effect generally demands that trail
    investigators collect post-randomization data on…:
    ・ adherence to the strategy
    ・ (time-varying) prognostic factors associated with adherence
    n Baseline randomization makes us expect baseline exchangeability for
    the assigned treatment strategy, not sequential exchangeability for the
    strategy that is actually received.
    Hernán MA, Robins JM. N Engl J Med 2017

    View full-size slide

  26. 22.2 Causal effects in randomized trials
    p.280
    Per-protocol effect
    “In our example above, a couple would have an interest in
    knowing the effect of the contraception method if all trial
    participants had used it as indicated in the protocol. This effect,
    which is referred to as the per-protocol effect, is what would have
    been observed if all patients had adhered to the trial protocol. ”
    Hernán MA, Robins JM. N Engl J Med 2017

    View full-size slide

  27. 22.2 Causal effects in randomized trials
    p.280
    n Contraindication
    n Toxicity etc.
    Discontinuation from Study Protocol
    Discontinuation of the originally assigned treatment or initiation
    of other treatments cannot possibly be considered a deviation
    from protocol

    View full-size slide

  28. 22.2 Causal effects in randomized trials
    p.281
    Intention-to-treat effect でも Per-protocol effect でもない Effect
    The effect of receiving the interventions other than
    the ones specified in the study protocol
    Strategy g0
    は g1
    よりも劣っているという意⾒が出始めた
    → 医師のよっては、臨床経過が思わしくない場合に治療を勧めようになった
    (例えば CD4 cell count がはじめて 200 cells/μL を下回った場合)
    → 開始時に g0
    に割り振られた被験者が実際は modified g’0
    に従うようになった
    “receive treatment Ak
    = 0 continuously during the follow-up
    but, after Lk
    < 200, switch to treatment Ak
    = 1”
    3つの causal effect
    〜その 3〜
    Ex

    View full-size slide

  29. 22.2 Causal effects in randomized trials
    p.281
    n receive treatment strategy g1
    continuously
    between baseline k = 0 and end of follow-up
    n receive treatment strategy g'0
    continuously
    between baseline k = 0 and end of follow-up
    Per-protocol effects that do not correspond to the
    original per-protocol effect
    Per-protocol effects in target trials that can be
    emulated using the randomized trial data
    g0
    ではない

    View full-size slide

  30. 22.2 Causal effects in randomized trials
    p.279-281
    █ receive treatment strategy g1
    continuously between baseline k = 0 and end of follow-up
    █ receive treatment strategy g'0
    continuously between baseline k = 0 and end of follow-up
    The effect of receiving the interventions other than the ones specified in the study protocol
    █ receive treatment strategy g1
    continuously between baseline k = 0 and end of follow-up
    █ receive treatment strategy g0
    continuously between baseline k = 0 and end of follow-up
    Per-protocol effect
    █ be assigned to strategy g1
    at baseline and remain under study until the end of follow-up
    █ be assigned to strategy g0
    at baseline and remain under study until the end of follow-up
    Intention-to-treat effect
    22.2で解説した3つの causal effect

    View full-size slide

  31. 22.3 Causal effects in observational analyses
    that emulate a target trial
    Section 22.1 Section 22.2 Section 22.3 Section 22.4 Section 22.5

    View full-size slide

  32. 22.3 Causal effects in observational analyses
    that emulate a target trial
    p.281
    n Randomized trial (←前述)
    n Observational analyses that emulate a target trial (←これから解説)
    以下の2つの Causal effect は同様に定義することができる

    View full-size slide

  33. 22.2 Causal effects in randomized trials
    p.278
    Ak
    1: if the individual receives therapy at time k
    0: otherwise
    Ck
    1: if the individual remains uncensored at time k
    0: otherwise
    k 0, 1, 2...K with K =59 (5-year follow-up)
    Z
    Assigned to g1
    or g0
    1: g1
    “receive treatment Ak = 1 continuously during the follow-up
    unless a contraindication or toxicity arises”
    0: g0
    “receive treatment Ak = 0 continuously during the follow- up”
    Dk Indicator for death (1: yes, 0: no) by month K = 1, 2….K +1
    Key variables (再掲)

