DLHS raw data versus aggregated reports for district analysis

On the ground, however, the same figure may conceal several decisions about who was eligible for the indicator, how missing responses were treated, which households received greater sampling weight, and whether the district boundary still means the same thing in another survey round.

That is the central choice in DLHS raw data vs aggregated reports for district analysis. Published district summaries are usually the quickest route to a defensible descriptive comparison. Raw or micro-level files are the route to custom questions, subgroup analysis, and statistical modelling—but only if the researcher can reconstruct the survey design rather than treating the records as a simple spreadsheet.

I have seen the practical consequences of this distinction in maternal and reproductive health work. A published number can help a programme team identify where antenatal care appears weak. It cannot, by itself, explain whether the gap is concentrated among younger women, rural households, poorer families, or communities that were least connected to a functioning facility. For that next question, the level at which the data are available matters enormously.

The architecture of DLHS-3: from household records to district summaries

DLHS-3 was conducted mainly in 2007–08 and covered 601 districts across 34 Indian states and union territories. Its household component included 720,320 surveyed households according to the Government of India’s data catalogue, alongside interviews with 643,944 ever-married women aged 15–49 and 166,260 unmarried women aged 15–24.

Those figures describe much more than the size of a dataset. They show the layered structure of the District Level Household Survey. DLHS-3 was not a single questionnaire producing one universal table. It brought together several units of observation:

• Households, with information on socioeconomic characteristics, living conditions, and household composition.
• Individual women, including maternal health, reproductive health, fertility-related experiences, and selected child-health information.
• Villages or communities, capturing local infrastructure and contextual characteristics.
• Health facilities, collected through a separate component with its own coverage rules and unit of observation.

This distinction is clinically meaningful. If I am asking whether women received antenatal care, the relevant record may be an individual woman’s interview. If I am asking whether the village has a functioning sub-centre nearby, I may need community information. If I am examining staffing, equipment, or service readiness, the facility file becomes relevant. These are connected care pathways, but they are not interchangeable datasets.

Aggregated reports compress these layers into indicators. A district fact sheet may present the proportion of women receiving a specified number of antenatal visits, the proportion delivering in a health facility, or selected immunization measures. That summary is valuable because it has already passed through a process of indicator definition, tabulation, and publication.

It is also limited in a very specific way: the published table usually gives you the result, not the full analytical route that produced it.

A report may not give enough information to recreate the exact denominator, identify all excluded observations, distinguish “not applicable” from “missing,” or determine whether the displayed percentage is weighted. Nor should a reader assume that a district percentage can be reverse-engineered simply by multiplying it by the apparent sample size. The numerator, denominator, eligibility rules, missing-value treatment, and weighting method all matter.

An aggregated district indicator is a conclusion supplied by the survey team; raw data are the evidence trail from which a new conclusion may be built.

What the two formats are designed to do

The most useful way to compare the formats is not to ask which one is “better.” They answer different kinds of questions.

For a district planning meeting, an official fact sheet may be exactly the right starting point. For a thesis asking whether maternal-care uptake differs by education, residence, and household socioeconomic position, the summary report will usually not be enough.

Navigating access: what “DLHS raw data” actually means

Researchers often use the phrase “DLHS raw data” to describe several different things. That can create confusion before the analysis even begins.

A district fact sheet is not raw data. A table of indicators is not raw data. A downloadable report containing district percentages is still a published summary, even if it is available as a PDF or spreadsheet. Raw or unit-record data refer to records at the household, individual, village, or facility level, from which the researcher can construct or re-tabulate indicators.

The current IIPS data-access information describes demographic datasets as available through an online portal for registered users, with identity documentation and a legitimate academic research purpose required. In practical terms, researchers should plan for controlled or request-based access rather than assuming that every DLHS-3 unit-record file is an unrestricted public download.

That affects project planning. Access is not a final administrative step after the research design is complete; it is part of the design.

Before requesting files, I would write down the exact analytical question and the minimum data structure needed to answer it. For example:

1. Define the population of interest.

Are you studying all households, ever-married women aged 15–49, unmarried women aged 15–24, mothers of children in a specified age range, or another eligible group?

2. Identify the unit of analysis.

A household-level question should not be analysed from an individual women’s file without a clear linkage strategy. A facility-level question cannot be answered by treating household records as a proxy for service readiness.

3. List the indicators and their likely denominators.

“Institutional delivery” and “antenatal care” are not self-defining. Their denominators may be restricted to women with a live birth in a particular reference period or to another eligibility group specified by the questionnaire and tabulation rules.

4. Request the supporting documentation, not only the data file.

A file without a questionnaire, codebook, merge instructions, and survey-design information is not a ready-to-analyse dataset. It is an archive that still requires reconstruction.

