Statistical methods in geography

Geography (BA) · Techniques · free study notes for exams, placements and interviews

In short

Statistical methods in geography are the quantitative tools used to summarise, compare, test and model geographical data — rainfall, population, literacy, crop area, settlement spacing, migration flows. They came into the discipline with the "quantitative revolution" of the 1950s-60s (Burton, Garrison, Haggett), which turned description into measurement and hypothesis testing. Univariate (mean, standard deviation), bivariate (correlation, regression) and spatial (nearest neighbour, mean centre, Moran's I) techniques together form the "Techniques" paper in almost every Indian BA/MA Geography syllabus. They are examined in UGC-NET Geography Paper II, UPSC Geography optional (Paper I, Part B), State PSC and university practical records, and are used daily by planners, census and NSO statistical staff, market-research and GIS analysts.

Explained

Geographical data, variables and scales of measurement

Why quantification: it allows precise comparison between places, testing of hypotheses, and statements of probability instead of impression.
Individual vs grouped data: raw values per unit (district, village, station) vs a frequency distribution with class intervals.
Discrete vs continuous: number of villages (discrete) vs rainfall or slope angle (continuous).
Scales of measurement (Stevens):
- Nominal — categories only (soil type, land use class); only mode, counts, chi-square apply.
- Ordinal — ranked (city hierarchy, Likert "very poor-very good"); median, quartiles, Spearman's rho.
- Interval — equal units, arbitrary zero (temperature in °C); mean, SD, Pearson's r.
- Ratio — equal units, true zero (rainfall mm, population, area km²); all operations, including CV and ratios.
Spatial data types: point (settlements, wells), line (roads, rivers), area/areal (districts), surface (elevation, rainfall field), flow (migration, commuting).
Data sources in India: Census of India (decennial; last completed 2011, the next Census announced with house-listing in 2026 and population enumeration with reference date 1 March 2027 — dates may be revised), National Statistics Office / NSS and the Periodic Labour Force Survey, NFHS, Agricultural Statistics at a Glance, India Meteorological Department records, Forest Survey of India, data.gov.in, Bhuvan (NRSC). Official statistics are collected under the Collection of Statistics Act, 2008 (amended 2017), which makes returns compulsory and protects individual confidentiality.
Primary vs secondary data: field survey, schedules and questionnaires vs published/official records; always state the base year, the areal unit and the definition used (e.g. "urban" = statutory towns + census towns).

Describing a distribution: central tendency and dispersion

Mean (arithmetic): sum ÷ number of observations; uses every value, but is pulled by extremes (one Mumbai distorts a state mean).
Median: middle value of an ordered series (or ½(n/2 + n/2+1 terms) for even n); resistant to outliers — preferred for income, landholding size, village population.
Mode: most frequent value or modal class; the only average for nominal data (dominant crop, modal house type).
Other averages: weighted mean (district means weighted by population), geometric mean (growth rates, index numbers), harmonic mean (average speed, density ratios).
Partition values: quartiles, deciles, percentiles — used for ranking districts in indices such as the NITI Aayog SDG India Index.
Dispersion: range, quartile deviation (semi-interquartile range), mean deviation, standard deviation (σ) and variance (σ²).
Coefficient of variation (CV = σ/mean × 100): unit-free, so it compares variability of different things or places — the standard measure of rainfall variability and rainfall reliability in India (high CV in Rajasthan and Gujarat, low in the Western Ghats and the Northeast).
Shape: skewness (Sk = (mean − mode)/σ; positive for income, landholding, settlement size) and kurtosis (peakedness). Normal curve: symmetrical, mean = median = mode, with ~68%, ~95% and ~99.7% of values within 1σ, 2σ and 3σ.
Standardisation: z-score z = (x − mean)/σ puts variables of different units on one scale; the basis of composite indices (e.g. a development index combining literacy, income and infrastructure).