    View full-size slide

  34. p.281
    n initiate treatment A0
    = 1 at baseline and remain under study until the end of follow-up
    n initiate treatment A0
    = 0 at baseline and remain under study until the end of follow-up
    22.3 Causal effects in observational analyses
    that emulate a target trial
    Observational analog of the intention-to-treat effect
    Pr D
    $
    %
    &
    '(,*
    +
    $
    '*
    , = 1 −Pr D
    $
    %
    &
    ',, ̅
    +
    $
    '*
    , = 1
    ... corresponds to the intention-to-treat effect in a target trial in which assignment to
    and the initiation of the strategies occurs simultaneously
    Observational analog of the intention-to-treat effect at time k

    View full-size slide

  35. 22.3 Causal effects in observational analyses
    that emulate a target trial
    p.282
    n 治療群に割り振られたのに治療を開始しなかった個⼈が含まれる trial の intention-to-
    treat effectとは少し異なる。
    <理由(例の場合)> Baselineで antiretroviral therapyを開始したか否かで⽐較している
    Observational analog of the intention-to-treat effect
    Observational analog of the per-protocol effect
    n Target trial の per-protocol effect と同様に定義される
    Randomized trial では以下の2つを区別する必要があるが、Observational study では不要
    ・Original per-protocol effect
    ・Per-protocol effects in alternative target trials
    <理由> pre-specified protocol が存在しないため、それぞれの per protocol effect は特定
    のtarget trail に⼀致する

    View full-size slide

  36. Fine Point 9.4
    Effectiveness vs Efficacy
    p.239 (in Chapter 19)
    There is a wide spread view that:
    p.122 (in Chapter 9)
    To measure... Loosely defined
    Intention-to-treat effect Effectiveness The effect of treatment that would be
    observed under realistic conditions
    Per-protocol effect Efficacy The effect of treatment that would be
    observed under perfect conditions

    View full-size slide

  37. 22.3 Causal effects in observational analyses
    that emulate a target trial
    p.283
    There is a wide spread view that:
    To measure Loosely defined
    Intention-to-treat effect Effectiveness The effect of treatment that would be
    observed under realistic conditions
    Per-protocol effect Efficacy The effect of treatment that would be
    observed under perfect conditions
    しかし effectiveness, efficacy というラベルは sustained strategy においては曖昧である
    以下のように主張することが難しいから
    △ (Realistic condition で) per-protocol effect は efficacy を評価している
    △ (治療効果が判明した後で)intention-to-treat effect は effectiveness を評価している
    “especially problematic
    when interested in sustained treatment strategies”

    View full-size slide

  38. 22.3 Causal effects in observational analyses
    that emulate a target trial
    p.283
    Example: Discrepancy between the estimates in
    observational study and randomized trial
    ホルモン補充療法が閉経後⼥性の冠動脈疾患のリスクを下げるか?
    Hernán MA, Robins JM. Epidemiology 2008
    ⽐較対象 冠動脈疾患のリスク(*)
    Observational study
    (N = 59,337)
    F Grodstein. N Engl J Med. 1996
    Never users
    vs
    Current users
    (estrogen with progestin, or estrogen alone)
    Randomized trial
    (N = 16,608)
    JE Rossouw. JAMA. 2002
    Receive Placebo
    vs
    Receive Estrogen with Progestin
    *ハザード⽐による
    (ぜひ原著を)
    “Never users” vs “Current users” の⽐較が Randomized trial でなされることは稀。
    ∵ “prevalent users” の背景が時間経過で変化していくのに
    “prevalent users” vs ”non-users”で⽐較しても介⼊に⼀致しない、selection biasを受けやすい