5. Plan for identity and research-use requirements.

The access pathway may require registration, valid identification, and a description of the academic purpose. That is not the same as commercial or unrestricted redistribution permission.

6. Keep an access log.

Record the dataset version, request date, files received, documentation supplied, and any decisions made when a variable or district identifier is unclear. This becomes part of the study’s reproducibility record.

The phrase “accessing DLHS microdata files” can therefore be misleading if it suggests a single button and a single universal package. The practical route may involve a portal request, review, documentation checks, and careful confirmation of which survey components are actually supplied.

The files may not map neatly onto one another

A household file, women’s file, village file, and facility file may have different record counts and identifiers. A facility survey is not simply a second version of the household survey. Its sampled facilities, geographic coverage, and questionnaire are different.

This matters when a researcher wants to connect service availability with service use. A district-level facility measure and a district-level household measure may be compared descriptively, but that does not automatically establish that the same households were linked to the same facilities. A district average can hide substantial variation within the district, and the two survey components may not share the same sampling frame.

The safest approach is to treat each component as its own analytical layer and document any aggregation or linkage explicitly.

Statistical precision: what aggregated reports conceal

A district percentage feels concrete because it is easy to read. Its uncertainty is less visible.

DLHS-3 used a complex multistage design. In rural areas, sampling involved two-stage stratified random selection; in urban areas, the design used three stages. Primary sampling units were selected using probability proportional to size, with the 2001 Census used as the sampling-frame reference for PSU selection.

This is not a technical footnote. It determines how much information a record contributes and how much statistical independence can reasonably be assumed among records from the same sampled unit.

The district sample sizes were deliberately not uniform. IIPS describes nominal allocations of:

• 1,000 households in districts classified as performing well;
• 1,200 households in medium-performing districts;
• 1,500 households in low-performing districts.

The classification was based on earlier indicators, including antenatal care, institutional delivery, and immunization measures from DLHS-2. The purpose was to provide more observations in districts where weaker performance required closer measurement.

A larger nominal sample can improve the stability of a district estimate, but it does not make all district estimates equivalent. Equal numbers of households would not guarantee equal precision, and unequal numbers certainly do not. The effective sample size is influenced by clustering, stratification, weighting, nonresponse, and the distribution of the outcome.

Why the apparent denominator can mislead

Suppose a report presents an indicator for a district with 1,200 surveyed households. That does not mean the indicator was calculated from 1,200 households. The relevant denominator might instead include only eligible women, recent births, children in a specified age range, or respondents with complete information for the item.

A household survey contains many denominators. They are created by the questionnaire’s skip patterns and the indicator definition.

For a maternal-health measure, the analytical population might be:

• all women aged 15–49;
• ever-married women aged 15–49;
• women who had a birth within a reference period;
• women with a reported pregnancy outcome;
• children eligible for a vaccination question;
• households with a child of a specified age.

These populations are not substitutes for one another. A percentage calculated among all women cannot be compared casually with one calculated among recent mothers. When comparing survey summary reports with raw datasets, the denominator is often the first place where an apparently small discrepancy becomes an important methodological difference.

Clustering changes the uncertainty

People living in the same village or urban sampling unit often share transport barriers, local service availability, social norms, and exposure to the same health system. Their responses are therefore not equivalent to observations drawn independently from entirely separate places.

If a researcher analyses records as though every respondent were independent, standard errors may be too small. The result can look more precise than it really is. That is particularly concerning when comparing districts with modest differences or when fitting models with several predictors.

Aggregated reports may not publish confidence intervals, design effects, denominators, nonresponse adjustments, or the exact weighting procedures for every indicator. Their purpose is often communication and comparison, not a full statistical audit. A published district table remains useful, but its apparent simplicity should not be confused with complete uncertainty information.

Reconstructing indicators: weights and design variables are part of the data

The most common analytical mistake with survey microdata is to begin counting before understanding how the sample was drawn.

A raw file may contain one row per household or one row per respondent, but the rows do not necessarily represent equal numbers of people in the population. Sampling probabilities can differ by geographic area and survey component. Nonresponse and post-sampling adjustments may also affect the appropriate weight.

For district-level work, the researcher should locate and understand, at minimum:

• the final sampling weight for the relevant record type;
• the strata or stratification variable;
• the primary sampling unit identifier;
• the district identifier and its coding scheme;
• the household or respondent identifier needed for linking files;
• the survey round and geographic reference;
• missing-value and “not applicable” codes;
• the questionnaire wording and skip patterns.

The exact variable names cannot be assumed without the complete DLHS-3 codebook and file layout. That is one reason a methodological plan should not promise a particular code or software command before the documentation has been obtained.