Sampling and hypothesis testing

Why sample: complete enumeration of all villages or households is costly; a probability sample gives estimates with measurable error.
Types: simple random (lottery/random numbers), systematic (every kth unit, every 10th house), stratified (strata by relief, caste, income — proportionate or disproportionate), cluster/multi-stage (NSS design: villages/urban blocks → households), spatial sampling (point, line/traverse, quadrat; random, systematic and stratified-systematic unaligned grids).
Non-probability: purposive, quota, snowball (migrant studies) — convenient but not generalisable.
Hypothesis testing steps: (1) state null H₀ (no difference / no association) and alternative H₁; (2) choose the test and the significance level α (usually 0.05 or 0.01); (3) compute the statistic; (4) compare with the critical (table) value at the right degrees of freedom; (5) reject or fail to reject H₀ and interpret geographically.
Errors: Type I — rejecting a true H₀ (α); Type II — accepting a false H₀ (β). One-tailed tests for directional hypotheses.
Common tests: z-test (large samples, n > 30), t-test (small samples, comparing two means), chi-square (nominal counts: goodness of fit and contingency/independence), ANOVA / F-test (three or more group means, e.g. yields on three soil types), Kolmogorov-Smirnov (comparing distributions), Mann-Whitney U and Kruskal-Wallis (non-parametric, ordinal data).
Parametric vs non-parametric: parametric tests assume normality and interval data and are more powerful; non-parametric tests suit ranks, small samples and skewed geographical data.

Association: correlation and regression

Correlation measures the strength and direction of covariation; r ranges −1 to +1, with 0 = no linear relation.
Pearson's product-moment r: interval/ratio data, linear relation, both variables roughly normal. Spearman's rank rho: ordinal data or non-linear monotonic relations, and robust to outliers. Kendall's tau for many tied ranks.
Significance of r: test with t = r√(n−2)/√(1−r²) at (n−2) degrees of freedom; a high r from n = 5 may be meaningless.
Interpretation cautions: correlation is not causation; a third variable may drive both (spurious correlation); ecological fallacy — a district-level correlation cannot be read as a statement about individuals; modifiable areal unit problem (MAUP) — r changes when the areal units are redrawn or aggregated.
Regression: fits the best straight line Y = a + bX by least squares (minimising Σ of squared vertical deviations). b = slope (change in Y per unit X), a = intercept. Y is the dependent/response variable, X the independent/predictor.
Coefficient of determination r²: proportion of variance in Y explained by X; the rest are residuals. Mapping residuals is a classic geographical step — positive residuals show places performing better than the model predicts.
Multiple and stepwise regression: several predictors (Y = a + b₁X₁ + b₂X₂ …) with an adjusted R²; watch for multicollinearity between predictors.
Other multivariate methods named in syllabi: principal component and factor analysis (reducing many census variables to a few dimensions), cluster analysis and regionalisation, discriminant analysis.
Time series: trend (linear or moving average), seasonal and cyclical components, residual/irregular — used for rainfall series, population growth and crop output.

Spatial statistics: point patterns, centrography, autocorrelation

Nearest neighbour analysis (Clark-Evans Rn): classifies a point pattern as clustered (Rn → 0), random (Rn ≈ 1) or uniform/regular (Rn → 2.149 maximum, hexagonal). Sensitive to the boundary chosen and to edge effects.
Quadrat analysis: divide the area into equal cells, count points per cell, and compare the variance-mean ratio (VMR) with 1 — VMR > 1 clustered, = 1 random (Poisson), < 1 uniform. Depends strongly on quadrat size.
Centrographic measures: mean centre (average of x and y coordinates — the "centre of population" of India shifts each census), weighted mean centre, median centre, and standard distance — the spatial equivalent of σ, giving a circle of dispersion around the mean centre.
Spatial autocorrelation: nearby places resemble each other, so observations are not independent (Tobler's first law). Moran's I (−1 to +1; positive = clustering of like values) and Geary's C (≈1 = no autocorrelation) measure it; LISA gives local hot and cold spots, as do Getis-Ord G*.
Spatial interaction: gravity model Tij = k PiPj / dij^b, Reilly's law of retail gravitation, and potential models — used for market area and migration studies.
Surface and interpolation statistics: trend surface analysis, isopleth fitting, and inverse-distance or kriging interpolation of rainfall and groundwater data.

