Research files & reproducibility

Download the full report, Institute-authored analysis code and two complete summary datasets. All files are hosted on DigitalOcean Spaces.

📄PDFFull research report ↗16 pages · 501 KB🗜️ZIPCode & derived results ↗Methods, scripts, aggregates · 225 KB📊CSV125-geography results ↗Rank, model score & coverage · 9.4 KB📋CSV24 source families ↗Provenance & source limitations · 7 KB

Research status: exploratory secondary analysis, not peer reviewed. Country model scores are neither observed national smile percentages nor measures of happiness. Source images and underlying restricted records are not redistributed; some scripts require the original private input archives.

An audit of 125 geographies, 24 source families, and the limits
of inferring happiness from faces

The Null Institute | Research report | 31 October
2026
Study type: Original secondary analysis of previously
published observations, derived aggregates, and source records. No new
people were photographed or recruited. This report has not undergone
external peer review.

Abstract

Question. Can heterogeneous observations of facial
smiling support a reproducible ordering of countries, and does that
ordering represent happiness? Data. We assembled an
inclusive numerical registry of 475 source entries, organized into 24
provenance families. Nineteen families yielded 10,740 derived
within-study pair comparisons spanning 125 countries and territories;
four contained only one geographically identifiable setting, and one
could not reliably identify the country of photographed subjects.
Methods. The exploratory Synthetic Observed Smiling
Index (SOSI) transforms within-study differences into weighted pairwise
preferences, fitted with a ridge-regularized Bradley–Terry model. We
independently reproduced its computational outputs, fitted an
alternative pairwise estimator, varied acquisition-family weights, and
audited concordance with fixed country-pair overlap.
Results. Among 17 sufficiently overlapping comparisons
involving eight families, median agreement on which country smiled more
was 59.1%, with five comparisons below chance level. An exact
whole-family sign-orientation exercise gave a conditional p =
0.03125, but stricter overlap thresholds yielded p = 0.0625 and
0.125, and four of eight leave-one-family-out tests exceeded 0.05. An
alternative estimator using original weights produced Spearman rank
correlation ρ = 0.982 with the archived index, while preserving only six
of its top ten positions. Giving acquisition families equal total
weight, while retaining the original overall weight mass, produced ρ =
0.785 and moved 58 of 125 geographies by at least ten ranks. Fifty-one
geographies had only one pair-comparable family, including 50 covered
solely by Instagram-based photographs. Interpretation.
These records document country-associated differences within their
selected situations, not a uniquely identifiable,
population-representative global frequency of smiling. Neither a
definitive national-smiling league table nor a rank of experienced
happiness has been validated.

Key distinction: The numerical index is
reproducible. Its interpretation as a country-level psychological trait
is not established.

1. Why this question
is harder than it looks

A smile is a visible event. But a national rate of smiling is not
automatically defined by counting smiles in any available collection of
faces. A tourist’s Instagram photo, a corporate executive portrait, a
politician’s campaign image, an employee greeting a customer, and a
pedestrian walking alone answer different questions. Each setting
selects different people, different moments, and often a different
meaning of facial expression.

Our motivating hypothesis was deliberately strong: if the
mean frequency of spontaneous positive-expression smiling is a useful
behavioural indicator of long-run experienced happiness, sufficiently
repeated and comparably sampled smiles might reveal a national pattern
that life-evaluation questionnaires do not measure. This is an
operational research hypothesis, not a proven definition of happiness. A
smile can also be courtesy, embarrassment, nervousness, social
obligation, or a posed photographic convention. Conversely, a contented
individual may show no smile at the observed moment.

Two distinct questions follow:

1. Behavioural comparability: do countries retain a
   similar order across independently acquired smiling datasets?
2. Psychological validity: after accounting for
   context, does the measured frequency predict long-run experienced
   positive emotion among the same people?

Our analysis primarily tests the first. The present source material
cannot decisively test the second. This distinction sets the limits of
every conclusion below.

