joshua project joshua project countries

Overview

Source: /home/coolhand/html/datavis/data_trove/joshua-project/joshua_project_countries.json

Saturn profiled 238 rows across 39 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).

PoplPeoplesLR numeric feature

Likely a population count by some geographic or organisational unit ('PoplPeoples'), spanning from 50 to ~1.39B with a median of 532,500. The distribution is extremely right-skewed (skew 12.1, kurtosis 156) with 31 outliers (15.8%) and a std (~104M) far exceeding the mean (~18M), suggesting a few massive entities dominate. Also notable: 17.65% of rows are null.

Treatment: Log-transform and impute the ~17.65% nulls before modelling.

PercentUnknown numeric feature

PercentUnknown is a numeric proportion ranging from 0.0 to 2.679, with a mean of 0.218 and median of 0.176. About 34% of the 238 rows are exactly zero (q1 is also 0.0), yet the column is heavily right-tailed with skew 3.92 and kurtosis 32.16, plus a max above 1.0 that is unusual if this was meant to be a 0-1 share. Three outliers (1.3%) sit far above the bulk of the distribution.

Treatment: Verify whether values >1 are valid, then log1p-transform given the heavy right skew and zero-inflation.

ReligionPrimary categorical feature

Primary religion of each record across 238 rows with 6 distinct values and no nulls. Christianity dominates at 159/238 (top_rate 0.668), followed by Islam at 55, with Buddhism, Ethnic Religions, Non-Religious, and Hinduism sharing the long tail under 10 each. Entropy ratio of 0.54 confirms the heavy concentration in one category.

Treatment: One-hot encode, optionally collapsing the four smallest categories into 'Other' to handle imbalance.

SecurityLevel numeric feature

SecurityLevel takes only 3 distinct integer values across 238 rows (min 0, max 2) with no nulls, suggesting an ordinal tier code rather than a continuous measure. The distribution is heavily weighted toward the lowest tier: 67.6% of rows are zero, the median is 0, and the mean is just 0.55, producing a right skew of 1.01. No outliers were flagged, which is consistent with a bounded categorical scale.

Treatment: Treat as an ordinal category (0/1/2) rather than a continuous numeric.

PercentNonReligious numeric feature

This column reports the percentage of a population that is non-religious across 238 rows, with 228 unique values and no nulls. The distribution is heavily right-skewed (skew 2.61, kurtosis 8.00): the median is just 2.85% while the mean is 7.31% and the max reaches 68.81%, with 30 outliers (12.6% of rows) and 4.6% of values exactly zero. The std (11.39) dwarfs the IQR (7.08), confirming a long upper tail rather than a symmetric spread.

Treatment: Log1p- or rank-transform before modelling to tame the heavy right tail.

ISO2 categorical identifier

This column holds ISO2 country codes (AF, AL, DZ, AS, AD...), serving as a unique key with 238 distinct values across 238 rows and zero nulls. Cardinality equals row count and entropy_ratio is 1.0, meaning every code appears exactly once — top_rate is just 0.0042. The long_tail alert is expected here since the column is effectively a primary identifier.

Treatment: Use as the join key to merge country-level attributes.

JPScaleImageURL categorical feature

This column holds URLs to one of five Joshua Project 'gauge' images (gauge-1.png through gauge-5.png), almost certainly a visual encoding of an ordinal progress/status score on a 1-5 scale. With only 5 unique values across 238 rows and no nulls, gauge-5 leads at 37.4% (89 rows) while gauge-2 is rarest at 11 rows, suggesting a skew toward the high end of the scale. The URL itself carries no information beyond the trailing digit.

Treatment: Extract the trailing digit (1-5) and treat as an ordinal feature; drop the URL string.

PercentEthnicReligions numeric feature

Numeric share (0–75.16) representing the percentage of ethnic-religion adherents per row, likely one country or region per record across 238 entries with 199 unique values. The distribution is heavily right-skewed (skew 3.29, kurtosis 12.38) with a median of just 1.12% but a mean of 5.59% and a long tail producing 27 outliers (11.3% of rows); 16.8% of rows are exact zeros. The IQR (5.52) is far smaller than the std (11.21), confirming most values cluster near zero while a few cases dominate.

Treatment: Apply a log1p or similar transform before modelling to tame the heavy right tail and zero mass.

RegionName categorical feature

RegionName is a categorical geographic grouping with 12 distinct values across 238 rows and no nulls. The distribution is remarkably even — entropy ratio 0.96 and the top bucket 'America, North and Caribbean' accounts for only 12.6% (30 rows) — suggesting these are world regions assigned to countries or similar entities. The Asian regions (Southeast: 11, Central: 10) are notably smaller than the African and European groupings.

Treatment: one-hot or target-encode for modelling; safe to use as a grouping key.

