Author

Mamadou S. Diallo, Ph.D.

Modified

August 29, 2026

DHS Ethiopia 2024–25 Case Study Outline — svy

Validates svy’s Taylor linearisation (and, in a second phase, jackknife) against the sampling error appendix of the 2024–25 EDHS final report (FR399). Unlike the LFS Canada study, the producer here publishes the results (Appendix B: estimate, SE, DEFT, CI for ~130 indicators × 17 domains), not reference code — so the validation target is the published tables themselves.

Design facts and target numbers below were transcribed from FR399.pdf (Appendix A pp. 451–457, Appendix B pp. 463–468) on 2026-08-29. svy API calls verified against svy 0.26.0. Microdata pending DHS Program approval — dataset filenames and design-variable choices marked TBD until the files arrive.

0. Scope decisions (state up front in the case study)

Decision Choice Why
Validation target Appendix B, Tables B.2 (national), B.3–B.4 (urban/rural), regional tables The producer’s own SAS/Taylor output. No reference code published, so the tables are the ground truth.
Phase 1 estimators Proportions and means, Taylor linearisation Matches ICF’s method for means/proportions/ratios. Direct one-to-one comparison.
Phase 2 (maybe separate post) Jackknife (delete-one-cluster, 797 reps) ICF uses JRR only for complex rates (TFR, mortality). svy has create_jk_wgts but no TFR/mortality-rate estimators — reproducing those estimators is out of scope; reproducing the JK variance machinery on a proportion is feasible.
Datasets IR (women) + PR/HR (household population) first; KR (children) for vaccination if phase 1 goes fast Table B.2 splits into HOUSEHOLDS AND POPULATION (PR/HR base) and WOMEN (IR base). Two files cover most rows.
Weight v005/1e6 (IR), hv005/1e6 (HR/PR) Standard DHS. Weights are normalised — valid for means/proportions/ratios, not totals. The report says this itself (A.4), which makes a clean callout.
Design variables psu = v021, strata = v022 — TBD: confirm v022 vs v023 against the file Sampling strata should be region × urban/rural = 27 (Addis Ababa has no rural stratum). Acceptance check: the strata variable must have exactly 27 levels and cross-tab cleanly with v024 × v025.
Subpopulations where= everywhere, never filter Same footgun as NHANES: subsetting rows treats domain size as fixed and biases the variance. Appendix B bases are all subpopulations (currently married women, women with a live birth in last 2 years, …).
Domains by= for urban/rural and the 14 regions One pass reproduces Tables B.3–B.20’s structure.
FPC None (pop_size not set) ICF’s formula drops the (1−f) factor explicitly: “f … is so small that it is ignored.” svy without pop_size matches.
HIV prevalence Out of scope Needs the AR recode merged in, separate weights (A.4 lists 7 weight sets), extra approval sensitivities. Mention, don’t do.
GPS / spatial Out of scope Not requested; different abstract.

1. The survey (Background section material)

  • 2024–25 EDHS, fifth in the series (2000, 2005, 2011, 2016). Frame: 2019 Population and Housing Census cartographic frame from the Ethiopian Statistical Service — 149,215 residential EAs. 14 regions.
  • Representative for national, urban/rural, and each of the 14 regions.
  • Fieldwork disruption worth narrating: 8 of the 75 Amhara clusters were not interviewed for security reasons (1 urban, 7 rural). Weights were adjusted for cluster nonresponse at the stratum level. 805 selected clusters → 797 non-empty clusters, which is also the JRR replicate count.
  • Report interpretation example to reuse: mean children ever born 2.432 ± 2 × 0.041 → CI (2.351, 2.513). Average women’s-survey DEFT is 2.282.

2. The design (Appendix A → svy)

Two-stage stratified:

  1. 27 strata = 14 regions × urban/rural, minus rural Addis Ababa.
  2. Stage 1: 805 EAs by PPS (households from the 2019 PHC as MOS); implicit stratification via frame sorting. Large EAs (urban >300 hh, rural >200 hh) segmented; one segment kept PPS. A “cluster” = EA or segment.
  3. Stage 2: fixed 28 households per cluster, equal-probability systematic, from a fresh listing. No replacements.
  4. Weights: inverse selection probability × nonresponse adjustments (cluster, household, individual — all at stratum level), then normalised so weighted cases = unweighted cases nationally. Seven weight sets (household, women, men’s-survey household, men, domestic violence, HIV-w, HIV-m).

The svy setup is the whole point of the post — three lines:

import svy

women = svy.read_stata("ETIR8?FL.DTA")  # TBD: actual filename on approval

sample = svy.Sample(
    data=women.with_columns((svy.col("v005") / 1e6).alias("wgt")),
    design=svy.Design(stratum="v022", psu="v021", wgt="wgt"),
)
cpr = sample.estimation.prop(
    "using_modern", where=svy.col("v502") == 1, by="v025"
)

(Exact column expressions TBD against codebook; where semantics and method="taylor"-by-default verified in 0.26.0.)

3. Validation targets (Table B.2, total sample)

Transcribed 2026-08-29; re-verify against the PDF before publication.

