Agricultural Statistics & Data Science · Sample surveys and official agricultural statistics · also Design and Organisation of Surveys, Survey Methodology, Census and Survey Operations
Turning a sampling design into data: pilot surveys, questionnaire and schedule design, enumerator training and supervision, CAPI and digital data capture, total survey error, non-sampling error models and post-enumeration checks, as used in large-scale agricultural and household surveys.
In the exams: AICE-JRF Ag Stat Unit 5 'non-sampling errors', 'design and organization of pilot and large scale surveys' (heavy); UPSC ISS Statistics III (present); national statistical offices (applied)
A sampling design is only the skeleton of a survey. Turning it into reliable data requires clear objectives, a good frame, a tested questionnaire, trained and supervised field staff, controlled data processing and a responsible release policy, and in large surveys the errors that arise in these operations often exceed the sampling error. This subject covers the planning of surveys; pilot surveys and pre-tests; questionnaire and schedule design and recall periods; field organisation, training, supervision and scrutiny; CAPI and paradata; the sources and classification of non-sampling errors; response errors, interviewer effects and Mahalanobis's interpenetrating sub-samples; the total survey error framework; post-enumeration surveys and coverage evaluation; data processing, editing, coding and tabulation; and confidentiality and disclosure control.
It assumes the sampling designs and estimators of sampling-theory-basic and sampling-theory-advanced (including design effects and the Hansen-Hurwitz treatment of non-response), and it refers to crop-estimation-surveys for crop cutting and agricultural statistics systems, to small-area-estimation for model-based estimates from survey and auxiliary data, and to mixed-models-variance-components for variance components such as the intra-interviewer correlation. Those topics are not repeated here.
In the ICAR AICE-JRF/SRF Agricultural Statistics syllabus, non-sampling errors and the design and organisation of pilot and large-scale surveys are part of the sampling unit; the UPSC Indian Statistical Service and state statistical services include survey organisation and non-sampling errors. Questions test definitions and classifications, the interviewer design effect, interpenetrating sub-samples, non-response bias, the dual-system estimator and disclosure-control methods.
Every survey begins with a statement of what is to be estimated, for which population, at what level of detail and with what precision, and by when. A vague objective such as 'study the condition of farmers' cannot be turned into a design; a usable one names the parameters (for example average net income per agricultural household, the proportion of households indebted, the area under each crop), the domains for which they are needed (state, district, farm-size class), the reference period and the precision required (for example a relative standard error of no more than 5% at state level). The objectives drive every later decision: the frame, the sampling design (covered in sampling-theory-basic and sampling-theory-advanced), the questionnaire, the field schedule and the tabulation plan. Writing the tabulation plan, with dummy tables, before the questionnaire is a classical discipline: every item asked should feed a table, and every table should be fed by items that are actually asked.
The target population is the set of units about which conclusions are wanted; the survey population is the set that can actually be reached through the frame and field procedures. Differences between the two, such as excluded remote villages, institutional households or migrant labourers, should be stated explicitly. The units must be defined unambiguously: a holding, a household, a plot, a field, an animal, a market. Agricultural surveys meet particular definitional problems, such as operational versus ownership holdings, mixed and successive cropping on the same field, and common grazing lands. Definitions and concepts should follow national and international standards (for example the FAO World Programme for the Census of Agriculture) so that results can be compared.
A sampling frame is the list or map from which the sample is drawn: a list of villages from the population census, a list of holdings from the agricultural census, an area frame of segments drawn on maps or satellite images, or a register of farmers. Frames have four classical defects: under-coverage (units missing), over-coverage (units not in the population, such as closed enterprises), duplication and outdated or wrong auxiliary information. Area frames are complete by construction but expensive to build; list frames are cheaper but age quickly. Multiple-frame designs combine a complete area frame with a list frame of large or special units. Where a frame is old, a listing operation in the sampled first-stage units brings it up to date before second-stage selection, which is standard in the National Sample Survey.
The budget and the precision targets must be reconciled. A simple cost function is C = C0 + c n, where C0 covers fixed costs (design, training, processing) and c is the cost per sampled unit; with a total budget of Rs 5,00,000, fixed costs of Rs 1,00,000 and Rs 800 per unit, the affordable sample is (500000 - 100000)/800 = 500. More realistic cost functions separate the cost per first-stage unit from the cost per household and lead to the optimal number of households per village derived in the multistage chapters of the sampling subjects. A sample size computed for simple random sampling, n0, must be inflated for the design effect and for expected non-response: n = n0 x deff / (response rate). For n0 = 400, a design effect of 1.5 and an expected response rate of 80%, n = 400 x 1.5 / 0.8 = 750. Time is also a resource: a survey whose results arrive after the decision they were meant to inform has failed regardless of its precision.
