Weight gain from quitting smoking

1 The gist · for everyone

If every smoker in this study had quit, how much more weight would they have gained over about ten years than if none had quit?

+3.44

kg more weight gain from quitting, averaged over everyone in the study

Plausible range +2.41 to +4.47

We followed 1,566 adult smokers from a health check in 1971–75 to a second visit in 1982. We compare weight change in people who quit with people who did not, after so the two groups match on nine traits.

Without the reweighting the gap is 2.54 kg. Non-quitters gained about 2 kg on their own.

Picture two groups of 100 smokers who are alike in every measured way. Over about ten years the group that quits gains about 340 kg more in total than the group that keeps smoking: roughly 3.4 kg each.

What kind of claim: Points toward a cause, with assumptions. The design removes some alternative explanations, not all. Read it as a cause only if its assumptions hold.

Compare the two reweighted groups. Because both are made to resemble the whole study, the gap estimates what quitting would do on average across all of these smokers.

Safe to say

"Under the study's assumptions, quitting smoking added about 3.4 kg (2.4 to 4.5) of weight gain over roughly ten years, on average across these smokers."

Not this

"Quitters gained 3.4 kg," or "every smoker who quits will gain 3.4 kg."

Before you repeat it

This is only causal if the nine traits capture everything that made people both quit and gain weight. The study also kept only people alive and measured in 1982, which the authors flag as a possible source of bias they chose not to fix.

2Who exactly is counted, and when?Follow one person through the rules 3What could make this misleading?4 caveats · 3 common misreadings 4How would I rebuild it?Exact rules, data, code · 4 doc-vs-code gaps

2 How it's built · for the curious

How the number is made

Follow one person: Ruth

Ruth is a 45-year-old smoker at her 1973 health check, weighing 68 kg. Women like her, with her smoking history, quit less often than average.

1day 0 First health check: smokes, weighs 68 kg
2day 1460 Quits smoking (around 1977)
3day 3255 1982 visit: weighs 75 kg, a 7 kg gain counts

Result: Ruth adds 7 kg to the quitters' average. Because people like her rarely quit, she counts for more than one person in the quitter group. Had she missed the 1982 weigh-in, she would not count at all.

Now the general rule, one step at a time. Each step has the plain version, with the exact version one click away.

Step 1 · Who is counted

Start with people, then apply each rule

One adult who smoked at the first health check.

Exactly

One row per seqn in nhefs.csv (1,629 rows).
S3S1

All people

Keep if Adults aged 25–74 who smoked at the first health check and had the basic facts recorded (sex, age, race, weight, height, education, alcohol use, smoking intensity).1,629 remain

Keep if Had their weight measured at the 1982 visit.1,566 remain

Counted in the denominator
Step 2 · When the clock starts

Time zero: first health check

Each person's first health check, held some time between 1971 and 1975.
- Time to quit: Any time between the first health check and the 1982 visit. Quitting counts if the person reported having quit by 1982. (from 1 to 73 months (up to, not including, 73))
- 1982 visit: The second visit in 1982, roughly 7 to 11 years after the first, depending on when that was. (from 73 to 144 months (up to, not including, 144))
Exactly

Baseline visit 1971–75 (§12.1 margin note). Visit dates are not in nhefs.csv.
S1
Step 3 · What counts

Weight change

Body weight at the 1982 visit minus body weight at the first health check, in kilograms.

Also measured the same way:
- Quit smoking: Reported in 1982 having quit smoking.
Exactly

wt82_71 = wt82 − wt71 (kg); matches the difference of the two columns to 1e-14.
S1S3
Step 4 · Combine

Average across people

Average weight change in each group, with each person counted according to their weight.

Exactly

Weighted least squares of wt82_71 on qsmk; the intercept is the weighted mean for non-quitters and intercept + slope the weighted mean for quitters.
S2S4

weighted sum of weight change in the group

sum of the weights in the group403

= 5.22Quitters, reweighted
Step 5 · Compare

Compare with a matched group

Compare the two reweighted groups. Because both are made to resemble the whole study, the gap estimates what quitting would do on average across all of these smokers.
mean if all quit − mean if none quit
With the real numbers: 5.221 − 1.78 = 3.441
Whose effect? Average effect if everyone did it.

Exactly

θ₁ in E[Y | A] = θ₀ + θ₁A fitted to the pseudo-population (§12.2–12.3).
S1S2
Step 6 · Make it fair

Reweight the comparison group

People who were more likely to quit, given their traits, count for less; people unlikely to quit who did quit count for more. This balances the nine traits across the two groups.

