joshua project joshua project unreached

Overview

Source: /home/coolhand/html/datavis/data_trove/joshua-project/joshua_project_unreached.parquet

Saturn profiled 7,124 rows across 109 columns. The stats below are deterministic and machine-readable; the prose is a language-model interpretation of those stats (opt-in, added after the fact, never sees raw rows).

PeopleID3ROG3 text identifier

PeopleID3ROG3 is almost certainly a person-level identifier: every one of the 7,124 rows holds a unique 7-character, all-caps, single-token code (n_unique equals n, len_min=len_max=7, allcaps_rate=1.0, one_word_rate=1.0). Sample values like '10208ng' and '10375su' suggest a 5-digit numeric prefix followed by a 2-letter suffix. There are no nulls, duplicates, or empties, so the key looks clean.

Treatment: Use as a primary key for joins; exclude from modelling features.

ROG3 categorical feature

ROG3 holds two-letter country codes across 7,124 rows with 202 distinct values and no nulls. India (IN) dominates at 28.5% of records, followed by PK (767) and CH (442); the top 10 codes account for the bulk of mass while a long tail of ~190 other codes shares the remainder, giving an entropy ratio of 0.66.

Treatment: Group rare codes into an 'other' bucket before one-hot or target encoding.

PeopleID3 numeric foreign_key

PeopleID3 is an integer key spanning 10120 to 22661 with 4614 unique values across 7124 rows, suggesting a person identifier that recurs (about 1.5 rows per id on average). The distribution is mildly left-skewed (-0.23) and platykurtic (-0.95) with no nulls, no zeros, and no outliers, consistent with a dense allocated id range rather than a measured quantity.

Treatment: Treat as a join key; do not use as a numeric feature.

ROP3 numeric identifier

ROP3 is a numeric column with 4608 unique values across 7124 rows, ranging tightly from 100005 to 119619 with a mean of 111443.68 and median of 112533. The narrow ~19k span sitting well above zero, combined with integer-looking bounds, suggests a coded identifier or sequence number rather than a measured quantity. Mild left skew (-0.47) and no outliers indicate a fairly uniform spread within that band, and the null rate is negligible at 0.001.

Treatment: Treat as a categorical code or key; do not feed raw into numeric models.

PeopNameInCountry text label

This column names a people group as it appears within a given country (e.g., 'Turk', 'Persian', 'Arab, Moroccan'), likely from a Joshua Project-style ethnographic registry. Values are short (len_mean 12.5, word_mean 1.78) with 4,722 uniques across 7,124 rows, and 33.7% are duplicates because the same people group recurs across countries — 'Deaf' alone appears 151 times. Frequent qualifiers like '(Hindu traditions)' and '(Muslim traditions)' in top_words show religion-tagged variants are baked into the label.

Treatment: Treat as a categorical label; pair with country to form a unique key, and consider stripping parenthetical religion tags for cleaner grouping.

ROG2 categorical feature

ROG2 is a low-cardinality categorical with 7 region codes (ASI, AFR, EUR, NAR, LAM, AUS, SOP) and no nulls across 7,124 rows, consistent with a continental/region-of-origin grouping. The distribution is highly imbalanced: ASI accounts for 75.1% of records while AUS and SOP together contribute fewer than 80 rows, yielding an entropy ratio of just 0.45. Any model conditioned on this field will be dominated by the ASI bucket.

Treatment: One-hot encode and consider grouping AUS/SOP/LAM into an 'other' bucket given the severe imbalance.

Continent categorical feature

Continent is a low-cardinality geographic categorical with 7 distinct values and no nulls across 7,124 rows. The distribution is heavily concentrated: Asia alone accounts for 75.1% of records, with Africa a distant second at 986. Notably, both 'Australia' (39) and 'Oceania' (36) appear as separate categories, which is a labeling inconsistency since Australia is part of Oceania.

Treatment: Reconcile Australia/Oceania into a single category, then one-hot encode.

RegionName categorical feature

RegionName is a categorical geographic grouping with 12 distinct regions and no nulls across 7,124 rows. The distribution is heavily concentrated: 'Asia, South' alone accounts for 47.0% of records, followed distantly by 'Asia, Southeast' at 726 and 'Asia, Northeast' at 521, leaving the Americas and Europe sparsely represented. Entropy ratio of 0.76 confirms meaningful but uneven coverage across the 12 buckets.