A survey plan normally records the objectives, concepts and definitions, the frame and sampling design, the sample size and allocation with the precision expected, the schedules, the field organisation and calendar, the processing and tabulation plan, the budget and the release policy. A timeline that works backwards from the required release date shows whether the plan is feasible: for an agricultural survey, enumeration must fit the crop season, processing must follow fieldwork, and training must precede both. Revisions to any of these during the survey should be documented, because later users need to know what was actually done rather than what was planned.
A pilot survey is a small-scale rehearsal of the main survey. Mahalanobis in the Indian Statistical Institute's early crop surveys in Bengal stressed pilot surveys as the way to obtain the variances and costs needed for design, and to test procedures before committing large resources. A pilot provides estimates of the variability of key variables (and therefore of required sample sizes), estimates of time and cost per unit, response rates, the adequacy of the frame, and evidence on how the questionnaire and field procedures work in practice. It is not a substitute for the main survey and its estimates are rarely published.
Pre-testing refers more narrowly to testing the instrument. Methods include expert review, cognitive interviews (respondents think aloud as they answer, revealing misunderstanding of terms and reference periods), behaviour coding of recorded interviews (counting how often interviewers change wording or respondents ask for clarification), and field pre-tests in conditions like the main survey. Typical findings are ambiguous terms (for example 'land owned' versus 'land operated'), local units of measurement that need conversion factors, questions that are sensitive, skip patterns that fail, and an interview that is too long. A split-ballot experiment, in which random halves of the pre-test sample receive alternative wordings, can compare versions quantitatively.
Before a census or a very large survey, a dress rehearsal runs the full chain on a small scale: listing, training, enumeration, supervision, transmission, processing and tabulation. It reveals bottlenecks in logistics and data flow that a questionnaire pre-test cannot. The results of pilots and rehearsals should be documented and the decisions they led to recorded, so that later users understand why the design took its final form.
Suppose a pilot in 20 villages finds that the coefficient of variation of paddy yield per hectare between households is about 40% and that an investigator completes about four household schedules a day, with half a day lost travelling to each village. If the main survey must estimate state average yield with a relative standard error of 2% under simple random sampling, the required sample is (0.40/0.02) squared = 400 households before allowing for clustering and non-response. The pilot's time records then convert this into investigator-days and a field budget. Pilot estimates of variance are themselves imprecise, so the main survey sample is usually rounded upwards to provide a margin.
In Indian survey practice a schedule is a form filled by the enumerator through interview or observation, while a questionnaire may be completed by the respondent. Large agricultural and household surveys use schedules canvassed by trained investigators, which allows complex concepts to be applied consistently. A schedule is organised in blocks: identification of the sample unit, particulars of field operations, household characteristics, and then subject blocks (land, crops, livestock, expenditure, income), with remarks blocks for the investigator and supervisor.
Good questions are short, specific and in the respondent's language; they ask one thing at a time (avoiding double-barrelled questions such as 'Do you use fertiliser and pesticide?'), avoid leading or loaded wording, define the reference period and the unit, and offer exhaustive and mutually exclusive response categories. Closed questions are easier to code; open questions capture unexpected answers but are costly to process. Question order matters: early questions set context for later ones, and sensitive questions (income, debt, illegal activities) are better placed after rapport has been built. Filter and skip instructions direct the interviewer through relevant sections and must be tested carefully, especially on paper. Translation into regional languages should use translation and back-translation by independent translators.
Retrospective questions rely on memory, and two opposite errors arise. Recall loss (omission) increases with the length of the recall period, especially for small, frequent or unremarkable events such as minor purchases. Telescoping is the misdating of events: forward telescoping brings events that happened before the reference period into it, inflating reports, and is common for salient events. Short recall periods reduce omission but raise sampling variance, because fewer events fall in the window, and increase the share of telescoped events at the boundaries. Bounded recall, in which a previous interview marks the start of the reference period, reduces telescoping (Neter and Waksberg). The choice of reference period is therefore item-specific: in Indian consumer expenditure surveys the NSS has used uniform, mixed and modified mixed reference periods, with shorter periods for frequently bought food items and a year for infrequent durable purchases, and the choice changes the estimates. In agricultural surveys the crop season is the natural reference period for production and input use.