Indices of concentration, specialisation and hierarchy

Lorenz curve: cumulative % of a variable (income, landholding, crop area) plotted against cumulative % of units; the further from the 45° line of equal distribution, the greater the concentration.
Gini coefficient: the Lorenz gap expressed as 0 (perfect equality) to 1 (total concentration); standard for income and land inequality.
Location quotient (LQ): a region's share of an activity compared with the national/state share; LQ > 1 = specialisation and (in base theory) an export or "basic" activity.
Index of concentration / dissimilarity and Gibbs-Martin index of diversification; Herfindahl index for industrial concentration.
Weaver's crop combination method: ranks crops by percentage area and finds the number of crops whose deviation from the theoretical equal-share distribution is least (least-squares deviation per combination) — the standard Indian agricultural-geography technique; Doi's method is a refinement.
Crop concentration, crop diversification and agricultural efficiency indices (Bhatia, Kendall's ranking coefficient) are examined in Indian agricultural geography.
Rank-size rule: Pr = P1 / r^q — the rth city's population equals the largest city's divided by its rank; deviations indicate primacy. Primacy index = P1/P2 or P1/(P2+P3+P4).
Demographic measures that use these tools: arithmetic/physiological density, decadal growth rate, sex ratio, dependency ratio, literacy rate, index of urbanisation, Whipple and Myers indices for age-data quality.
Tools and software: Excel/LibreOffice, SPSS, R, Python (pandas, statsmodels), QGIS/ArcGIS with spatial statistics tool-boxes, GeoDa for Moran's I. Report means with the same number of decimals as the data and always state the source, year and unit.

Key points

Quantitative revolution - the 1950s-60s shift to measurement, models and hypothesis testing in geography.
Scales of measurement - nominal, ordinal, interval, ratio; they decide which statistic is legitimate.
Mean - uses all values but is distorted by outliers; median is preferred for skewed data such as income or landholding.
Mode - the only average usable for nominal data.
Standard deviation - absolute dispersion in the data's own unit; variance is its square.
Coefficient of variation - σ as a % of the mean; the standard measure of Indian rainfall variability because it is unit-free.
z-score - (x − mean)/σ; standardises unlike variables for composite development indices.
Null hypothesis - the assumption of no difference/association that a test tries to reject at a chosen α.
Chi-square - tests observed counts against expected; df = (r−1)(c−1) for a contingency table, (k−1) for goodness of fit.
Pearson's r - linear correlation for interval data; Spearman's rho for ranks and non-linear monotonic data.
r² - share of variance in Y explained by X; residual mapping reveals unexplained geography.
Least squares regression - minimises squared vertical deviations to predict Y from X.
Ecological fallacy and MAUP - area-level results must not be applied to individuals, and change when units are redrawn.
Nearest neighbour index Rn - 0 clustered, 1 random, 2.149 perfectly hexagonal/uniform.
Mean centre and standard distance - the spatial mean and spatial standard deviation of a point set.
Moran's I - measures spatial autocorrelation; positive values mean like values cluster together.
Location quotient > 1 - the region is specialised in that activity relative to the larger area.
Weaver's method - identifies crop combination regions by minimum squared deviation from theoretical shares.
Rank-size rule - city size ∝ 1/rank; large deviation indicates urban primacy.

Formulas and examples

Mean: x̄ = Σx / n (grouped: Σfm / Σf, where m = class mid-value) — unit = unit of x
Median (grouped): M = L + [(n/2 − cf)/f] × h; L = lower limit of median class, cf = cumulative frequency before it, f = its frequency, h = class width
Standard deviation: σ = √[Σ(x − x̄)² / n] (sample s uses n − 1); Variance = σ²
Coefficient of variation: CV = (σ / x̄) × 100 (%)
z-score: z = (x − x̄) / σ (dimensionless)
Pearson's r = Σdxdy / √(Σdx² × Σdy²), where dx = x − x̄, dy = y − ȳ (dimensionless, −1 to +1)
Spearman's rho: rs = 1 − [6Σd² / (n³ − n)]; d = difference of ranks of the same unit
Regression of Y on X: Y = a + bX; b = Σdxdy / Σdx²; a = ȳ − b x̄; r² = coefficient of determination
Chi-square: χ² = Σ[(O − E)² / E]; df = k − 1 (goodness of fit) or (r − 1)(c − 1) (contingency)
t-test for r: t = r√(n − 2) / √(1 − r²), df = n − 2
Sample size for a proportion: n = z²pq / e² (p = expected proportion, q = 1 − p, e = allowable error)
Nearest neighbour index: Rn = D̄obs / D̄exp, where D̄exp = 0.5 × √(A/n); A = area (km²), n = number of points
Mean centre: (x̄, ȳ) = (Σxi/n, Σyi/n); Standard distance SD = √[Σ(xi − x̄)²/n + Σ(yi − ȳ)²/n]
Location quotient: LQ = (eᵢ/e) / (Eᵢ/E); e = regional, E = national employment (or area/output)
Gini (from Lorenz): G = 1 − Σ(Xk+1 − Xk)(Yk+1 + Yk), X and Y as cumulative proportions
Moran's I = (n/ΣΣwij) × [ΣΣwij(xi − x̄)(xj − x̄) / Σ(xi − x̄)²]; wij = spatial weight (1 if neighbours)
Rank-size rule: Pr = P1 / r^q (q ≈ 1); Primacy index = P1 / P2

Example 1 - Rainfall variability of seven stations (mm): 620, 740, 810, 900, 1050, 1180, 1300.