Exactly

Logistic model for qsmk on the nine confounders, with squared terms for age, intensity, duration and weight, no interactions (Program 12.3, lines 93–97). Numerator: intercept-only logistic model (403/1566).
S1S2S4
Made similar on:
- sex
- race
- age (linear and squared)
- education (5 levels)
- cigarettes per day (linear and squared)
- years smoking (linear and squared)
- exercise (3 levels)
- how active day to day (3 levels)
- baseline weight (linear and squared)
Stabilized weights range 0.33 to 4.30 with mean 1.00 (book: 0.33 to 4.30, mean 1.00). Nonstabilized weights range 1.05 to 16.7 with mean 2.00. Program 12.2 tabulates one stratum (white women, by age) as a positivity check; the book recommends checking that stabilized weights average 1.
Step 7 · How sure

Show the range of plausible values

The range 2.4 to 4.5 kg uses a standard formula that allows for the weighting. The authors describe it as slightly wider than necessary.
Not covered by this range: Treats the propensity model as correct and the weights as known. It does not cover unmeasured confounding, the selection of people who survived to 1982, or the 63 people with no 1982 weight.

Exactly

geeglm(wt82_71 ~ qsmk, weights = sw, id = seqn, corstr = "independence"); β ± 1.96 SE. Reproduced with weighted least squares and HC0 standard errors: SE 0.525.
S1S2S4

Each dot is one of 20 equally likely values, given sampling noise alone (normal approximation to the interval).

The same definition, assembled mechanically from the spec

Generated from the structured definition, not written by hand. If it disagrees with the summary above, the summary is wrong.

The unit is one person. One adult who smoked at the first health check.
Keep people that meet every rule, in order: adults aged 25–74 who smoked at the first health check and had the basic facts recorded (sex, age, race, weight, height, education, alcohol use, smoking intensity); had their weight measured at the 1982 visit.
Time zero is the first health check: Each person's first health check, held some time between 1971 and 1975.
Time to quit (exposure): from 1 months after the first health check up to, but not including, 73 months after.
1982 visit (observation): from 73 months after the first health check up to, but not including, 144 months after.
For each person, record weight change (headline) as a value within the 1982 visit: Body weight at the 1982 visit minus body weight at the first health check, in kilograms.
For each person, record quit smoking as yes/no within the time to quit: Reported in 1982 having quit smoking.
Average weight change in each group, with each person counted according to their weight.
Do this separately for each group: Quitters, reweighted, Non-quitters, reweighted.
Compare the two reweighted groups. Because both are made to resemble the whole study, the gap estimates what quitting would do on average across all of these smokers.
People who were more likely to quit, given their traits, count for less; people unlikely to quit who did quit count for more. This balances the nine traits across the two groups.
The range 2.4 to 4.5 kg uses a standard formula that allows for the weighting. The authors describe it as slightly wider than necessary.

3 Keep in mind · before you act on it

What could make this misleading

Completeness not stated. The card does not state when the data were complete, so we cannot confirm every window was fully observed.

Caveats

↑ May make the number look bigger than it really is↓ May make the number look smaller than it really is± The true uncertainty is larger than shown↕ Could push the number either way◎ Limits what the number can be used to conclude

↕

Only as good as the nine traits

removes differences in the nine measured traits. If something else, such as health worries, made people both quit and gain weight, it is still mixed into the number.

Could push the number either way · high importance S1

↕

Survivors only

Only people who lived to 1982 and came back could report quitting. If quitting affects survival, the groups are selected after the fact. The authors flag this and leave it unfixed.

Could push the number either way · medium importance S1

◎

Quitting happened at different times

Someone who quit in 1972 and someone who quit just before 1982 both count as quitters. The number averages over very different amounts of time without smoking.

Limits what the number can be used to conclude · medium importance S1

◎

The interval is wider than it needs to be

The robust formula behind 2.4 to 4.5 kg covers the true value more than 95% of the time, according to the authors. It still assumes the model for quitting is right.

Limits what the number can be used to conclude · low importance S1

Common misreadings

Tempting, but wrong	Accurate
Quitters gained 3.4 kg.	Quitters gained about 4.5 kg on average. 3.4 kg is the extra gain attributed to quitting after balancing the groups.
This is the effect for people who actually quit.	It is the average effect if everyone in the study had quit, compared with nobody quitting (the ), not just the effect among quitters.
The raw gap of 2.5 kg is the effect of quitting.	The raw gap mixes in age and smoking history differences. Reweighting moves it to 3.4 kg.

Same question, other reasonable choices

Each row reruns the analysis with one choice changed. If the rows disagree a lot, the headline depends on that choice.

No adjustment (raw difference): Program 12.1: quitters 4.53 kg, non-quitters 1.98 kg.
Nonstabilized weights: Program 12.2; same estimate here.
Also weighting for missing 1982 weights: Program 12.7; book: 3.5 (2.5, 4.5).

Judgment calls

Which traits to balance

Chose: Nine traits measured at the first check, with squared terms for the continuous ones.

Could have chosen: More traits, interactions, or flexible models.

Why: A standard, transparent model for a teaching example.

What it changes: If an important cause of both quitting and weight gain is missing, the estimate is biased by an unknown amount.

Stabilized or plain weights

Chose: Stabilized.

Could have chosen: Plain inverse-probability weights.

Why: Usually narrower intervals.

