DHS recode files, South Asia

Starter code in Stata, R and Python for the Demographic and Health Surveys of India, Bangladesh, Nepal, Pakistan, Maldives, Afghanistan and Sri Lanka. Twenty-eight surveys, one loader: the recode structure is the same in all of them.

This folder does not contain data and cannot. The DHS Program licence lets you use the files, not pass them on, so everyone downloads their own copy.

What the loaders do

They read only the columns you name from a recode such as IAIR7EFL.DTA (724,115 women, about five thousand variables), attach the weight and design variables for that recode, apply the documented scale factors, and stop with an explanation when something is off.

Getting the files

1. Register at dhsprogram.com, describe your project, and request access. Access is granted per country, so India and Bangladesh are two separate requests.
2. Download the recode you need in Stata (.DTA) or SPSS (.SAV) format. The flat .DAT files need the accompanying dictionary and are more work for no gain here.
3. Keep the file under its DHS name. The name carries the country, recode type and version, and these scripts read all three from it.

IAIR7EFL.DTA is India, Individual Recode, phase 7, version E, flat. The version character changes when DHS reissues a file, and it can be a digit: NFHS-4’s individual recode is IAIR74FL.DTA. If you rename the file, pass the recode type explicitly instead.

Which file you want

KR covers the last five years and BR the woman’s complete birth history, so under-five mortality rates come from BR. Both include children who have died, which is why anthropometry work on KR has to filter on b5 == 1 first. Child anthropometry appears in both KR and PR, under different variable names.

Quick start

Stata:

do load_dhs.do
dhs_vars, file("IAIR7EFL.DTA") vars(v012 v106 v190 sdist)   // what exists here?
dhs_load, file("IAIR7EFL.DTA") vars(v012 v106 v190)
svy: mean v012

R:

source("load_dhs.R")
ir  <- read_recode("IAIR7EFL.DTA", c("v012", "v106", "v190"))
des <- dhs_design(ir)
survey::svymean(~v012, des)

Python:

from load_dhs import read_recode, clean_anthropometry
ir = read_recode("IAIR7EFL.DTA", ["v012", "v106", "v190"])
kr = clean_anthropometry(read_recode("IAKR7EFL.DTA", ["hw70", "b5", "v190"]), "KR")

Or from the shell:

python load_dhs.py IAKR7EFL.DTA --vars v190 hw70 hw71 b5 --anthro --out stunting.csv

What goes wrong silently

Each of these produces a plausible number rather than an error.

Weights are stored with six implied decimals. v005 averages about a million and has to be divided by 1,000,000. Leaving it out does not change a weighted mean, because the scale cancels, so the error survives every sanity check you are likely to run. It multiplies every weighted count and total by a million. The loaders divide it and refuse to run on a file whose weights average under 1,000, which is what an already-divided file looks like.

The design variables change with the recode. The men’s recode weights on mv005 and clusters on mv021. Analysing men with v005 does not fail; those variables simply are not there, and if you have merged files they may be there and wrong. Use v021 for the PSU rather than v001, and v022 for strata, falling back to v023 in older rounds.

Anthropometry is stored times 100, with flags above it. hw70 of -162 means a height-for-age z-score of -1.62. Values of 9996 and up are flags and missing. Divide before you drop them and every flagged child becomes a plausible 99.96, then a plausible 0.9996, and your stunting rate falls.

The anthropometry columns are named differently in different recodes. hw70 to hw73 in the children’s recode, hc70 to hc73 in the household member recode. Ask for the wrong pair and you get nothing rather than an error.

Dead children stay in the birth history. BR and KR include children who have died. Filter on b5 == 1 for analyses of living children, and do not filter on it for mortality.

Region codes are not comparable across rounds. v024 in NFHS-3 (2005-06), NFHS-4 (2015-16) and NFHS-5 (2019-21) uses three different state code sets: Telangana was created in 2014 and Ladakh separated from Jammu and Kashmir in 2019. Nothing in the file tells you this. Map to names round by round before pooling.

Wealth quintiles are relative to their own survey. v190 is constructed within each survey, so quintile 1 in Bangladesh 2022 and quintile 1 in India 2019-21 are not the same standard of living, and neither are quintile 1 in NFHS-4 and NFHS-5. Use v191, the underlying score, if you need something continuous, and remember that it is still survey-specific.

Century month codes. v008 and b3 are months since January 1900, so January 2020 is 1441. cmc_to_year_month in each script converts them.

Checking your answer

benchmarks/nfhs5_stunting.csv holds the figures DHS published for India’s NFHS-5 child stunting: 35.5 percent nationally, 46.1 in the poorest wealth quintile falling to 22.9 in the richest, with breakdowns by residence and mother’s education. NFHS-4’s national figure of 38.4 is there for comparison.

Two worked examples reproduce the table, one in R and one in Python:

• FieldStack survey_tools/dhs_stunting.R
• EquityStack survey_estimation/dhs_stunting.py

Both consume the CSV this loader writes, so the repositories are coupled through a file rather than a dependency.

Across rounds

variables.csv maps a canonical name to the variable in each recode, with its scale factor and the caveats that apply. The loaders read the scale factors from that file at runtime, so the documented scale and the applied scale cannot drift apart.

Before pooling rounds, run the variable check first. It tells you what is missing from the older file rather than letting a merge fill it with nothing:

dhs_vars, file("IAIR52FL.DTA") vars(v190 v191 sdist)

v190 is absent from rounds before roughly 2000. sdist exists for India from NFHS-4 onward only, which is why NFHS-5 supports district estimates and NFHS-3 does not.

The surveys

Pulled from the DHS Program API on 2026-09-08. Sample sizes are the counts DHS publishes for the survey, not the row count of any particular recode file.

Sri Lanka appears once. The 1987 survey is the only Sri Lankan round distributed as DHS recode files; the later Demographic and Health Surveys of 2006-07 and 2016 were run by the Department of Census and Statistics and are obtained from them directly, in a different format. Afghanistan has a single round, 2015.

Files here

Testing

No real data is needed. make_fixture.py writes small synthetic files carrying the structure that matters, raw weights near a million, anthropometry times 100 with flags at 9990 and above, century month codes, the mv prefix in the men’s recode. Every value comes from a seeded generator and none of it is data about anybody.

python make_fixture.py --outdir fixtures
python test_load_dhs.py     # 30 checks
Rscript test_load_dhs.R     # 33: the same ones, plus the design object

The R suite mirrors the Python one and adds three checks for the survey design object. Verified on Python 3 with pandas and pyreadstat, and on R 4.3.3 with survey 4.2.1 and haven 2.5.4; both return the same numbers from the same fixture.

Stata has no free runtime, so load_dhs.do is the one file here that has not been executed. It carries the same logic and the same guards, and the fixture check at the bottom of it takes about a minute. Run that once before trusting it on real data.

Sources

• Survey list, fieldwork dates, release dates and sample sizes: DHS Program API, api.dhsprogram.com/rest/dhs/surveys, retrieved 2026-09-08.
• Recode structure, variable naming and scale factors: Guide to DHS Statistics and the recode manual for each phase, published at dhsprogram.com/publications.
• Anthropometry plausibility bounds: WHO Child Growth Standards, 2006.

Where this folder and your survey’s codebook disagree, the codebook is right.