CntPeoples numeric feature

Numeric count of people per record, fully populated across 238 rows with 96 distinct values ranging from 1 to 2262. The distribution is severely right-skewed (skew 8.20, kurtosis 85.91): the median is 24.5 and Q3 is 56.75, yet the mean is 68.83 and the max reaches 2262, with 24 outliers (10.1% outlier rate). Std (184.37) far exceeds the IQR (48.75), confirming a long heavy tail.

Treatment: log-transform (or winsorize the top decile) before any distance- or variance-based modelling.

PoplPeoplesFPG numeric feature

Likely a population count of 'people groups' (PoplPeoplesFPG), with values ranging from 50 to roughly 1.09 billion and a median of 217,000. The distribution is extremely right-skewed (skew 11.6, kurtosis 138.7) with 19% of values flagged as outliers, and 31.5% of rows are null. The mean (~12.26M) sits far above the median, confirming a handful of massive groups dominate.

Treatment: log-transform and impute the 31.5% nulls before any modelling.

ROG3 categorical identifier

ROG3 looks like a country/region code identifier — every one of the 238 rows holds a unique two-letter value (AF, AL, AG, AQ, ...), giving cardinality 238 and entropy_ratio 1.0. With top_rate at 0.0042 and no nulls, this column carries no predictive signal on its own and behaves as a primary key for the row.

Treatment: Use as a join key to country-level attributes; drop from any model as a feature.

PercentEvangelical numeric feature

Numeric share (0–53.4%) of an evangelical population across 238 rows, almost all unique (232 distinct values), suggesting one observation per geographic or demographic unit. The distribution is right-skewed (skew 1.25) with mean 10.46% well above the median 6.92% and an IQR spanning 1.38–17.69%, so a long tail of high-evangelical units pulls the average up. About 2.5% of rows are null and 4 outliers sit beyond the upper whisker; no zero values were recorded.

Treatment: Consider a log or sqrt transform before modelling to tame the right skew.

ROL3OfficialLanguage categorical feature

ISO 639-3 language codes denoting each entity's official language, with 88 distinct values across 238 rows and no nulls. English dominates at 26.5% (63 rows) followed by French (25), Spanish (21), and Arabic (20), but the long tail is heavy — entropy ratio 0.74 against cardinality 88 means most codes appear only once or twice (e.g. aln, smo at 2). Worth noting some codes like 'arb' and 'cmn' are macro-language specific variants, so consistency of coding granularity should be checked.

Treatment: Group the long tail into an 'other' bucket or target-encode before modelling.

TranslationUnspecified numeric feature

A heavily right-skewed count of 'TranslationUnspecified' occurrences per row, with 21 distinct integer values across 238 rows. Roughly 42% of rows are zero and the median is 1, yet the maximum reaches 80 against a Q3 of 2, producing extreme skew (6.02) and kurtosis (45.1). About 11% of values (26 rows) flag as outliers, so a small tail dominates the mean (2.80) versus median (1).

Treatment: Log1p- or rank-transform before modelling, and consider winsorising the long tail.

TranslationNeeded numeric feature

TranslationNeeded is a numeric count column where 68% of rows are zero and the median and Q1-Q3 sit at 0-1, suggesting it tracks how many items required translation per record. The distribution is extremely right-skewed (skew 8.56, kurtosis 79.6) with a max of 104 against a mean of 2.24, and 31 rows (13%) flag as outliers. With only 18 unique values across 238 rows, this behaves more like a sparse event counter than a continuous metric.

Treatment: Log1p-transform or binarise (zero vs non-zero) before modelling to tame the heavy tail.

TranslationStarted numeric feature

A numeric counter named TranslationStarted, likely the number of translation jobs initiated per row/entity. Over half the rows are zero (zero_rate 0.517) and the median is 0 with q3=3, yet the max reaches 261, producing extreme skew (7.81) and kurtosis (64.98) plus 35 outliers (14.7%). The mean of 6.46 is pulled far above the median by these heavy-tail cases.

Treatment: Apply a log1p (or zero-inflated) transform before modelling to tame the heavy right tail.

RegionCode numeric feature

RegionCode holds 12 distinct integer values from 1 to 12 across 238 rows with no nulls, which strongly suggests a categorical region identifier rather than a true numeric measure. The distribution is mildly left-skewed (skew -0.47) with a median of 8 and no outliers, indicating fairly even coverage across the higher-numbered regions.

Treatment: Cast to categorical and one-hot or target-encode before modelling.

Ctry categorical identifier

Despite the abbreviated header `Ctry`, this column holds full country names (Afghanistan, Albania, Algeria, …) and acts as a unique key: all 238 rows have distinct values, with entropy_ratio of ~1.0 and a top_rate of just 0.0042. There are no nulls, and the alphabetical run in top_values suggests the dataset is a one-row-per-country reference table.