What it changes: None here; both give 3.44 (2.41, 4.47).

People with no 1982 weight

Chose: Left out of the headline.

Could have chosen: Add weights for being measured (Program 12.7).

Why: Simpler first analysis; the book adds them in §12.6.

What it changes: 3.44 becomes 3.50.

4 Reproduce it · for specialists

The exact specification

Download the spec (YAML) Methods section

schema: metric-card/1
id: smoking-cessation-ipw
title: Weight gain from quitting smoking
subtitle: Inverse-probability-weighted difference in 10-year weight change, quitters vs non-quitters (NHEFS)
status: provisional
owner: data-visual-explainer worked example (public data)
updated: "2026-10-07"
question: >
  If every smoker in this study had quit, how much more weight would they have gained over about
  ten years than if none had quit?

gist:
  sentence: >
    We followed 1,566 adult smokers from a health check in 1971–75 to a second visit in 1982.
    We compare weight change in people who quit with people who did not, after reweighting so
    the two groups match on nine traits.
  headline:
    value: 3.44
    low: 2.41
    high: 4.47
    unit: other
    decimals: 2
    label: kg more weight gain from quitting, averaged over everyone in the study
    context: "Without the reweighting the gap is 2.54 kg. Non-quitters gained about 2 kg on their own."
  frequency: >
    Picture two groups of 100 smokers who are alike in every measured way. Over about ten years the group that
    quits gains about 340 kg more in total than the group that keeps smoking: roughly 3.4 kg each.
  reading:
    direction: neutral
    say: >
      "Under the study's assumptions, quitting smoking added about 3.4 kg (2.4 to 4.5) of weight
      gain over roughly ten years, on average across these smokers."
    dont_say: >
      "Quitters gained 3.4 kg," or "every smoker who quits will gain 3.4 kg."
  key_caveat: >
    This is only causal if the nine traits capture everything that made people both quit and
    gain weight. The study also kept only people alive and measured in 1982, which the authors
    flag as a possible source of bias they chose not to fix.
  claim_strength: quasi-causal

grammar:
  unit:
    singular: person
    plural: people
    plain: One adult who smoked at the first health check.
    exact: "One row per seqn in nhefs.csv (1,629 rows)."
    src: [S3, S1]
  population:
    - id: eligible
      plain: >
        Adults aged 25–74 who smoked at the first health check and had the basic facts recorded
        (sex, age, race, weight, height, education, alcohol use, smoking intensity).
      exact: >
        Margin note to §12.1. Applied before distribution: nhefs.csv already holds only these
        1,629 people, and no code restricts on these fields.
      n_after: 1629
      src: [S1, S3]
    - id: weighed_1982
      plain: Had their weight measured at the 1982 visit.
      exact: "nhefs.nmv <- nhefs[which(!is.na(nhefs$wt82)),] (chapter12.R line 12). Drops 63 people."
      n_after: 1566
      src: [S2, S1]
  anchor:
    kind: per_unit_event
    label: First health check
    plain: Each person's first health check, held some time between 1971 and 1975.
    exact: "Baseline visit 1971–75 (§12.1 margin note). Visit dates are not in nhefs.csv."
    src: S1
  windows:
    - id: quit_period
      label: Time to quit
      start: 1
      end: 73
      unit: month
      role: exposure
      plain: >
        Any time between the first health check and the 1982 visit. Quitting counts if the person
        reported having quit by 1982.
      exact: >
        qsmk = 1 if the person reported in 1982 having quit (§12.1). The quit date is not used;
        the authors note it is really a time-varying treatment they ignore (Fine Point 12.1).
        The window shown is the shortest possible gap (1975 → 1982).
      src: S1
    - id: followup
      label: 1982 visit
      start: 73
      end: 144
      unit: month
      role: observation
      plain: The second visit in 1982, roughly 7 to 11 years after the first, depending on when that was.
      exact: "Follow-up visit in 1982 (§12.1); per-person gap between 73 and 143 months, not recorded in the file."
      src: [S1, S3]
  measures:
    - id: weight_gain
      label: Weight change
      kind: value
      window: followup
      primary: true
      plain: Body weight at the 1982 visit minus body weight at the first health check, in kilograms.
      exact: "wt82_71 = wt82 − wt71 (kg); matches the difference of the two columns to 1e-14."
      src: [S1, S3]
    - id: quit
      label: Quit smoking
      kind: indicator
      window: quit_period
      plain: Reported in 1982 having quit smoking.
      exact: "qsmk (0/1); 403 of 1,566 = 1."
      src: [S1, S3]
  aggregate:
    kind: mean
    measure: weight_gain
    numerator: weighted sum of weight change in the group
    denominator: sum of the weights in the group
    plain: Average weight change in each group, with each person counted according to their weight.
    exact: >
      Weighted least squares of wt82_71 on qsmk; the intercept is the weighted mean for
      non-quitters and intercept + slope the weighted mean for quitters.
    src: [S2, S4]
  groups:
    - id: quitters
      label: Quitters, reweighted
      role: treated
      plain: The 403 people who quit, reweighted to look like the whole study.
      exact: "qsmk = 1, stabilized weights Pr[A=1] / Pr[A=1 | L]."
      src: [S2, S4]
    - id: non_quitters
      label: Non-quitters, reweighted
      role: adjusted_control
      plain: The 1,163 people who did not quit, reweighted to look like the whole study.
      exact: "qsmk = 0, stabilized weights (1 − Pr[A=1]) / (1 − Pr[A=1 | L])."
      src: [S2, S4]
  comparison:
    kind: adjusted_difference
    target: ATE
    plain: >
      Compare the two reweighted groups. Because both are made to resemble the whole study, the
      gap estimates what quitting would do on average across all of these smokers.
    exact: "θ₁ in E[Y | A] = θ₀ + θ₁A fitted to the pseudo-population (§12.2–12.3)."
    formula: "mean if all quit − mean if none quit"
    src: [S1, S2]
  adjustment:
    method: Inverse-probability weighting with stabilized weights, average treatment effect
    balances_on:
      - sex
      - race
      - age (linear and squared)
      - education (5 levels)
      - cigarettes per day (linear and squared)
      - years smoking (linear and squared)
      - exercise (3 levels)
      - how active day to day (3 levels)
      - baseline weight (linear and squared)
    diagnostics: >
      Stabilized weights range 0.33 to 4.30 with mean 1.00 (book: 0.33 to 4.30, mean 1.00).
      Nonstabilized weights range 1.05 to 16.7 with mean 2.00. Program 12.2 tabulates one stratum
      (white women, by age) as a positivity check; the book recommends checking that stabilized
      weights average 1.
    plain: >
      People who were more likely to quit, given their traits, count for less; people unlikely
      to quit who did quit count for more. This balances the nine traits across the two groups.
    exact: >
      Logistic model for qsmk on the nine confounders, with squared terms for age, intensity,
      duration and weight, no interactions (Program 12.3, lines 93–97). Numerator: intercept-only
      logistic model (403/1566).
    src: [S1, S2, S4]
  uncertainty:
    method: Robust (sandwich) variance, as from a generalized estimating equations fit with independence
    level: 0.95
    plain: >
      The range 2.4 to 4.5 kg uses a standard formula that allows for the weighting. The authors
      describe it as slightly wider than necessary.
    exact: >
      geeglm(wt82_71 ~ qsmk, weights = sw, id = seqn, corstr = "independence"); β ± 1.96 SE.
      Reproduced with weighted least squares and HC0 standard errors: SE 0.525.
    known_gaps: >
      Treats the propensity model as correct and the weights as known. It does not cover
      unmeasured confounding, the selection of people who survived to 1982, or the 63 people with
      no 1982 weight.
    src: [S1, S2, S4]

