Start Here
Welcome. This is a plain-language companion to the thesis, built so that anyone — with or without a background in economics — can follow what the study asked, how it was done, and what it found. Read the one-sentence summary, then pick a learning path below.
The whole study in one sentence
The three questions it answers
- How do land fragmentation and irrigation vary across Nepal's districts?
- How does production sufficiency vary?
- What is the relationship between them — and does irrigation soften the fragmentation disadvantage?
The three conclusions it reaches
What the study actually claims
Two careful contributions: (1) the first nationwide, district-level mapping of these three conditions from the newest census; and (2) a robust irrigation association that survives every specification, control, and diagnostic the data allow. The fragmentation×irrigation interaction is an honest, power-limited null — reported as such, not dressed up as a finding.
Guided learning paths
Not sure where to begin? Choose the route that fits you. Each step links straight to a lesson, and the site remembers which lessons you have opened.
The big picture, without the technical detail.
The complete story, in the order it is meant to be read. Best for a first proper study.
For readers who want to see exactly how the analysis was built and stress-tested.
Look up a figure, or find a plain answer to a common question.
The Problem
Nepal's paradox: farming is the main livelihood for most rural households, yet most farms cannot feed their own household from what they grow.
Why this is worth a whole thesis
Agriculture is about a quarter of GDP and the main income for roughly 70% of the country's 4 million-plus holdings. But only about 45% of holdings say their own production is enough to eat — leaving roughly 55% reporting insufficiency. So the structural condition of farming is not just an output question, it is a welfare question.
Three tangled conditions
- Land fragmentation — one farm split into many scattered plots (national average ~2.8 parcels per holding).
- Irrigation access — water beyond rainfall (national ~54.5% of holding area, but from 3% in Rukum East to 91% in Bara).
- Production sufficiency — whether the farm feeds its household (32% insufficient in Rupandehi up to 87% in Dolpa).
Three things that are easy to confuse
Why study it at the district level?
The census only reports these variables together at the district level, and under federalism it is provincial and local governments who now run irrigation and extension — so district evidence is directly usable for targeting. The three objectives follow the three questions: (1) map the patterns of fragmentation and irrigation; (2) map the patterns of production sufficiency; and (3) analyse the relationship between them, including the moderating role of irrigation.
Key terms — tap to flip
Key figures from this chapter
- 77districts of Nepal, the complete unit-of-analysis population; also 7 provinces and 3 ecological belts
- ~25% (about one-fourth)agriculture's share of GDP (MoF 2024)
- ~70% (more than two-thirds)holdings reporting agriculture as their main income source
- >4 milliontotal agricultural holdings recorded in the NSCA 2021/22
- 0.54 hanational average holding size
- 2.8national average parcels per holding
- ~45% / ~55%share of holdings reporting production sufficient vs insufficient for household consumption
- 54.5%national irrigated share of total holding area
- 2.9% (Rukum East) to 91.1% (Bara)district range of irrigated-area share — a factor of more than thirty
- 1.5 (Lalitpur) to 6.3 (Humla)district range of average parcels per holding
- ~32% (Rupandehi) to ~87% (Dolpa)district range of the insufficiency share — a factor of almost three
- Karnali and Gandakiprovinces with the highest insufficiency averages
- 2021/22census reference year, partly overlapping COVID-19 disruptions (a stated limitation)
Literature & Theory
The theory in one idea, then four studies worth knowing, then the gap this study fills.
The theory: why "grow enough to eat" is a real welfare question
The agricultural household model (Singh, Squire & Strauss 1986) treats the farm family as producer, consumer, and worker at once. Its crucial idea is non-separability: if markets for food, labour and credit worked perfectly, a household would just farm for profit and buy its food — whether it grew "enough" would be irrelevant. But in rural Nepal those markets are thin or missing, so own production directly determines what the family eats. That is exactly what makes the census sufficiency question economically meaningful.
The two structural conditions in theory
- Fragmentation mostly hurts — travel between plots, supervising labour, boundary losses, and it blocks lumpy machines. Its one upside is spreading weather and pest risk across micro-environments. The net effect is therefore theoretically ambiguous, though expected to be negative for labour-scarce Nepal.
- Irrigation is a risk-reducing, productivity-raising input — it loosens the farm's dependence on the monsoon.
- The interaction sign is left open on purpose: irrigation could substitute for the coordination scattered plots frustrate (negative), or only pay off on consolidated, mechanisable plots (positive).
Four studies worth knowing
- Hao 2023 (China): 1% more fragmentation → 20–22% lower labour productivity — the penalty runs through blocked machinery, relevant to out-migration Nepal.
- Ali 2019 (Rwanda): consolidation's yield gain only ~7% in an unmechanised setting — foreshadows this study's fragmentation null.
- Kafle 2022 (Niger): irrigation → ~10% less food insecurity — the benchmark this study compares its magnitude against.
- Joshi & Joshi 2017 vs Karki 2015 (Nepal): the structure-vs-management debate this study weighs in on — coming down, qualifiedly, on the management and water side.
Key terms — tap to flip
Key figures from this chapter
- 20–22%fall in labour productivity per 1% more fragmentation, Chinese grain farmers, IV estimates (Hao 2023) — the key international fragmentation penalty
- +4.8%rise in probability of the severest food insecurity per 1% rise in a fragmentation index, Vietnam (Phan 2022)
- ~10% / +9%reduction in food insecurity and increase in food-consumption spending from irrigation access in Niger (Kafle & Balasubramanya 2022) — the benchmark the thesis compares its magnitude to
- ~55%gains in both income and food expenditure for small-scale irrigators in Ethiopia (Ebrahim & Toy 2024)
- ~7%potential yield gain from consolidation in unmechanised Rwanda (Ali 2019) — the caution that consolidation alone pays little without machinery
- ~USD 1,200/hanet income gain for irrigating households in Burkina Faso (Nyamba & Zidouemba 2025)
- ~65%share of Nepalese farmers preferring consolidation policy, especially bundled with roads and irrigation (Pun 2024)
- ~76%share of potential output achieved by Nepalese paddy farmers, with fragmentation a key inefficiency source (Choudhary 2022)
- ~81%mean technical efficiency of Nepalese wheat farmers; irrigation raises it (Adhikari 2021)
- ~24 ppincrease in probability of joining collective irrigation management when a monetary fine exists (Khanal 2021)
- +17 ppcrop-productivity gain from improved organic-manure practices (Dhakal & Escalante 2022)
- ~40% households fallowing land, ~28% plots uncultivated, food for only ~7 monthsKaski District (Khanal 2018) — the closest Nepali precedent linking plots, irrigation, and sufficiency
Data & Design
One official data source, rebuilt from scratch by scripts, covering every district.