Weighted and unweighted estimates answer different questions

An unweighted tabulation describes the records in the file. A weighted tabulation is intended to represent the target population under the survey’s sampling and adjustment procedures.

Neither is automatically “the real number.” They serve different purposes.

An unweighted count is useful for understanding how many observations are available for a subgroup and how sparse a cell may be. A weighted percentage is generally the appropriate basis for population-level estimation, provided the correct weight is used. In a district comparison, the researcher may need both: the weighted estimate for interpretation and the unweighted denominator for transparency about the underlying sample.

This is where the distinction between district level household survey data interpretation and simple spreadsheet work becomes clear. The software can calculate a percentage in seconds. It cannot decide whether the numerator belongs in the denominator, whether a missing code has been mistaken for “no,” or whether the estimate should account for clustering.

Survey weights are not a technical decoration added at the end; they are part of the meaning of the estimate.

Rebuilding a published indicator is a validation exercise

When raw files are available, I would not begin by creating new indicators. I would first attempt to reproduce a small set of published district figures.

That process can reveal:

• an incorrectly interpreted eligibility condition;
• a mismatch between household and women’s records;
• an overlooked “don’t know” response;
• a missing or misapplied weight;
• an incorrect district code;
• a difference between the report’s reference period and the researcher’s assumption;
• a distinction between household-level and individual-level tabulation.

If the reconstructed value does not match the published value, the response should not be to adjust the result until it “looks right.” The discrepancy is a research finding about the data pathway. It needs to be traced and documented.

A sensible reconstruction sequence is:

1. Select the relevant survey component and eligible population.

2. Apply the questionnaire’s skip logic and indicator definition.

3. Separate valid responses from missing, not applicable, and inconsistent records.

4. Produce the unweighted count and crude percentage.

5. Apply the documented survey weight.

6. Recalculate using the survey’s strata and PSU information.

7. Compare the result with the published district table.

8. Record every difference in definitions, exclusions, and rounding.

Only after this baseline is stable should a researcher proceed to custom subgroup comparisons or multivariable modelling.

Data integrity challenges: records are not automatically error-free

Large surveys carry the strengths and weaknesses of fieldwork at scale. DLHS-3 data-quality research identified age misreporting, response inconsistency, and incomplete records as potential sources of distortion.

Age is especially consequential in reproductive and child-health analysis. Misreported age can shift a respondent into or out of an eligibility group, alter age-specific fertility calculations, and affect comparisons between adolescents, younger adults, and older women. In a setting where age may be reported approximately, copied from another document, or recalled under difficult circumstances, a single variable should not be treated as an unquestionable biological fact.

The same principle applies to parity, date of birth, pregnancy history, and child age. A response may be plausible but internally inconsistent with another part of the questionnaire. That does not mean the respondent has done something wrong. It means the researcher needs a respectful and transparent rule for handling the record.

A practical quality review before analysis

Before estimating district indicators from microdata, I would examine five areas.

1. Completeness

Measure the proportion of missing or nonresponse values for each variable used in the indicator. A low overall missingness rate can still conceal serious gaps in a particular district or subgroup.

2. Logical consistency

Check relationships that should generally make sense within the questionnaire. For example, a respondent cannot be assigned a maternal-care episode that the rest of the record indicates did not occur. The exact checks depend on the documentation and should not be invented from assumptions alone.

3. Age and date plausibility

Review unusual ages, impossible dates, and concentrations of rounded ages. These patterns may not invalidate the survey, but they should be reported when they affect eligibility or age-specific estimates.

4. Geographic coding

Confirm that district identifiers, state identifiers, and names are aligned. District names and boundaries are not stable across all DLHS rounds, Census years, and later health surveys. A geographic merge based only on text labels can silently attach an estimate to the wrong administrative unit.

5. Sparse subgroups

A district may have enough total records for a broad indicator but too few observations for a subgroup such as adolescents, unmarried women, or a specific birth-order category. A precise-looking percentage from a very small unweighted cell is not reassuring simply because the software produced it.

In community clinics, we are trained to ask whether a finding is clinically meaningful, not merely whether it is measurable. Population analysis needs the same discipline. A subgroup estimate can be technically calculable and still too fragile to carry the weight of a policy conclusion.

Comparing DLHS rounds and other surveys without creating false trends

DLHS-3 sits within a sequence that includes DLHS-1 in 1998–99, DLHS-2 in 2002–04, DLHS-3 in 2007–08, and DLHS-4 in 2012–13. That sequence invites trend analysis, but the dates alone do not make the indicators comparable.

Before comparing values across rounds, I would review:

• whether the indicator definition changed;
• whether the questionnaire wording or skip pattern changed;
• whether the target population and denominator remained the same;
• whether the sampling frame and design were comparable;
• whether district boundaries or names changed;
• whether the estimate came from a household or facility component;
• whether the reporting convention and missing-value treatment changed;
• whether the later survey provides the same underlying measure rather than a similarly named one.