results:
  as_of: "Hernán & Robins, What If (version of 19 Aug 2026); recomputed 2026-10-07"
  unit: other
  cells:
    - {group: quitters, value: 5.221, n: 403}
    - {group: non_quitters, value: 1.780, n: 1163}
  estimate: {value: 3.44, low: 2.41, high: 4.47}
  sensitivity:
    - {label: "No adjustment (raw difference)", value: 2.54, low: 1.66, high: 3.43, note: "Program 12.1: quitters 4.53 kg, non-quitters 1.98 kg.", src: [S1, S4]}
    - {label: "Stabilized weights (headline)", value: 3.44, low: 2.41, high: 4.47, primary: true, src: [S1, S4]}
    - {label: "Nonstabilized weights", value: 3.44, low: 2.41, high: 4.47, note: "Program 12.2; same estimate here.", src: [S1, S4]}
    - {label: "Also weighting for missing 1982 weights", value: 3.50, low: 2.47, high: 4.53, note: "Program 12.7; book: 3.5 (2.5, 4.5).", src: [S1, S4]}

example:
  name: Ruth
  intro: >
    Ruth is a 45-year-old smoker at her 1973 health check, weighing 68 kg. Women like her, with
    her smoking history, quit less often than average.
  anchor_date: "1973-06-01"
  events:
    - {at: 0, label: "First health check: smokes, weighs 68 kg"}
    - {at: 1460, label: "Quits smoking (around 1977)"}
    - {at: 3255, label: "1982 visit: weighs 75 kg, a 7 kg gain", counts: true}
  outcome: >
    Ruth adds 7 kg to the quitters' average. Because people like her rarely quit, she counts for
    more than one person in the reweighted quitter group. Had she missed the 1982 weigh-in, she
    would not count at all.