The design in one line
Quantitative, cross-sectional, all 77 districts: first describe the patterns, then estimate multivariate associations. All results are read as district-level associations, never household causal effects.
The data
Everything comes from the NSCA 2021/22 district tables. Each census table was reduced to a district series, merged on harmonised district names, and every share was computed from the underlying counts and areas (never transcribed), with zero missing values — all through documented Python scripts, so any number can be regenerated.
(1) "It's a census, so there is no sampling error, right?" Not quite — the NSCA is a sample census: district figures are estimated from a large sample, so every share is itself an estimate.
(2) "If all 77 districts are covered, why report standard errors at all?" Because of super-population inference: the districts are treated as draws from an underlying data-generating process whose parameters we want to learn; the standard errors describe uncertainty about that.
Why secondary data is a strength here, not an apology
A uniform national method makes districts comparable in a way no single-researcher survey could; full coverage removes site-selection bias; and public data makes every step verifiable. The cost — a fixed set of variables at district resolution — is acknowledged and shapes interpretation.
The variables (Table 3.1) and the hypotheses
Outcome: insufficiency share. Stars: fragmentation (parcels/holding) and irrigation (irrigated-area share), plus their interaction. Six controls: holding size, loan, subsidy, mechanisation, market access, climate impact. Hypotheses: H1 irrigation lowers insufficiency (β₂<0); H2 fragmentation raises it (β₁>0); H3 irrigation moderates fragmentation (β₃≠0, sign open).
Key terms — tap to flip
The Methods Toolbox
The guiding philosophy: believe a result only if it survives every reasonable way of looking at the data. That is why there are so many tools.
Each card below answers the four things worth knowing about any method: what it is, why it was used, what it outputs, and what that output was in this study. Click a card to open it.
Method cards — why it’s done & what it outputs
Weighted OLS (the benchmark)
What it is. An ordinary straight-line regression, with each district counted in proportion to its number of holdings.
Why we use it. It is the transparent baseline: its coefficients read directly in percentage points, so anyone can interpret them without extra translation.
What it outputs. A change in the outcome share per one-unit change in a predictor. Multiply a change in the share by 100 to express it in percentage points.
In this thesis. Irrigation = −0.468 (p<0.001): a 10-point higher irrigated share goes with about 4.7 points lower insufficiency. It explains 83% of the weighted variation (R² = 0.832).
Fractional logit (the preferred model)
What it is. A model built on an S-shaped (logistic) curve, made for an outcome that is a proportion trapped between 0 and 1.
Why we use it. The outcome is a share. A straight line could predict impossible values like −5% or 130%; the S-curve can never leave the 0–1 range.
What it outputs. Index coefficients (not percentages) that are then translated into readable average partial effects.
In this thesis. Irrigation = −1.993 (p<0.001) — the headline estimate of the thesis. Fragmentation and the interaction are not significant.
Quasi-maximum likelihood (QML)
What it is. Fitting the fractional logit as if the share were a yes/no outcome (a Bernoulli likelihood), even though it is a fraction.
Why we use it. Papke & Wooldridge (1996) proved the estimates stay correct (consistent) as long as the average relationship — the conditional mean — is specified correctly, even with the 'wrong' likelihood.
What it outputs. Consistent coefficients paired with robust (HC1 sandwich) standard errors.
In this thesis. Valid inference for a share outcome without pretending it is binary — the standard modern approach for proportions.
Mean-centring the interaction
What it is. Subtracting each variable's (holdings-weighted) average before multiplying fragmentation by irrigation.
Why we use it. The interaction is mechanically correlated with its own parts, which tangles the model. Centring reduces that mechanical collinearity and makes the main effects read 'at the average farm' instead of at an impossible zero-parcel farm.
What it outputs. Identical model fit and predictions, but interpretable, low-collinearity main-effect coefficients.
In this thesis. VIFs collapse from 17.7 / 13.9 to 2.5 / 2.9, and irrigation sharpens to −1.993 (p<0.001). This is the standard fix for interaction collinearity.
Fixed effects (province / belt)
What it is. A dummy variable for each province (main models) or each ecological belt (robustness), soaking up everything districts in the same group share.
Why we use it. Irrigation and fragmentation line up with geography. Fixed effects force the comparison to happen within a province or belt, so a raw Mountain-vs-Terai gap cannot masquerade as an irrigation effect.
What it outputs. Coefficients identified only from differences among districts inside the same group.
In this thesis. Province FE absorb shared provincial policy/extension/weather; belt FE absorb the agro-ecological gradient itself — a stricter test the result still survives.
VIF (multicollinearity check)
What it is. A variance-inflation-factor score, 1/(1−R²), from regressing one variable on all the others.
Why we use it. To measure whether variables move together so tightly the model cannot separate their individual roles.
What it outputs. One number per variable; above 10 is the conventional danger level.
In this thesis. After centring every VIF is below 10 (2.5 irrigation, 2.9 interaction, 5.4 fragmentation) — none exceeds the thesis’s conventional threshold of 10.
Parsimonious spec + stability rule
What it is. A stripped-down model (fragmentation, irrigation, interaction, holding size, province FE only), fitted alongside the full one.
Why we use it. The full model spends 16 parameters on 77 districts — fewer than five observations per parameter — so no single p-value is decisive.
What it outputs. A second set of estimates to compare against the full model.
In this thesis. Signs and magnitudes are stable across specifications, which is what the conclusions rest on rather than any one p-value.
Pairs (district) bootstrap
What it is. Redraw 77 districts at random with replacement, refit the model, and repeat 999 times.
Why we use it. To get standard errors and intervals that do not lean on the large-sample normal approximation, which is shaky at n=77.
What it outputs. Distribution-free standard errors and 95% intervals (the 2.5th–97.5th percentiles of the 999 estimates).
In this thesis. The irrigation coefficient's interval [−2.64, −1.30] excludes zero, confirming significance without the normal assumption.