The published literature offers reasons to take facial behaviour
seriously without treating it as a mind reader. Ekman, Davidson and
Friesen (1990) linked a specific smile involving eye and mouth muscles
to positive emotional experience during film viewing [1]. Gardiner and
colleagues (2022) found that people reporting greater overall happiness
smiled more frequently in two observed social settings, with
correlations of r = 0.18 (interviews, n = 281) and
r = 0.26 (conversation, n = 186) [2]. These are
participant-level associations within restricted situations, not
between-country validation of a happiness index. Sheldon and colleagues
(2021) argued that frequent Duchenne smiling may signal chronic positive
mood [3]. Krumhuber and Kappas (2022) challenged whether Duchenne smiles
reliably and uniquely communicate genuine positive affect, noting
evidence of deliberately produced smiles [4]. Both the empirical support
and its counterarguments matter.

2. What
was assembled, and what the numbers actually count

We collected sources reporting facial behaviour, not a worldwide
survey of randomly selected residents. The original registry contains
475 numerical source entries drawn from pre-existing
publications, repositories, released aggregates, historical reports and
derived transcriptions. A source entry might be an AU12 (lip-corner
movement) intensity, a fraction of faces classified as smiling, a
photograph-based smile score, a service-interaction percentage, or an
ordinal rank. It is not one observed person. The term
family indicates a likely shared acquisition programme,
platform or photo corpus, so that correlated reanalyses are not counted
as independent studies.

An exact accounting
of the source universe

An earlier count of 23 families referred to those
with a defensible geographic mapping (19 plus four single-location
families). The complete inclusive inventory has 24, because a separate
set of rated photographs cannot be mapped to the nationality or
residence of the people pictured. Neither count corresponds to 24
independent national samples.

The actual selection
mechanisms

The eight families in the primary cross-source concordance comparison
are: Affectiva advertising-response videos; Jetpac and Selfiecity
Instagram photo material (one broad platform family); European election
political images analysed by Haim and colleagues; CEO photographs
analysed in the Fujiwara-related archive; DatingScout selected dating
portraits; Szarota’s Messenger profile photographs; Tsai’s selected
public headshots; and MSPA service mystery-shopping reports. Six
are predominantly photographic, one is elicited advertising behaviour
and one is service interaction. No nationally representative repeated
everyday-smiling dataset enters these 17 principal
overlaps.

Several other sources are valuable for specific contrasts but do not
create global identification. Talhelm, Oishi and Zhang observed
pedestrians in China and the United States; Chuang and colleagues used
particular East Asian and European social contexts; Patterson and
colleagues observed conditional reactions to strangers; and some sources
contain only two countries or even two public figures [5–7]. Small or
restricted results were retained in the positive-weight registry where
geographically identifiable, but a few observations cannot acquire the
information content of a representative national study merely through a
mathematical weighting rule.

Coverage is not evenly
distributed

Only six geographic codes have any directly comparable multi-country
conversational, pedestrian or other live-interaction material in this
ledger: China, Japan, Spain, Sweden, Taiwan and the United
States. These observations are not a common, representative
six-country sample. Conversely, the large photo-based Jetpac dataset
connects approximately 124 locations to the index, but it does not
provide the underlying country-specific face denominators, original
photograph corpus, or reliable residence of photographed subjects.

This is why 125 should be read as geographical
coverage of an exploratory compilation, not as 125 equivalently
observed national populations.

3. The synthetic
index: exactly what we calculated

Simple averaging would silently treat incompatible quantities as if
all were percentages of spontaneous smiles. We instead used a
within-source ordinal synthesis: each study says which of its
countries has a higher observed measure under that study’s protocol. The
model then tries to find one latent ordering that reconciles those
comparisons.

Step
1: turn each source into bounded pairwise preferences

For countries i and j within source s,
with kₛ represented countries and within-source standard
deviation SDₛ, the implemented pair preference is:

[ q_{ijs}=0.5+0.45(). ]

Here v is the recorded outcome in its original
source-specific units. The square-root term shrinks comparisons from
few-country studies towards an uninformative 0.5. The 0.45 cap keeps the
transformed values away from certainty. This is an
analyst-designed transformation, not a probability
experimentally calibrated against citizens’ happiness or independent
smile outcomes.