Farmers' reports of area and production are subject to rounding, local units and deliberate understatement. Objective measurement (area by GPS or chain survey, yield by crop cutting) is preferred where affordable; the crop estimation surveys subject treats crop cutting in detail. Where reports must be used, conversion factors for local units and probes for each plot reduce error. Diaries kept by respondents can reduce recall error for frequent transactions but suffer from fatigue and declining completeness over the diary period.
Questions on income, debt, land disputes, use of banned pesticides or receipt of subsidies may be answered untruthfully. Remedies include assurance of confidentiality, self-completion of sensitive sections, indirect questions, and randomised response designs such as Warner's, treated in sampling-theory-advanced, which protect the respondent by making individual answers uninformative while allowing estimation of the population proportion. Wording that normalises the behaviour ('many farmers find it necessary to borrow from moneylenders') can reduce under-reporting but must not lead the respondent.
On paper, the layout affects error: clear numbering, boxes for codes, units printed next to each entry, and instructions placed where they are needed reduce mistakes. Columns for plots or household members allow the same questions to be asked for each with little space. The same principles apply on screens in CAPI, where one question per screen or grouped rosters can be used.
Large surveys are organised in a hierarchy: headquarters (design, instructions, processing), regional offices, supervisors and field investigators. Each level has defined responsibilities and work loads. The field schedule fixes when each first-stage unit is to be visited, often with the sample divided into sub-rounds so that seasonal variation is spread evenly. In agricultural surveys, the timing of field visits must follow crop calendars, for example crop cutting at harvest and area enumeration after sowing.
An instruction manual defines every concept, explains every item of the schedule with examples, and specifies procedures for listing, selection and contacting units. Training combines lectures, demonstrations, mock interviews, field practice and tests, and is repeated or refreshed for each round. Training of trainers ensures consistency when many regional sessions are held. Poor training is a principal source of non-sampling error because concepts such as 'operational holding' or 'principal activity status' are applied inconsistently by different investigators.
Supervisors accompany investigators on a proportion of interviews, re-interview a sub-sample of households, check listing and selection procedures, and inspect crop-cutting plots. Scrutiny is the manual and computer checking of completed schedules for completeness, consistency (for example area under crops not exceeding operated area) and plausibility, with unresolved items referred back to the field. Re-interview checks provide data on response error, discussed in chapter 7. The independence of supervision from the investigator's own reporting line encourages honest reporting of problems. Records of supervision visits and of corrections made during scrutiny are themselves useful quality indicators.
Progress reports track the number of units completed, response rates, and time per unit by investigator and region. Unusual patterns, such as an investigator with exceptionally short interviews or a very low rate of household members reported, may indicate shortcuts or fabrication and lead to re-interviews. With electronic data collection these indicators are available daily, as described in the next chapter.
The number of investigators follows from the total work load and the duration of fieldwork. If 6,000 households must be interviewed in 60 working days and an investigator completes about five households a day, about 20 investigators are needed, before allowing for travel, leave and re-visits. Fewer investigators with larger workloads are cheaper to train and supervise, but they increase the interviewer design effect discussed in chapter 7 and lengthen the field period, which matters for seasonal items. Balancing these considerations is part of the total survey error trade-off.
Agricultural surveys combine interviews with physical measurement. Area enumeration may involve plot-by-plot recording of crops by the village accountant, as in the land-record-based systems described in the crop estimation surveys subject, or field measurement by GPS. Crop cutting requires the location of a randomly selected plot within a field, harvesting a marked area and weighing the produce, with supervisors attending a share of the experiments. Each of these operations needs instructions, equipment and checks specific to it, and timing tied to the crop calendar.
Computer-assisted personal interviewing (CAPI) replaces the paper schedule with a program on a tablet or phone. Its advantages are automatic routing through skip patterns, range and consistency checks during the interview, no separate data entry, GPS coordinates and timestamps, photographs, and rapid transmission to headquarters. Disadvantages include the cost of devices and programming, the need for power and connectivity (handled by offline operation and later synchronisation), the risk that a programming error affects every interview, and a reduced overview of the whole questionnaire for the interviewer. Common platforms include Survey Solutions (World Bank), CSPro and ODK-based tools; Indian national surveys have increasingly moved to tablet-based collection.