Key terms — tap to flip
Key figures from this chapter
- n = 77districts, complete population; full specification estimates 16 parameters (fewer than 5 observations per parameter)
- 2.81 parcels and 0.525 irrigated-area sharethe holdings-weighted means at which fragmentation and irrigation are centred before forming the interaction
- 999bootstrap replications for the pairs (district) bootstrap; 95% CI from 2.5th and 97.5th percentiles
- 7provinces — too few clusters for reliable cluster-robust standard errors, hence robust SEs + province FE instead
- 16 Mountain / 40 Hill / 21 Teraithe belt classification of the 77 districts
- 10conventional VIF threshold; without centring the VIFs would be 17.7 (irrigation) and 13.9 (interaction)
- 6control variables — holding size, loan share, subsidy share, power-tiller (mechanisation) share, 30-minute market-access share, climate-impact share; all holdings-based shares on a comparable scale
- NSCA source tables35 (sufficiency outcome), 36.2 (severe duration), 3 (parcels and holding size), 4.1 (irrigated area), 17 (loans), 19 (subsidy), 14 (power tiller), 38.1 (market access), 22 (climate impact)
Descriptive Patterns
Before any regression: what the raw map of Nepal looks like.
The average holding
Holdings-weighted district means are 0.537 ha per holding, 2.81 parcels per holding, a 52.5% irrigated-area share, and a 55.0% production-insufficiency share (Table 4.1). The 52.5% irrigation mean weights district shares by holdings; the 54.5% national area aggregate in Chapter I weights them by area. For a majority of holdings, own production does not cover consumption.
The Mountain → Terai gradient
| Belt | Parcels | Irrigated | Insufficient | Severe |
|---|---|---|---|---|
| Mountain (16 districts) | 3.99 | 0.284 | 0.711 | 0.018 |
| Hill (40) | 2.84 | 0.300 | 0.662 | 0.042 |
| Terai (21) | 2.57 | 0.767 | 0.422 | 0.060 |
Province highlights
Madhesh is best (74% irrigated, 43% insufficient); Karnali worst (29% / 70%, most fragmented); Bagmati is the oddball — low irrigation yet middling insufficiency, because the Kathmandu valley buys food with off-farm income (this reappears as the only significant province effect later).
Correlations: irrigation stands out, but bivariate isn't enough
Irrigation's raw correlation with insufficiency is a striking −0.86; fragmentation's is +0.40. But the regressors overlap heavily (fragmentation and irrigation correlate −0.41), so no single correlation can be trusted alone — which is exactly why the multivariate models in Module 6 carry the conclusions.
Key terms — tap to flip
The Main Results
Table 4.7, read in plain language.
The headline
Irrigation carries the expected negative sign and is estimated precisely in both models: −0.468 (p<0.001) in weighted OLS and −1.993 (p<0.001) in the fractional logit. Fragmentation is positive but far from significant; the interaction is null.
A worked example to make it concrete
In a hypothetical district of 50,000 holdings at the average profile, a 10-point higher irrigated share maps to roughly 2,250 fewer holdings reporting they cannot feed themselves; moving across the full middle-half of irrigation coverage maps to about 10,000 fewer.
The ~20-point scenario gap — and an important caveat
Predicted insufficiency falls from ~0.68 to ~0.48 moving from low- to high-irrigation profiles — about a 20-percentage-point gap, with a bootstrap interval that excludes zero.
Two things that look odd but aren't
- Fragmentation's +0.40 raw correlation becomes ~0 in the regression. Both are true: fragmented districts are more insufficient, but they are overwhelmingly mountain/hill districts — once you hold location fixed, parcel count adds nothing.
- Better market access → more insufficiency (+0.42). This is a district-composition story: fast-market districts are urban, full of tiny part-time farms that buy food rather than grow it. It is the thesis's own worked example of why a district pattern is not a household effect.
Method cards — why it’s done & what it outputs
Average partial effect (APE)
What it is. The model's marginal effect, averaged over every observed district.
Why we use it. Fractional-logit coefficients (like −1.993) are not readable directly; the APE converts them into plain 'one point more X goes with this much change in the outcome'.
What it outputs. A percentage-point-per-percentage-point number.
In this thesis. Irrigation APE = −0.453: one point more irrigation goes with about 0.45 points less insufficiency. Fragmentation APE = +0.005 (essentially zero).
Scenario predictions
What it is. Predicted insufficiency at a low-irrigation profile (0.230) versus a high one (0.668) — the middle half of the data, so no extrapolation.
Why we use it. To express the size of the pattern in a way a policymaker can picture.
What it outputs. Predicted shares at each profile and the gap between them.
In this thesis. About a 20-percentage-point gap, bootstrap interval excluding zero. Crucially: a descriptive contrast across existing districts, NOT an estimate of what building irrigation would achieve.
Robustness & the Severe Twist
Six robustness lessons, each answering a natural challenge to the result — then the most interesting twist in the study.
The six lessons of the robustness section
- Irrigation survives everything. Negative and significant at 1% in all 14 insufficiency-outcome specifications.
- Centring beats the collinearity. The irrigation and interaction VIFs drop from 17.7/13.9 to 2.5/2.9; fragmentation’s VIF is 5.4. Centring changes the reference point for main effects, not model fit or predictions.
- The fragmentation proxy matters. The alternative parcels-per-hectare proxy gives a positive coefficient (p=0.073), significant at 10% but not at 5%. The thesis reads this as support for the measurement-error conjecture, without treating it as proof.
- Household size is not the driver. Adding average family size leaves irrigation essentially unchanged; family size itself is insignificant (a natural question, answered before it is asked).
- The interaction null is "no detectable moderation". At n=77 the design lacks power; the point predictions even lean the right way. Honest, not embarrassing.
- Weighting is innocuous. Weighted and unweighted give almost the same answer.
The diagnostics section (§4.5) reports the model checks. Moran's I does not reject no spatial autocorrelation (p = 0.14 and 0.17); this does not prove spatial dependence is absent. The Breusch–Pagan test detects unequal error spread, supporting the use of robust standard errors. The Jarque–Bera test does not reject normality. Excluding the three Kathmandu-valley districts also leaves the main irrigation association similar. The method cards below explain each test.
The severe-outcome twist (the best part)
Switch the outcome to severe 10–12-month food insufficiency and both key signs flip: irrigation turns positive (+0.967) and fragmentation negative (−0.450).
Method cards — why it’s done & what it outputs
Belt-FE attenuation check
What it is. Re-running the model with ecological-belt dummies replacing (and then supplementing) province effects.
Why we use it. To see how much of the irrigation gradient is really just Mountain–Hill–Terai geography.
What it outputs. The irrigation coefficient under the stricter geographic control.
In this thesis. It shrinks from −1.993 to −1.300 but stays negative and significant (p<0.01); −1.633 with both sets of effects. The sign and policy-relevant magnitude survive.
Household-size control (a natural robustness check)
What it is. Adding average family size per holding (from the census farm-population table) to the main model.