    View full-size slide

  39. ちなみ『疫学⽤語辞典第5版』では
    EFFECTIVENESS 効果
    Archibald L. Cochrane (1909-88) らにより,疫学者間の標準的⽤法とされているのは,特定の介
    ⼊,処置,治療法あるいはサービスを⽇常的な状況でフィールドに適⽤した場合に,対象集団に対
    して実現しようと意図したことをどの程度実現できるかの指標。ヘルスケア介⼊が現実の⽬標をど
    の程度実現できるかの指標。EFFICACY有効性や EFFICIENCY効率 とは区別される。
    EFFICACY 有効性
    特定の介⼊,⼿順,治療法あるいはサービスが理想的な状況でもたらす便益の程度。理想的には,
    有効性は RANDOMIZEDCONTROLLEDTRIAL 無作為化⽐較試験の結果に基づいて決定される。

    View full-size slide

  40. 22.4 Time zero
    Section 22.1 Section 22.2 Section 22.3 Section 22.4 Section 22.5

    View full-size slide

  41. 22.4 Time zero
    p.283
    A crucial component of target trial emulation
    Start of follow-up
    Baseline
    Time zero
    Eligibility criteria need to be met at that point but not later
    n In randomized experiment:
    the time when individuals are assigned to a treatment strategy
    n In observational analysis:
    the time follow-up would have started in the target trial

    View full-size slide

  42. 22.4 Time zero
    p.283
    Randomization
    Not reasonable
    ・Treatment strategies had yet to be assigned.
    ・Eligibility criteria had not been defined.
    Potentially biased
    -2 years +2 years
    ・Deaths during the first 2 years of the trial would be excluded from the analysis.
    ・Short-term effect on mortality would be missed.
    ・Differential proportion of susceptible individuals (→ selection bias)
    (Antiretroviral therapy)
    Ex
    Time zero

    View full-size slide

  43. 22.4 Time zero
    p.283-284
    Start of the follow-up
    in the observational analysis
    Time the follow-up would have
    started in the target trial
    How to emulate the start of follow-up of the
    target trial is not always obvious
    Time zero

    View full-size slide

  44. 22.4 Time zero
    1. Eligibility criteria can be met at a single time
    Immediate initiation of antiretroviral therapy when the CD4 cell count first drops
    below 500 cells/μL vs Delayed initiation (in HIV-positive individuals)
    2. Eligibility criteria can be met at multiple times
    Initiation vs No initiation of hormone therapy
    (among postmenopausal women no history of chronic disease and no use of
    hormone therapy during the previous two year) → At age 51, 52, 53 ...?
    p.284
    How many times
    the eligibility criteria can be met?
    Ex
    Ex

    View full-size slide

  45. 22.4 Time zero
    a) The first eligible time
    b) A randomly chosen eligible time
    c) Every eligible times
    ・Fixed schedule for data collection at pre-specified times
    → emulate a new trial at each pre-specified time
    every two years, like in many epidemiologic cohorts
    ・Subject-specific schedule for data collection
    → emulate a new trial starting at each time unit
    a day, week or month in electronic medial records
    p.284
    Time zero in settings with multiple eligibility times
    Ex
    Ex
    ・More efficient because of more available data
    ・Requiring adjustment of the variance of the effect estimate
    (ex. bootstrapping)

    View full-size slide

  46. Fine Point 22.1
    Grace periods (1/3)
    p.285
    In the real world, antiretroviral therapy cannot be started exactly on the
    same day the CD4 cell count is measured.
    →Investigators need to define a grace period
    Delayed initiation
    Immediate initiation of treatment when
    CD4 cell count first drops below 500
    cells/μL
    vs
    Redefine …
    Immediate Initiate therapy within 3 moths
    after CD4 cell count first drops below 500
    cells/μL
    Immediate Initiate therapy more than 3 moths
    after CD4 cell count first drops below 500
    cells/μL
    vs
    Ex
    被験者が 2 monthsで死亡してしまったらどちらのarmに割り付けるのか?
    Question:
    Answer 1: その被験者をどちらかのarmにランダムに割り付ける