fine_print:
  judgment_calls:
    - decision: Which traits to balance
      choice: Nine traits measured at the first check, with squared terms for the continuous ones.
      alternatives: More traits, interactions, or flexible models.
      why: A standard, transparent model for a teaching example.
      impact: If an important cause of both quitting and weight gain is missing, the estimate is biased by an unknown amount.
      src: S1
    - decision: Stabilized or plain weights
      choice: Stabilized.
      alternatives: Plain inverse-probability weights.
      why: Usually narrower intervals.
      impact: None here; both give 3.44 (2.41, 4.47).
      src: [S1, S4]
    - decision: People with no 1982 weight
      choice: Left out of the headline.
      alternatives: Add weights for being measured (Program 12.7).
      why: Simpler first analysis; the book adds them in §12.6.
      impact: 3.44 becomes 3.50.
      src: [S1, S4]
  caveats:
    - title: Only as good as the nine traits
      text: >
        Reweighting removes differences in the nine measured traits. If something else, such as
        health worries, made people both quit and gain weight, it is still mixed into the number.
      effect: unknown_direction
      severity: high
      src: S1
    - title: Survivors only
      text: >
        Only people who lived to 1982 and came back could report quitting. If quitting affects
        survival, the groups are selected after the fact. The authors flag this and leave it unfixed.
      effect: unknown_direction
      severity: medium
      src: S1
    - title: Quitting happened at different times
      text: >
        Someone who quit in 1972 and someone who quit just before 1982 both count as quitters.
        The number averages over very different amounts of time without smoking.
      effect: limits_scope
      severity: medium
      src: S1
    - title: The interval is wider than it needs to be
      text: >
        The robust formula behind 2.4 to 4.5 kg covers the true value more than 95% of the time,
        according to the authors. It still assumes the model for quitting is right.
      effect: limits_scope
      severity: low
      src: S1
  misreadings:
    - wrong: "Quitters gained 3.4 kg."
      right: >
        Quitters gained about 4.5 kg on average. 3.4 kg is the extra gain attributed to quitting
        after balancing the groups.
    - wrong: "This is the effect for people who actually quit."
      right: >
        It is the average effect if everyone in the study had quit, compared with nobody quitting
        (the average treatment effect), not just the effect among quitters.
    - wrong: "The raw gap of 2.5 kg is the effect of quitting."
      right: >
        The raw gap mixes in age and smoking history differences. Reweighting moves it to 3.4 kg.
  discrepancies:
    - topic: Who is in the study
      described: >
        Book margin note: restricted to people with known sex, age, race, weight, height,
        education, alcohol use and smoking intensity, who answered the medical history questionnaire.
      implemented: >
        No code applies these rules. The downloadable nhefs.csv already contains only the 1,629
        eligible people; the code's only restriction is a non-missing 1982 weight.
      impact: >
        The eligibility rules cannot be checked or varied from the distributed files.
      src: [S1, S2, S3]
    - topic: Treatment model in the censoring analysis
      described: >
        §12.6 says the model for being measured in 1982 was fit to all 1,629 people. It does not
        say the model for quitting was refit.
      implemented: >
        Program 12.7 refits the quitting model on all 1,629 people (data = nhefs), not the 1,566
        used for the headline.
      impact: Small. The censoring-weighted estimate is 3.50 vs 3.44.
      src: [S1, S2]
    - topic: Input file
      described: "The book's site provides nhefs.csv, NHEFS.xls and nhefs.dta."
      implemented: >
        chapter12.R reads "XXXX/nhefs.xls" with readxl; other chapters point to a contributor's
        local Windows path.
      impact: None for the numbers; reproducers must edit the path.
      src: [S2, S5]
    - topic: Year labels
      described: "Baseline visits ran 1971–75 (§12.1)."
      implemented: >
        Columns are named wt71 and wt82_71, and chapter15.R labels groups "Quit Smoking 1971-1982".
      impact: Cosmetic, but readers of the code may assume a fixed 11-year gap.
      src: [S1, S3, S5]