Treatment: One-hot encode and consider grouping rare regions, given the dominance of 'Asia, South'.

ISO3 categorical foreign_key

ISO3 looks like a country code in standard 3-letter ISO 3166-1 alpha-3 format, with 202 distinct values across 7,124 rows and zero nulls. The distribution is heavily concentrated on India (IND) at 28.5% of rows (2,032), followed by PAK (767) and CHN (442), so South and East Asia dominate. Entropy ratio of 0.66 confirms the imbalance is material rather than uniform across countries.

Treatment: Use as a join key to country reference tables; consider grouping long-tail codes or stratifying by ISO3 to control for the IND-heavy skew.

LocationInCountry text free_text

Free-text geographic descriptions of where a people group lives within a country, ranging from terse tags like "Widespread." (56 occurrences) to multi-sentence paragraphs up to 939 characters. 64.23% of rows are null and 14.6% of the non-null values are duplicates, so usable signal is concentrated in roughly a third of the dataset. Content is overwhelmingly English (1714 of 1729 detected) with a long tail of place names producing a 10,936-token vocabulary across 2,176 unique strings.

Treatment: Normalize boilerplate phrases like "Widespread" into a categorical flag, then tokenize and embed the residual prose before modelling.

PeopleID1 numeric feature

PeopleID1 is stored as numeric but only takes 16 distinct integer values between 10 and 26, with a tight IQR of 1.0 (Q1=20, Q3=21) around a median of 21. The distribution is heavily left-skewed (skew -1.34) and 32.9% of rows fall outside the Tukey fence, so the 'outliers' alert reflects the column being near-categorical rather than truly continuous. No nulls and no zeros, but the name suggests an identifier despite the low cardinality.

Treatment: Cast to categorical (16 levels) rather than treating as a continuous numeric.

ROP1 categorical feature

ROP1 is a low-cardinality categorical code (16 distinct values, all following an 'A0xx' pattern) with no nulls across 7,124 rows. The distribution is heavily concentrated: 'A012' alone accounts for 51.7% of records, and entropy ratio of 0.67 confirms the imbalance. This looks like a controlled vocabulary or lookup code rather than free input.

Treatment: One-hot or target-encode; consider grouping rare codes given the dominance of A012.

AffinityBloc categorical feature

AffinityBloc is a categorical grouping of peoples/ethnolinguistic blocs, with 16 distinct values across 7124 rows and no nulls. The distribution is heavily concentrated: 'South Asian Peoples' alone accounts for 51.7% of records (3681), followed distantly by 'Sub-Saharan Peoples' (632) and 'Arab World' (475). Entropy ratio of 0.67 confirms the imbalance, and the inclusion of 'Deaf' (151) as a bloc alongside geographic/ethnic categories is a notable taxonomy quirk.

Treatment: One-hot or target-encode, and consider grouping rare blocs given the dominance of South Asian Peoples.

PeopleID2 numeric foreign_key

PeopleID2 is a numeric identifier-like field with only 205 distinct values across 7,124 rows, ranging 101 to 475 with no nulls or zeros. The distribution is left-skewed (skew -0.50) and platykurtic (kurtosis -1.24) with median 412 well above the mean 339, suggesting a few low-id clusters pull the mean down while most rows concentrate near the upper end. The low cardinality relative to row count indicates this is a repeated key rather than a per-row unique id.

Treatment: Treat as a categorical foreign key and left-join to the people dimension rather than using as a numeric feature.

ROP2 categorical feature

ROP2 is a categorical code field with 155 distinct values following an A####/C#### pattern, suggesting a classification or routing code. One value, 'A012', dominates at 51.6% of the 7,124 rows, while the remaining categories are long-tailed C-codes each below 2.4%. Entropy ratio of 0.56 confirms the distribution is heavily concentrated rather than uniform.

Treatment: Collapse rare C-codes into an 'other' bucket and one-hot or target-encode before modelling.

