MAGA Weight Calibration

Module 2: Economic Impact

Why Calibrate to MAGA?

Survey data and administrative records often produce different estimates of agricultural activity. ENCOVI 2023 captures detailed farmer-level information but uses a general population sampling frame that may under- or over-represent certain agricultural regions. MAGA (Ministerio de Agricultura, Ganadería y Alimentación) publishes official departmental statistics based on comprehensive agricultural monitoring.

By calibrating ENCOVI survey weights to match MAGA totals, we achieve:

  1. Administrative consistency — Weighted estimates align with official government statistics used in policy planning
  2. Improved representativeness — Correct for differential sampling rates across agricultural departments
  3. Credibility for stakeholders — Results match published figures that donors and policymakers reference

This calibration step is particularly important for economic impact projections, where departmental production volumes directly affect estimated benefits from biofortified maize adoption.

Calibration Approach

The calibration process adjusts individual survey weights so that weighted departmental totals for production (quintales) and cultivated land (manzanas) match MAGA 2023 official statistics.

Data Sources

Source Purpose
ENCOVI 2023 Farmer-level characteristics, original survey weights
MAGA 2023 Departmental targets for production and land area
INE 2011 Total agricultural household reference (1,299,377)

Calibration Variables

  • Production (total_production) — Total white maize harvested in quintales
  • Land (land_wtcorn) — Cultivated white maize area in manzanas

Both variables are calibrated simultaneously within each department using linear calibration, ensuring weighted totals match MAGA targets exactly.

Initial Weight Scaling

Before departmental calibration, we apply a scaling factor to align ENCOVI’s total agricultural household estimate with official INE census figures.

Table 1: Initial weight scaling to match INE agricultural census
Farmer Population Scaling Factor. Alignment of ENCOVI estimates with INE agricultural census
Farmer Population Scaling Factor
Alignment of ENCOVI estimates with INE agricultural census
Metric Value
INE reference (2011) 1,299,377
ENCOVI weighted total 664,971
Scaling factor 1.9540
Source: INE (2011) — Distribución de hogares agropecuarios según tipología
NoteScaling Factor

At 1.954, the factor places ENCOVI’s original weights at around half the agricultural population the INE census estimates. ENCOVI draws on a general population sampling frame, not an agriculture-specific one, and the factor closes that gap before the departmental calibration runs.

Pre-Calibration Diagnostic

Before applying calibration, we compare ENCOVI weighted estimates against MAGA departmental targets to quantify the magnitude of discrepancies.

Table 2: ENCOVI vs MAGA comparison before calibration
Pre-Calibration: ENCOVI vs MAGA Comparison. Weighted estimates before departmental calibration
Pre-Calibration: ENCOVI vs MAGA Comparison
Weighted estimates before departmental calibration
Department N (sample)
Production (qq)
Land (mz)
ENCOVI MAGA Ratio ENCOVI MAGA Ratio
Alta Verapaz 99 12,521,541 5,296,634 2.36 338,833 252,979 1.34
Baja Verapaz 47 7,294,761 720,542 10.12 124,530 40,233 3.10
Chimaltenango 84 9,976,272 1,065,264 9.37 240,782 34,835 6.91
Chiquimula 97 9,367,005 1,518,844 6.17 201,754 65,717 3.07
El Progreso 28 1,129,470 367,087 3.08 37,795 20,283 1.86
Escuintla 28 942,767 528,828 1.78 31,456 10,216 3.08
Guatemala 14 2,937,732 816,778 3.60 71,871 28,939 2.48
Huehuetenango 64 3,493,937 4,023,703 0.87 128,316 156,526 0.82
Izabal 85 6,386,107 1,254,753 5.09 149,271 30,116 4.96
Jalapa 54 3,851,500 1,095,143 3.52 77,082 40,300 1.91
Jutiapa 116 5,869,034 2,646,775 2.22 152,504 86,098 1.77
Petén 73 5,612,033 10,658,142 0.53 147,838 356,518 0.41
Quetzaltenango 33 1,287,887 2,580,710 0.50 60,794 62,296 0.98
Quiché 74 9,399,415 5,171,702 1.82 219,059 224,720 0.97
Retalhuleu 22 2,924,853 1,129,300 2.59 72,074 29,346 2.46
Sacatepéquez 22 363,951 321,895 1.13 25,534 12,495 2.04
San Marcos 46 4,081,664 2,538,994 1.61 105,051 74,815 1.40
Santa Rosa 70 1,757,133 886,464 1.98 88,869 23,071 3.85
Sololá 21 749,294 810,483 0.92 23,154 24,550 0.94
Suchitepéquez 27 2,545,635 555,180 4.59 73,293 11,689 6.27
Totonicapán 40 721,686 999,141 0.72 36,995 42,234 0.88
Zacapa 55 3,822,529 774,130 4.94 85,968 23,755 3.62
Total — 1,199 97,036,208 45,760,490 — 2,492,824 1,651,732 —
Ratio = ENCOVI / MAGA. Values far from 1.0 indicate large discrepancies requiring calibration.
Red highlight: ratio < 0.5 or > 2.0
WarningSubstantial Discrepancies