Σx = 6,600; n = 7 → x̄ = 6,600/7 = 942.86 mm. Ordered series, 4th value → median = 900 mm.
Σ(x − x̄)² = 104,238.5 + 41,152.2 + 17,651.8 + 1,837.0 + 11,479.6 + 56,235.4 + 127,549.0 = 360,143.5.
σ = √(360,143.5/7) = √51,449.1 = 226.8 mm; CV = (226.8/942.86) × 100 = 24.1% — moderate variability, far below the 40%+ CV typical of western Rajasthan.
z-score of the 1,300 mm station = (1,300 − 942.86)/226.8 = +1.57 (about 94th percentile on a normal curve).

Example 2 - Rainfall and crop yield: Pearson's r and regression.

Rainfall X (cm)	50	60	70	80	90
Yield Y (q/ha)	18	22	23	28	29

x̄ = 350/5 = 70; ȳ = 120/5 = 24. dx = −20, −10, 0, 10, 20; dy = −6, −2, −1, 4, 5.
Σdxdy = 120 + 20 + 0 + 40 + 100 = 280; Σdx² = 1,000; Σdy² = 82.
r = 280/√(1,000 × 82) = 280/286.36 = +0.978 — very strong positive correlation.
b = 280/1,000 = 0.28 q/ha per cm; a = 24 − (0.28 × 70) = 4.4 → Y = 4.4 + 0.28X.
Predicted yield at 75 cm = 4.4 + 21 = 25.4 q/ha. r² = 0.956 → 95.6% of yield variation is explained by rainfall.
Significance: t = 0.978√3/√(1 − 0.956) = 1.694/0.2098 = 8.07, df = 3; table t₀.₀₅ = 3.182 → significant.

Example 3 - Spearman's rank correlation between literacy and per capita income (6 districts).

District	A	B	C	D	E	F
Literacy rank	1	2	3	4	5	6
Income rank	2	1	4	3	6	5
d	−1	+1	−1	+1	−1	+1
d²	1	1	1	1	1	1

Σd² = 6, n = 6 → rs = 1 − (6 × 6)/(216 − 6) = 1 − 36/210 = 1 − 0.171 = +0.829 — strong positive rank agreement.

Example 4 - Chi-square goodness of fit: are villages evenly spread over four relief zones? Observed 40, 30, 20, 10 villages (n = 100); H₀: equal distribution → E = 25 each.

χ² = (15²+5²+5²+15²)/25 = (225+25+25+225)/25 = 500/25 = 20.0; df = 4 − 1 = 3; critical χ² at 5% = 7.815.
20.0 > 7.815 → reject H₀: settlement is significantly concentrated on the lower relief zones.

Example 5 - Three spatial indices.

(a) Nearest neighbour: 30 settlements in 1,200 km², observed mean nearest-neighbour distance 3.2 km. D̄exp = 0.5√(1,200/30) = 0.5 × 6.325 = 3.162 km → Rn = 3.2/3.162 = 1.01 → random pattern. Had D̄obs been 2.0 km, Rn = 0.63 → clustered (typical of a canal-irrigated tract).
(b) Location quotient: district textile employment 12,000 of 80,000 total; state 3,00,000 of 50,00,000. LQ = (12,000/80,000)/(3,00,000/50,00,000) = 0.15/0.06 = 2.5 → strongly specialised, an exporting (basic) sector.
(c) Weaver's crop combination for crop shares 45, 25, 15, 10, 5 (%): squared deviations of the ranking crops in each combination from their theoretical share, ÷ number of crops in that combination →

Combination	Σ(deviation)²	÷ n
1-crop (100)	3,025	3,025.0
2-crop (50,50)	650	325.0
3-crop (33.3 each)	541.7	180.6
4-crop (25 each)	725	181.3
5-crop (20 each)	1,000	200.0

Lowest value is 180.6 → the unit belongs to a three-crop combination region (the 4-crop value, 181.3, is very close, so marginal cases must be reported as such).