Treatment: Use as the primary key; left-join other country-level data on this column.

BibleNewTestament numeric feature

Numeric counts of New Testament references per row, ranging from 0 to 274 with a median of just 3. The distribution is extremely right-skewed (skew 6.08, kurtosis 49.4) with 23.1% zeros and 10.5% outliers, so a small number of rows dominate the totals while most carry few or no references.

Treatment: Apply a log1p transform and consider winsorising before modelling.

BibleComplete numeric feature

A numeric count, plausibly the number of times a respondent has read the Bible cover-to-cover or a similar completion tally, ranging from 0 to 162 across 238 rows with no nulls. The distribution is heavily right-skewed (skew 3.32, kurtosis 17.0): median is 9 with an IQR of 15, yet 18 values (7.6%) qualify as outliers and the max of 162 sits far above q3=19. Only 1.7% are zero, so non-engagement is rare; the long tail of high counts is the dominant surprise.

Treatment: Apply a log1p transform before modelling to tame the heavy right tail.

Window1040 categorical feature

Binary Y/N flag with no nulls across 238 rows. The distribution is imbalanced toward 'N' at 71.4% (170 of 238) versus 'Y' at 68, giving an entropy ratio of 0.86. The 'Window1040' name suggests a windowed indicator tied to event 1040, but the evidence does not clarify what that event represents.

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

JPScaleText categorical label

JPScaleText is a 5-level ordinal label describing reach status (Unreached, Minimally, Superficially, Partially, Significantly Reached) across 238 complete rows. The distribution is fairly balanced with high entropy ratio (0.887) and a modal class of 'Significantly Reached' at 37.4%, though 'Minimally Reached' is sparse at only 11 records. No nulls and tight cardinality make this a clean categorical feature.

Treatment: Encode as an ordered ordinal (Unreached → Significantly Reached) before modelling.

CntPrimaryLanguages numeric feature

CntPrimaryLanguages is a numeric count (likely the number of primary languages associated with each row, e.g., a country or region) ranging from 1 to 827 across 238 rows with no nulls. The distribution is heavily right-skewed (skew 5.42, kurtosis 36.5): the median is just 20 while the mean is 45.6 and the max reaches 827, with 20 outliers (8.4%) sitting well above the Q3 of 46.5. Most entities have modest language counts but a small tail dominates the variance (std 90.1).

Treatment: log-transform (or winsorize the upper tail) before any distance- or regression-based modelling.

OfficialLang categorical feature

This column lists the official language(s) of 238 entities, almost certainly countries or territories, with one near-null. English dominates at 63 occurrences (26.6% top rate), followed by French (25), Spanish (21), and Standard Arabic (20), but the long tail spans 87 distinct values including narrow entries like Gheg Albanian and Samoan, yielding entropy 4.78 (ratio 0.74). The high cardinality relative to 238 rows means many languages appear only once or twice.

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

BiblePortions numeric feature

Numeric count of Bible portions, likely per language or region, across 238 rows with no nulls and only 35 unique values. The distribution is severely right-skewed (skew 5.57, kurtosis 36.36) with a median of 2 against a max of 161, and 26.05% of rows are zero. Roughly 10.9% of values flag as outliers, so a few entries dominate the tail.

Treatment: Apply a log1p transform before modelling to tame the heavy right tail.

ROG2 categorical feature

ROG2 is a low-cardinality categorical with 7 region-like codes (AFR, EUR, ASI, NAR, SOP, LAM, AUS) across 238 rows and no nulls. The distribution is fairly even — entropy ratio 0.893 and the top value AFR holds just 24.4% — though AUS is a tiny tail with only 2 records. The codes look like geographic groupings (Africa, Europe, Asia, North America, South Pacific, Latin America, Australia).

Treatment: One-hot encode; consider merging AUS (n=2) into a neighbouring region or 'Other' to avoid sparse dummies.

PercentIslam numeric feature

Numeric share (0–99.47) of Muslim population per row, almost certainly one row per country or territory given n=238. The distribution is heavily right-skewed (skew 1.35) with median just 2.51% but mean 22.31%, and 17.2% of rows are exactly zero while 16.8% flag as outliers — a bimodal world where most places have negligible Muslim populations and a minority are overwhelmingly Muslim.

Treatment: Consider a log1p or logit transform, or bucket into low/medium/high bands, before modelling.

RLG3Primary numeric feature

RLG3Primary is a small-cardinality numeric code with only 6 unique values spanning 1 to 7 across 238 rows and no nulls. The distribution is bottom-heavy: median is 1.0 and Q1 equals the minimum, yet Q3 reaches 5.75, producing a wide IQR of 4.75 and right skew of 0.98. This looks like an ordinal category (e.g., a primary-rating or grade code) masquerading as a number rather than a continuous measurement.