Why we use it. The outcome is judged against household needs, so a district of bigger families could look more insufficient for that reason alone.
What it outputs. The irrigation coefficient with the family-size channel removed.
In this thesis. Irrigation barely moves (−1.993 → −2.073) and family size itself is insignificant (+0.057, p=0.63). This check does not support household size as the explanation for the irrigation association.
Parcels-per-hectare proxy
What it is. A richer fragmentation measure combining parcel count with holding size, replacing the plain parcel count.
Why we use it. To test whether the fragmentation null is just measurement error in a crude proxy.
What it outputs. The fragmentation coefficient under a better proxy.
In this thesis. Its coefficient is positive (+0.050, p=0.073), significant at 10% but not at 5%, while irrigation stays negative (−2.013). This supports, but does not demonstrate, the thesis’s measurement-error conjecture.
Moran's I (spatial diagnostic)
What it is. A test of whether neighbouring districts' model errors cluster together, using a queen-contiguity map of the 77 districts.
Why we use it. Positive spatial clustering would make the reported standard errors too optimistic.
What it outputs. A statistic compared to its no-clustering expectation (−0.013), with a permutation p-value.
In this thesis. −0.104 (OLS) and −0.098 (logit), p = 0.14 and 0.17 — no statistically significant evidence of spatial autocorrelation under this diagnostic.
Breusch–Pagan test (heteroscedasticity)
What it is. A formal check of whether the errors have a constant spread, run in the studentized (Koenker) form that does not itself assume normal errors.
Why we use it. If the spread of the errors changes with the regressors, ordinary standard errors would be misleading, so it is worth testing directly.
What it outputs. A chi-square statistic and p-value; a small p-value signals unequal error variance (heteroscedasticity).
In this thesis. LM = 30.55 (p = 0.010): heteroscedasticity is present. That is exactly why the thesis reports heteroskedasticity-robust (HC1) standard errors everywhere — the test confirms the choice rather than undermining it.
Jarque–Bera test (normality)
What it is. A check of whether the residuals follow a normal, bell-shaped distribution, based on their skewness and kurtosis.
Why we use it. Roughly normal residuals reassure that the linear benchmark behaves sensibly at this small sample size.
What it outputs. A chi-square statistic and p-value; a large p-value means normality is not rejected.
In this thesis. JB = 1.42 (p = 0.491), skewness −0.21, kurtosis 3.51 — the residuals are close to normal. The fractional logit does not need this, but it is reassuring for the OLS benchmark.
Key terms — tap to flip
Key terms — tap to flip
Key figures from this chapter
- 0.537 ha / 2.81 parcels / 0.525 irrigated / 0.550 insufficientthe weighted-mean profile of the average Nepalese holding
- -0.856 (weighted)correlation of irrigated-area share with insufficiency — the headline bivariate fact (-0.751 for the irrigated-holdings share)
- +0.396weighted fragmentation-insufficiency correlation, which dissolves to ~0 once geography is controlled; other correlations: loans -0.643, climate impact +0.466, holding size -0.355, market access +0.327 (the counter-intuitive positive)
- -0.41weighted correlation between fragmentation and irrigation (structural disadvantages overlap); loans-irrigation +0.64; climate-irrigation -0.52; climate-fragmentation +0.41
- -0.468 (p<0.001)irrigation coefficient, weighted OLS with province FE — a 10 pp higher irrigated share goes with ~4.7 pp lower insufficiency
- -1.993 (p<0.001)irrigation coefficient, main fractional logit with province FE — THE headline estimate
- p=0.76 / p=0.70fragmentation p-values (OLS / fractional logit) — the informative null; interaction p=0.58 / p=0.44
- +0.419 (p=0.093)market-access coefficient in OLS, positive and only marginal; insignificant in the logit (p=0.11); market x size test: +3.48 (p=0.02) main term, -3.97 (p=0.19) interaction
- ~-0.13 (p=0.002)Bagmati province fixed effect (OLS), the only significant province effect relative to Koshi
- R² = 0.832 (weighted OLS); McFadden pseudo-R² = 0.058 (mechanically small for fractional outcomes)
- -0.453APE of irrigation — 1 pp more irrigation goes with ~0.45 pp less insufficiency; APE of fragmentation +0.005 (essentially zero)
- 0.230 and 0.668low and high irrigation scenario values (the IQR of the data); fragmentation scenarios 2.3 and 3.5
- 0.679 → 0.481 and 0.699 → 0.480predicted insufficiency moving low-to-high irrigation at low / high fragmentation; gaps 19.7 and 21.9 pp (the '~20-point' claim); bootstrap 95% CIs [-0.266, -0.131] and [-0.296, -0.128]
- [-2.640, -1.299]bootstrap 95% CI for the irrigation coefficient itself — excludes zero
- ~2 pp vs ~0.1 ppfragmentation gap at low vs high irrigation — descriptively consistent with moderation, statistically null
- 2,250 and 10,000fewer insufficient holdings implied in a hypothetical 50,000-holding district for a 10 pp irrigation rise and the full IQR move
- -1.300 (belt-only) / -1.633 (province+belt)attenuated irrigation coefficients under belt controls; all 14 insufficiency-outcome specifications keep p<0.01
- -2.100 vs -2.144parsimonious irrigation coefficient without vs with the interaction — nothing rests on the interaction term
- -2.110 (p<0.001, n=74)irrigation coefficient after excluding the three Kathmandu-valley districts
- -2.073 with household size added; household-size term +0.057 (p=0.63); weighted mean household size 4.71, range 3.1–6.5
- -2.013 with parcels-per-hectare proxy; fragmentation term +0.050 (p=0.073) — alternative proxy is significant at 10%, but not at 5%
- -0.455 / -1.953unweighted OLS / fractional logit irrigation coefficients (vs -0.468 / -1.993 weighted) — weighting is innocuous; -1.830: irrigated-holdings-share variant
- Moran's I = -0.104 (OLS residuals) and -0.098 (logit residuals) vs expectation -0.013; permutation p = 0.14 and 0.17 (9,999 permutations) — no positive spatial autocorrelation
- +0.967 (p=0.028) and -0.450 (p<0.001)irrigation and fragmentation coefficients for the SEVERE outcome — both signs flip versus the main outcome
- 0.163 (Kaski), 0.161 (Kathmandu)highest severe shares; Dolpa has severe share 0.001 despite the highest overall insufficiency 0.869
- Belt profileMountain 3.99 parcels / 0.284 irrigated / 0.711 insufficient / 8.6% of holdings; Terai 2.57 / 0.767 / 0.422 / 48.3% of holdings; Hill irrigation only 0.300; severe share 0.018 Mountain vs 0.060 Terai (reversed gradient); Terai insufficiency 24 pp below Hill
- Province profileMadhesh 0.743 irrigated / 0.430 insufficient (best); Karnali 0.289 / 0.695 (worst), fragmentation 3.35; Gandaki insufficiency 0.650 (second worst); Koshi least fragmented at 2.20; Bagmati the outlier: 0.393 irrigated yet 0.559 insufficient
Conclusions & Policy
What it all means, and the discipline that keeps the policy talk honest.