Where it is used

Government statistics: NSO/MoSPI, Census of India and State Directorates of Economics and Statistics — survey design, estimation, tabulation; recruitment through SSC JSO and ISS examinations.
Urban and regional planning: town and country planning departments, Smart City SPVs and metropolitan authorities use population projections, LQ-based economic base studies, rank-size and gravity models for master plans.
Agriculture and rural development: crop concentration, diversification and Weaver combination analysis for cropping-pattern policy, crop insurance (PMFBY) yield estimation and watershed planning.
Climate and disaster management: IMD rainfall variability (CV), return periods and trend analysis; NDMA/state SDMA vulnerability indices built from standardised (z-score) indicators.
GIS, market research and consulting: GIS analyst, data analyst and research associate roles at NRSC/Bhuvan projects, ISRO centres, RMSI, Esri India, Nielsen, Kantar and development think-tanks (NCAER, IIHS, CSTEP) — spatial autocorrelation, hot-spot and catchment analysis.
Academia, NGOs and teaching: M.Phil/PhD field surveys, UGC-NET and B.Ed/school teaching of practical geography, baseline and impact evaluations for NGOs and CSR programmes.

Interview questions

Q: When would you prefer the median to the mean for geographical data? A: When the distribution is strongly skewed or has outliers — landholding size, household income, village or city population — because the median is not pulled by extreme values.
Q: Why is the coefficient of variation, not the standard deviation, used to compare rainfall variability between stations? A: CV expresses dispersion as a percentage of the mean, so it is unit-free and comparable between a 3,000 mm Ghat station and a 300 mm desert station.
Q: What does a nearest neighbour index of 1 mean? A: The observed mean nearest-neighbour distance equals that expected from a random (Poisson) pattern, so the settlements are randomly distributed; 0 indicates clustering and 2.149 a perfectly hexagonal, uniform pattern.
Q: Difference between Pearson's r and Spearman's rho? A: Pearson's r measures linear covariation of interval/ratio data and assumes near-normality; Spearman's rho correlates ranks, so it suits ordinal data, non-linear monotonic relations and outlier-prone data.
Q: What is the ecological fallacy? A: Inferring something about individuals from area-level statistics — concluding that literate people earn more because literate districts are richer; the relationship may not hold at household level.
Q: What is the modifiable areal unit problem? A: Statistical results such as correlations and densities change when boundaries or levels of aggregation are changed, so conclusions depend partly on the units chosen.
Q: Interpret r² = 0.64 in a regression of yield on rainfall. A: 64% of the variation in yield is explained by rainfall; the remaining 36% is due to soil, inputs, management and unmeasured factors shown by the residuals.
Q: Why is spatial autocorrelation a problem in regression? A: Neighbouring observations are not independent, which violates an assumption of least squares, understates standard errors and inflates apparent significance; spatial regression models are used instead.
Q: What is a location quotient of 0.4 telling you? A: The region's share of that activity is only 40% of the national share, so it is under-represented and likely to be a net importer of that good or service.
Q: Which test would you use to check whether land use differs significantly across three soil types using plot counts? A: Chi-square test of independence on a contingency table, with df = (rows − 1)(columns − 1).
Q: How large a random sample is needed for a ±5% margin at 95% confidence on an unknown proportion? A: n = z²pq/e² = (1.96² × 0.5 × 0.5)/0.05² ≈ 384 respondents.
Q: Name the chief limitations of statistical methods in geography. A: They need reliable, comparable data; they treat space as a set of units and can ignore place, meaning and process; significance is not importance; and poor variable or boundary choices give precise but wrong answers.

Common mistakes

Applying the mean or standard deviation to nominal or ordinal data (soil class, Likert ranks) where only mode, counts or rank methods are valid.
Dividing by n instead of (n − 1) for a sample standard deviation, or mixing the two within one answer without saying which was used.
Reading correlation as causation, and quoting a high r from four or five observations without testing its significance.
Confusing the regression of Y on X with that of X on Y, or extrapolating the line far outside the observed range of X.
Forgetting degrees of freedom or using the wrong table value in chi-square and t-tests; applying chi-square to percentages instead of actual frequencies.
In nearest neighbour analysis, measuring the "nearest" distance along roads instead of straight line, or ignoring edge effects and the arbitrariness of the study-area boundary.
Unit and base errors: mixing cm with mm, hectares with km², or comparing Census 2011 figures with later survey estimates without noting the different definitions and years.