spec:
  estimand:
    population: >
      NHEFS participants aged 25–74 who smoked at the 1971–75 baseline visit, met the book's
      completeness restrictions, and had weight measured at the 1982 follow-up.
    exposure: Reported having quit smoking by the 1982 follow-up (qsmk = 1) vs not.
    time_zero: Baseline visit, 1971–75.
    variable: Weight at follow-up minus weight at baseline, kg (wt82_71).
    intercurrent_events: >
      Death or loss to follow-up before 1982 excludes the person from treatment classification
      (Fine Point 12.1); 63 eligible people lack a 1982 weight and are excluded from the headline.
    summary: >
      Difference in IP-weighted means, E[Y^{a=1}] − E[Y^{a=0}], kg; robust 95% interval.
  parameters:
    - {name: weights, value: stabilized, meaning: "Pr[A]/Pr[A|L] for quitters, (1−Pr[A])/(1−Pr[A|L]) otherwise"}
    - {name: confounders, value: 9, meaning: "sex, race, age, education, intensity, duration, exercise, activity, weight"}
    - {name: variance, value: robust, meaning: "GEE independence working correlation (R), HC0 (Python)"}
  lineage:
    - name: nhefs.csv
      role: One row per eligible person; baseline traits, qsmk, wt71, wt82
      grain: person
      src: S3
    - name: chapter12.R
      role: Authors' code for Programs 12.1–12.7
      grain: person → two weighted means
      src: S2
    - name: examples/nhefs-ipw/analysis.py
      role: Python reproduction of Programs 12.1–12.3 and 12.7
      grain: person → estimates
      src: S4
  code:
    - purpose: Stabilized weights and the weighted model (authors' R, Program 12.3)
      lang: r
      src: S2
      excerpt: |
        pd.qsmk <- predict(denom.fit, type = "response")
        numer.fit <- glm(qsmk~1, family = binomial(), data = nhefs.nmv)
        pn.qsmk <- predict(numer.fit, type = "response")
        nhefs.nmv$sw <- ifelse(nhefs.nmv$qsmk == 0, ((1-pn.qsmk)/(1-pd.qsmk)),
                            (pn.qsmk/pd.qsmk))
        msm.sw <- geeglm(wt82_71 ~ qsmk, data=nhefs.nmv, weights=sw, id=seqn,
                        corstr="independence")
    - purpose: The same estimate in Python
      lang: python
      src: S4
      excerpt: |
        sw = np.where(unc.qsmk == 1, pn / p, (1 - pn) / (1 - p))
        fit = smf.wls("wt82_71 ~ qsmk", df, weights=weights).fit(cov_type="HC0")
  reproduce:
    steps:
      - "python examples/fetch_data.py nhefs-ipw (downloads and checks the data, R code and book)."
      - "python examples/nhefs-ipw/analysis.py (writes results/estimates.csv and results/weights.csv)."
    receipts:
      - "2026-10-07: stabilized IPW 3.441 (2.411, 4.470); book 3.4 (2.4, 4.5). Weights 0.331–4.298, mean 0.999."
      - "2026-10-07: unadjusted 2.541 (1.656, 3.425), book 2.5 (1.7, 3.4); Program 12.7 3.496 (2.466, 4.527), book 3.5 (2.5, 4.5)."
    tests: "Reproduced figures are compared with the book's text by hand; see the receipts."

glossary:
  - term: inverse-probability weighting
    aliases: [IP weighting, IPW, reweighting, reweighted]
    plain: >
      Counting each person more or less so that the groups being compared look alike on measured
      traits. People whose choice was unusual for someone like them count more.
  - term: average treatment effect
    aliases: [ATE]
    plain: The average difference if everyone had quit compared with if nobody had quit.
  - term: stabilized weights
    plain: A version of the weights scaled to average about 1, which usually gives a tighter range.
  - term: 95% range
    aliases: [95% interval, confidence interval]
    plain: >
      The values that fit the data well, given random noise and the model. It does not cover
      traits nobody measured.
  - term: robust variance
    aliases: [sandwich]
    plain: A way of computing the range that stays valid when people are counted with different weights.

quiz:
  - kind: verbatim
    q: "The headline is 3.4 kg. Which statement is right?"
    options:
      - "Quitters gained 3.4 kg on average."
      - "Quitting added about 3.4 kg of gain, averaged over everyone in the study, after balancing nine traits."
      - "Every smoker who quits gains 3.4 kg."
    answer: 1
    why: Quitters gained about 4.5 kg; 3.4 kg is the gap after reweighting, and it is an average.
  - kind: gist
    q: "What would make this number wrong as a causal effect?"
    options:
      - "A trait that drives both quitting and weight gain but is not among the nine."
      - "Using stabilized instead of plain weights."
    answer: 0
    why: Reweighting only balances measured traits. Stabilized and plain weights give the same answer here.
  - kind: transfer
    q: "A participant quit in 1980 but missed the 1982 weigh-in. Is she in the headline?"
    options: ["Yes, as a quitter", "No, she is left out", "Yes, as a non-quitter"]
    answer: 1
    why: The headline uses only the 1,566 people weighed in 1982; Program 12.7 reweights to account for the 63 who were not.

changelog:
  - {date: "2026-08-19", change: "Book version used for the cited text and figures."}
  - {date: "2026-10-07", change: "Metric card written from nhefs.csv and chapter12.R; four discrepancies recorded."}

sources:
  - {id: S1, path: "examples/nhefs-ipw/data/hernanrobins_WhatIf_19aug26.pdf", note: "Hernán & Robins, Causal Inference: What If, chapter 12 (§12.1–12.3, 12.6, Fine Point 12.1)"}
  - {id: S2, path: "examples/nhefs-ipw/data/chapter12.R", lines: "9-16, 51-76, 92-122, 231-292"}
  - {id: S3, path: "examples/nhefs-ipw/data/nhefs.csv", note: "1,629 rows; downloaded from miguelhernan.org/whatifbook"}
  - {id: S4, path: "examples/nhefs-ipw/analysis.py", lines: "22-77", note: "Outputs in examples/nhefs-ipw/results/"}
  - {id: S5, path: "examples/nhefs-ipw/data/chapter15.R", lines: "11, 82-83"}