Several departments show ratios far from 1.0, indicating that ENCOVI weighted estimates diverge substantially from MAGA administrative data. Calibration corrects these discrepancies departmentally.

Post-Calibration Validation

After applying departmental calibration, we verify that weighted totals match MAGA targets exactly.

Table 3: Validation of calibrated weights against MAGA targets
Post-Calibration Validation: ENCOVI vs MAGA. Calibrated weighted estimates should match MAGA targets (ratio ≈ 1.00)
Post-Calibration Validation: ENCOVI vs MAGA
Calibrated weighted estimates should match MAGA targets (ratio ≈ 1.00)
Department N (sample)
Production (qq)
Land (mz)
Calibrated MAGA Ratio Calibrated MAGA Ratio
Alta Verapaz 99 5,296,634 5,296,634 1.0000 252,979 252,979 1.0000
Baja Verapaz 47 720,542 720,542 1.0000 40,233 40,233 1.0000
Chimaltenango 84 1,065,264 1,065,264 1.0000 34,835 34,835 1.0000
Chiquimula 97 1,518,844 1,518,844 1.0000 65,717 65,717 1.0000
El Progreso 28 367,087 367,087 1.0000 20,283 20,283 1.0000
Escuintla 28 528,828 528,828 1.0000 10,216 10,216 1.0000
Guatemala 14 816,778 816,778 1.0000 28,939 28,939 1.0000
Huehuetenango 64 4,023,703 4,023,703 1.0000 156,526 156,526 1.0000
Izabal 85 1,254,753 1,254,753 1.0000 30,116 30,116 1.0000
Jalapa 54 1,095,143 1,095,143 1.0000 40,300 40,300 1.0000
Jutiapa 116 2,646,775 2,646,775 1.0000 86,098 86,098 1.0000
Petén 73 10,658,142 10,658,142 1.0000 356,518 356,518 1.0000
Quetzaltenango 33 2,580,710 2,580,710 1.0000 62,296 62,296 1.0000
Quiché 74 5,171,702 5,171,702 1.0000 224,720 224,720 1.0000
Retalhuleu 22 1,129,300 1,129,300 1.0000 29,346 29,346 1.0000
Sacatepéquez 22 321,895 321,895 1.0000 12,495 12,495 1.0000
San Marcos 46 2,538,994 2,538,994 1.0000 74,815 74,815 1.0000
Santa Rosa 70 886,464 886,464 1.0000 23,071 23,071 1.0000
Sololá 21 810,483 810,483 1.0000 24,550 24,550 1.0000
Suchitepéquez 27 555,180 555,180 1.0000 11,689 11,689 1.0000
Totonicapán 40 999,141 999,141 1.0000 42,234 42,234 1.0000
Zacapa 55 774,130 774,130 1.0000 23,755 23,755 1.0000
Total — 1,199 45,760,490 45,760,490 — 1,651,732 1,651,732 —
Ratio = Calibrated / MAGA. Values of 1.0000 confirm exact calibration.
Green highlight: both ratios within ±0.1% of target.
NoteCalibration Result

The 22 departments reach a ratio of 1.0000 against their MAGA targets for both production and land. The calibration is exact by construction: it solves for the departmental weight adjustment that reproduces both totals. What the diagnostics below report is the size of the adjustment this required.

Record Expansion with Jitter

Calibration can produce very large individual weights, which create artificial “steps” in cumulative distributions. These steps interfere with threshold-based farmer segmentation, where cutpoints may fall on weight discontinuities.