The three conclusions (again, because they matter)
How the study keeps its policy talk honest
Recommendations, keyed to the objectives
- Targeting: the Mountain belt — under 9% of holdings but the highest insufficiency — is the natural priority; provinces can read their own districts straight from Annex I.
- Irrigation, realistically: uplands can't copy Terai canals, so the instruments are rehabilitating farmer-managed systems and small solar/lift schemes, targeted so they don't bypass marginal farmers.
- Consolidation: only ever bundled with irrigation and mechanisation — supported by the null, by Ali (2019), and by farmers' own stated preferences (Pun 2024).
- Chronic deprivation: production indicators point to the mountains, but severe deprivation lives among near-landless Terai/peri-urban households — so targeting needs landlessness and market-dependence indicators too.
The honest limits
Cross-sectional (association, not causation); district unit (ecological inference); self-reported outcome; count-based fragmentation proxy; missing covariates (remittances the key suspect); unpropagated sampling error; a COVID-overlapping reference year. None reverses a conclusion — each only bounds how far it travels.
Key terms — tap to flip
Key figures from this chapter
- -0.86the weighted irrigation-insufficiency correlation quoted in the summary
- ~20 percentage pointspredicted insufficiency gap between low- and high-irrigation district profiles, bootstrap intervals excluding zero
- 55%holdings whose own production does not cover consumption — insufficiency is the national norm
- 0.54 ha in 2.8 parcelsthe average holding recapped
- 1 percent levelsignificance of the irrigation association in every specification of the insufficiency outcome
- <9% of the country's holdingsthe Mountain belt's share, yet the highest insufficiency — the targeting recommendation
- n=77the sample size at which no fragmentation-irrigation moderation is detectable
Key Numbers at a Glance
The figures that anchor the study, gathered in one place. Flip a card to reveal what each number means; shuffle to see them in random order.
Questions & Answers
Common questions about the study, grouped by topic — each with a short, plain answer and the point readers most often misread. Browse by topic, or shuffle the cards to see them in a different order.
Motivation and Statement of the Problem 5
Easy to miss: Answering only 'both are important for agriculture' without the entanglement argument.
Answer: The most fragmented, least irrigated, and most insufficient districts overlap heavily, so their separate associations can only be distinguished in a multivariate framework that holds location constant; no nationwide study examines them jointly, including their interaction, with the NSCA 2021/22.
Easy to miss: Assuming the study's aims are vague and generic rather than three objectives matched one-to-one to three research questions.
Answer: Three research questions, matched one-to-one by three objectives. Objective 1: document the district-level patterns of land fragmentation and irrigation; Objective 2: document the district-level patterns of production sufficiency; Objective 3: analyse their multivariate district-level relationship, including whether irrigation moderates the fragmentation-insufficiency association.
Easy to miss: Saying 'more data is always better' rather than pointing to the scale and entanglement of the variation.
Answer: Irrigation coverage differs across districts by a factor of more than thirty and insufficiency by almost three; problems of this spatial magnitude cannot be characterised from single-district studies, and the entangled geography requires holding location constant.
Easy to miss: Calling it a poverty or food-security measure.
Answer: Both, carefully: it measures production relative to household needs and is a practical indicator of agricultural strain, but explicitly not poverty or food security as such — a distinction the thesis maintains throughout.
Easy to miss: Assuming the holdings shares and area shares are interchangeable figures.
Answer: Over four million holdings averaging about 0.54 ha in roughly 2.8 parcels; about 54.5 percent of holding area irrigated nationally; roughly 55 percent of holdings report own production insufficient, ranging from 32 percent (Rupandehi) to 87 percent (Dolpa).
Key Terms and Definitions 6
Easy to miss: Confusing fragmentation with small holding size — size is a separate control variable.
Answer: The division of a single holding into spatially separated parcels, measured as average parcels per holding (NSCA Table 3), ranging from 1.5 in Lalitpur to 6.3 in Humla; average holding size enters separately as a control.
Easy to miss: Mentioning only one measure.
Answer: Availability of water beyond rainfall; the main measure is the irrigated share of total holding area (Table 4.1), and the share of holdings reporting any irrigation is used as an alternative measure in robustness checks.
Easy to miss: Treating the outcome as a food-security measure.
Answer: Sufficiency is the holder's self-report (Table 35) of whether own production covers household consumption; a holding can be production-insufficient yet food-secure through purchases, which is exactly why the thesis reserves 'food insufficiency' for the duration-based census measure.
Easy to miss: Treating it as merely a more extreme version of the main outcome.
Answer: The share of holdings reporting food insufficiency for 10-12 months (Table 36.2); it is duration-based, comes from a different census table, and has a nearly opposite geography, so it is treated as a separate dimension of deprivation.
Easy to miss: Implying the choice changes anything substantively.
Answer: The insufficiency share is the exact complement, and signs are interpreted accordingly; re-estimating with the sufficiency share as outcome reverses all key coefficients exactly, which is reported as a robustness check.
Easy to miss: Treating them as interchangeable geographic dummies.
Answer: Belts (Mountain/Hill/Terai) summarise agro-ecological geography — terrain, climate, cropping potential — while the seven provinces share policies, extension systems, and budgets; they absorb different confounds, so the thesis estimates specifications with each and with both.
Literature Review 6
Easy to miss: Assuming the literature is just a catalogue of studies rather than a source of design implications.
Answer: Higher irrigation coverage should mean lower insufficiency with context-dependent magnitudes; fragmentation gains come largely through complements like mechanisation and irrigation; food security reflects structural endowments plus assets and market access — motivating the interaction term and the control set.
Easy to miss: Citing efficiency studies (e.g. Manjunatha) as the direct precedent.
Answer: Zheng (2023), Phan et al. (2022), and Tran and Vu (2021), because they link fragmentation to food-insecurity or abandonment outcomes rather than to technical efficiency; no reviewed study estimates the fragmentation-irrigation interaction directly.
