GBD smoking attributable deaths: modeled burden, delivered as rows

Not counted one by one. Nobody tallies headstones marked "smoking"; epidemiologists estimate what mortality and disease burden would remain if smoking disappeared, then publish the difference as formal measures. IHME's Global Burden of Disease program is the machinery behind almost every attributable-death figure quoted in public health, and the difference arrives as four related quantities: deaths, DALYs, and their two components - years of life lost (YLLs) and years lived with disability (YLDs). Each ships with a 95% uncertainty interval, because a modeled quantity without its range is a press release wearing a lab coat.

The scale anchor most quotations reach for comes from IHME's own smoking topic record: 1.14 billion smokers in 2019, 7.7 million deaths and 200 million DALYs attributable to smoked tobacco in that reference year, and 82.6% of current smokers reporting they started between ages 14 and 25. Behind those three headlines sits the full cube: estimates spanning 204 countries and territories plus 660 subnational locations, every single year from 1990 through 2023 in the GBD 2023 release.

A vocabulary trap distorts searches before any data lands. Attributable burden covers fatal and non-fatal harm together, so filtering a query to deaths alone understates the total - which is why health-economics teams pull DALYs in the same delivery and reconstruct the fatal share themselves. See Global Burden of Disease for the program behind the estimates.

Three answer the question directly, and they divide by job rather than by publisher:

1. IHME GBD Results Tool - Tobacco/Smoking Query Interface - the cube. Deaths, YLLs, YLDs, DALYs, prevalence, summary exposure values and population attributable fractions for the Smoking and Secondhand smoke risk factors, addressable by location, age band, sex, year, measure and metric. It scores 9 on Datadory's rubric, and its subnational depth - 660 locations beneath the 204 national units - is unique in the tobacco pool.
2. Our World in Data - Smoking - the quotation layer. About 38 chart-backed tables including death rates from smoking and from secondhand smoke, the share of deaths attributed to smoking, lung-cancer deaths and tobacco-attributable cancer deaths, harmonized onto one entity-year frame from providers including IHME's GBD study. Also 9.
3. HealthData.org - Tobacco Disease & Injury Factsheets Data - the headline shelf. Dozens of topic statistics - the 7.7 million deaths, the 1.14 billion smokers - pinned to the GBD 2019 reference year with their Lancet citations one field away. Narrative-grained rather than row-grained, which caps it at 5.

Beside them sits the exposure side of the ledger: the IHME GHDx GBD 2019 smoking prevalence record packages age-standardized prevalence, smoker counts and cigarette-equivalents per capita for the same 204 countries, deliberately excluding attributable deaths so the two halves stay distinct. Sixteen primary records make up the tobacco pool overall, averaging 8.4 on Datadory's rubric against a catalog-wide mean of 7.81 across 1,744 datasets.

One estimate row, addressed down to a single cell of the cube:

# smoking-attributable DALYs - one location x age x sex x year cell
measure        : DALYs            # Deaths, YLLs, YLDs, DALYs, Prevalence,
                                 # Summary Exposure Value, Attribution Fraction...
location_name  : Japan            # 204 national units or 660 subnational ones
sex_name       : Both             # Male / Female / Both
age_group_name : All ages         # standard bands, early neonatal to 95-plus
rei_name       : Smoking          # Secondhand smoke is its own risk factor
metric         : Number           # Number / Rate / Percent
year           : 2019             # continuous annual series, 1990-2023
val            : <point estimate>
lower_val      : <95% UI lower>   # the interval rides the row, always
upper_val      : <95% UI upper>

Three things arrive pre-decided in that shape. The measure is explicit on every row, so "deaths" and "DALYs" never blur into one column. The interval travels beside the point estimate instead of living in a methods appendix, which turns sensitivity analysis into a column operation. And the metric convention is declared - Number, Rate or Percent - so a rate for Japan and a count for Brazil never share an axis by accident. Cause-level rows use the same spine with cause_name replacing rei_name, which lets one delivery hold both "burden attributed to smoking" and "burden from tracheal, bronchus and lung cancer" on identical keys.

Depends who is asking, and the honest rule splits cleanly. Our World in Data when one number with provenance has to be on screen in minutes: pick the chart - share of deaths attributed to smoking, say - read the harmonized entity-year rows, cite the provider named on the face of the series. Point values only, no intervals, and coverage varies by chart, generally running from about 1990 out to 2022-2023. The GBD cube when a reviewer will ask how the attribution was constructed: any age band, any subnational unit, intervals included, one set of assumptions holding the whole extract together.