A CAPI instrument is software and must be specified, tested and version-controlled. Hard checks prevent impossible entries; soft checks warn the interviewer but allow an override with a comment, because genuine extreme values exist and forcing the interviewer to change them introduces error. Pre-loading information from a listing or a previous round, called dependent interviewing, reduces burden and spurious change but can encourage confirmation of wrong values. CAPI also allows randomisation of question order and of experimental versions.
Paradata are data about the data collection process: timestamps of each question, interview duration, number of contact attempts, keystroke records of changed answers, GPS location of the interview, and interviewer observations. The term was introduced by Couper. Paradata are used to monitor fieldwork (for example detecting very short interviews or interviews recorded far from the sampled village), to model non-response and adjust weights, to identify problematic questions with long response times or many corrections, and to manage responsive designs, in which fieldwork effort is redirected during collection on the basis of indicators. They are also a sensitive source because they reveal locations and behaviour, so access to them should be controlled.
Computer-assisted telephone interviewing (CATI) and web self-completion are cheaper but have coverage problems in rural populations and mode effects on responses. Mixed-mode designs combine them, at the cost of having to assess differences between modes. Remote sensing and administrative data increasingly supplement field collection, as discussed in the crop estimation and small area estimation subjects.
With data arriving daily, headquarters can run automated checks on each batch: completeness of rosters, distributions of key variables by interviewer, interview duration, the proportion of 'don't know' answers, and the distance between recorded GPS points and the sampled village. Interviewers whose indicators differ markedly from the others receive feedback, retraining or re-interviews of their households. Audio recording of selected questions, with consent, is another check on interviewer behaviour. These controls are effective only if they lead quickly to action in the field.
Sampling error arises because only a sample is observed; it decreases as the sample grows and is estimated from the sample by the methods of the sampling subjects. Non-sampling errors arise in every stage of a survey or census, including complete enumerations, and do not necessarily decrease with sample size; indeed, a larger survey with less well trained staff may have larger non-sampling errors. In large surveys non-sampling errors are often larger than sampling errors, which is why their control is part of the AICE syllabus on the design and organisation of large-scale surveys.
A common classification distinguishes: (1) specification errors, when the concept measured differs from the concept intended; (2) coverage or frame errors, under- and over-coverage and duplication; (3) non-response errors, unit non-response (no interview) and item non-response (missing items); (4) measurement or response errors, from the respondent, the interviewer, the instrument or the mode; and (5) processing errors, in editing, coding, data entry, imputation and tabulation. Errors can be variable (random, averaging out over repetitions but adding variance) or systematic (bias, the same direction on average). Sukhatme and Murthy's textbooks treat these errors in the context of Indian agricultural surveys.
Unit non-response arises from refusals, non-contact and inability to respond. The bias of the respondent mean is approximately (1 - RR)(Y-bar_r - Y-bar_nr), where RR is the response rate and Y-bar_r and Y-bar_nr are the means of respondents and non-respondents. With a response rate of 80%, a respondent mean of 50 and a non-respondent mean of 40, the bias is 0.2 x 10 = 2. A high response rate limits the bias but does not guarantee it is small if non-respondents differ greatly. Remedies are prevention (call-backs, advance letters, trained interviewers), sub-sampling of non-respondents (Hansen and Hurwitz, treated in sampling-theory-advanced), weighting adjustments within classes, calibration to known totals and imputation for missing items.
Because no single number captures all errors, statistical offices publish quality reports covering relevance, accuracy (sampling and non-sampling), timeliness, accessibility, coherence and comparability. Response rates, coverage rates, edit failure rates and re-interview results are the usual accuracy indicators.
Agricultural frames have particular coverage problems: holdings operated by people living outside the village, tenancy arrangements that are not recorded, newly brought-under-cultivation land, and livestock kept by landless households. Over-coverage arises when holdings that have been sold or merged remain on a list. A listing operation in the sampled villages, with clear rules for which holdings belong to which village, reduces both, and comparisons with census totals reveal remaining gaps.
Item non-response is concentrated in sensitive or difficult items such as income, outstanding loans and quantities of produce sold. Reporting the item response rate for each important variable, and examining whether missingness depends on observed characteristics, informs the choice of imputation method. Where the probability of missingness depends on the value itself (for example the richest households refusing income questions), no method based only on observed variables fully removes the bias.