Step 2: assign limited,
positive weights

Study-level comparison weights are distributed over the study’s
country comparisons, using a factor proportional to source weight
divided by kₛ − 1. Where effective subject counts can be
inferred for selected new sources, low-n comparisons receive a further
factor (). Related outcomes share the same family weight budget rather
than appearing as independent replications. One-source and n = 1 items
are never automatically erased, but are given very small positive
influence; isolated single-location items require a visibly arbitrary
weak reference anchor.

The weights reflect documented analyst judgement
about design, context, dependence and missing information. They are
not inverse-variance weights derived from a common
sampling model. Large numbers of images do not necessarily outweigh a
well-designed smaller field observation: 150 million voluntary geotagged
photographs are not 150 million random residents.

Step 3: fit a
regularized comparative ordering

The model estimates one latent parameter (_i) per geography by
minimizing a weighted pairwise Bradley–Terry cross-entropy plus an (L_2)
penalty with λ = 0.075. A monotone logistic
transformation (100/[1+(-2_i)]) produces a score from 0 to 100. That
score is a model coordinate, not a percentage of
smiling persons, the fraction of time happy, or a known national
probability.

The original v0.2 model involved 14 source streams; v0.3 added nine
multi-geography acquisition families and several weak single-location
contributions. The displayed conservative v0.3
multiplies all newly introduced family budgets by 0.25
while leaving earlier budgets fixed. Crucially, this “conservative”
setting was chosen after examining sensitivity, not
pre-registered as an externally validated best weighting. Other strictly
positive variants are therefore essential to interpreting it.

4. An
illustrative ranking, and why it must not be read literally

The model’s conservative v0.3 output places Egypt 1st, Panama
2nd, El Salvador 3rd, South Africa 4th, Chile 5th, the United Kingdom
6th, Brazil 7th and the United States 8th. Finland appears
114th; Japan 122nd. These are
positions inside our heterogeneous model, not observed
national population smiling rates. Egypt’s position relies heavily on
its archived advertising-response measurements and supporting selected
photo scores. Finland has much less direct naturalistic observational
coverage than the rank number suggests. Neither “Egyptians are the
happiest people” nor “Finns rarely smile in everyday life” follows.

The underlying world lists have different numbers of
geographies and different measurement targets. A rank
difference is not an effect size, a causal estimate, or a comparison of
the same people. See the accompanying 125-row data table for the actual
SOSI ranking, source-family counts and photographic-only coverage. It is
presented as a transparent output audit, not as a final list of
the world’s happiest nations.

5.
First falsification: do independent sources agree on which country
smiles more?

A transferable national characteristic predicts that two independent
protocols comparing the same two countries will usually point in the
same direction, allowing for noise. A pure context-specific explanation
predicts much less convergence. We measured the sign agreement between
acquisition-family pairs using only country pairs available in
both sources. This avoids the invalid shortcut of treating two
publications from the same photo corpus as independent confirmation.

Across all 19 pair-comparable families there are 69
pairs of families sharing at least one country comparison. But small
overlaps are unstable: our primary descriptive subset requires at least
eight common countries and 28 shared country-pair
comparisons, yielding 17 inter-family comparisons from
eight families. Their median directional agreement was
59.09%; the mean was 58.75%, and a
common-edge-weighted mean was 59.00%. Five of
the 17 agreed on fewer than half of their overlapping country
pairs.

This is a small positive tendency in selected overlapping sources,
not universal agreement. For example, commercial ad response versus
Instagram photos agreed on the direction of roughly
64.2% of common country pairs in a 33-geography
overlap; European politician photography versus service encounters
agreed on approximately 42.9% for their particular
overlap. The source-specific rank orders are not interchangeable.

Fixed-overlap sensitivity
tests

We then preserved each observed family-pair overlap structure and
reversed whole-family ranking orientations under an artificial symmetric
null. With eight participating families, there are 128
distinguishable relative orientation configurations; the
observed statistic has one-sided conditional p =
0.03125. This is an exact computation for that
constructed null, not a sampling-design-valid probability that
a universal national smiling trait exists.