Estimand

Population	NHEFS participants aged 25–74 who smoked at the 1971–75 baseline visit, met the book's completeness restrictions, and had weight measured at the 1982 follow-up.
Exposure / contrast	Reported having quit smoking by the 1982 follow-up (qsmk = 1) vs not.
Time zero	Baseline visit, 1971–75.
Variable	Weight at follow-up minus weight at baseline, kg (wt82_71).
Intercurrent events	Death or loss to follow-up before 1982 excludes the person from treatment classification (Fine Point 12.1); 63 eligible people lack a 1982 weight and are excluded from the headline.
Summary measure	Difference in IP-weighted means, E[Y^{a=1}] − E[Y^{a=0}], kg; robust 95% interval.
Target	ATE: Average effect if everyone did it

Where the description and the code disagree

Who is in the study

Described

Book margin note: restricted to people with known sex, age, race, weight, height, education, alcohol use and smoking intensity, who answered the medical history questionnaire.

Implemented

No code applies these rules. The downloadable nhefs.csv already contains only the 1,629 eligible people; the code's only restriction is a non-missing 1982 weight.

Impact: The eligibility rules cannot be checked or varied from the distributed files.

S1S2S3

Treatment model in the censoring analysis

Described

§12.6 says the model for being measured in 1982 was fit to all 1,629 people. It does not say the model for quitting was refit.

Implemented

Program 12.7 refits the quitting model on all 1,629 people (data = nhefs), not the 1,566 used for the headline.

Impact: Small. The censoring-weighted estimate is 3.50 vs 3.44.

S1S2

Input file

Described

The book's site provides nhefs.csv, NHEFS.xls and nhefs.dta.

Implemented

chapter12.R reads "XXXX/nhefs.xls" with readxl; other chapters point to a contributor's local Windows path.

Impact: None for the numbers; reproducers must edit the path.

S2S5

Year labels

Described

Baseline visits ran 1971–75 (§12.1).

Implemented

Columns are named wt71 and wt82_71, and chapter15.R labels groups "Quit Smoking 1971-1982".

Impact: Cosmetic, but readers of the code may assume a fixed 11-year gap.

S1S3S5

Rules as implemented

Stage	Exact rule	Source
Eligibility 1 1,629 remain	Margin note to §12.1. Applied before distribution: nhefs.csv already holds only these 1,629 people, and no code restricts on these fields.	S1S3
Eligibility 2 1,566 remain	nhefs.nmv <- nhefs[which(!is.na(nhefs$wt82)),] (chapter12.R line 12). Drops 63 people.	S2S1
Time zero	Baseline visit 1971–75 (§12.1 margin note). Visit dates are not in nhefs.csv.	S1
Time to quit [1, 73) months	qsmk = 1 if the person reported in 1982 having quit (§12.1). The quit date is not used; the authors note it is really a time-varying treatment they ignore (Fine Point 12.1). The window shown is the shortest possible gap (1975 → 1982).	S1
1982 visit [73, 144) months	Follow-up visit in 1982 (§12.1); per-person gap between 73 and 143 months, not recorded in the file.	S1S3
Weight change value	wt82_71 = wt82 − wt71 (kg); matches the difference of the two columns to 1e-14.	S1S3
Quit smoking indicator	qsmk (0/1); 403 of 1,566 = 1.	S1S3
Aggregate	Weighted least squares of wt82_71 on qsmk; the intercept is the weighted mean for non-quitters and intercept + slope the weighted mean for quitters.	S2S4
Comparison	θ₁ in E[Y \| A] = θ₀ + θ₁A fitted to the pseudo-population (§12.2–12.3).	S1S2
Adjustment	Logistic model for qsmk on the nine confounders, with squared terms for age, intensity, duration and weight, no interactions (Program 12.3, lines 93–97). Numerator: intercept-only logistic model (403/1566). Balances on: sex, race, age (linear and squared), education (5 levels), cigarettes per day (linear and squared), years smoking (linear and squared), exercise (3 levels), how active day to day (3 levels), baseline weight (linear and squared) Diagnostics: Stabilized weights range 0.33 to 4.30 with mean 1.00 (book: 0.33 to 4.30, mean 1.00). Nonstabilized weights range 1.05 to 16.7 with mean 2.00. Program 12.2 tabulates one stratum (white women, by age) as a positivity check; the book recommends checking that stabilized weights average 1.	S1S2S4
Uncertainty	geeglm(wt82_71 ~ qsmk, weights = sw, id = seqn, corstr = "independence"); β ± 1.96 SE. Reproduced with weighted least squares and HC0 standard errors: SE 0.525.	S1S2S4

Numbers per group

Group	Definition	Value	n
Quitters, reweighted	qsmk = 1, stabilized weights Pr[A=1] / Pr[A=1 \| L].	5.221	403
Non-quitters, reweighted	qsmk = 0, stabilized weights (1 − Pr[A=1]) / (1 − Pr[A=1 \| L]).	1.78	1,163

Check: 5.221 − 1.78 = 3.441 (headline 3.44).