Hansen, Hurwitz and Bershad formulated a model in which the observed response of unit i in a particular survey trial is y_it = Y_i + b + e_it, where Y_i is the true value, b a constant bias and e_it a response deviation. The total variance of the survey mean then has components: sampling variance, simple response variance (from uncorrelated response deviations) and correlated response variance (from response deviations correlated within an interviewer's workload). The simple response variance falls with sample size like sampling variance; the correlated component does not fall in the same way when each interviewer takes more interviews.
If each interviewer completes m interviews and responses in the same workload have intra-interviewer correlation rho_int, the variance of a mean is inflated by the interviewer design effect 1 + (m - 1) rho_int, analogous to the cluster design effect. With m = 50 interviews per interviewer and rho_int = 0.02, the inflation is 1 + 49 x 0.02 = 1.98, so the effective sample size is roughly halved. With m = 20 the factor is 1.38. Even small correlations matter when workloads are large, which argues for more interviewers with smaller workloads, better training and standardised interviewing. Interviewer correlations are typically larger for sensitive, attitudinal or complex items than for simple facts.
Mahalanobis introduced interpenetrating sub-samples in the Bengal crop surveys of the 1930s and 1940s. The sample is divided into two or more independent sub-samples, each drawn by the same design and each enumerated by a different set of investigators. Each sub-sample yields its own estimate, so the agreement between them provides a simple, design-based estimate of the total variance including variable non-sampling errors, and a large disagreement signals problems in a particular team. For two sub-sample estimates t1 = 100 and t2 = 110, the combined estimate is 105 and its standard error is |t1 - t2|/2 = 5. With k sub-samples, the variance of the mean of the estimates is the variance among the k estimates divided by k. The National Sample Survey has used two independent sub-samples, often canvassed by different agencies, in many rounds. Interpenetration estimates total variance but does not detect a bias common to all investigators.
A re-interview repeats some questions with a sub-sample, using a better interviewer or method. If the re-interview is meant to be a replicate of the original procedure, the gross difference rate g (the proportion of cases whose answers differ) estimates twice the simple response variance for a binary item, and the index of inconsistency I = g / (p1 q2 + p2 q1) expresses response variance relative to total variance, where p1 and p2 are the proportions with the attribute in the two interviews and q = 1 - p. For example, with p1 = 0.30, p2 = 0.32 and g = 0.10, I = 0.10 / (0.30 x 0.68 + 0.32 x 0.70) = 0.10/0.428 = 0.234. If instead the re-interview is a reconciled, best-possible measurement, the net difference between the two estimates estimates the response bias.
The intra-interviewer correlation rho_int is a variance component: the ratio of the between-interviewer variance to the total variance. It can be estimated by fitting a random-effects model with interviewer as a random factor, as described in the mixed-models subject, provided interviewers are assigned randomly to sample units, which is exactly what interpenetration ensures. Without random assignment, interviewer effects are confounded with area effects because interviewers usually work in particular areas.
The interviewer design effect reduces the effective sample size: 1,000 interviews with an inflation factor of 1.98 carry about the information of 505 independent interviews. When the interviewer effect is combined with the cluster design effect, the two factors multiply approximately if interviewers work across clusters, so the total inflation can be large even when each correlation is small.
Total survey error (TSE) is the accumulation of all errors that make a survey statistic differ from the true population value. The framework, developed by Groves and colleagues, links survey design decisions to errors on two sides: representation (coverage error, sampling error, non-response error and adjustment error, which together concern which units are measured) and measurement (validity, measurement error and processing error, which concern how well each unit is measured). Its central quantity is the mean squared error of an estimate, MSE = variance + bias squared, with both terms including non-sampling components.
Because resources are limited, the framework emphasises trade-offs between error sources and costs. Spending on a larger sample reduces sampling error but may leave less for training or call-backs; cutting the questionnaire may reduce measurement error but lose content. Groves's Survey Errors and Survey Costs (1989) argued that design should minimise total error for a fixed cost, not sampling error alone. A useful corollary is that once bias dominates the MSE, increasing the sample yields little improvement in accuracy: with bias 2 and sampling variance 4 at n = 400, quadrupling n reduces the MSE only from 8 to 5.
The TSE paradigm has been extended to total survey quality, which adds relevance, timeliness, accessibility, interpretability and coherence to accuracy, as in the quality frameworks of national statistical offices and international organisations. In India the National Statistical Office and the National Statistical Commission have issued guidance on quality and on releasing survey results with metadata on design and errors.