Excluding one family at a time produces conditional p values
above 0.05 in four of eight cases. The resulting
evidence is therefore sensitive both to source composition and to how
much overlap is required. An earlier country-label permutation had
produced p ≈ 0.0015, but that procedure sometimes altered the
number of shared pairs below the original inclusion cutoff and is
not a robust confirmatory finding. A separate
10,000-draw fixed-mask random sign perturbation produced an even smaller
conditional p but destroyed the within-family transitivity of
country rankings; it too is an artificial sensitivity diagnostic, not
proof of a national trait.

Result: The source records contain descriptive
partial agreement. We cannot yet distinguish a common underlying
country component from shared photo-selection mechanisms, national
display conventions, publication choices, or the structure of country
overlaps. A statistically secure, context-invariant country
characteristic has not been demonstrated.

6.
Second falsification: do the rankings survive a different mathematical
model?

We reproduced all 12 archived derived outputs
byte-for-byte in a clean execution from the archived inputs.
All 21 files in the original ZIP passed its recorded
size and SHA-256 checks. This establishes computational
reproducibility, not validity of the observational sampling
frames.

Then we replaced the pairwise Bradley–Terry cross-entropy objective
with regularized weighted least squares on transformed pair preferences.
The inputs and nominal pair weights were held fixed. Spearman agreement
with the conservative SOSI ranking was ρ = 0.982, with
a median absolute movement of five places and
11/125 geographies moving by ten or more positions.
Despite high global rank agreement, only six of the top
ten remained the same. A global correlation can look impressive
while a headline top-ten list is unstable.

We also gave the 19 pair-comparable acquisition families
equal total comparison-weight mass, keeping every weight
positive and holding the total mass of weights equal to
the original to avoid changing the effective strength of the ridge
penalty at the same time. This resulted in:

The equal-family specification is a stress test, not
a claim that all sources should scientifically receive identical weight.
A previous uncorrected stress test changed the sum of weights from 28.31
to 19.00 while holding λ fixed; it reported 66 locations moving at least
ten places and ρ = 0.767. We report the
weight-mass-matched comparison as the cleaner result.
Its central lesson survives: precise national positions depend
materially on analyst decisions.

A separate earlier influence check within the original model found
that reducing the Instagram-photo family’s weight to 1%
of its previous positive value moved 78/125 geographies
by at least ten ranks; similarly reducing the Affectiva ad-response
family moved 27/125 by that amount. These are model
scenarios, not 95% confidence intervals or corrected true ranks.

7. Why
smiling cannot yet replace self-reported happiness

The World Happiness Report (WHR) 2026 ranks countries by nationally
weighted, three-year mean responses to the Cantril
Ladder, a 0–10 evaluation of one’s life. Its
headline ranking is not an average of recorded facial smiles, nor a
direct time-weighted record of pleasant emotions. The WHR explicitly
describes this distinction in its methods and FAQ, and separately
discusses positive and negative affect [8]. The six commonly discussed
socioeconomic variables are used to explain variation in life
evaluations; they are not the arithmetic ingredients of the published
country ranks.

For 109 geographies matching our conservative SOSI
and WHR life evaluation for 2023–2025, the rank
correlation was Spearman ρ = +0.081. Under a different
pre-existing all-positive model-weight configuration, it was
approximately +0.148. With the WHR’s separately
published positive-affect measure, the corresponding correlations were
approximately +0.174 and +0.217. The
positive-affect composite itself includes a question about smiling or
laughing yesterday, together with other experiences: it is not
an independent long-term objective happiness reference. Results
vary with the source model and geographic composition. The reported
comparisons are ecological and non-causal; country-aligned historical
comparisons also yielded weak effects.

Thus one defensible inference is clear: the rank ordering
derived from our selected smiling evidence differs substantially from
the WHR ordering of life evaluations. That is not proof that
WHR authors falsified their data, nor proof that our index has uncovered
the true happiness ranking. The two operational definitions differ, and
ours has markedly weaker representativeness.

At the individual level, the face-emotion literature supports some
relationship between positive expression and emotion, not exclusive
equivalence. In Talhelm and colleagues’ two-campus pedestrian
observations, 29.1% of people walking alone at one US
setting smiled versus 3.0% in the comparable Chinese
setting; overall rates in the sampled campus areas were
40.4% versus 20.2% [5]. This is a
substantial observed contextual behavioural difference. It does
not imply that Americans are ten times happier than Chinese people, and
even smiles when alone do not exclude culturally learned habits or
selection by place and time. Conversely, the original study demonstrates
why a hypothesis that all smiles are merely directed at
strangers is too simple.

