ICEQC-R-2006-06 — Household Disadvantage and Participation in Education: A Disaggregation Protocol

The public question is whether children in different circumstances have equal practical opportunity to enter, attend and complete education of acceptable quality. A statistical difference is one part of that answer.

The inquiry should identify which institution can affect the barrier. Household circumstance may describe exposure while responsibility lies in fees, distance, admission, teaching, safety or social protection.

A universal entitlement applies to every child; its implementation can differ systematically. Disaggregation tests that distribution without converting group membership into a lesser entitlement.[REF-08]

The overall rate remains relevant for scale. Group results explain distribution and should be published beside, not instead of, the national result.

Equality may refer to the same formal rule, while equity concerns whether circumstances beyond the child’s control create unequal opportunity. Statistical analysis should state the operational meaning used.[REF-04]

Neither term is an indicator by itself. The measurable result may be an attendance gap, delayed entry, lower completion or unequal learning exposure.

Disadvantage is relative to a specified outcome, population, place and time. A household may have limited monetary resources but strong local access; another may have resources but face discrimination or an inaccessible school.

The comparison should therefore name the characteristic and result rather than label a household permanently disadvantaged in all respects.

Exclusion is produced through decisions, rules, conditions and conduct. Statistical categories should support inquiry into agency: who charges, locates, admits, teaches, protects, accommodates or withholds information.[REF-03]

Attributing the result solely to household choice can conceal public and institutional responsibility.

Household circumstances and school participation change. Income loss, illness, migration, work, conflict or school closure can alter a child’s pathway after a survey interview.

Cross-sectional status should not be represented as permanent. Repeated data or retrospective histories are needed to examine movement.

Economic, geographic, social, cultural, health and protection conditions can reinforce each other. An asset index can describe one dimension but cannot serve as a complete exclusion index without justified construction.[REF-03]

Reporting several dimensions separately often preserves more policy meaning than one composite score.

Household variables belong to a shared unit; participation belongs to the child. The analysis should not assume that all children in one household have the same education status or resource access.

Child-level sex, age, relationship, work and disability may modify the association between household circumstances and participation.

A rural–urban participation difference can reflect school access, household resources, age structure, migration or survey coverage. These are composition and context questions.

Adjustment may describe how the gap changes under measured characteristics, but it does not make one context equivalent to another.

Entry, attendance, progression, completion and learning form related but separate outcomes. Household disadvantage may be strongest at one stage and weak at another.

A protocol should identify the earliest divergence and should not use current enrolment as a complete education result.

A complete claim states: among a defined population, under a stated education measure and survey period, group A has a result different from group B by a stated amount, with stated uncertainty and coverage.

Explanation and policy follow in separate sentences with their own evidence.

The protocol does not permit “the poor do not value education”, “rural families choose non-attendance” or similar causal descriptions from a cross-tabulation. It does not use household-head sex as a proxy for women’s poverty without analysis.

It does not treat a non-significant sample difference as proof of equality.

A group difference may be material because of size, severity, persistence, age, complete exclusion or compounding barriers. Statistical significance is relevant but not the only public-interest test.

An imprecise estimate concerning a severe, poorly observed population may justify better evidence and immediate protective inquiry without supporting an exact prevalence claim.

The target population is the full group about which the study intends to conclude, such as all children of primary-school age usually resident in a country. It should be defined before the sample frame is examined.

The achieved survey population may be narrower. The difference is part of the result.

The sampling frame lists or organises units from which the sample is selected. Census enumeration areas, household listings and institutional records have different coverage and age.

Frame omissions cannot be corrected by survey weights unless the omitted population is represented elsewhere under defensible assumptions.

A conventional household is commonly defined through shared residence, food or living arrangements, but national definitions vary. The rule determines which adults, children and resources are analysed together.

The report should reproduce the survey definition rather than assume a universal family unit.

Usual-resident rules determine whether temporarily absent students, seasonal workers, boarders and mobile children belong to the household roster. A de facto rule based on presence at interview produces a different population.

Education variables should follow the same membership rule or explain exceptions.

The reported head may reflect age, income, authority, ownership, gender convention or respondent choice. “Female-headed household” therefore has no uniform economic or decision-making meaning.

Analysis should examine household composition, resources and education directly where possible and avoid a deficit assumption based on headship alone.

Relationship codes identify children, grandchildren, foster children, relatives, non-relatives and domestic workers only as far as the questionnaire permits. Misclassification can affect parental education and orphanhood variables.

The biological, legal and caregiving relationship should not be inferred beyond the recorded question.

Household members may live across locations for work, schooling or migration and share resources. A single residence-based roster may assign income and children to different units.

Survey analysis should describe the treatment of transfers and absent contributors. Household wealth measured only at the child’s current residence may not represent available support fully.

Children at boarding schools may be counted at the institution, at their family household or excluded, depending on frame and residence rules. Their participation is likely high by construction, but their household characteristics may be missing.

The protocol should identify their treatment and avoid assigning institutional assets to household welfare.

Children in residential care, detention, health facilities or other institutions are often outside household surveys. Their education opportunities may differ substantially.

The national disaggregation should state the omission and use administrative or specialised evidence.[REF-03]

Children without conventional residence can be absent from both household frames and school registers. Their exclusion from the estimate should not be mistaken for a small or zero population.

Specialised enumeration and service records can support a bounded description, with protection and ethical safeguards.

Fixed frames may miss mobile populations or reach them only seasonally. Residence, school attendance and household composition can change across the reference period.

The survey design may require adapted listing, timing or supplementary samples. Standard weights should not conceal systematic non-coverage.

Conflict or disaster can make frames and population projections obsolete. Camps, host households and dispersed populations require distinct coverage assessment.

Area exclusion for security should be prominent. A national point estimate based only on accessible areas is not national coverage.

High survey cost can lead to exclusion of remote areas. Because remoteness may itself affect school access, the omission is directly relevant to the education result.

The sample report should quantify the excluded population where possible and describe likely direction without inventing an adjustment.

Unit non-response occurs when a selected household is not interviewed; item non-response affects particular variables. Substitution of an easier household can bias results and should follow only an authorised design, if permitted at all.

Response rates should be reported overall and by strata or area relevant to participation.

An adult respondent may report for a child away at school, work or another household. Proxy response can affect attendance, grade and work information.

The questionnaire should record respondent relationship, and analysis should test whether proxy status changes missingness or results.

Education participation should be analysed by age in relation to the national education structure. Broad age groups can include children with different entitlements and expected grades.

Age heaping, missing birth dates and interview timing require validation before fine disaggregation.

Before any wealth, residence or household comparison, the report should state who is represented and who is not. This statement belongs in the main findings when omitted populations are likely to have different participation.

A weighted estimate is representative only of the population and response process supported by the design.

Economic position may be used to describe unequal participation, identify populations for further inquiry or evaluate the reach of a policy. The purpose determines whether current resources, longer-term living standards or eligibility under a national rule is required.

One measure should not be selected merely because it is present in the data set.

Income records money and in-kind receipts during a stated period. It can reflect immediate command over resources but varies seasonally, is difficult to capture in informal activity and may be underreported.