Easy to miss: Citing it as evidence that fragmentation does not matter at all.
Answer: With a terrain-adjusted travel-cost measure, they confirm fragmentation penalties but estimate the consolidation yield gain at only about 7 percent and warn that in unmechanised settings consolidation may not pay — anticipating the thesis's informative null for unmechanised hill agriculture.
Easy to miss: Claiming the thesis settles the debate definitively.
Answer: Structure (Joshi and Joshi 2017: landholding size matters) versus management (Karki 2015; Kumar 2020; Morioka and Kondo 2017); the district evidence sides qualifiedly with the management view — what distinguishes high-insufficiency districts is water access and location, not parcel structure per se.
Easy to miss: A generic 'no one has studied this in Nepal'.
Answer: No existing study provides nationwide, district-level evidence examining fragmentation and irrigation together, including their interaction, linked to production sufficiency; the nearest precedents (Khanal 2018, Joshi and Joshi 2017, Choudhary et al. 2022) each fall short on coverage, variables, or outcome.
Easy to miss: Conceding a contradiction or dismissing the Chinese evidence.
Answer: The contrast is informative rather than contradictory: their estimate is household/plot-level with an IV design and a labour-productivity outcome; the thesis's district parcel-count proxy is coarse, ecological inference cautions against equating levels, and the penalty shrinks in unmechanised settings like the Nepali hills.
Theoretical Framework 6
Easy to miss: Naming only a generic production function.
Answer: The agricultural household model (Singh et al. 1986; Taylor and Adelman 2003), in which the household is simultaneously producer, consumer, and labour supplier; the census sufficiency question is meaningful precisely under market non-separability.
Easy to miss: Assuming own-production sufficiency always matters for welfare.
Answer: With perfect markets, production and consumption decisions separate and whether own production covers consumption is economically irrelevant; with thin or missing food, labour, credit, and insurance markets, as in rural Nepal, own production directly determines food consumption.
Easy to miss: Presenting fragmentation as unambiguously harmful.
Answer: Costs: travel time, labour supervision, boundary losses, and obstruction of lumpy investments like machinery; benefit: risk diversification across micro-environments; the net effect is theoretically ambiguous, though the expectation for labour-constrained Nepal is negative.
Easy to miss: Asserting it must obviously be negative (irrigation always helps).
Answer: On the water-management margin irrigation substitutes for the coordination that scattered plots frustrate (negative interaction); on the labour-supervision margin irrigation only pays on consolidated mechanisable plots (positive); the estimate is read as a test of which margin dominates.
Easy to miss: Stating them as causal claims.
Answer: H1: beta2 < 0 (irrigation lowers insufficiency); H2: beta1 > 0 (fragmentation raises it); H3: beta3 not equal to 0 with sign open; all are hypotheses about conditional district-level associations, not household-level causal effects.
Easy to miss: Treating fixed effects as a mechanical add-on rather than theoretically motivated.
Answer: Agro-ecological geography shapes both the structural conditions and the outcome (the dashed arrows), which is exactly why province and belt fixed effects are needed to absorb this common influence before any structural coefficient is interpreted.
Data and Variable Construction 7
Easy to miss: Vague 'downloaded from NSO' without the construction steps.
Answer: NSCA 2021/22 district tables from the NSO open-data portal; each table reduced to a district series, merged on district names harmonised to the 77-district structure, every share computed from underlying counts and areas rather than transcribed, verified complete with no missing values, all via documented Python scripts.
Easy to miss: Assuming a census means complete enumeration with no sampling error.
Answer: District figures are estimated from a large probability sample of holdings, so every district share in the study is itself an estimate subject to sampling error — acknowledged as a limitation but not propagated through the models.
Easy to miss: Assuming the standard errors are meaningless because every district is observed.
Answer: Under model-based, super-population inference the districts are treated as draws from a data-generating process whose parameters are the objects of interest; the sampling variability embedded in the census estimates is a further reason not to treat the shares as exact constants.
Easy to miss: Treating it as an inconsistency or error.
Answer: Neither: 52.5 is the holdings-weighted mean of the district irrigated-area shares, while 54.5 is the census's national area aggregate; the thesis flags the distinction explicitly.
Easy to miss: Assuming the controls are an arbitrary list rather than each tied to a mechanism or the literature.
Answer: Holding size captures land endowment; loan and subsidy shares proxy credit and public support; the power-tiller share proxies mechanisation, the margin where fragmentation bites hardest; the 30-minute market-access share captures market integration; the climate-impact share captures the risk irrigation buffers; all are holdings-based shares on a comparable scale.
Easy to miss: Apologising for secondary data as second-best.
Answer: A deliberate design choice: uniform national methodology makes district figures comparable in a way no single-researcher survey could match, full 77-district coverage eliminates site-selection bias, and public availability makes every step verifiable; the costs (fixed variables, district resolution) are acknowledged and shape interpretation.
Easy to miss: Denying any relevance, or conceding the results are unusable.
Answer: It is an acknowledged limitation: a single cross-section cannot separate period effects, and structural conditions may have evolved; per the conclusions chapter it bounds how far the conclusions travel but does not reverse any of them.
Methods and Estimator Choice 10
Easy to miss: Saying 'because the outcome is binary' or that OLS is invalid.
Answer: The outcome is a proportion bounded in [0,1]; the Papke-Wooldridge (1996) fractional logit respects the bounded support and yields fitted values inside the unit interval by construction, while weighted OLS is retained as a transparent benchmark interpretable in percentage points.
Easy to miss: Claiming the likelihood must be correctly specified.
Answer: It is quasi-maximum likelihood: Papke and Wooldridge show the estimator is consistent whenever the conditional mean E[y|x] = Lambda(x'beta) is correctly specified, even though y is a share, and inference uses a robust sandwich covariance built from the weighted scores.
Easy to miss: Saying weighting corrects heteroskedasticity or improves efficiency.
Answer: Weights are total holdings normalised to mean one, so the estimand is the association experienced by the average agricultural holding rather than the average district; inference remains district-level.
Easy to miss: Assuming weighting must change the conclusion.
Answer: Unweighted estimates, reported as a robustness check following Solon et al. (2015), are close to the weighted ones: -0.455 vs -0.468 in OLS and -1.953 vs -1.993 in the fractional logit, all p<0.001.
Easy to miss: Claiming centring 'fixes' multicollinearity substantively or changes the model's fit.