We address this by expanding high-weight farmers into multiple records with controlled random variation (jitter). The following variables are jittered in the expanded records (the original base record of each farmer is left untouched):

  • qty_sold_qq, qty_household_qq, qty_animal_seed_qq — Destination quantities (sales, household consumption, animal feed and seed reserves)
  • total_production — Recalculated as the sum of the jittered destination quantities, so that allocation coherence is preserved
  • land_wtcorn — Cultivated land (manzanas)
  • exp_seeds — Seed expenditure (Q)
  • revenue — Sales revenue (Q)
  • farmer_age — Age of the farmer

Jitter is applied with the following parameters:

  • Intensity: 3% of each variable’s standard deviation
  • Bounds: non-negative values enforced post-jitter
  • Allocation coherence: total_production is derived from the jittered destination quantities rather than jittered on its own, which keeps production equal to the sum of its uses in every expanded record
Table 4: Weight diagnostics and expansion parameters
Calibration Diagnostics and Expansion Plan. Weight validation metrics and jitter expansion parameters
Calibration Diagnostics and Expansion Plan
Weight validation metrics and jitter expansion parameters
Metric Value Threshold Status
Calibration Metrics
Weight CV (initial) 64.1% — —
Weight CV (calibrated) 89.5% — —
Design Effect (DEFF) 1.801 < 2.0 ✓ Pass
Adjustment ratio range [0, 10.28] [0.3, 3.0] Warning
Ratios < 0.3 (extreme low) 198 (16.5%) < 10% Warning
Ratios > 3.0 (extreme high) 7 (0.6%) < 10% ✓ Pass
Expansion Parameters
Expansion threshold 500 — —
Target weight per record 100 — —
Farmers to expand 621 (51.8%) — —
Final record count 7,429 (6.2x expansion) — —
DEFF = Design Effect (1 + CV²). Values > 2 indicate efficiency loss.
Blue rows: Expansion parameters to smooth distribution steps.
Histogram of calibrated farmer weights along the x-axis against the number of farmers on the y-axis. The distribution is right-skewed, with most farmers concentrated at low weights and a long tail of high weights. A dashed vertical line marks the expansion threshold of 500; farmers above it are split into multiple records so that oversized weights, which create artificial steps in the distribution, are smoothed out.
Figure 1: Distribution of calibrated weights with expansion threshold
ImportantExpansion Strategy

Farmers with calibrated weight above the expansion threshold are split into multiple records, each with a target weight of approximately one fifth of the threshold. This brings the maximum weight down by an order of magnitude and smooths the large steps out of the distribution. A 3% standard deviation jitter is applied to base variables to introduce controlled variation among copies.

ImportantWhat the Adjustment Costs

The departmental targets are met exactly, and the adjustment needed to meet them is large. Individual weights are multiplied by factors spanning [0, 10.28], against the conventional working range of [0.3, 3.0]: 198 (16.5%) of records fall below 0.3 and 7 (0.6%) above 3.0.

The dispersion of the weights rises accordingly, with their coefficient of variation moving from 64.1% to 89.5% and a design effect of 1.801. This is the arithmetic consequence of the pre-calibration gap: the ENCOVI-to-MAGA production ratio runs from 0.5 to 10.12 across departments, so the weights carry the whole of that correction.

Final Validation

After expansion, jitter application, and re-calibration (to correct small jitter-induced bias), we verify the final dataset still matches MAGA targets.