Household size, composition, prices and production for own use affect comparison. Gross income and disposable income should not be mixed.

Consumption or expenditure can provide a more stable account of material welfare where income fluctuates. It requires detailed recall periods, valuation of home production, treatment of durable goods and price adjustment.

The measure’s apparent precision depends on extensive questionnaire and processing decisions. Education analysis should use the official welfare aggregate and document its basis rather than reconstruct it partially from selected items.

Household spending on fees, books, uniforms, transport, meals, tutoring and other charges can reveal the private cost of participation. Zero spending can indicate free access, non-attendance or omitted expenditure.

Expenditure among enrolled children should not be used to infer affordability for children already excluded by cost.

Asset ownership and housing characteristics can proxy longer-term economic position when income or consumption is unavailable. Electricity, water, sanitation, flooring, vehicles, communication devices and durable goods may enter an index.

The meaning and prevalence of each asset differ by place. An item reflecting public infrastructure may classify geography as much as household resources.

An asset index assigns weights to observed items and produces a relative score. The method, item coding, treatment of missing values and population used to derive weights should be recorded.

The score has no natural monetary unit. Small numerical distances should not be interpreted as equal differences in welfare.

Items should vary sufficiently, be measured consistently and represent economic position rather than the education outcome itself. School attendance, books required by school or programme benefits should not be included where the index will explain participation.

Items owned by very few or almost all households may contribute little to ranking but can remain substantively important.

Piped water or electricity may depend on settlement infrastructure, while livestock or agricultural equipment may be relevant in rural livelihoods. One national index can rank rural households low because of location-specific asset patterns.

The analyst should examine item distributions and compare a national ranking with context-specific indices without selecting the version that minimises the education gap.

Weighted households or persons are ordered by the welfare measure and divided into five groups of approximately equal population under the stated unit. Household quintiles and child-population quintiles are not necessarily the same because household size differs.

The protocol should state whether each child inherits the household score and whether cut-points are based on households, persons or children.

The lowest quintile is the lowest fifth within the covered survey population. It does not indicate a common absolute living standard across countries or survey years.

If general living standards rise, the bottom fifth continues to exist. Trend analysis should therefore distinguish relative rank from absolute improvement.

Asset indices can contain many identical scores. Forcing equal-sized quintiles may split households with the same observed assets across groups. The tie rule should be stated.

Sensitivity analysis can compare quintiles, broader groups and the continuous score where its properties permit.

An official absolute poverty classification applies a defined resource aggregate and line. It answers a different question from relative quintiles. Prices, household needs and survey method determine its comparability over time.

The education report should cite the competent statistical definition and should not create an unofficial poverty label from an asset score.

Total household resources do not represent resources per person. Per-capita measures are simple but assume equal needs and no economies of scale. Equivalence scales introduce assumptions about adults, children and shared goods.

The selected adjustment should follow the welfare measure and be reported. Education-specific costs can vary by number and ages of children.

Households with many children, older persons or persons requiring care may face different resource demands. A dependency ratio is a demographic summary, not a direct measure of actual work or support.

Composition variables should be analysed separately from welfare rank to avoid embedding a normative judgement in the index.

A household welfare score is assigned to every member, but food, time, fees and learning materials may not be distributed equally. Differences by sex, birth order, disability or relationship can remain within the same quintile.

Child-level outcomes and expenditure should therefore be examined alongside household position where data permit.

Illness, death, crop loss, employment loss, displacement or price change can reduce participation even where longer-term assets remain. A cross-sectional asset index may classify the household above its current capacity.

Shock variables require a period and should not be interpreted as exogenous without examining their relation to residence and prior poverty.

Asset ownership can coexist with limited cash for school charges. Debt may support attendance or indicate financial stress. Many surveys do not measure liabilities fully.

The absence of debt information should qualify claims about immediate affordability drawn from asset rank.

Highest adult or parental education may reflect social and economic position, ability to navigate institutions, language, expectations and the education available to an earlier generation. It should not be described as a measure of parental concern.

Biological parents may not reside in the household. The selected adult and missing-parent rule should be stated.

Occupation can describe resources, working time, seasonality and social position. Classification should follow a consistent system and distinguish current inactivity, unemployment and work outside the survey reference.

Child participation may be associated with employment patterns through several mechanisms. One category does not establish which applies.

Validation examines item frequencies, missingness, household-size distribution, urban–rural composition, stability under alternative coding and relationship with independent welfare evidence. It does not require perfect correlation with income or consumption because the concepts differ.

Unexpected results should prompt method review before a substantive claim.

Every education table using welfare groups should name the welfare concept, reference period, ranking population, weighting, number of groups and whether the grouping is relative or absolute.

The phrase “poor children” should be reserved for a stated poverty definition; otherwise the measured group should be named.

Current enrolment indicates formal registration in an education programme under the survey question. It can include children who attend irregularly or not at all during the reference period.

The question should identify the school year and provider types. Registration reported by a proxy may differ from school records.

Attendance concerns actual participation during a stated day, week or longer interval. Short periods are sensitive to illness, closure, holidays and seasonality; long recall increases reporting error.

The survey should record whether instruction was available. Absence during official closure is not pupil non-attendance.

Ever attendance distinguishes children who have entered education from those reported never to have attended. It does not describe duration, quality or current status.

Young children below or near entry age require separate treatment because future entry remains expected.

Age at first attendance can identify delayed entry but depends on recall and the definition of school. Early childhood and non-formal provision may be reported inconsistently.

The questionnaire should ask programme and grade with age where feasible. Retrospective age should be grouped when exact recall is weak.

Current grade supports age-for-grade and progression analysis. National programme names require a mapping to level and grade. Repeated, ungraded and multi-grade provision need explicit codes.

The highest grade attended is different from the grade successfully completed.

Highest grade completed should record a grade for which requirements were satisfied. Respondents may report the current or last attended grade instead.

Consistency checks compare attendance status, current grade, highest grade and age without overwriting plausible exceptions.

Completion of primary or basic education depends on national structure and the survey’s wording. A certificate or examination may be required in some systems.

Cross-country analysis should use a documented programme mapping and should not assume equal years of schooling.

Regularity may be measured through days attended, absence threshold or reported usual attendance. The threshold should be justified and should remain distinct from formal enrolment.

Attendance quality is affected by register and respondent accuracy and by days when teaching was actually provided.

Out-of-school measures should state age group and whether children attending another education level count as in school. Household attendance and administrative enrolment approaches can produce different estimates.[REF-02]

The status may include never entered, left, temporary interruption and expected later entry. These pathways require further classification.

Transition identifies movement from one level to another. Household cross-sectional data may infer transition from current and completed grade but cannot always establish timing.

Retrospective questions or linked observations strengthen the measure. Eligibility and available places should be considered.

Self-reported literacy, reported ability to read a simple statement and direct assessment are different. The 2006 monitoring report emphasises the importance and complexity of literacy measurement.[REF-06]

Literacy should not be used as a substitute for school participation or vice versa.