Answer: Centring at the holdings-weighted means (2.81 parcels, 0.525 irrigated share) leaves fit, predictions, and the interaction coefficient unchanged; what it changes is that beta1 and beta2 are evaluated at the average holding's profile instead of the out-of-sample zero point, and it removes mechanical collinearity (VIFs fall from 17.7 and 13.9 to 2.5 and 2.9).
Easy to miss: Assuming standard errors should always be clustered at the geographic-aggregation level.
Answer: With only seven provinces, cluster-robust asymptotics are unreliable (far too few clusters); HC-robust SEs are retained, the province fixed effects absorb the province-common disturbance component directly, and a wild cluster bootstrap (Cameron et al. 2008) is explicitly noted as future work.
Easy to miss: Claiming province FE automatically solve it without a diagnostic.
Answer: A direct diagnostic: Moran's I on the residuals using a queen-contiguity matrix with permutation inference gives -0.104 (OLS) and -0.098 (flogit) against an expectation of -0.013, p = 0.14 and 0.17 — no positive autocorrelation, the pattern that would make robust SEs optimistic; a full spatial model is left for future work.
Easy to miss: Describing a residual bootstrap or forgetting what was bootstrapped.
Answer: A pairs (district) bootstrap: draw 77 districts with replacement keeping outcome, regressors, and weight; re-estimate the fractional logit; record the irrigation coefficient and scenario gaps; 999 replications; report the SD as the SE and the 2.5th/97.5th percentiles as the CI — distribution-free confirmation that inference does not rest on the normal approximation.
Easy to miss: Denying the problem.
Answer: Acknowledged directly — fewer than five observations per parameter — which is why a parsimonious specification (fragmentation, irrigation, interaction, holding size, province FE) is estimated alongside, and conclusions rest on stability of signs and magnitudes across specifications rather than any single p-value.
Easy to miss: Assuming multicollinearity must be inflating the irrigation result.
Answer: Variance inflation factors: after centring, every VIF in the main specification is below the conventional threshold of 10 — 2.5 for irrigation, 2.9 for the interaction, 5.4 for fragmentation — whereas without centring irrigation and the interaction reach 17.7 and 13.9.
Main Results 7
Easy to miss: Assuming the fractional-logit coefficient is itself a marginal effect.
Answer: Irrigation is the dominant correlate: -0.468 in weighted OLS and -1.993 in the fractional logit, both p<0.001; taking the OLS at face value, 10 percentage points more irrigation is associated with about 4.7 points lower insufficiency, holding controls and province effects fixed.
Easy to miss: Interpreting the logit index coefficient directly.
Answer: The APE is -0.453: averaged over the observed district profiles, a one-percentage-point higher irrigated share is associated with insufficiency about 0.45 points lower; the fragmentation APE is essentially zero (+0.005).
Easy to miss: Reading it as the predicted effect of an irrigation intervention.
Answer: It compares predicted insufficiency at the observed IQR endpoints of irrigation (0.230 vs 0.668): 0.679 to 0.481 at low fragmentation, 0.699 to 0.480 at high, gaps of 19.7 and 21.9 points with bootstrap CIs excluding zero; the thesis states it is a descriptive contrast across observed profiles, not an intervention effect, and never uses it as one.
Easy to miss: Reading the district coefficient as a household-level behavioural effect.
Answer: The contradiction is apparent, not real: at district level fast market access proxies urban and peri-urban composition, where small part-time holdings buy rather than grow food; a market-access-by-holding-size interaction test supports this (market term +3.48, p=0.02; interaction -3.97), and the coefficient is not robust across estimators — the thesis's own illustration of why district associations are not household effects.
Easy to miss: Conceding poor fit or comparing pseudo-R2 to OLS R2.
Answer: No: pseudo-R2 is mechanically small for fractional outcomes because the quasi-likelihood is bounded away from zero; the relevant evidence is that fitted values track actual district shares closely over their full range (Annex III), and the weighted OLS explains 83 percent of the weighted variation.
Easy to miss: Dismissing them as nuisance parameters.
Answer: Yes: relative to Koshi only Bagmati's effect is significant (about -0.13 in OLS, p=0.002), indicating lower insufficiency than its structure would predict — plausibly the Kathmandu valley's market production and off-farm income, echoing the province-profile anomaly.
Easy to miss: Choosing one as 'the' answer instead of explaining the conditioning.
Answer: Both, at different conditioning: fragmented districts do report more insufficiency, but they are overwhelmingly mountain and hill districts; once location is held fixed the independent association is statistically undetectable — the bivariate disadvantage is absorbed by geography.
Robustness and Specification Checks 8
Easy to miss: Naming only one or two checks.
Answer: Belt FE replacing and supplementing province FE, alternative irrigation measure, sufficiency and severe outcomes, parsimonious and no-interaction variants, unweighted estimation, Kathmandu-valley exclusion, household-size control, parcels-per-hectare proxy, VIF monitoring, Moran's I diagnostic, and pairs-bootstrap inference.
Easy to miss: Denying the attenuation or conceding the result is fragile.
Answer: The attenuation is expected and interpreted: part of the raw gradient reflects agro-ecology, since surface irrigation concentrates in the Terai for terrain and canal-history reasons; the coefficient stays negative and significant (p<0.01) within belts and lies at -1.633 with both sets of effects, so the sign and the policy-relevant magnitude survive, read as conditional on geographic controls.
Easy to miss: Treating the two as redundant.
Answer: Province effects hold constant shared provincial policies, extension, and budgets while letting the within-province Mountain-to-Terai gradient identify the coefficients; belt effects absorb the agro-ecological gradient itself and ask only whether irrigation distinguishes districts within the same belt.
Easy to miss: Claiming the null proves there is no moderation.
Answer: The thesis says so explicitly: with 77 observations the design lacks power for moderate interactions, the term also flips sign in the irrigated-holdings and unweighted variants, so the conclusion is 'no detectable moderation' rather than 'no moderation' — even though the point predictions (a 2-point fragmentation gap at low irrigation shrinking to 0.1 at high) are descriptively consistent with moderation.
Easy to miss: Reading it as proof that fragmentation is irrelevant.
Answer: It cautions against expecting large sufficiency gains from consolidation alone in unmechanised hill agriculture, echoing Ali et al. (2019); but the count-based proxy ignores dispersion and inter-plot distance, so the null is read as a lower bound on the true relationship, not proof of irrelevance.
Easy to miss: Assuming the richer proxy would prove fragmentation matters.