HealthData.org's topic record serves a third job - the framing sentence. Its statistics stay pinned to the GBD 2019 reference year even as newer rounds publish, which is a feature for briefing decks that need stable citable numbers and a liability for anyone modeling the present. Treat it as citation scaffolding, not as modeling input.

Teams doing sustained work usually keep both postures: a curated chart series as the screenshot, a shaped GBD extract as the audit trail. Deliveries pin the source release per extract, so a figure quoted this quarter reproduces next quarter unless a value genuinely moved - revisions land as identified rounds, never as silent overwrites.

Burden gains meaning beside behavior and policy, and the pool supplies all three layers on compatible keys.

On exposure, the World Bank prevalence panel reports the share of adults using any tobacco product for roughly 265 economies annually from 2000 to 2024 - 6,625 observations on the headline series, age-standardized to the WHO Standard Population, with male and female variants near 5,000 rows apiece riding sibling codes on identical economy-year keys. On policy, the WHO Report on the Global Tobacco Epidemic 2023 grades every member state 1-to-4 on each MPOWER measure back to 2008 - the 2023 edition reported smoke-free legislation covering roughly a third of the world's population - while the WHO Global Health Observatory collection carries about 300 tobacco indicators with bounds embedded in every observation; Malaysia's modeled male prevalence for 2022 arrives as 34.5 within a 22.0-to-47.0 interval. Europe adds a surveyed mirror: Eurostat sdg_03_30 counts daily smokers of cigarettes, cigars, cigarillos or a pipe by sex across 35 entities in seven survey waves from 2006 to 2023.

Three caveats keep such joins honest. These are different estimators - modeled intervals, self-reported survey shares, policy grades - so splicing their levels into one series draws a line that does not exist; compare them, label them, cite them separately. Regional composites such as EU27_2020 or Africa Eastern and Southern ride the same schema as sovereign states, and a ranking that fails to filter them ranks Germany against an entire continent. And Eurostat's timeline is a wave grid, not a clock - 2006, 2009, 2012, 2014, 2017, 2020, 2023 - so interpolating between observed years is a modeling decision to declare, not a default to hide. Quoting attributable burden with its interval beside the point estimate is what separates a defensible analysis from a headline.

Five jobs this evidence settles outright:

1. Health-economics cases - price attributable DALYs and YLLs by country or by state, intervals feeding sensitivity analysis instead of getting trimmed for convenience.
2. Insurer and pharma diligence - read long-run morbidity and mortality trajectories as demand-side fundamentals, with subnational granularity where the market-level question demands it.
3. Citation-grade journalism and scholarship - every quoted figure resolves to location, age band, sex, year, measure and release.
4. Regulatory and ESG screening - MPOWER grades beside burden trends show where policy intensity moved and whether burden followed.
5. Modeling targets - a dense, uncertainty-tagged location-age-sex-year panel gives forecasting work ground truth instead of repackaged press numbers.

API, files, or your warehouse. Daily, weekly, or hourly.

Name the geographies, age bands, sexes, years and measures, and the sample lands cut to that scope with the field dictionary attached - typed rows, parsed years, uncertainty columns intact. The historical backfill arrives first; the feed continues at whatever pace the question needs, with the source release pinned per delivery.

Persona fit has edges worth naming: market researchers get market-level burden sizing; journalists and academics get citable figures with intervals attached; quant teams get slow-moving fundamentals for sector screens. What these records do not measure is behavior at the till or volume through the supply chain - pair cigarette consumption per capita and US cigarette production statistics for the demand and supply sides this evidence deliberately leaves alone.

This page is the burden thread of the tobacco stack. Open the tobacco data hub for every record in the industry with coverage and field dictionaries, or the ranked best tobacco datasets shortlist for the top of the pool. The trade-offs run deeper on the dedicated comparisons: Eurostat sdg_03_30 versus the GBD Results Tool sets observed European behavior against modeled global consequences, and GBD 2019 smoking prevalence versus OWID Smoking puts the exposure archive beside the quotation layer.

Practical order of operations: pull the burden spine first - deaths plus DALYs for the Smoking risk factor, all ages beside the working-age bands, your markets plus the benchmarks - then append exposure and policy layers on the shared country keys once the outcome series is clean. When you want rows rather than reading, request a sample cut to the countries, years and measures on your desk this quarter; the schema in the sample is the schema you ship against.