Household surveys may collect reasons for non-attendance, satisfaction, expenditure, travel time or perceptions. These provide valuable user evidence but do not directly measure teaching quality or learning.

Question wording and respondent experience shape the result. School and assessment evidence are complementary.

Reason lists may include cost, distance, work, illness, disability, safety, lack of interest, poor quality, family responsibility or no school. One response can force a complex pathway into a single category.

Multiple responses and an open category may be appropriate, with careful coding. “Not interested” should not be treated as explanation without examining the conditions producing it.

Official school-age groups facilitate system monitoring; single-year ages reveal delayed entry and progression. Broad child age bands improve precision but can mix entitlement and expected status.

Every group comparison should use the same age definition and population denominator.

Age, entry, current attendance, grade and completion should form a plausible sequence. An apparent inconsistency may be a reporting error, acceleration, interruption, programme change or misunderstanding.

Editing rules should flag and document rather than force every record into a conventional path.

Public, private, religious, community and non-formal systems may use different grades or calendars. Survey categories should enable inclusion and mapping without implying equal legal status or quality.

Participation should be reported by provider where sample permits and where the distinction informs access or cost.

Attendance can vary with work, weather, migration and school operations. A survey conducted once may misrepresent annual exposure.

Fieldwork dates and regional seasons should accompany the result. Repeated or calendar questions can improve interpretation.

Binary attendance treats one day and regular participation alike. Where data permit, days or sessions attended relative to those offered provide an intensity measure.

The measure requires accurate calendars and should be interpreted with illness, closure and authorised absence.

Attendance is an opportunity condition, not a learning result. Children with similar attendance may experience different teaching, language, materials and support.

Household disaggregation of learning requires a valid assessment linked to the same sampled population and appropriate consent.

The published title should name enrolment, attendance, entry, grade or completion exactly. It should state age group, period, provider scope and household population.

No table should use “school participation” as an undefined umbrella while comparing groups.

A probability sample gives each covered unit a known, non-zero selection probability under the design. This permits design-based population estimates and measures of sampling error.

Convenience or service samples can describe participants but should not be weighted to a national household population as if selection were probabilistic.

Household surveys commonly select geographic clusters before households. Children within a cluster may share schools, transport and socioeconomic conditions, reducing the independent information in the sample.

Variance estimation should incorporate clustering. Treating every child as independently sampled usually understates uncertainty.

Strata may ensure representation by region, residence or another design domain. Estimates and variance should use the actual stratification variables and design.

Post hoc analytical groups are not necessarily design strata. Confusing the two can produce incorrect standard errors.

Small regions or populations may be selected at higher rates to support reliable estimates. Survey weights restore their population contribution in national results.

Unweighted counts remain important for precision and disclosure. A large weighted population can be based on a small achieved sample.

Within selected clusters, the household listing and selection interval determine coverage. Outdated listings, inaccessible dwellings and informal settlements can create systematic omission.

The field report should identify listing date, additions, deletions and replacement rules.

Some surveys collect education data for every child; others select one person or use age-specific modules. The selection probability and respondent rules should be incorporated.

Comparisons across children in the same household are not available when only one was selected, and household-level clustering remains relevant.

The base weight is the inverse of the probability of selection across all sampling stages. It represents how many covered population units a sampled unit stands for before later adjustments.

Errors in selection probabilities cannot be corrected by normalising weights to a familiar total.

Weights may be adjusted within classes assumed to share response propensity. The classes should use information available for respondents and non-respondents and should be specified before outcome analysis where possible.

Adjustment reduces bias only if the assumptions are credible. It does not recover children outside the frame.

Weights may be calibrated to known population totals by age, sex, region or other variables. This can improve coherence and precision but transfers the quality of the control totals into the survey estimate.

Calibration should not force education results to match administrative counts unless that event is deliberately used as a control and is known to be more accurate.

Very large weights can make results unstable. Trimming reduces variance but can introduce bias if high weights identify genuinely underrepresented populations.

The rule, affected cases and sensitivity should be documented. Convenience is not sufficient reason to suppress the influence of remote or difficult-to-reach households.

Normalised weights sum to the achieved sample or another fixed total and can estimate proportions under certain analysis. They do not produce population counts unless rescaled appropriately.

Tables should identify whether weighted counts are population estimates or normalised analytical totals.

When every eligible child in a sampled household is included, the household selection weight may form the base for child estimates with relevant adjustments. When child selection differs, a child-specific weight is required.

Household weights should not be applied automatically to women, children or submodules with different eligibility and response.

Welfare cut-points should be constructed using the intended weighted population. If the analytical unit is children, quintiles based on weighted households can contain unequal numbers of children.

The report should state whether the grouping describes the household distribution, person distribution or child distribution.

Analytical files should retain weight, stratum and primary sampling unit identifiers in a protected form. Variance software requires these variables and any finite-population correction used.

Dropping design variables before analysis converts a complex sample into an incorrectly treated simple sample.

The standard error describes sampling variability under the design and estimator. It does not include coverage, questionnaire, processing or model error.

A small standard error should not be described as overall data accuracy.

A confidence interval provides a range under stated statistical assumptions. The level and method should be consistent across comparisons. Intervals for proportions near zero or one may require methods beyond a simple symmetric approximation.

The interval should be calculated from unrounded values and displayed with sensible precision.

The design effect compares the variance under the actual design with that under a simple random sample of the same size. It varies by indicator and domain.

One general design effect should not be assumed for every education and household characteristic without evidence.

Weight variation and clustering reduce effective sample size. A large raw sample can support a less precise group estimate than its count suggests.

The release may report the unweighted denominator, weighted population and relative standard error or interval together.

A domain is a subgroup for which an estimate is required, such as rural girls in the lowest welfare group. Correct domain analysis retains the full sample design rather than deleting all other records before variance estimation.

Domains planned in the design may be more precise than unplanned intersections.

Direct survey estimates may be unreliable below the designed reporting level. Model-based small-area estimates require additional assumptions, covariates and validation and should be identified as modelled.

A survey should not publish direct district ranks merely because software can calculate them.

Testing many groups and outcomes increases the chance of apparently unusual differences arising through sampling variation. The analysis should identify primary comparisons and interpret exploratory results cautiously.

Public-interest relevance and consistency across evidence matter, but they do not remove the need to disclose extensive searching.

A release rule may use minimum unweighted counts, maximum relative standard errors or interval widths. The rule should reflect consequence and should not be manipulated after seeing the estimate.

Suppression for low precision should remain distinguishable from zero and confidentiality suppression.

Statistical precision and policy materiality answer different questions. A large, precisely estimated gap is a strong descriptive finding. A severe but imprecisely measured exclusion may justify urgent evidence collection and protection while the numerical claim remains qualified.

Neither concern should be used to erase the other.

The record should contain sample design, weights, eligibility, exclusions, estimator, variance method, software or calculation procedure, precision rule and reviewer. Weighted and unweighted denominators are retained.

An independent calculation should reproduce principal group estimates and intervals.

Participation should be reported by the sex variable collected, with age and overall levels. Parity at one stage does not establish parity in attendance, grade, completion or learning.