Answer: Replacing parcels per holding with parcels per hectare leaves irrigation intact (-2.013, p<0.001) while fragmentation turns positive at the 10 percent level (+0.050, p=0.073) — moving exactly in the direction the measurement-error argument predicts, so the conjecture is supported but not demonstrated.
Easy to miss: Ignoring that the concern was directly testable.
Answer: Average household size (computed from the census farm-population table; weighted mean 4.71, range 3.1-6.5) was added to the main fractional logit: the irrigation coefficient moves only from -1.993 to -2.073 and household size itself is small and insignificant (+0.057, p=0.63), so the composition effect is not what the irrigation association picks up.
Easy to miss: Assuming a few urban districts could be driving the result.
Answer: No: excluding the three Kathmandu-valley districts gives -2.110 (p<0.001, n=74), and dropping the interaction altogether leaves the irrigation coefficient essentially unchanged (-2.100 vs -2.144), so nothing rests on the interaction term or on a few urban observations.
The Severe Food-Insufficiency Outcome 4
Easy to miss: Assuming the severe-outcome signs match those of the main outcome.
Answer: Both key signs flip: irrigation becomes positive and significant (+0.967, p=0.028) and fragmentation negative and significant (-0.450, p<0.001) — the opposite of the main outcome in both cases.
Easy to miss: Treating the severe model as a failed confirmation check.
Answer: No — the two constructs have almost disjoint geographies: Dolpa has 0.869 insufficiency but a severe share of only 0.001, while the severe share peaks in Kaski (0.163) and Kathmandu (0.161) among near-landless urbanised holdings; the thesis states that using severe insufficiency as a confirmation check would be methodologically wrong in both directions.
Easy to miss: Conflating own-production shortfall with hunger.
Answer: Through purchases funded by non-agricultural income, echoing Gartaula et al. (2017): Table 35 measures whether own production covers needs, while the extreme duration category of Table 36.2 concentrates among near-landless, purchase-dependent holdings in urbanised districts.
Easy to miss: Expecting the mountains to top both measures.
Answer: The severe share rises from 0.018 in the Mountain belt to 0.060 in the Terai — the mirror image of overall insufficiency — because the irrigated belts contain the urbanising districts where near-landless, purchase-dependent holdings concentrate.
Policy Implications and Conclusions 5
Easy to miss: Presenting the irrigation conclusion as causal.
Answer: Water: irrigation is the clearest structural correlate of sufficiency, conditional on geography, not causal; land structure: an informative null — no support for consolidation-alone gains in unmechanised hills; measurement: overall and severe insufficiency capture almost disjoint populations, so conflating them misdirects targeting.
Easy to miss: Either overclaiming a causal payoff or retreating into 'no policy relevance'.
Answer: The recommendations are phrased conditionally — 'the pattern is consistent with prioritising' irrigation in low-coverage hill and mountain districts; the finding motivates targeting, identifying where deprivation and low coverage coincide, not an impact estimate of an intervention.
Easy to miss: Recommending consolidation as a stand-alone remedy, or ruling it out entirely.
Answer: Consolidation should be pursued, if at all, as a complement bundled with irrigation and mechanisation — supported by the fragmentation null, Ali et al. (2019), farmers' own preference for bundled consolidation (Pun et al. 2024, about 65 percent), and the sequencing implied by the National Land Policy 2019 and Land Use Act 2019.
Easy to miss: Directing them to the mountain districts where production insufficiency peaks.
Answer: Production indicators alone would point to the mountains; the severe-insufficiency evidence shows chronic deprivation lives among near-landless holdings in peri-urban and Terai districts (Kaski, Kathmandu, eastern Terai), so targeting must add indicators of landlessness and market dependence.
Easy to miss: Recommending Terai-style canal systems for the hills.
Answer: Extension and most local irrigation functions now rest with provincial and local governments, which can read their districts' position from the Annex I dataset; since uplands cannot replicate large gravity-fed Terai canal networks, the realistic instruments are rehabilitating farmer-managed systems and small-scale lift or solar irrigation targeted so they do not bypass marginal holdings (Shrestha et al. 2023).
Limitations and Identification 8
Easy to miss: Answering only 'correlation is not causation' without the specific confounds.
Answer: Irrigation placement is not random: canal investment tracks the same Terai terrain and fertility that independently raise sufficiency, and public irrigation programmes may respond to observed deprivation (reverse causality); fixed effects absorb some but not all of this, so selection cannot be excluded in a single cross-section and the estimate is read as a conditional association.
Easy to miss: Claiming no better design exists.
Answer: Instrumenting irrigation with a determinant of feasibility that does not otherwise affect sufficiency — slope-based gravity-fed feasibility or proximity to canal command areas; constructing such an instrument requires geographic data beyond the census and is explicitly beyond the thesis's scope.
Easy to miss: Generalising district coefficients to household behaviour.
Answer: Associations across areal units need not equal household-level relationships and are sensitive to how the units are drawn (King 1997; Openshaw 1984); the thesis confines all claims to the district level, and the positive market-access coefficient is its own worked example of an aggregate compositional effect.
Easy to miss: Naming a variable without a bias direction.
Answer: Remittance income: it is plausibly correlated with fragmentation (out-migration leaves land underused and finances subdivision through inheritance) and with the outcome (remittances fund food purchases), so its omission tends to bias the fragmentation coefficient away from zero and load sufficiency variation onto the included structural variables.
Easy to miss: Either dismissing self-reports as worthless or defending them as objective.
Answer: It is a perception-based measure of production relative to needs that may differ from objective gaps; under non-separability it is a meaningful welfare indicator, the household-size composition concern was directly tested and found null, and the thesis interprets it strictly as an indicator of agricultural strain rather than food security.
Easy to miss: Claiming it invalidates the results, or ignoring it.
Answer: The district shares are census estimates rather than error-free observations; the thesis acknowledges this uncertainty without propagating it and accordingly reads the reported standard errors as indicative rather than exact.
Easy to miss: Naming only one or two; the study is careful to list several distinct limitations.
Answer: Cross-sectional single-round design (associations, not causal effects); district unit and ecological inference; self-reported perception-based outcome; count-based fragmentation proxy ignoring dispersion and distance; unavailable covariates (rainfall, soil, crop mix, remittances, roads); unpropagated sampling error; COVID-19 overlap in the reference year; focus confined to production sufficiency.
Easy to miss: Overclaiming robustness or conceding collapse.
Answer: The thesis's own formulation: none reverses a conclusion, but each bounds how far it travels — the irrigation finding motivates targeting rather than an impact estimate, and the fragmentation null cautions against consolidation-first policy rather than proving consolidation worthless.