Consider a survey estimating average fertiliser use per hectare. Sampling error can be computed from the design. Coverage error arises if the frame omits recently created holdings, which may use more fertiliser. Non-response error arises if large commercial farmers refuse more often. Measurement error arises if farmers report in bags of different sizes that are converted wrongly, or recall poorly over a season. Processing error arises if a unit code is mis-keyed. TSE thinking asks which of these is largest relative to its cost of reduction, for example whether a conversion-factor study would improve accuracy more than an increase in sample size.
A post-enumeration survey (PES) is an independent re-enumeration of a sample of areas soon after a census or a large survey, carried out by separate and usually more experienced staff, to measure coverage error (omission and duplication) and content error (wrong answers to questions). The Census of India and agricultural censuses have used post-enumeration checks, and many countries conduct a PES after each population census.
Coverage is estimated by matching the census and PES records in the sample areas. Treating the census and the PES as two independent captures, the dual-system (capture-recapture or Chandrasekar-Deming) estimator of the population is N-hat = (n1 n2)/m, where n1 is the census count, n2 the PES count and m the number matched in both. With n1 = 900, n2 = 800 and m = 720, N-hat = 900 x 800 / 720 = 1000, so census coverage is 900/1000 = 90% and the net undercount is 10%. The estimator assumes independence of the two enumerations, homogeneous capture probabilities within post-strata, accurate matching and a closed population; violations (for example people missed by both because they are hard to reach) bias it, typically towards underestimating the undercount. Post-stratification by area type, age and sex reduces heterogeneity bias.
Content error is studied by comparing answers in the census and the PES for matched units, using gross difference rates and the index of inconsistency described in chapter 7. In agricultural censuses, re-measurement of holdings and areas in a sample of villages estimates errors in reported operated area and land use. Results feed back into the design of the next census through better maps, training and questions.
Matching census and PES records is laborious and error-prone. False non-matches (records of the same person not recognised as such) inflate the estimated undercount; false matches deflate it. Clerical review of doubtful cases, clear matching rules based on name, age, sex and relationship, and a search area wide enough to find people recorded in a neighbouring block are standard. The PES must also determine whether census records that cannot be matched are erroneous enumerations, such as duplicates or fictitious records, since these offset omissions in the net coverage figure.
The PES should follow the census closely, usually within weeks, so that the population has changed little through births, deaths and migration, but it must be operationally independent: different staff, separate listing, and no access to census records during enumeration. If PES enumerators know which households the census found, they may concentrate on them, and the independence assumption of the dual-system estimator fails. Results are usually published as net omission rates by state, rural and urban areas, age and sex.
Paper schedules are captured by keyboard entry, often with double entry by two operators and comparison of the two files to catch keying errors, or by scanning. Coding converts open answers (occupation, industry, crop names, items of consumption) into classification codes such as the National Industrial Classification or the National Classification of Occupations; automated coding with dictionaries handles common answers, and coders deal with the remainder. Coding quality is checked by independent recoding of a sample, giving coder error rates.
Editing detects and treats invalid, inconsistent and implausible values. Validity edits check ranges and codes; consistency edits check relationships (crop areas summing to net sown area, age of a child below that of the mother); statistical edits flag outliers relative to the distribution. Fellegi and Holt (1976) formalised automatic editing: given a set of edit rules, the minimum number of fields should be changed to make a record pass all edits, and imputation should preserve the joint distribution as far as possible. Over-editing, spending large resources on changes that have negligible effect on estimates, is a common problem; selective or significance editing concentrates review on records whose correction would change the published estimates most.
Item non-response and values rejected by editing are filled by imputation: deductive imputation from other answers, mean or ratio imputation within classes, hot-deck imputation from a similar donor record, regression imputation and multiple imputation. Single imputation understates variance if imputed values are treated as observed, so variance estimation must account for it. Imputation flags should be kept in the file so that users can identify imputed values.
Weights are computed from the design (inverse selection probabilities), adjusted for non-response and calibrated to known totals; then estimates and their standard errors are produced by the methods of the sampling subjects, often by replication. Tables follow the tabulation plan, report the number of sample units behind each estimate and its relative standard error, and suppress or flag cells with too few observations. Tables should be checked against previous rounds and external sources before release, and unexpected changes explained.