Household and school mechanisms should be investigated rather than attributed to sex itself.

Single-year age is central because enrolment expectations and grade placement change rapidly during childhood. Broad age groups can conceal late entry or early departure.

Age should be calculated at a controlled reference date and examined for heaping and missingness.

Birth order may be associated with care, work and resource allocation, but it is correlated with family size and parental age. The definition should distinguish biological birth order from order among resident children.

A cross-sectional association should not be treated as a fixed family preference without supporting evidence.

Multiple school-age children can increase costs and care demands while also providing support. Counts should use a defined age range and household membership rule.

The effect of family size cannot be inferred without considering resources, spacing and composition.

Co-residence with biological or legal parents may be measured where relationships are available. Absence can reflect death, migration, work, fostering, separation or schooling arrangements.

The category should not be labelled orphanhood unless the survey establishes parental survival.

Maternal, paternal and double orphanhood require direct survival questions or reliable records. Missing parental information is not death.

Education analysis should account for age and living arrangement and should avoid assuming uniform disadvantage among all orphaned children.

Children living apart from parents may receive support and schooling or may face unequal work and expenditure within the host household. Relationship, duration and reason may be incompletely measured.

Comparisons should examine participation within household welfare groups and not assign the host household’s resources as proof of equal access.

Work measures should state age, activity, hours and reference period and should include unpaid family and domestic work where collected. The Worst Forms of Child Labour Convention provides a protective context for harmful work and education access.[REF-11]

Work and non-attendance can influence each other. Time sequence and household need require analysis before causal judgement.

Water collection, care, cooking and other household tasks may compete with attendance or study and may be distributed by sex and age. Surveys often measure these activities less completely than economic work.

The absence of a recorded labour activity should not be treated as absence of time burden.

Disability may be underreported or defined through impairments, diagnoses or activity limitations. One household question cannot establish educational need or school accessibility fully.

Analysis should state the exact question and acknowledge children outside the household frame. The inclusion guidelines support examining barriers in systems and schools rather than locating exclusion solely in the child.[REF-12]

Recent illness can affect attendance; chronic conditions can affect access and continuity. Self or proxy reports depend on recognition and reference period.

Health variables should be handled confidentially and should not be used to justify exclusion from education.

Nutrition can affect attendance, concentration and development, while schooling and household resources influence nutrition. Anthropometric data require age, measurement quality and appropriate reference.

Associations with participation should be interpreted within this two-way relationship.

Household language, mother tongue and language usually spoken are different variables. They may relate to the language of instruction, access to information and assessment.

Categories should preserve meaningful languages where sample permits and should not be reduced to a deficit label.

Ethnic, indigenous, caste, racial or other identity categories are nationally specific and sensitive. Collection requires legitimate purpose, self-identification where appropriate, protection and consultation.

International aggregation may erase meaning. The analysis should examine discrimination and service conditions rather than present identity as causal.

Religious affiliation can relate to provider choice, calendar, location or social treatment. Questions and categories require sensitivity and lawful purpose.

An observed difference should not be interpreted as doctrine or preference without direct evidence.

Recent move, place of birth, nationality and citizenship are distinct. Migrant children may face documentation, language, residence and transfer barriers.

The household frame can miss mobile populations and should report duration of residence where available.

Status may be legal, administrative or self-reported and can be sensitive. Education access may occur through temporary, public or community provision not mapped in ordinary categories.

Coverage and protection take precedence over an overly precise rate based on an incomplete frame.

Urban and rural follow the national statistical definition. They do not measure distance, remoteness, school quality or poverty directly.

The definition and survey stratum should be distinguished from respondent address coding and school location.

Distance may be reported, mapped or measured as travel time. Travel mode, terrain, safety and season affect practical access.

Self-reported time is useful but can be rounded. The nearest school may not offer the required grade or language.

Administrative regions support policy responsibility but may contain large internal differences. Boundary changes and small samples affect trends.

The report should not rank regions without precision, population and service context.

Crowding, construction materials, water, sanitation and electricity can describe living conditions and enter asset measures. They may also affect study time and health.

Using a housing item both within a welfare index and as a separate explanatory variable requires attention to mechanical overlap.

Shock categories should distinguish event, timing, severity and reported consequence. Recall can be selective, and households may have changed behaviour before the survey.

Policy analysis should examine support received and school response, not only the occurrence of the shock.

Receipt of a transfer, food, fee waiver or scholarship identifies programme reach. Recipients are selected and therefore differ from non-recipients before support.

Participation differences between the groups do not estimate programme effect without a suitable design.

Headship is retained where policy or comparability requires it, but the analysis should present composition, welfare and adult education alongside it. The category should not imply that one sex of head causes disadvantage.

Differences in how respondents designate the head limit comparison.

Adult literacy may be self-reported or directly assessed and can influence household access to written school information. It is related to, but not identical with, years of education.

The report should identify the measure and selected adult rather than use “literate household” without definition.

The primary caregiver may differ from the household head or parent. Caregiver education, language, work and health can be relevant if the survey identifies the relationship reliably.

The child’s participation should not be presented as the caregiver’s sole responsibility.

Household disadvantage interacts with the number, level, provider and accessibility of nearby schools. Community or facility data can be linked geographically with appropriate protection.

The presence of a school does not establish places, teaching or a complete grade cycle.

Transport costs, local wages, harvest timing and prices can influence participation and the meaning of household income. Community data strengthen interpretation but may not align exactly with the household interview period.

Area averages should not be assigned as if they were directly observed for every household.

Dimensions should be selected because they are relevant to rights, policy or an evidenced mechanism, not because the data contain many variables. Pre-specification reduces selective reporting.

Exploratory analysis remains useful when labelled and followed by confirmation.

Group labels should reproduce the variable and avoid pejorative or essentialising language. “Children in households in the lowest asset-index quintile” is more accurate than a broad label of deprived families where no absolute threshold was used.

The wording should remain understandable without sacrificing precision.

Household characteristics identify unequal distribution and possible pathways. They do not transfer the obligation to provide accessible, acceptable and adaptable education from public authorities to families.[REF-09]

Every group finding should lead back to the service or protection condition that can be examined and changed.

The overall participation estimate uses the full covered target population and survey design. It provides context for group levels and should not be calculated as an unweighted average of group percentages.

The unweighted sample denominator and weighted population estimate should be retained.

A group estimate applies the same participation definition to the domain defined by the household or child characteristic. Weights and design remain in the calculation.

The group denominator excludes only records under the stated eligibility and missing-data rule, not inconvenient or extreme observations.

The reference group determines the sign and meaning of a gap or ratio. It should be selected for policy clarity and stated in the table.

Changing the reference across outcomes makes a profile difficult to interpret and can create selective emphasis.

Subtracting group B’s participation percentage from group A’s produces an absolute difference in percentage points. It is intelligible across levels but does not show proportional disparity.

The estimate and its standard error should account for covariance between group estimates under the design.

Dividing one group rate by another produces a relative measure. It becomes unstable where the denominator is low and can appear close to one when both groups have very low or high non-participation depending on event orientation.

The two rates and counts should accompany the ratio.