The same caution applies when placing DLHS-3 beside NFHS, Census, Annual Health Survey, or administrative health-system data. A difference may represent a genuine population change, but it may also reflect a different reference period, eligibility rule, sampling design, or mode of data collection.

District boundaries are an analytical variable

A district label looks like a stable geographic fact. In longitudinal population analysis, it is closer to a variable that requires documentation.

Administrative reorganisation can split a district, rename it, or change which population is included. If a DLHS-3 district is compared with a later district under the same or a similar name, the researcher needs to establish geographic comparability rather than relying on the label.

This is particularly important for fertility-rate analysis and demographic trend work. A change in the population base can appear as a change in fertility, maternal-care use, or child-health coverage. When the boundary has shifted, the trend may be partly administrative rather than demographic.

For a clean comparison, retain the original survey district code and name, create a documented crosswalk to any later geography, and state clearly where an exact match is not possible. Sometimes the correct conclusion is not a district trend but a comparison of available administrative units with a limitation attached.

Choosing the right format for the question in front of you

The choice between raw data and aggregated reports should follow the decision the analysis needs to support.

If a state team wants to identify districts with lower published coverage of institutional delivery, a standardized fact-sheet comparison may be efficient. If a researcher wants to ask whether the pattern differs by rural residence and household socioeconomic characteristics, raw data are much more appropriate.

A useful decision path looks like this:

• Use aggregated reports when the task is descriptive benchmarking using indicators already defined and published.
• Use raw data when the question requires new subgroups, alternative denominators, record-level quality checks, or multivariable analysis.
• Use both when the aim is rigorous reconstruction: start with published indicators, reproduce them from the microdata, and then extend the analysis.
• Pause before combining datasets when one source is a facility survey and the other is a household survey, or when district boundaries and indicator definitions are not aligned.
• Avoid unsupported precision when documentation is incomplete, the subgroup is small, or the survey design variables cannot be identified.

The comparison is not only about convenience. It is also about accountability. A published report gives the reader a standardized result that can be compared with other published results. A raw-data analysis gives the researcher more freedom, but with that freedom comes responsibility for every analytical decision between the respondent’s answer and the final percentage.

A grounded route for district analysis

For most projects, I would recommend a staged workflow rather than choosing one format and ignoring the other.

Begin with the aggregated district reports. They provide the map: the broad distribution of indicators, the districts that appear to stand out, and the published language used to define the measures. This first pass can prevent a researcher from constructing a technically elaborate analysis that answers a different question from the one decision-makers are asking.

Then move to the raw data if the research question requires more detail. Obtain the questionnaire and documentation alongside the files, identify the relevant unit of analysis, and establish how records are linked. Do not calculate pooled or national estimates from unweighted records unless the documentation confirms the correct weight and design variables.

Next, reproduce a limited set of published results. This is the point at which many hidden assumptions become visible. It is also the best protection against producing a beautifully formatted table that cannot be explained.

Finally, build the custom analysis: subgroup estimates, district comparisons, regression models, or geographic assessments. Report the weighted estimate, the unweighted sample information, the handling of missing records, the survey-design approach, and the limits of geographic or cross-survey comparability.

For practitioners, this discipline has a human purpose. When a district is described as having low maternal-health coverage, the label can influence programme funding, outreach priorities, and the way communities are discussed. A poorly specified denominator can make a group appear underserved when the indicator was not designed to measure that group. A boundary mismatch can make one district look better or worse because its population changed. A model that ignores clustering can turn a tentative pattern into an overconfident claim.

The statistical validity of district health indicators is therefore not separate from patient-centred care. It shapes which women are seen as being missed, which services are considered reachable, and where health workers are sent.

The final choice: speed, flexibility, or a defensible combination

Aggregated DLHS reports are not an inferior form of evidence. They are often the most practical and transparent source for published district benchmarking, especially when the indicator definitions are suitable and the analysis does not require new cross-tabulations.

Raw DLHS-3 data are not automatically superior either. They offer flexibility, but they demand a careful reconstruction of the survey’s sampling design, weights, denominators, geographic identifiers, and data-quality limitations. Without that work, microdata can provide an illusion of precision while quietly discarding the safeguards built into the survey.

For district analysis, the strongest route is usually a combination: use reports to understand the established indicator landscape, use microdata to answer questions the reports cannot answer, and use the published values as a validation point rather than a decorative citation.

The practical lesson I would carry into any research or service-planning team is simple: choose the data format according to the decision, then preserve the route from record to result. In reproductive and child health, that route is not an academic detail. It is how a district statistic remains connected to the women, children, households, and care pathways it is meant to describe.