    View full-size slide

  47. Fine Point 22.1
    Grace periods (2/3)
    p.285
    Create two exact copies of the individual (clones)
    l If the individual starts therapy in month 3
    l The clone assigned to “start after 3 month” would be censored at that time
    LE Cain, Hernán MA,et al. Int J Biostat 2010
    Ex
    l If the individual had died in month 2
    l The both clones would have died
    → the death would have been assigned to both arms
    Ex
    Double allocation of events
    To prevent the bias that could arise if events occurring during the grace
    period were systematically assigned to one of the two arms only
    Question: 被験者が 2 monthsで死亡してしまったらどちらのarmに割り付けるのか?
    Answer 2:

    View full-size slide

  48. Fine Point 22.1
    Grace periods (3/3)
    p.285
    When using grace periods with cloning and censoring, ...
    Because each individual is assigned to all strategies at baseline,
    a contrast based on baseline assignment will compare groups with
    essentially identical outcomes.
    Intention-to-treat effect Per-protocol effect
    cannot be estimated

    View full-size slide

  49. 22.5 A unified analysis for causal inference
    Section 22.1 Section 22.2 Section 22.3 Section 22.4 Section 22.5

    View full-size slide

  50. 22.5 A unified analysis for causal inference
    p.284‒
    Randomized trial を
    Follow up study without baseline randomization と捉えることができる
    Randomized experiment
    (i) No baseline confounding
    (ii) Known randomization probability
    (iii) Known assignment to a treatment
    strategy for each individual at baseline
    Observational analysis
    (i) If one measures and appropriately
    adjusts for a sufficient set of covariates
    (ii) If the model for treatment assignment
    is correctly specified
    (iii) ( Not necessary for estimating
    the per-protocol effect )
    How to emulate randomized experiments
    Only three things distinguish the data from randomized experiments and observational studies
    emulate

    View full-size slide

  51. 22.5 A unified analysis for causal inference
    p.286
    “intention-to-treat analysis”
    n Lack of adjustment for baseline confounding is justified by randomization
    n No adjustment for post-randomization confounding is required
    ∵ There cannot be post-randomization confounding for the effect of baseline assignment
    “per-protocol analysis”
    intention-to-treat effect per-protocol effect
    In a typical intention-to-treat analysis …
    Validly estimate?
    In randomized trial

    View full-size slide

  52. 22.5 A unified analysis for causal inference
    p.286
    “intention-to-treat analysis”
    n To eliminate selection bias from loss to follow-up, an “intention-to-treat analysis”
    adjusted for post-randomization (time-varying) prognostic factors is required
    n When the time-varying prognostic factors are affected by prior treatment, an
    appropriate adjustment will require the use of g-methods
    intention-to-treat effect
    For valid estimation of the intention-to-treat, ...
    Validly estimate?
    The probability of dropping out may be influenced by the onset of symptoms
    ( = a consequence of treatment itself)
    Ex
    Post-randomization selection bias
    ・differential lost to follow-up
    ・influence from prognostic factors
    In randomized trial

    View full-size slide

  53. 22.5 A unified analysis for causal inference
    p.286-287
    “per-protocol analysis”
    per-protocol effect
    Validly estimate?
    (“on treatment analysis”)
    Only includes individuals who adhered to the
    instructions specified in the study protocol
    Problematic for two reasons
    ■Postrandomization selection bias due to differential loss to follow-up
    (like “intention-to-treat” analysis)
    n Partly disregards the randomized groups
    Naive per-protocol analysis:
    per-protocol analysis does not even attempt to adjust for confounding
    In randomized trial

    View full-size slide

  54. 22.5 A unified analysis for causal inference
    p.286
    “intention-to-treat analysis”
    Adjustment for baseline and time-varying prognostic factors using g-methods
    “per-protocol analysis”
    intention-to-treat effect per-protocol effect
    To estimate the observational analog of
    the intention-to-treat effect and the per-protocol effect ...
    Validly estimate?
    In Observational Analysis

    View full-size slide

  55. 22.5 A unified analysis for causal inference
    p.287
    Randomized trial
    n Adjustment for post-baseline (time-varying) factors:
    Generally necessary for per-protocol analyses for both randomized trails and observational
    studies
    Observational study
    ・Post-randomization confounding
    ・Post-randomization selection bias
    ・Time-varying confounding
    ・Selection bias
    Similar

    View full-size slide

  56. 22.5 A unified analysis for causal inference
    p.287
    Randomized trial
    “When we cannot conduct the randomized experiment that would answer our
    causal question, we resort to observational analyses. It is therefore important to
    use a sound approach to design and analyze observational studies.”
    Observational study
    ・expensive
    ・infeasible
    ・unethical
    ・untimely
    far superior
    Historically, ...
    “Making the target trial explicit is one step in that direction.”