A participation ratio and an out-of-school ratio can give different impressions from the same data. For example, 90 versus 80 per cent attendance is a 1.125 participation ratio, while 10 versus 20 per cent non-attendance is a 0.5 ratio under the same group order.

The chosen orientation should reflect the policy question and should not be switched to magnify or minimise disparity.

Odds compare the probability of an event with its complement. Odds ratios are common in models but are less intuitive than percentage-point differences for public reporting.

They should not be described as risk ratios, especially where the outcome is common.

The estimated number of non-participating children equals the weighted population in the group multiplied by its non-participation rate under compatible estimates. Counts inform service scale.

Their uncertainty includes rate and population estimation. Rounding should reflect that uncertainty.

A group’s share of all out-of-school children depends on both its population size and rate. A small group can face severe risk while contributing a small share of the national count; a large group can account for many excluded children with a modest rate difference.

Severity and scale should be reported separately.

An ordered welfare measure allows examination across all groups rather than only the extremes. A monotonic gradient can strengthen description but does not establish a linear relationship or cause.

Group levels, confidence intervals and population shares should remain visible.

Concentration measures summarise how an education outcome is distributed across an ordered welfare population. They require a clear ranking variable and assumptions that may be difficult for a general audience.

The summary should not replace quintile or decile levels and should be used only where it adds policy meaning.

Regression-based inequality measures can use the full welfare distribution and account for group population shares. Their interpretation depends on model form and ranking.

Technical results should be accompanied by predicted group levels or an equivalent accessible display.

Groups with different age structures may have different participation because expected schooling varies by age. Direct or model-based age standardisation can support comparison if a common standard population is stated.

Standardised rates should accompany observed rates. Policy still concerns the actual children and age distribution.

A model may adjust a wealth gap for age, residence or adult education to describe conditional association. Adjustment does not reveal what the gap would be under a feasible intervention and can remove variables on the causal pathway.

The unadjusted distributional finding should not disappear from the public account.

A confounder is related to both the household characteristic and participation and is not a consequence of the characteristic under the causal question. Identifying it requires subject knowledge, not only statistical significance.

Cross-sectional household surveys often cannot establish time order sufficiently for strong causal adjustment.

Cost, distance, work or school availability may mediate an association between household position and participation. Adjusting for a mediator estimates a different relationship and may conceal the mechanism of policy interest.

Reports should state the purpose of each model rather than call the most adjusted estimate the true effect.

Interaction occurs when a relationship differs across another characteristic, such as a wealth gap varying between urban and rural areas. It should be assessed through group predictions or a specified model, with adequate sample.

Separate significance tests within groups do not prove that effects differ between them.

A precisely estimated small gap may be statistically distinguishable but limited in practical consequence; a larger imprecise gap may require more evidence. Context, number affected and stage of education inform materiality.

The report should not equate a probability threshold with the public-interest decision.

Group trends require stable participation and household measures. Relative asset quintiles can change composition even when the construction method remains constant.

Absolute welfare measures or repeated item distributions may be needed to interpret whether the same kind of household improved.

Older and younger age groups can show how attainment differs across cohorts, but household composition, migration and survival also vary. A cross-sectional cohort comparison is retrospective, not longitudinal.

The education structure and group definition applicable to each cohort should be considered.

Decomposition can allocate an overall difference into components associated with measured characteristics under a model. The result is sensitive to reference, order and omitted variables.

Components should not be called causes or policy contributions without a design that supports those claims.

If one group has lower participation across every age or welfare threshold examined, the descriptive conclusion is stronger than a difference at one arbitrary cut-point. Sampling uncertainty and coverage still apply.

Dominance does not identify why the difference persists.

Confidence intervals can appear as table bounds, error bars or notes. Where space is limited, a quality flag should link to the exact precision criteria and values.

Stars alone are inadequate because they do not show effect size or uncertainty range.

Rates identify unequal risk; counts identify scale. Both should be presented when the population estimates support them.

An intervention prioritised only by count may overlook a small severely excluded group, while one prioritised only by rate may fail to reach most excluded children.

Statements that a group “would attend if” a household characteristic changed are counterfactual and require causal evidence. A cross-tabulation supports only the observed difference.

Policy scenarios should state assumptions and should not be presented as survey results.

Country comparisons require compatible survey populations, years, school ages, participation questions, welfare measures and group construction. National quintiles are relative within each country.

The protocol favours country profiles under a common method over a rank that implies identical living standards.

Regional group estimates should weight country populations and identify coverage. Pooling microdata requires harmonised variables and survey design treatment.

A regional poorest quintile can mean pooled regional rank or the aggregation of each country’s lowest fifth; these are different populations and should be named.

The preferred public result is an overall level followed by group levels, gaps, counts and uncertainty. Complex summaries may support analysis but should not displace the underlying distribution.

Every comparison should remain traceable to people represented by the survey.

Missingness can arise from frame exclusion, household non-response, an absent respondent, skipped module, refusal, “do not know”, processing loss or an invalid response. These mechanisms have different consequences.

One generic missing code prevents diagnosis and appropriate treatment.

When an entire household does not respond, weights may adjust within response classes. The adjustment depends on information about selected non-respondents and an assumption about similarity after conditioning.

Response rates should be reported by region and other design variables relevant to education access.

Item non-response can affect income, assets, parental education, disability, work or attendance. Rates should be reported for numerator, denominator and grouping variables.

Records missing the outcome and records missing a group characteristic have different analytical roles.

A question may be inapplicable, such as parental education where no parent is identified under the roster. This is not ordinary non-response.

The analysis should create a substantively meaningful category or define the population rather than impute a parent mechanically.

Modules may be administered to a subsample or only certain ages. The design is known and should be represented through eligibility and appropriate weights.

Combining module and full-sample variables without adjustment produces an incorrect denominator.

Complete-case analysis uses records with all required variables. It is simple but can reduce precision and bias results if completeness relates to participation or household circumstances.

The retained and excluded populations should be compared before use.

Treating missing as a category preserves records and makes absence visible, but it does not recover the true characteristic and can mix several mechanisms.

It is often appropriate for descriptive release alongside a sensitivity analysis.

Single imputation fills one value and can understate uncertainty if treated as observed. Deterministic household rules may be justified for clear derived fields but require an audit trail.

Outcome values should not be filled merely to complete a table.

Multiple imputation reflects uncertainty by creating several plausible values under a model. It requires variables predictive of missingness and value, appropriate survey design treatment and combined variance.

The result remains conditional on assumptions that should be accessible to reviewers.

For material missingness, analysts should examine plausible extreme or model-based scenarios and identify whether the substantive conclusion changes.

Sensitivity bounds should be defensible and should not be chosen to preserve the preferred finding.

The visibility gap is the difference between the target child population and those represented with usable education and group data. It includes frame exclusions and unresolved missingness.

It should be described in counts or ranges where possible and should accompany claims about the most disadvantaged.

A child can be absent from a survey but known to education or protection services, or present in a household survey but absent from school records and programmes. The two forms should be distinguished.

Complementary sources can improve the public account without disclosing identity.