Treatment: Treat as an ordinal categorical and one-hot or ordinal-encode before modelling.

Capital categorical identifier

This column lists capital cities, with 233 unique values across 238 rows and a null rate of 1.68%. Cardinality is essentially one-per-row (entropy ratio 0.9997), and the only repeat is Kingston appearing twice — likely Jamaica and Norfolk Island sharing the name. Effectively a near-unique label tied to the country/territory record.

Treatment: Treat as a near-unique label; drop or use as a join key rather than a model feature.

Population numeric feature

This is a country/region population count, with 238 rows and 230 unique values, no nulls. The distribution is extremely right-skewed (skew 9.10, kurtosis 89.06): the median is 5,606,500 but the max reaches 1,463,866,000, dwarfing the Q3 of 23,200,000. About 10.9% of values (26 rows) flag as outliers, consistent with a handful of population giants like China/India-scale entities.

Treatment: log-transform before any modelling or aggregation.

CntPeoplesLR numeric feature

CntPeoplesLR is a numeric count of people (likely a left/right group size or attendance metric) with 57 distinct values across 238 rows and a 15.13% null rate. The distribution is severely right-skewed (skew 10.79, kurtosis 129.0): the median is 6.5 and Q3 is 24.75, yet the max reaches 2032 and the mean is 35.27 with std 157.42. 17 outliers (8.42%) pull the tail dramatically, and no zeros are recorded.

Treatment: log-transform (or winsorize the upper tail) and impute the ~15% nulls before modelling.

CntPeoplesFPG numeric feature

A numeric count column (likely 'count of people FPG') with only 44 unique values across 238 rows and 28.57% nulls. The distribution is extremely right-skewed (skew 10.07, kurtosis 109.5): median is 4 and Q3 is 12.75, yet the max reaches 1700, producing 18 outliers (10.6% rate) and inflating the mean to 28.04 against a std of 143.69.

Treatment: Impute the 28.57% nulls and apply a log1p transform before modelling to tame the heavy right tail.

PercentHinduism numeric feature

Country-level share of population identifying as Hindu, expressed as a percentage. The distribution is dominated by zeros (zero_rate 0.559, median 0) with a long right tail to 82.4, producing extreme skew (6.90) and kurtosis (53.0); 43 of 238 rows (18.1%) flag as outliers, presumably the few Hindu-majority countries.

Treatment: Apply a log1p or zero-inflated transform before modelling, since most values are 0 with a heavy right tail.

PercentOtherSmall numeric feature

A numeric share/percentage feature called PercentOtherSmall, with 238 rows and 199 unique values, no nulls, but 16.4% zeros and a long right tail (median 0.29, max 12.84). Skew of 5.05 and kurtosis 39.4 with 15 outliers (6.3%) signal an extremely heavy-tailed distribution. Despite the name, values exceed 1, so this is not bounded as a 0–1 proportion.

Treatment: Apply a log1p transform and consider a zero-inflation indicator before modelling.

PercentBuddhism numeric feature

This column appears to be the percentage of Buddhists per country (or similar geographic unit), with 238 rows and 125 unique values. The distribution is extremely right-skewed (skew 4.59, kurtosis 20.4): nearly half the rows are zero (zero_rate 0.48), the median is 0.004%, yet the max reaches 88.74%. Saturn flagged 38 outliers (16% of rows), reflecting the handful of Buddhist-majority countries dominating the tail.

Treatment: Apply a log1p or similar transform before modelling to tame the heavy right tail.

JPScaleCtry numeric feature

JPScaleCtry holds an integer 1–5 rating with only 5 unique values across 238 rows and no nulls, consistent with a Likert-style country-level scale. The distribution leans high (mean 3.62, median 4, Q1=3, Q3=5) with a left skew of -0.78, indicating most respondents cluster at the upper end. No outliers are flagged.

Treatment: Treat as an ordinal Likert feature; consider ordered encoding rather than raw numeric use.

PercentChristianity numeric feature

This column reports the percentage of Christians per row (likely country or region), spanning the full 0.02% to 100% range across 238 nearly unique values. The distribution is strongly bimodal-feeling: the median (75.3%) sits far above the mean (58.2%), with a wide IQR from 11.7% to 90.9% and negative skew (-0.54), suggesting many heavily Christian populations alongside a substantial cluster of very low-share rows. No nulls, no zeros, and no statistical outliers despite the extreme spread.

Treatment: Use as-is or rescale to 0-1; consider pairing with other religion-share columns since values are bounded percentages.

ISO3 categorical identifier

ISO3 is the standard three-letter country code, with all 238 values unique (AFG, ALB, DZA, ASM, ...) and zero nulls. Maximum entropy ratio (≈1.0) and a top_rate of 0.0042 confirm one row per country, making this a clean primary key for the table.

Treatment: use as the join key to merge country-level data; do not feed into models as a feature.