PeopleCluster categorical feature

PeopleCluster is a high-cardinality categorical taxonomy of ethno-religious groupings, with 205 distinct values across 7124 rows and no nulls. The distribution is dominated by South Asian categories — 'South Asia Hindu - other' alone accounts for 12.2% (869 rows), followed by 'South Asia Muslim - other' (586) and 'South Asia Dalit - other' (352). Entropy ratio of 0.80 indicates a fairly spread distribution despite the South Asian skew, and the appearance of 'Deaf' alongside ethnolinguistic labels signals a mixed taxonomy worth flagging.

Treatment: Group the long tail and target- or frequency-encode before modelling.

PeopNameAcrossCountries text label

This column holds people-group / ethnolinguistic names spanning countries (e.g. 'Turk', 'Persian', 'Kurd, Kurmanji', 'Arab, Moroccan'), with frequent religious-tradition qualifiers like '(Hindu traditions)' and '(Muslim traditions)' appearing in 985 and 424 rows respectively. Values are short (mean 12.4 chars, median 2 words) and 47.5% are single-word labels, yet 35.4% (2,520 rows) are duplicates across 4,604 unique strings out of 7,124 — the same group recurs across countries. The most surprising entry is 'Deaf' at 151 occurrences, which sits oddly alongside ethnic categories.

Treatment: Treat as a categorical people-group label; normalize qualifiers in parentheses and join on (group, country) for cross-country aggregation.

Population numeric feature

Population counts per record, ranging from 10 to 135,533,000 with a median of just 30,000. The distribution is extraordinarily right-skewed (skew 21.1, kurtosis 607) — the mean of ~502,570 sits far above Q3 of 129,000, and ~14.9% of rows flag as outliers, suggesting a mix of small localities with a few country- or megacity-scale entries. Null rate is negligible (0.21%) and there are no zeros.

Treatment: Apply a log1p transform before any modelling to tame the extreme skew.

Category categorical label

Three-level categorical with values "1", "2", "3" and no nulls across 7124 rows. Distribution is imbalanced: "3" dominates at 60.3% (4299), "1" follows at 2330, and "2" is rare at 495. Entropy ratio of 0.78 confirms the skew toward the majority class.

Treatment: Treat as a categorical label; consider class weighting or stratified sampling to handle the minority class "2".

ROL3 text feature

ROL3 holds three-letter ISO 639-3 language codes — every value is exactly 3 characters and one word, with top entries like 'hin', 'ben', 'urd', 'guj', and 'tel' pointing to South Asian languages. The distribution is heavily skewed toward Hindi (662 of 7124) and only 1565 unique codes appear across 7124 rows, giving a 78% duplicate rate. No nulls, no empties, no formatting noise.

Treatment: treat as a categorical language code; one-hot or target-encode, and consider grouping rare codes.

PrimaryLanguageName text feature

Categorical language label, almost always a single word (one_word_rate 0.70, word_mean 1.33) drawn from a vocabulary of 1,641 tokens across 1,563 distinct values. Hindi (662), Bengali (357), and Sindhi (191) dominate the 7,124 rows, and 78.1% of values are duplicates of an earlier row — expected for a controlled language taxonomy. Compound names like 'Pashto, Northern' and 'Punjabi, Eastern' indicate ISO-style subvariant naming rather than free text.

Treatment: Treat as a categorical factor; normalise the comma-separated subvariants before one-hot or target encoding.

PrimaryLanguageDialect categorical metadata

Free-text or controlled-vocabulary field naming a primary language dialect, with 303 distinct values across 7124 rows but populated in only 5.5% of records (null_rate 0.945). Distribution is essentially flat — entropy_ratio 0.97 and the modal value 'Punjabi' covers just 3.1% of non-nulls (12 occurrences) — so no dialect dominates. The mix spans South Asian, Middle Eastern, African, and European dialects, suggesting a global but extremely sparse roster.

Treatment: Drop or collapse into a coarser language grouping; too sparse and high-cardinality to use directly as a feature.