Intersections test whether combinations of circumstances correspond to different participation beyond single-group averages. They should be selected through policy relevance and evidence, not an unrestricted search over every variable.

The result remains a measured subgroup, not a fixed identity of disadvantage.

Each additional category divides the sample. A five-level welfare group crossed with sex, residence and region may produce many cells with few observations.

Collapsing categories should preserve meaning and should be specified before seeing which grouping produces the largest difference.

An intersection table presents combined group levels. Statistical interaction tests whether a relationship differs across another variable under a model.

The table can be policy-relevant without a significant interaction, and separate within-group significance does not establish interaction.

An additive model assumes each characteristic contributes the same difference regardless of others. Compounded barriers may violate that assumption.

Predicted group results and residual checks can reveal limitations, but complex models require adequate data.

Regression coefficients for categorical interactions depend on the reference categories. Public reporting should translate them into predicted probabilities or group differences.

Changing the reference should not change the substantive joint predictions.

Some categories are nested, such as districts within regions, while others cross, such as welfare and sex. The analytical structure should reflect this to avoid treating an administrative hierarchy as independent household traits.

Multilevel models may assist but do not remove coverage and measurement limits.

Rare or geographically concentrated groups may not be captured reliably by a general national sample. Oversampling, dedicated surveys, census evidence or qualitative and administrative sources may be necessary.

The inability to publish a precise rate should not be read as low importance.

Small intersection cells can identify households or children, particularly when combined with place and sensitive characteristics. Suppression, aggregation or controlled access may be required.

Complementary suppression should prevent totals from revealing the hidden cell.

If every estimate for a small group is suppressed, public reporting can reproduce invisibility. A broader safe category, qualitative finding or secure aggregate may preserve public meaning.

The protection decision should be recorded and reviewed with the affected information need.

An observed zero in a sample does not establish that no person or event exists in the population. Weighted zero, structural zero and suppressed small count are different.

The table symbol and note should distinguish them.

Ordering many intersection groups by point estimate exaggerates random variation. Intervals may overlap widely, and small group ranks can change under minor revisions.

Profiles should use substantive ordering and avoid league tables.

Combining surveys can increase sample size if questions, frames and methods are compatible. Pooling should retain survey-year and design information and should not erase time change.

An average over several years may be inappropriate for a rapidly changing or displaced population.

Census data can support small-group and geographic description but may contain fewer education and welfare variables. Survey data provide depth and sampling uncertainty.

Combining them through models or calibration requires validation and separate identification of modelled results.

Interviews, observation and community inquiry can identify mechanisms, categories and omitted populations. Selection and analysis should be systematic and ethical.

Qualitative evidence explains experience and process but should not be converted into a population prevalence without a suitable design.

The strongest claim about household disadvantage is only as inclusive as the frame, response and usable group variables. Survey sophistication cannot compensate for systematic absence of the children most difficult to enumerate.

The public record should place these visibility limits beside the group findings and assign responsibility for complementary evidence.

An observed association is a difference in participation across measured household or child groups under the survey design. It is a valid descriptive result when definitions, coverage and uncertainty are sound.

The statement should not insert a causal verb such as produces, prevents or determines.

A plausible mechanism connects the characteristic to participation through cost, distance, time, information, health, discrimination, safety, language or service quality. Its plausibility should be supported by relevant household, school or qualitative evidence.

More than one mechanism may fit the same pattern. The report should identify what would distinguish them.

A causal conclusion requires a design and evidence capable of addressing alternative explanations and time order. Cross-sectional adjustment alone is usually insufficient.

The strength of the claim should match the evaluation, not the urgency of the policy concern.

Household characteristics describe circumstances; they do not define who must act. Authorities and providers remain responsible for available, accessible, acceptable and adaptable education within their mandates.[REF-09]

The response record should name the institution able to change the barrier.

A welfare gradient can justify examining official and unofficial fees, uniforms, materials, transport, meals and foregone work. Household expenditure and reported reasons can strengthen the inquiry.

Spending only among enrolled children cannot show the price that excluded households would face or the maximum they could pay.

A residence gap should be examined with school locations, grade availability, travel time, transport, terrain and safety. Rural status alone does not identify the binding constraint.

Planning should distinguish access to any school from access to the required programme and grade.

Where working children attend less, analysis should examine hours, type, timing, household shocks and school schedules. Harmful work requires protective action under applicable law.[REF-11]

Education policy should also examine whether poor quality, cost or exclusion preceded work or made departure more likely.

Language differences require evidence on school information, admission, instruction, materials and assessment. A household language variable should not be used as a proxy for ability.

Consultation should include families and teachers in the relevant language communities.

A lower participation result among children identified under a disability question should lead to examination of physical, communication, curricular, attitudinal and policy barriers.[REF-12]

The survey may undercount disabled children or capture only some functional difficulties. Service planning should use complementary evidence.

Sex, age or location differences may relate to travel safety, violence, harassment or inadequate facilities. These topics require confidential, safe methods and referral procedures.

Absence of disclosure in a general household survey is not proof of absence of risk.

Birth registration, residence, citizenship or prior school records may be requested at entry. The policy and actual school practice should be examined separately.

Where documentation is not a lawful condition, public information and correction should remove the barrier. Where records are needed, an accessible provisional route should be considered.

Households may withdraw or not enrol children where teaching is unavailable or perceived as poor. Staff attendance, instructional time, materials, language, treatment and learning evidence should be assessed.

The explanation should not frame household response as lack of demand while institutional conditions remain unexamined.

Removing fees may improve access but can increase enrolment faster than staffing and materials. Monitoring should include group entry, attendance, class conditions, learning and remaining indirect costs.

The policy effect requires a comparison beyond a post-announcement rise.

A scholarship or transfer can reduce financial constraint if eligibility, amount, timing and access align with the barrier. Recipient–non-recipient comparison is selected and does not estimate impact directly.

Coverage among the intended group and exclusion errors should be reported before outcome claims.

School meals may address hunger and participation and can reduce household cost. Programme days, meal receipt, attendance, nutrition and learning should be measured separately.

Schools without capacity or remote excluded children may be missed, potentially widening a gap if expansion is uneven.

Transport, satellite schools, boarding or expansion of grades can address distance. Each option has safety, cost, staffing and family implications.

Evaluation should use travel and grade-access evidence rather than residence category alone.

Flexible schedules can accommodate seasonal work or mobility but should preserve instructional time and avoid legitimising harmful work or separate inferior provision.

Attendance and learning should be reviewed by group and season.

Accessibility, adapted teaching, support personnel and reasonable changes can improve participation of disabled children. The response should be planned through system and school barriers, consistent with inclusion principles.[REF-12]

Placement in a programme should not be counted as inclusion without participation and learning evidence.

Information and instruction in languages learners understand may improve entry and participation. Programme design requires qualified teachers, materials and a clear transition where another instructional language is used later.

The outcome should include learning and continuation, not only initial enrolment.

Outreach can identify unregistered or never-enrolled children and explain entitlements. It should connect households to a real place, support and follow-up rather than end with a referral count.

Records should protect personal information and should not expose families to unrelated enforcement.