Processing should be reproducible: edit rules, imputation procedures, weighting steps and tabulation programs should be written as documented scripts, version-controlled, and run on the raw captured data to produce the published tables. This allows errors found after release to be corrected and the corrections to be traced. Releasing the questionnaire, instructions and processing documentation with the data enables users to understand derived variables and their limitations.
Respondents give information on the promise that it will be used only for statistical purposes and that they will not be identified. In India the Collection of Statistics Act, 2008 provides for the collection of statistics and restricts the use and disclosure of information collected under it, and the Fundamental Principles of Official Statistics of the United Nations make confidentiality a core principle. Breaches destroy trust and future response rates. Statistical disclosure control (SDC) is the set of methods that limit the risk of identifying respondents or learning their attributes from released tables or microdata while keeping the data useful.
A table cell is sensitive if it allows a respondent's value to be estimated closely. Threshold rules flag cells with fewer than a minimum number of contributors (for example 3). Dominance rules flag cells where a few contributors account for most of the total: under the (n, k) rule, a cell is sensitive if its n largest contributors account for more than k% of the total. Sensitive cells are suppressed (primary suppression), and further cells are suppressed (secondary or complementary suppression) so that the primary cells cannot be recovered from row and column totals. Alternatives include collapsing categories, rounding and adding controlled noise to cell values.
Microdata (record-level files) carry a higher risk because combinations of quasi-identifiers such as district, age, sex, caste, occupation and land size can single out a person. Direct identifiers (names, addresses, phone numbers, exact GPS coordinates) are removed. Methods that reduce the risk from quasi-identifiers include global recoding (coarser categories), top-coding of extreme values such as large holdings or incomes, local suppression of rare values, data swapping between records, adding random noise, and microaggregation (replacing values by averages of small groups). k-anonymity requires that every combination of quasi-identifier values in the released file be shared by at least k records. Each method reduces risk at some cost in analytical value, which should be measured. Differential privacy offers a formal guarantee by adding calibrated noise to outputs and has been adopted for some census products elsewhere, though it remains debated because of its effect on small-area accuracy.
Statistical offices balance access and risk through tiers: public-use files with strong protection, licensed or scientific-use files with less protection released to researchers who sign agreements, and secure data enclaves or remote access where approved researchers work with detailed data under supervision and outputs are checked before release. The National Statistical Office releases unit-level data from NSS rounds and other surveys with documentation for researchers. Good metadata (questionnaires, instructions, sampling design variables and weights, and a description of the disclosure controls applied) are essential for correct analysis.
Disclosure control is a trade-off. Measures of disclosure risk, such as the number of sample-unique records on combinations of quasi-identifiers, or the probability of re-identification estimated by matching against external files, are compared with measures of information loss, such as changes in means, variances and correlations or in the results of standard analyses. Agricultural microdata pose particular risks because large holdings, unusual crops or rare livestock types in a named district may identify a farmer. Geographical detail is often the most identifying variable, so files may give district but not village codes, or coarsen GPS coordinates by displacement.
A PhD student who collects primary data for a thesis must write objectives and a tabulation plan, design and pre-test a schedule, train enumerators and check data, and report non-response and data-quality measures alongside the sampling error.
A five-year statistician in a statistical office, an ICAR institute or a development project designs and pre-tests schedules, programs CAPI instruments, trains and supervises field staff, monitors paradata, edits and imputes data, and prepares quality reports and anonymised files.
A senior official responsible for national surveys sets standards for concepts, training and quality, decides trade-offs between sample size and field quality under fixed budgets, commissions post-enumeration checks, and approves microdata release policies under the Collection of Statistics Act.
A PhD with long experience researches measurement error, interviewer effects, responsive and adaptive designs, mixed-mode collection and disclosure control, and advises statistical systems on total survey error and modernisation.
Specific NSS and census procedures change between rounds and are described in general terms only; readers should consult the instructions for the round they use. Section numbers of the Collection of Statistics Act are not quoted. Response-error models are stated without full derivation. Disclosure-risk measures for microdata and the mathematics of differential privacy are outlined only. Software for CAPI and disclosure control is named but not taught.
Original practice questions written to the syllabus and pattern of the exams named in each question's examTags. They are not past papers and do not reproduce any board's questions. Every numerical answer was computed with a script; the values needed are stated in the question.
These notes are free to read here, with no account.
What the Laisha app adds:
Free for ten days, then a subscription. We would rather write that here than have you find it later.