NumberLanguagesSpoken numeric feature

Counts the number of languages spoken, with values ranging from 1 to 120 across 7,124 rows and no nulls. The distribution is heavily right-skewed (skew 4.87, kurtosis 36.3): the median is 1 and Q3 is 5, yet the mean is 4.33 and 597 rows (8.4%) flag as outliers, suggesting a long tail of implausibly high counts up to 120.

Treatment: Cap or log-transform before modelling, and audit the extreme tail (values up to 120) for data-entry errors.

OfficialLang categorical feature

Categorical column listing an official language per record, with 79 distinct values across 7124 rows and effectively no nulls (0.08%). Hindi dominates at 28.5% (2032 rows), followed by Urdu (767), Standard Arabic (657), Mandarin (475), and English (433), giving an entropy ratio of 0.66 — moderately concentrated rather than uniform. The South/Central Asian skew is notable: five of the top ten values are languages of that region, which may bias any downstream language-level analysis.

Treatment: Group rare languages into an 'Other' bucket and one-hot or target-encode before modelling.

SpeakNationalLang unknown other

Column 'SpeakNationalLang' was skipped by the profiler, so type, cardinality and value distribution are all unavailable. The only confirmed facts are that it has 7124 rows and zero nulls. The name suggests a flag or category for whether a respondent speaks the national language, but this cannot be verified from the evidence.

Treatment: Re-profile with type inference enabled before deciding on any downstream handling.

BibleStatus numeric feature

BibleStatus is an ordinal/categorical code stored as a small integer, taking just 6 distinct values from 0 to 5 across 7,124 rows with no nulls. The distribution is heavily concentrated at the top (median 5, Q1=4, Q3=5, mean 4.05) with strong negative skew (-1.51), and 13.5% of rows flagged as low-end outliers plus a 3.8% zero rate. This looks like a status/level code rather than a true numeric measurement.

Treatment: Treat as an ordinal category (or one-hot encode) rather than a continuous numeric.

BibleYear categorical metadata

BibleYear appears to be a publication/edition year field for Bible translations, encoded mostly as date ranges like "1818-2022" rather than single years. Cardinality is high (163 distinct values across 7124 rows) and the column is missing for 45.79% of rows, which is a major coverage gap. The top value "1818-2022" covers 17.14% of non-nulls and most frequent entries are spans, while plain single years like "1954" (191 occurrences) are the exception.

Treatment: Parse into start_year and end_year integers, then decide imputation given the 45.79% null rate.

NTYear categorical feature

NTYear appears to hold year-range strings (e.g. "1811-1998", "1801-1984") indicating a span between two dates, though the presence of "Yes" as the third most common value (345 rows) signals encoding inconsistency. The column has 305 distinct values across 7124 rows with a high null rate of 24.65%, and the top value covers only 12.3% of non-nulls, so the distribution is fairly spread (entropy ratio 0.735).

Treatment: Parse year-range strings into start/end numeric fields and quarantine non-conforming values like "Yes" before modelling.

PortionsYear categorical metadata

PortionsYear appears to be a free-form field describing the year range covered by record portions, with 460 distinct values across 7124 rows and a 13.5% null rate. Most entries are date ranges like '1806-1962' or '1800-1980', but the single most common value is the literal string 'Yes' (821 rows, 13.3%), suggesting inconsistent data entry where a yes/no answer leaks into a date-range field. High entropy ratio (0.71) confirms values are spread across many ranges rather than concentrated.

Treatment: Parse date ranges into start/end year numerics and isolate the 'Yes' contamination as a separate boolean flag before use.

TranslationNeedQuestionable unknown other

The column "TranslationNeedQuestionable" was skipped by the profiler, so no type, uniqueness, or value statistics are available. All we know is that it has 7124 rows with a 0.0 null rate; nothing else can be inferred from the evidence.

Treatment: Re-profile or inspect manually before deciding on use; current evidence is insufficient.

JPScale numeric metadata

JPScale is a numeric column that is entirely constant: all 7124 rows hold the value 1.0 with zero nulls, zero variance, and a single unique value. It carries no information for any downstream model or comparison and was flagged as constant.

Treatment: Drop; constant column with no variance.