Universal action can improve the education system for all; targeted action can address a concentrated barrier. Household disaggregation helps select the balance.

Targeting based on a noisy or relative measure can exclude eligible children and should include review and appeal.

Area-based support can reach many households with common service constraints but may miss disadvantaged households in better-off areas. Area classifications require current population and boundary data.

Local rate and population count should both inform allocation.

Household eligibility may use an official poverty test, categorical rule or community process. The education survey’s asset quintile should not become an administrative eligibility rule without validation and authority.

Collection for benefits also creates incentives and privacy concerns different from research.

Schools serving high-need populations may require resources and support. School composition derived from household linkage should be robust and should not stigmatise pupils or encourage exclusion to improve indicators.

Allocation should recognise current need without lowering expectations.

Reach is the share and number of the intended population that receives the intervention. It should be estimated using an eligible denominator independent of programme records where possible.

High participation among recipients says little if most eligible children are not reached.

An inclusion error provides support outside the intended rule; an exclusion error omits an eligible person. Their social and fiscal consequences differ.

An eligibility system should permit correction and should not shift proof burdens beyond the capacity of disadvantaged households.

Fidelity describes whether the action occurred with intended content, timing, amount and quality. It should be measured before interpreting outcomes.

Local adaptation can be legitimate and should be distinguished from failure through the authorised design.

Evaluation should use the participation event the intervention seeks to change and should examine distribution, learning and possible adverse effects. Baseline and comparison should reflect programme selection.

A change in recipient enrolment alone may reflect eligibility or reporting rather than effect.

Targeting can stigmatise children, create conflict, raise administrative barriers or direct schools toward easily verified households. Fee removal can strain quality; transport can create safety risk.

These effects should be monitored and can justify revision even when the headline participation rate improves.

Authorities may act on a severe plausible barrier while commissioning stronger evidence. The decision should state uncertainty, risk of delay, reversibility and review date.

Urgent protection should not be communicated as proof of a final causal finding.

The disaggregated statistic identifies who experiences a different result. The policy inquiry identifies what can be changed. Keeping the two stages separate protects accuracy and leads to more proportionate action.

The following cases are constructed methodological tests and do not describe named countries or observed institutions. Each illustrates a different survey, denominator or policy problem.

A household survey reports primary-age attendance of 64 per cent in the lowest national asset quintile and 91 per cent in the highest. The index is based on housing materials, electricity, water and durable goods and uses the child population to set cut-points.

The 27-point difference is a strong descriptive inequality if the design and intervals support it. It does not establish an absolute poverty threshold, and the lowest group cannot be compared directly with another country’s lowest group. Infrastructure items may also embed an urban–rural dimension.

The next analysis should show every quintile, urban and rural composition, direct and indirect education costs, distance and age. A model may examine conditional patterns, but the unadjusted gap remains a public result.

Policy should not be selected from the asset index alone. Fee and transport evidence, school availability and household accounts are needed to determine the mechanism.

Children in households reported as female-headed have higher attendance than those in male-headed households. Female-headed households in the sample are smaller, more urban and more likely to receive remittances; head designation varies by region.

The result does not show that head sex causes higher attendance or that female-headed households are uniformly advantaged. The category combines household structure, migration and reporting convention.

The report should publish the observed levels and then examine resources, residence, composition and adult education. It should avoid replacing the household-head result with an adjusted coefficient without explaining the changed question.

Service policy should follow actual barriers and needs, not a general assumption about headship.

Rural attendance is lower than urban attendance. In a regression including reported travel time and welfare, the rural coefficient becomes small and imprecise.

This is consistent with travel and welfare accounting for the measured association under the model, but it does not prove that relocating schools will eliminate the gap. Travel time may be reported with error and may itself reflect school choice and attendance.

Observed rural levels, distance distributions and school grade availability should remain public. Facility and transport evidence can test the mechanism. Policy appraisal should compare school expansion, transport and staffing with safety and learning.

The adjusted result answers a conditional statistical question, not the distributional question of whether rural children currently participate equally.

Attendance is 18 percentage points lower among children reporting at least one hour of economic work in the previous week. The survey was conducted during harvest in some regions and later in others. Domestic work was collected only for girls.

The comparison is affected by season and unequal measurement. It cannot support a national causal estimate of work or a valid sex comparison of total work burden.

Analysis should align fieldwork and agricultural calendars, report activity and hours, and acknowledge incomplete domestic work. Household shocks, school schedules and prior participation should be examined.

Protective action may be justified for harmful work, but education policy also needs to address cost and school conditions that may precede departure.

Remote rural girls in the lowest wealth quintile have the lowest attendance estimate in a national table. The cell contains 27 unweighted children, has a wide confidence interval and comes from four clusters.

The point estimate is not suitable for a precise rank or national prevalence statement. The pattern is relevant enough to justify complementary evidence because the circumstances are plausibly compounded and the potential consequence is severe.

The release should show the estimate as low precision or provide a broader defensible grouping. A dedicated sample, census data or qualitative inquiry can establish access barriers. Confidentiality should be checked because geography and household characteristics may identify communities.

No policy should require the fragile cell to cross a significance threshold before examining whether schools and transport exist.

Parental education is missing for 4 per cent of all children but for 19 per cent of children not attending school. Many live with relatives and no parent is on the household roster.

Complete-case analysis would disproportionately remove non-attenders and could understate the association with adult education. Imputing a parent’s education from the household head would change the concept and may be wrong.

The report should retain a no-resident-parent or unavailable category where supported, show participation, and conduct sensitivity analysis. Caregiver or highest-resident-adult education may answer a separate question if defined.

The missingness pattern is also a substantive signal about living arrangements and should not be treated only as a technical nuisance.

Girls and boys have the same national attendance rate. In one large region girls attend less; in another boys attend less; other regions are near parity. Population weights cancel the national difference.

The national parity statement is numerically correct and substantively incomplete. Regional results, precision and age patterns should be published. The direction should not be generalised from one region to the country.

Inquiry may identify different mechanisms: safety or domestic work in one setting, economic work or school disengagement in another. A single gender programme may not address both.

The case demonstrates why group averages require geographic and pathway context without assuming that every small variation is meaningful.

The household survey estimates 16 per cent of primary-age children not attending, while administrative data report 8 per cent not enrolled using projected population. The survey excludes two insecure districts; school coverage of private providers is uncertain.

Neither value is a defensible national truth without reconciliation. Attendance and enrolment are different events; both sources have coverage problems; the population denominator may differ.

The authority should align age, period, geography and provider scope, review response and population projections, and publish both source-specific results. The insecure districts require complementary evidence and should remain explicit.

Selecting the smaller estimate would not resolve exclusion; selecting the larger would not resolve survey coverage.

Scholarship recipients attend less than non-recipients in a cross-sectional survey. Eligibility targets households with prior non-attendance and severe poverty, and some awards began shortly before interview.

The result likely reflects selection and timing and does not show harm from the scholarship. It is still important to examine delivery, amount, payment delay and current attendance.

Evaluation should compare eligible groups under a credible design and measure baseline status, receipt and exposure. Programme records alone may omit approved children who did not receive payment.