Other evidence also cuts both ways. Seder and Oishi (2012) associated
one selected college Facebook portrait with life satisfaction 3.5 years
later in small longitudinal samples, but an earlier, larger-scale
yearbook follow-up by Freese and colleagues (2007) failed to replicate
several claims from an older portrait-outcome study [9, 10]. One good
photograph is not a longitudinal behavioural measure. Nor can the
emotional validity of Duchenne (AU6+AU12) smiles be
transferred without testing to any classifier’s generic
AU12 or tooth-display output.

8. An
external check that did not rescue the ranking

An independent 2019 open celebrity-photograph dataset associated with
Girard and McDuff provides 30,601 people and 292,636
images across 131 geographic codes [11]. These
figures describe its own selected celebrity imagery and should not be
confused with the separate 2017 advertising-reaction corpus, its 866,726
participants, or later presentations with different people/photo
counts.

Across 98 overlapping geographies, the correlation
of the celebrity dataset’s person-mean smile measure with the
conservative SOSI model was ρ ≈ −0.099. Requiring at
least 20 people in that photo dataset gave ρ ≈ −0.281
among 53 geographies. This is negative evidence for simple
convergence of these photo-related measures, but it is not
definitive: celebrities are selected, public-image conventions vary, and
the threshold comparison is post hoc. The original measurements are not
population-based, and partial overlap between public-image corpora
cannot always be eliminated.

The failure to find reassuring external agreement matters. A credible
synthesis must retain disconfirming comparisons even when they make a
compelling country ranking less marketable.

9. The most
important source-quality problems

Geographic identity. Jetpac geographically tags
photographs, not the nationality, residence, age distribution or
identity of every person pictured. An advertisement panel’s country code
may reflect recruitment market rather than a random national sample. An
executive photograph’s country may be based on company headquarters
rather than the CEO’s citizenship.

Outcome mismatch. A video AU12 time fraction, a
face-classifier “happy” label, an ordinal dating-photo list and a
service-mystery-shop smile percentage do not share a measured real-world
unit. Cross-study ordinal transformation avoids pretending they are
identical, but cannot establish a common latent construct by itself.

Shared data and repeated reports. Four Haim variants
arise from the same election-image corpus; eleven MSPA annual
reports arise from one programme; smile/no-smile and tooth/no-tooth
categories in a professional portrait sample share photographs. Treating
all as independent votes would exaggerate evidence.

Documented data inconsistencies. Our Haim/Jungblut
source audit found an incompatibility between published ‘happy’ outcome
numerators and denominators for selected image populations, requiring an
alternative aligned construction. A separate source-geometry
investigation found face bounding rectangles incompatible with many
listed image dimensions. Its reproduced ‘close-up’ classification was
highly sensitive to assumed image scale; that does not
automatically invalidate Haim’s other facial-happiness
outcomes. The original image-processing coordinate frame is not
fully recoverable from deposited tables, so we cannot assert a unique
corrected value. These findings are reasons for transparent
downweighting and sensitivity analysis, not claims of misconduct.

Temporal mismatch. Archived sources span different
decades and situations; selected social-media imagery from 2014 need not
describe a nation’s 2026 residents. We should not infer year-specific
smiling trends from this compilation.

Uncertainty. The model’s scenario rank ranges are
ranges across chosen weight perturbations, not population
sampling confidence intervals. We lack many source-level denominator
counts, uncertainty estimates, demographic corrections, classifier
validation and selection probabilities. Confidence in arithmetic is not
confidence in the national ranking.

Researcher degrees of freedom. The conservative
scenario and several robustness thresholds were considered after looking
at earlier data. We report those decisions openly. The exploratory tests
are hypothesis-generating, not preregistered demonstrations.

10. What
the study establishes, and what remains unknown

Established within the available archives: There are
measurable differences among geographically labelled selected
contexts. Their derived numerical summary is independently
reproducible. Better-overlapping acquisition families agree on a little
more than half of country-pair directions, but often disagree. Ranking
details, including top-ten membership, shift under reasonable
mathematical and weighting alternatives. Source coverage is dominated by
selected photographs rather than repeated natural observations of
randomly sampled residents.