Parameters

Name	Value	Meaning
`weights`	stabilized	Pr[A]/Pr[A\|L] for quitters, (1−Pr[A])/(1−Pr[A\|L]) otherwise
`confounders`	9	sex, race, age, education, intensity, duration, exercise, activity, weight
`variance`	robust	GEE independence working correlation (R), HC0 (Python)

Data lineage

nhefs.csv
One row per eligible person; baseline traits, qsmk, wt71, wt82
Grain: person
chapter12.R
Authors' code for Programs 12.1–12.7
Grain: person → two weighted means
examples/nhefs-ipw/analysis.py
Python reproduction of Programs 12.1–12.3 and 12.7
Grain: person → estimates

Code that defines it

Stabilized weights and the weighted model (authors' R, Program 12.3)examples/nhefs-ipw/data/chapter12.R:9-16, 51-76, 92-122, 231-292

pd.qsmk <- predict(denom.fit, type = "response")
numer.fit <- glm(qsmk~1, family = binomial(), data = nhefs.nmv)
pn.qsmk <- predict(numer.fit, type = "response")
nhefs.nmv$sw <- ifelse(nhefs.nmv$qsmk == 0, ((1-pn.qsmk)/(1-pd.qsmk)),
                    (pn.qsmk/pd.qsmk))
msm.sw <- geeglm(wt82_71 ~ qsmk, data=nhefs.nmv, weights=sw, id=seqn,
                corstr="independence")

The same estimate in Pythonexamples/nhefs-ipw/analysis.py:22-77

sw = np.where(unc.qsmk == 1, pn / p, (1 - pn) / (1 - p))
fit = smf.wls("wt82_71 ~ qsmk", df, weights=weights).fit(cov_type="HC0")

Rebuild it

python examples/fetch_data.py nhefs-ipw (downloads and checks the data, R code and book).
python examples/nhefs-ipw/analysis.py (writes results/estimates.csv and results/weights.csv).

Receipts:

2026-10-07: stabilized IPW 3.441 (2.411, 4.470); book 3.4 (2.4, 4.5). Weights 0.331–4.298, mean 0.999.
2026-10-07: unadjusted 2.541 (1.656, 3.425), book 2.5 (1.7, 3.4); Program 12.7 3.496 (2.466, 4.527), book 3.5 (2.5, 4.5).

Tests: Reproduced figures are compared with the book's text by hand; see the receipts.

Sources

S1 examples/nhefs-ipw/data/hernanrobins_WhatIf_19aug26.pdf · Hernán & Robins, Causal Inference: What If, chapter 12 (§12.1–12.3, 12.6, Fine Point 12.1)
S2 examples/nhefs-ipw/data/chapter12.R lines 9-16, 51-76, 92-122, 231-292
S3 examples/nhefs-ipw/data/nhefs.csv · 1,629 rows; downloaded from miguelhernan.org/whatifbook
S4 examples/nhefs-ipw/analysis.py lines 22-77 · Outputs in examples/nhefs-ipw/results/
S5 examples/nhefs-ipw/data/chapter15.R lines 11, 82-83

Change log

2026-08-19 Book version used for the cited text and figures.
2026-10-07 Metric card written from nhefs.csv and chapter12.R; four discrepancies recorded.

Card quality checks: 0 errors · 1 warnings

info recompute estimate reproduced from cells: 5.221 − 1.78 = 3.441
warn maturity The card does not state when the data were complete, so we cannot confirm every window was fully observed.

Glossary

95% range (95% interval, confidence interval): The values that fit the data well, given random noise and the model. It does not cover traits nobody measured.
average treatment effect (ATE): The average difference if everyone had quit compared with if nobody had quit.
inverse-probability weighting (IP weighting, IPW, reweighting, reweighted): Counting each person more or less so that the groups being compared look alike on measured traits. People whose choice was unusual for someone like them count more.
robust variance (sandwich): A way of computing the range that stays valid when people are counted with different weights.
stabilized weights: A version of the weights scaled to average about 1, which usually gives a tighter range.

If every smoker in this study had quit, how much more weight would they have gained over about ten years than if none had quit?

✓ Safe to say

✗ Not this

! Before you repeat it

How the number is made

Follow one person: Ruth

Start with people, then apply each rule

Time zero: first health check

Weight change

Average across people

Compare with a matched group

Reweight the comparison group

Show the range of plausible values

What could make this misleading

Caveats

Only as good as the nine traits

Survivors only

Quitting happened at different times

The interval is wider than it needs to be

Common misreadings

Same question, other reasonable choices

Judgment calls

Which traits to balance

Stabilized or plain weights

People with no 1982 weight

The exact specification

Estimand

Where the description and the code disagree

Who is in the study

Treatment model in the censoring analysis

Input file

Year labels

Rules as implemented

Numbers per group

Parameters

Data lineage

Code that defines it

Rebuild it

Sources

Change log

Could you explain it to a colleague?

Explain it back

Glossary

Safe to say

Not this

Before you repeat it