JPScalePC categorical feature

JPScalePC is a 5-level categorical code (values "1" through "5") with no nulls across 7124 rows, likely an ordinal scale or rating. The distribution is heavily concentrated at "1" (70.2% of rows), with "3" the rarest at just 205 occurrences, yielding an entropy ratio of 0.59. The non-monotonic frequency order (1 > 4 > 2 > 5 > 3) is unusual for a true ordinal scale and worth checking.

Treatment: Treat as ordinal categorical; consider grouping minority levels (3, 5) given the dominance of "1".

JPScalePGAC categorical feature

JPScalePGAC is a 5-level categorical code (values "1" through "5"), likely a Japanese seismic intensity / PGA scale rating. The distribution is severely imbalanced: "1" accounts for 6910 of 7124 rows (top_rate 0.97), entropy_ratio is just 0.10, and the remaining four levels together hold under 220 records. No nulls are present.

Treatment: Collapse rare levels or binarise as "1" vs other before modelling.

LeastReached categorical metadata

This column is a single-valued categorical flag holding "Y" for all 7124 rows, with no nulls and zero entropy. Because cardinality is 1 and top_rate is 1.0, it carries no information for any downstream model.

Treatment: Drop; constant column with no variance.

LeastReachedPC categorical feature

Binary Y/N flag indicating whether some 'least reached' people-group condition is met. The column is fully populated across 7124 rows with only 2 distinct values, skewed toward 'Y' at 72.3% (5152) versus 'N' at 1972. Entropy ratio of 0.85 shows the split is uneven but still informative.

Treatment: Encode as a 0/1 boolean for modelling.

LeastReachedPGAC categorical feature

A binary Y/N flag (likely indicating whether the least-reached PGAC condition was met) with no nulls across 7124 rows. The distribution is severely imbalanced: 'Y' accounts for 6910 rows (97.0%) versus only 214 'N', yielding entropy_ratio of just 0.19. As a near-constant feature it carries little discriminative signal on its own.

Treatment: Encode as 0/1 but consider dropping or pairing with rare-class oversampling given the 97/3 imbalance.

GSEC categorical feature

GSEC is a low-cardinality categorical with 8 distinct values across 7124 rows and no nulls. The dominant value is the empty string at 51.08% (3639 rows), followed by '1' at 2767; the remaining codes ('0' through '6') together account for under 10% of rows. The mix of blanks and small integer codes suggests an optional categorical flag where 'missing' is encoded as '' rather than null.

Treatment: Recode '' as explicit missing and one-hot encode the remaining small-integer categories.

HasAudioRecordings categorical feature

Binary Y/N flag indicating whether a record has associated audio recordings. The class is heavily imbalanced toward 'Y' at 86.9% (6188 of 7124), with no nulls. Entropy ratio of 0.56 confirms the skew but the minority 'N' class still has 936 observations, enough to be usable.

Treatment: Encode as a 0/1 boolean; watch for class imbalance if used as a target.

NTOnline categorical feature

NTOnline is a categorical flag with only one observed value, 'Y', across 5528 non-null rows, while 22.4% of rows are null. With cardinality 1 and entropy 0, it carries no discriminative signal—presence vs. absence is the only information available.

Treatment: Drop, or replace with a binary is_present indicator if the null pattern is meaningful.

RLG3 numeric feature

RLG3 is a discrete numeric column with only 7 unique values spanning 2 to 9, suggesting an ordinal rating or Likert-style scale rather than a continuous measurement. The distribution is tight around the median of 6 (IQR=1, Q1=5, Q3=6) with mild left skew (-0.46), but 10.6% of rows (757) fall outside the IQR fences — an artifact of the narrow box rather than true anomalies.

Treatment: Treat as an ordinal categorical; the outlier flag is a side-effect of the compressed IQR, not bad data.

RLG3PC numeric feature

RLG3PC is an integer-coded ordinal feature with only 8 distinct values spanning 1-9 and a tight IQR of 1 (Q1=5, Q3=6). The distribution is left-skewed (-0.95) and concentrated around the median of 5, yet 14.3% of rows (1022) fall outside the IQR fence, suggesting a heavy lower tail rather than true anomalies. No nulls or zeros are present.

Treatment: Treat as an ordinal/categorical scale rather than continuous; the outlier rate reflects skew, not errors.