Working with NFHS data in R

The National Family Health Survey (NFHS) is the main source of population-level data on health, nutrition and family welfare in India. It is a household survey with a complex design. If you treat it as a simple random sample, your estimates and your confidence intervals will both be wrong. This post covers the steps I follow, from getting the files to a weighted estimate by state.

Getting the data

NFHS microdata are distributed by the DHS Program.

1. Register at dhsprogram.com and describe your project.
2. Request the India datasets for the survey round you need.
3. Download the files in Stata format (.dta).

You will see several files per round, one for each unit of analysis:

The file names depend on the round. Check the download page and the codebook that comes with the data.

A rule of the data agreement: you may not share the microdata. For a reproducible project, share your code and a script that tells readers how to download the files themselves. I keep the data in a folder that is excluded from Git.

Reading the file

library(haven)
library(dplyr)

br <- read_dta("data/your_births_file.DTA")

read_dta keeps the value labels from Stata. Use haven::as_factor() to turn a coded column into labelled categories, and haven::zap_labels() when you want plain numbers.

The four variables that define the design

Every DHS-style file has the same core design variables:

Divide v005 by 1,000,000 to get the weight.

br <- br |>
  mutate(wt = v005 / 1e6)

A warning about weights. The weights in a state-level file are made for estimates within a state. If you pool several states to get a national figure, read the survey’s final report and the DHS guidance on weights before you do it. Pooling state weights without adjustment can give a wrong national estimate.

Declare the survey design

library(survey)
options(survey.lonely.psu = "adjust")

des <- svydesign(
  ids     = ~v021,
  strata  = ~v022,
  weights = ~wt,
  data    = br,
  nest    = TRUE
)

nest = TRUE is needed because PSU numbers are only unique within a stratum. The lonely.psu option tells R what to do when a stratum has a single PSU, which happens when you subset to a small group.

A worked example: caesarean delivery by state

The births recode has m17 (delivery by caesarean) and b19 (age of the child in months). Restrict the analysis to births in the last five years.

br <- br |>
  mutate(
    csection = case_when(m17 == 1 ~ 1, m17 == 0 ~ 0),
    recent   = b19 < 60
  )

des <- svydesign(ids = ~v021, strata = ~v022, weights = ~wt,
                 data = br, nest = TRUE)

# subset the design, not the data frame
des_recent <- subset(des, recent)

est <- svyby(
  ~csection, ~v024, des_recent, svymean,
  na.rm = TRUE, vartype = c("se", "ci")
)

The point to notice is subset(des, recent). If you filter the data frame before you declare the design, R forgets that the full sample existed, and the standard errors come out wrong. Always subset the design object.

Multiply the proportions by 100 and report them with their confidence intervals. For a state with few births, check the number of unweighted cases too. A state estimate on 40 births is not one to put in a headline.

Mistakes I see often

• Using unweighted means. mean(br$csection) gives the average in the sample, not in the population.
• Treating “don’t know” as data. Codes such as 8, 9 or 99 often mean “don’t know” or “missing”. Recode them before you analyse.
• Comparing rounds without care. District boundaries and some definitions change between NFHS rounds. Check the codebook before you compare a district over time.
• Reporting a 95% interval from too few clusters. The effective sample size is closer to the number of PSUs than to the number of rows.
• Forgetting to cite the DHS Program. The data agreement asks for it.

Linking to maps

India files include a district code, usually sdist. Check the codebook for your round. You can match it to a district boundary file and use the resulting estimates in a map. District estimates from a survey designed for state-level precision have wide intervals, so smooth them with a model before you draw conclusions. My post on Bayesian disease mapping with INLA shows one way.

My own analysis of spatial clusters of caesarean section in NFHS data is described on the Research page.