Not established: That there exists a single
context-invariant national smile rate; that the conservative
125-geography order measures population prevalence; that the index is a
direct estimate of average experienced happiness; that a particular
country is objectively the world’s most or least happy; or that
divergence from the Cantril Ladder proves one instrument deceptive or
scientifically invalid.

A direct discriminating study would repeatedly observe the
same residents under a shared protocol in multiple countries,
sample both social and solitary contexts over weeks or longer, register
gender/age, encounter type and recruitment, and collect independent
brief measures of felt affect. It would then test whether an
individual’s averaged smiling predicts their future or
concurrent lived affect out of sample, and whether country
differences remain after matching contexts and demographics. A failure
of that held-out prediction would refute the stronger happiness-proxy
interpretation even if national smiling behaviour were genuinely
different.

We are not proposing another endless round of fishing through
available country rankings. With the present open data, the
honest empirical endpoint is a methodological result: we can
calculate a synthetic ordering, but cannot validate it as the singular
worldwide smiling or happiness ranking. More elaboration of the same
selected photo sources will not resolve that missing identification.

11. Reproducibility and data
access

The institute’s v0.3 source synthesis and two
independent numerical audit bundles preserve original
derived aggregates, full Python code, exact parameters, a 24-family
provenance ledger, a 125-geography source-coverage matrix, 69
inter-family overlaps, 17 qualifying comparisons, the fixed-overlap
sensitivity results and file-level SHA-256 records. The initial 21-file
model bundle has SHA-256:

ecaa91b202f765b26d3c882ca588a6ccca888a77b91480cfa59cdff96ac6216f

The latest fixed-overlap audit’s verified ZIP has SHA-256:

6807562ea18acfd4881b1c1d4241efd749306196ea86b2ab732edaa226a6a306

The source-reproducibility claim means that original archived
computational outputs were recomputed byte-for-byte and an audit ZIP was
run in a separate clean directory. It does not mean
independent facial images, proprietary original video, nationally
representative source microdata or every third-party copyright licence
were recovered. Further source-level audit is needed before distributing
third-party originals. Original source repositories and literature links
are below.

Access as of this report: The underlying working
archives are held privately by The Null Institute. The public
research-release package and source-code URL are not yet
available. We will only describe a public archive,
checksum-verified download or DOI once it has actually been released and
checked. The companion 125-row ranking and 24-family ledger are
publication-ready derived tables, but their eventual public
file addresses are intentionally not invented here.

How to read the
supplementary tables

Table S1 lists all 125 geography labels,
conservative model rank and number of pair-comparable families. A low
family count is a warning about evidence coverage, not a statistical
confidence level. Table S2 lists all 24 provenance
family labels, original source URLs, context and known selection/reuse
risks. Table S3 contains all 69 overlapping family
comparisons. Table S4 contains exact fixed-overlap
sign-orientation threshold and leave-one-family-out calculations. The
supplementary materials include no raw face photographs or proprietary
person-level data.