    View full-size slide

  57. Technical Point 22.1
    Controlled direct effects (1/3)
    p.278
    A B Y
    direct effect (A on Y)
    仮に intermediate variable である B が2値(0 or 1)の場合,
    Bがそれぞれ固定された条件での効果(A on Y)を考察することができる
    (controlled direct effect)
    E "#$%,'$% −E "#$),'$% E "#$%,'$) −E "#$),'$)

    (treatment) (outcome)

    View full-size slide

  58. Technical Point 22.1
    Controlled direct effects (2/3)
    p.278
    A B Y
    direct effect
    Baseline
    Randomly assigned to
    treatment A =1 or A =0
    1M
    Randomly assigned to
    treatment B =1 or B =0
    3M
    Outcome of interest Y
    is measured
    E "#,% = Pr "#,% = 1 Pr " = 1 | + = ,, - = .
    Counterfactual quantity Observed risk
    consistently estimated by
    Ex
    (treatment) (treatment) (outcome)

    View full-size slide

  59. Technical Point 22.1
    Controlled direct effects (3/3)
    p.278
    A B Y
    direct effect (A on Y)
    (treatment) (outcome)
    The controlled direct effect can also be validly estimated
    in observational studies as long as the identifiability
    conditions of ...
    ・Consistency
    ・Positivity
    ・Exchangeability
    ... hold for both A and B

    View full-size slide

  60. Technical Point 22.2 (1/2)
    Natural direct effects and Principal stratum direct effects
    p.279
    Natural direct effect
    Controlled direct effect 以外の direct effect
    E "#$%, −E "#$(,
    The average causal effect of A on Y if the value of B had been set
    to the value that B would taken if A had been set to 0
    a=0,a=1で⽰されるconterfactual outcomeを同時に考慮しなくてはないけない
    “cross-world quality”
    Randomized experimentではNatural direct effect を同定することはできない
    Observational data ではNatural direct effect estimate は⽴証されない

    View full-size slide

  61. Technical Point 22.2 (2/2)
    Natural direct effects and Principal stratum direct effects
    p.279
    Controlled direct effect 以外の direct effect
    Principal stratum direct effect
    The average causal effect of A on Y in the subset of the population whose value of
    B would have been equal to b regardless of the value of A ( !"#$ = !"#& = b )
    E *"#&,, | !"#$ = !"#& = b − E *"#$,, | !"#$ = !"#& = b
    E *"#& | !"#$ = !"#& = b − E *"#$ | !"#$ = !"#& = b
    do not involve joint counterfactuals *",,

    View full-size slide

  62. Technical Point 22.3
    Survival analysis with time-varying treatments (1/2)
    p.282
    何を⽬的に g-methods を使っているか?
    n Chapter 17 (“Causal survival analysis”)では
    Effect of point interventions on failure time outcomes を推定するため
    n Chapter 21 (“G-methods for time-varying treatments”)では
    Effect of sustained treatment strategies on non-failure time outcomes を推定するため
    n 実際にはしばしば
    Effect of sustained strategies on failure time outcomes を推定するため
    Pr D$%&
    '
    ( = 1 under the treatment strategy '
    +
    これを推定するためには g-formula と IP weighting にもとづく、2つのアプローチがある

    View full-size slide

  63. Technical Point 22.3
    Survival analysis with time-varying treatments (2/2)
    p.282
    数式は転載
    n g-formula
    n IP weighting
    Pr D$%&
    '
    ( = 1 under the treatment strategy '
    +
    Hernán MA,et al. JASA 2001
    Young JG, Hernán MA,et al. Stat Biosci 2011
    スライド終わり

    View full-size slide