Table 5: Final validation: Expanded dataset weighted totals vs MAGA targets
Final Validation: Expanded Dataset vs MAGA Targets. Weighted totals after expansion and re-calibration
Final Validation: Expanded Dataset vs MAGA Targets
Weighted totals after expansion and re-calibration
Department
Sample
Production (qq)
Land (mz)
Records Households Farmers Final MAGA Ratio Final MAGA Ratio
Alta Verapaz 1,090 99 106,782 5,296,634 5,296,634 1.0000 252,979 252,979 1.0000
Baja Verapaz 160 47 17,846 720,542 720,542 1.0000 40,233 40,233 1.0000
Chimaltenango 468 84 46,641 1,065,264 1,065,264 1.0000 34,835 34,835 1.0000
Chiquimula 493 97 50,738 1,518,844 1,518,844 1.0000 65,717 65,717 1.0000
El Progreso 33 28 5,848 367,087 367,087 1.0000 20,283 20,283 1.0000
Escuintla 179 28 18,216 528,828 528,828 1.0000 10,216 10,216 1.0000
Guatemala 199 14 19,316 816,778 816,778 1.0000 28,939 28,939 1.0000
Huehuetenango 851 64 82,011 4,023,703 4,023,703 1.0000 156,526 156,526 1.0000
Izabal 390 85 37,863 1,254,753 1,254,753 1.0000 30,116 30,116 1.0000
Jalapa 142 54 21,668 1,095,143 1,095,143 1.0000 40,300 40,300 1.0000
Jutiapa 391 116 54,742 2,646,775 2,646,775 1.0000 86,098 86,098 1.0000
Petén 376 73 42,446 10,658,142 10,658,142 1.0000 356,518 356,518 1.0000
Quetzaltenango 272 33 25,551 2,580,710 2,580,710 1.0000 62,296 62,296 1.0000
Quiché 893 74 86,830 5,171,702 5,171,702 1.0000 224,720 224,720 1.0000
Retalhuleu 102 22 10,577 1,129,300 1,129,300 1.0000 29,346 29,346 1.0000
Sacatepéquez 32 22 5,759 321,895 321,895 1.0000 12,495 12,495 1.0000
San Marcos 553 46 54,490 2,538,994 2,538,994 1.0000 74,815 74,815 1.0000
Santa Rosa 374 70 37,379 886,464 886,464 1.0000 23,071 23,071 1.0000
Sololá 64 21 9,459 810,483 810,483 1.0000 24,550 24,550 1.0000
Suchitepéquez 140 27 13,115 555,180 555,180 1.0000 11,689 11,689 1.0000
Totonicapán 149 40 19,718 999,141 999,141 1.0000 42,234 42,234 1.0000
Zacapa 78 55 16,478 774,130 774,130 1.0000 23,755 23,755 1.0000
Total — 7,429 1,199 783,474 45,760,490 45,760,490 — 1,651,732 1,651,732 —
Ratio = Final / MAGA. Values of 1.0000 confirm exact calibration.
Farmers = Weighted total of white maize farmers represented.
Green highlight: both ratios within ±0.01% of target.

Transformation Pipeline

Table 6: Complete transformation from ENCOVI sample to calibrated dataset
Dataset Transformation Pipeline. From ENCOVI sample to MAGA-calibrated expanded dataset
Dataset Transformation Pipeline
From ENCOVI sample to MAGA-calibrated expanded dataset
Stage Records Weight Range Description
A. Input data (02_01 output) 1,199 [76, 2988] Cleaned farmers from Script 02_01 with scaled ENCOVI weights (×1.95)
B. Weight calibration (MAGA) 1,199 [0, 6166] Weights adjusted to match MAGA production and land totals by department
C. Record expansion 7,429 [0, 498] High-weight farmers (>500) split into multiple records (target weight ≈100)
C. Jitter application 7,429 — 3% SD jitter applied to base variables; derived variables recalculated
C. Re-calibration 7,429 [0, 524] Second calibration pass to correct jitter-induced bias (~1%)
D. Final output 7,429 [0, 524] Final dataset ready for segmentation modeling
Pipeline progresses from cleaned ENCOVI input through calibration, expansion with jitter, and re-calibration to the final dataset ready for segmentation modeling

Summary

Key Findings

  1. Initial discrepancies: Before calibration the ENCOVI-to-MAGA production ratio ranges from 0.5 to 10.12, with 13 of the 22 departments outside the [0.5, 2.0] range.

  2. Calibration: Linear calibration brings both production and land area onto the MAGA departmental targets, with ratios at 1.0000.

  3. Cost in weight dispersion: Meeting those targets moves the weight coefficient of variation from 64.1% to 89.5%, with a design effect of 1.801 and adjustment ratios spanning [0, 10.28].

  4. Weight smoothing: Record expansion with controlled jitter removes the discontinuities from the distribution while holding the weighted totals, and a re-calibration pass corrects the small bias the non-negativity constraint introduces.

Application

The calibrated and expanded dataset provides:

  • Representative farmer population — Weighted totals match official MAGA statistics for economic impact projections
  • Smooth distributions — No artificial steps for threshold-based segmentation in farmer classification
  • Internal consistency — Allocation coherence between total production and destination quantities preserved through the jitter procedure
Back to top