Indicator Base R SE N (unwtd) DEFT
Electricity as primary source of lighting de jure household pop 0.402 0.019 95,594 5.061
Improved drinking water source de jure household pop 0.716 0.017 95,594 4.714
Urban residence women 15–49 0.346 0.010 21,395 3.164
No education women 15–49 0.325 0.009 21,395 2.927
Currently married/in union women 15–49 0.613 0.007 21,395 2.077
Currently using any contraceptive method currently married 15–49 0.354 0.012 13,007 2.638
Currently using any modern method currently married 15–49 0.345 0.011 13,007 2.640
Unmet need total currently married 15–49 0.154 0.006 13,007 1.902
Mean children ever born (15–49) women 15–49 2.432 0.041 21,395 2.243
Received ANC from skilled provider women w/ live birth, last 2 yrs 0.781 0.014 5,021 2.454
Delivered in a health facility live births, last 2 yrs 0.614 0.021 5,208 2.934

Success criterion: match R to the printed 3 decimals, SE/DEFT to printed precision. Exact float agreement is not expected against a PDF table — this is “published-precision” validation, framed like the LFS post’s Daily comparison, not like its Part A. Cross-check svy vs R survey (svydesign(ids=~v021, strata=~v022, weights=~wgt)) and/or Stata svyset at full float precision — that’s where the tight (1e-15-style) agreement claim can live.

Watch for: lonely PSUs. If any region×residence stratum ends up with a single cluster in a subpopulation table, note how ICF’s SAS handles it vs svy’s Singleton/SingletonHandling options. Could be its own callout.

4. Domain estimation

  • by="v025" (urban/rural) → Tables B.3–B.4; by="v024" (region) → regional tables. One call per indicator, 17 published domains covered.
  • Narrative point: DHS runs each domain as a separate SAS job; svy does it in one pass over the undamaged design. Variance still uses all strata (domain estimation, not subset estimation).
  • Regional mortality/fertility use 10-year windows vs 5-year national — a reason regional rate rows are NOT phase-1 targets.

5. DEFT section

DEFT = SE_design / SE_srs is reported for every cell and averages 2.282 for women. Checked svy’s implementation (svy-rs estimation/taylor.rs, verified 2026-08-29 against svy 0.26.0):

  • deff="wr" is Kish’s deft²: SRS variance = weighted S² / n, no FPC, weight-scale-invariant, with zero-weight (out-of-domain) rows excluded from
    1. S² uses the n/(n−1) factor, so a proportion’s SRS variance is effectively p(1−p)/(n−1) with weighted p and unweighted domain n.
  • deff="wor" uses sum(w) as N for the FPC when pop_size is absent. DHS weights are normalised (sum(w) ≈ n), so the FPC degenerates — svy deliberately returns a structured error rather than deff ≈ 1e16. Use deff="wr" and say why; the error itself is a teachable moment about normalised weights.
  • ICF’s DEFT also omits f (“so small that it is ignored”) and computes the SRS reference from the weighted estimate at unweighted n, so sqrt(deff_wr) is the right comparator. Residual conventions to confirm empirically: n vs n−1 in ICF’s SRS denominator (invisible at printed precision for n in the thousands) and the domain-n definition on subpopulation rows.

The lighting indicator’s DEFT of 5.06 is a nice hook: electrification is spatially clustered, so 95,594 people carry the information of ~3,700 SRS draws. Contrast with currently-married-in-union at 2.08 — attitudes and statuses cluster less than infrastructure.

6. Phase 2 sketch (probably a follow-up post, not this one)

  • JRR: 797 delete-one-cluster replicates, unstratified, k(k−1) scaling with r_i = k·r − (k−1)·r_(i). Map to create_jk_wgts(...) — check svy’s JK flavour (stratified delete-one is the default; DHS’s is unstratified) and the matching scale/df convention before promising anything.
  • TFR (4.049, SE 0.108) and U5MR (51.263, SE 4.304) need DHS-specific rate estimators (person-years exposure, synthetic cohort probabilities) that svy does not implement. Say so plainly; do not hand-wave.

6b. Scope decision: no GLM / chi-square in this piece (2026-08-29)

Keep this post strictly validation. GLM and Rao-Scott chi-square sections would have no published reference number to match (the report has no regressions or test statistics) and would dilute the one falsifiable claim. NHANES already owns the analytic-workflow showcase. Instead: a one-paragraph closer — the same svy.Design object drives regression/ChiSquare, teased as a follow-up case study (“determinants of modern contraceptive use”, validated against R svyglm/svychisq at float precision, same data request/abstract covers it).

7. Compliance / boilerplate

  • DHS data may not be redistributed — repo gets code + published-table comparisons only, never microdata. .gitignore the DTA files (mirror the LFS folder’s approach).
  • Citation: Ethiopian Statistical Service (ESS) and ICF. 2025(?). Ethiopia Demographic and Health Survey 2024–25. Verify exact citation page (FR399 p. 4).
  • Footer disclaimer mirroring LFS: use of the data does not imply endorsement by the DHS Program.

8. Proposed index.qmd skeleton

  1. Summary / TL;DR callout — design in 3 lines of svy; N indicators × 17 domains reproduced at published precision; cross-checked against R at float precision.
  2. Background — the survey, the security-related cluster loss, what Appendix B publishes.
  3. The design, from Appendix A to svy.Design — strata/PSU/weight table, normalised-weights callout (no totals!).
  4. Reproducing Appendix B — national table, then by= domains.
  5. Design effects — the DEFT story.
  6. What doesn’t match and why — printed rounding, any lonely-PSU rows, jackknife rows excluded.
  7. Cross-language check — R survey (and/or Stata) at float precision.
  8. References + DHS disclaimer + community-signal callout.

9. Open items

Back to top