References and primary
materials

1. Ekman P, Davidson RJ, Friesen WV (1990). The Duchenne smile:
   emotional expression and brain physiology II. Journal of
   Personality and Social Psychology 58:342–353.
   https://pubmed.ncbi.nlm.nih.gov/2319446/
2. Gardiner G, Sauerberger K, Lee D, Funder DC (2022). What happy
   people do: The behavioral correlates of happiness in everyday
   situations. Journal of Research in Personality 99:104236.
   https://doi.org/10.1016/j.jrp.2022.104236 ; data
   https://osf.io/bs36k/
3. Sheldon KM, Corcoran M, Sheldon M (2021). Duchenne smiles as
   honest signals of chronic positive mood. Perspectives on
   Psychological Science 16:654–666.
   https://doi.org/10.1177/1745691620959831
4. Krumhuber EG, Kappas A (2022). More what Duchenne smiles do,
   less what they express. Perspectives on Psychological Science
   17:1566–1575. https://doi.org/10.1177/17456916211071083
5. Talhelm T, Oishi S, Zhang X (2019, online 2018). Who smiles
   while alone? Rates of smiling lower in China than U.S. Emotion
   19:741–745. https://doi.org/10.1037/emo0000459
6. Girard JM, McDuff D (2017). Historical heterogeneity predicts
   smiling: Evidence from large-scale observational analyses.
   Proceedings FG 2017. https://doi.org/10.1109/FG.2017.135 ; archived
   country aggregates and code https://osf.io/4zxms/
7. McDuff D, Girard JM, el Kaliouby R (2017). Large-scale
   observational evidence of cross-cultural differences in facial
   behavior. Journal of Nonverbal Behavior.
   https://jmgirard.com/wp-content/uploads/2018/03/McDuff_2017_Large.pdf
8. World Happiness Report (2026). International evidence on
   happiness and social media; national rankings and methodological
   notes.
   https://www.worldhappiness.report/ed/2026/international-evidence-on-happiness-and-social-media/
   ; FAQ https://www.worldhappiness.report/faq/
9. Seder JP, Oishi S (2012). Intensity of smiling in Facebook
   photos predicts future life satisfaction. Social Psychological and
   Personality Science 3:407–413.
   https://doi.org/10.1177/1948550611424968
10. Freese J, Meland S, Irwin W (2007). Expressions of positive
    emotion in photographs, personality, and later-life marital and health
    outcomes. Journal of Research in Personality 41:488–497.
    https://doi.org/10.1016/j.jrp.2006.05.006
11. McDuff D, Girard JM (2019). Democratizing psychological insights
    from analysis of nonverbal behavior. Public image-based research
    and released derived country data.
    https://doi.org/10.1109/ACII.2019.8925503 ; preprint
    https://doi.org/10.31234/osf.io/2nvec
12. Haim M, Jungblut M. European politicians’ visual
    self-presentation. https://doi.org/10.1080/10584609.2020.1753869 ;
    released source files https://osf.io/wu2z5/
13. Jetpac (2014). Country-level smile scores derived from geotagged
    Instagram photographs; original person-level photographs and
    denominators unavailable. Contemporaneous coverage:
    https://www.fastcompany.com/3027822/the-happiest-countries-in-the-world-on-instagram
    ; archived complete transcriptions:
    https://www.el19digital.com/articulos/ver/17200-nicaragua-en-el-top-de-paises-mas-felices-y-sonrientes
    and
    https://geeksroom.com/2014/03/paises-mas-felices-instagram-fotografias/84056/
14. The Null Institute (2026). Synthetic Observed Smiling Index
    v0.3; Independent Numerical Challenge v0.1;
    Fixed-overlap and source-provenance audit v0.2. Original
    derived calculations and provenance, unpublished internal archives as of
    31 October 2026.

Research integrity note. This is a secondary,
exploratory methods investigation. Its visual and service data were
created for different original research questions. We make no allegation
of fabrication against any original researcher, and do not claim that a
125-geography model replaces probability-sampled wellbeing research. We
report the null, contradictory and underidentified results as part of
the substantive findings.

---

Technical
Appendix A: The complete 125-geography model output

These are model positions, not measured national smile rates
or national happiness ranks. The score is on an arbitrary
transformed 0–100 axis; the family count measures overlapping source
coverage, not statistical confidence. A star marks geographies with
only one pair-comparable family, which is
Instagram-based. Not all rows are sovereign countries. The source provenance and context classifications are given in Technical Appendix B.

Technical
Appendix B: Provenance families and their limits

The 24 family labels below are a provenance
taxonomy, not proof of 24 independent participant samples. Study
variants, repeated publication years and dataset overlap are shown
explicitly in the downloadable CSV. Four groups cannot establish
cross-country comparisons; one is not geographically identifiable.
Source links from the compiled ledger should be individually checked
before any third-party material is republished.

Data availability and reproducibility: The complete 125-geography model output and 24-family provenance table appear above and are available as downloadable CSV files in Research files & reproducibility near the beginning of this report. The full PDF and the Institute-authored code and aggregate diagnostics are available there too. The original source photo/video microdata and restricted study archives have not been rehosted; the public code package is not a fully self-contained copy of the source datasets.