The public release should state recipient attendance and selection rules without calling the difference a programme effect.

A general household survey reports no participation gap for children identified by one yes-or-no disability question. The identified group is very small, and specialised service records indicate many disabled children outside conventional households or unregistered.

The result applies only to children identified and represented under the question. Under-identification and frame exclusion can select children with less severe barriers, producing an apparently favourable estimate.

The release should state the question, sample, precision and omitted populations and should avoid a conclusion of equal national access. Specialised enumeration and school-access review are required.

The institutional response should examine barriers regardless of whether the general survey produces a statistically significant gap.

Across the cases, the principal risks are concept substitution, relative groups treated as absolute, weak intersection samples, informative missingness, selection into programmes and omission from the household frame.

The corrective discipline is to define the observed population and result, preserve the descriptive gap, test mechanisms with additional evidence and restrict policy claims to what the design supports.

The release should present the overall education result, population coverage, group levels, counts, gaps, uncertainty and limitations before policy interpretation. It should be possible to understand the main finding without consulting an inaccessible technical file.

Technical detail remains necessary for reproduction and specialist review.

Every table title or note identifies age, residence rule, covered household population, education event, school period and data source. The denominator for each group is visible.

A national label should not be used where selected areas or household types were excluded.

Welfare groups should follow their ordered measure; ages should follow age; regions may follow geography or an established administrative order. Ordering by the outcome can imply a rank beyond the precision of the estimates.

Reference categories and direction of gaps should remain consistent.

The overall estimate belongs beside group levels so that distribution and national scale can be read together. The overall value should be calculated from weighted microdata or compatible totals, not from an unweighted group average.

Where group categories do not exhaust the population because of missing values, the missing group is shown.

Weighted counts should be rounded according to sampling and population uncertainty. Unweighted sample counts support precision assessment and should be available without threatening confidentiality.

The table should distinguish estimated children from achieved respondents.

A percentage-point gap is preferred for the principal public comparison because it preserves the scale of participation. A ratio may accompany it where proportional difference matters.

Both components should remain visible. A summary column is not a substitute for the group rates.

Symbols may indicate wide relative error, low unweighted count, suppressed value or unavailable estimate. Each requires a different symbol and note.

A symbol should not conceal the interval or criterion from technical users.

Frame exclusions and non-response should be reported in the main methodological note. Where likely to alter the principal disadvantage finding, they belong in the executive summary.

Missingness within variables is shown in relevant tables rather than absorbed into a footnote.

The release should state items, method, ranking unit, weight and cut-point treatment for an asset index. It should use “asset-index quintile” rather than “income quintile” or “poverty group”.

Cross-country tables should state whether groups are national relative ranks.

Maps can show access and regional differences but may identify small communities or imply precision below the survey design. Boundaries and reference dates should be accurate.

Uncertainty and areas without estimates should remain visible. Colour scales should not exaggerate small differences.

Charts should display group levels and intervals. Truncated scales may be appropriate for small differences only when clearly marked and not used to overstate materiality.

Group populations and event orientation should be accessible through labels or tables.

The narrative should use “was lower”, “was associated with” or “differed” for descriptive findings. “Caused”, “led to” or “because” requires appropriate evidence.

Households and children should not be blamed or portrayed as deficient through group labels.

Possible mechanisms should be identified as hypotheses, supported explanations or established effects. The evidence for each status should be cited.

Policy urgency does not permit an unexplained movement between these categories.

International comparison should harmonise event, age, year, welfare grouping and survey design and should publish country-specific caveats. Relative wealth groups should not be treated as common resource thresholds.

Profiles and regional patterns are preferable to a simple inequality league table.

Repeated surveys require a method note on frame, question, season, weights, welfare index and education structure. A break should be shown rather than smoothed.

Changes in missing or excluded populations may explain apparent improvement and should be reported.

Statisticians should verify design and calculation; education specialists should verify event and programme mapping; subject experts should review group meaning; disclosure officers should review confidentiality; and the responsible authority should approve the public conclusion.

Review comments and changes should be logged.

A correction states original value, revised value, reason, affected tables and effect on the finding. Prior versions remain identifiable.

If weight, frame or group construction changes, the correction should identify whether the entire series was recalculated.

Users should have a route to ask about definitions, request precision details or provide evidence of omitted populations. Responses and recurring issues should inform later survey design.

Statistical authority remains responsible for deciding and explaining corrections.

De-identified files and documentation can support independent analysis under national law and confidentiality controls. Sensitive geography, identity and household variables may require restricted access.

Public reproducibility does not justify exposing children or households.

Group results should not be suppressed because they show an unfavourable distribution. Technical choices should be protected from programme incentives while remaining publicly accountable.

Independence includes willingness to correct and to state when the survey cannot support a requested conclusion.

The archive retains questionnaire, manuals, frame, sample design, weights, derived variables, welfare construction, analysis files, tables, review and corrections. Historical access follows confidentiality rules.

Without the archive, trend and independent reproduction cannot be assured.

Education ministries and national statistical offices should collaborate on questions, mappings, analysis and dissemination. Education expertise improves variable meaning; survey expertise protects design integrity.[REF-01]

Training should support ongoing national analysis rather than one external publication.

The authority of a disaggregated result depends on transparent population, design and group construction as much as on the numerical gap. Public release is complete only when users can identify who is represented and what action the finding can reasonably support.

Every analysis should begin with the target population, survey frame, achieved response and children outside conventional households or accessible areas. A weighted estimate cannot represent an omitted group by declaration.

Complementary evidence is required where the risk of invisibility is greatest.

Income, consumption, assets, adult education, headship, residence and household structure should retain their exact meaning. Broad labels should not replace the measured variable.

Relative welfare groups require particular caution across countries and years.

Enrolment, attendance, entry, grade, completion and learning are separate. The age, period and programme mapping determine each result.

Disaggregation is valid only when the same event is measured for every group.

Complex sample design, small domains, missingness and non-coverage constrain precision and inference. Counts, intervals and disclosure rules belong with the finding.

Uncertainty should direct better evidence and proportionate action, not erase serious possible exclusion.

A household difference identifies unequal participation. Service, longitudinal, qualitative or evaluation evidence is needed to establish why and what will change it.

Policy should address the evidenced barrier and should monitor reach, learning and adverse effects.

Poverty, work, language, disability or family structure should not be used to transfer responsibility for exclusion to the child or household. Public institutions should examine affordability, access, teaching, safety, accommodation and information.

The rights and inclusion frameworks place this responsibility within a universal education commitment.[REF-07] [REF-08] [REF-09] [REF-12]

This report does not prescribe national categories or an absolute welfare threshold. It does not estimate the prevalence of disadvantage in a country or region and does not assess a named household survey.

The protocol uses methods and evidence available by 15 August 2006. It requires adaptation to national law, statistical capacity and education structure.

Disaggregation is authoritative when it makes unequal participation visible without overstating who was observed, what the household measure means or why the difference arose. Its value lies in disciplined connection between population, education event, circumstance, comparison and public response.

The proper outcome is not a catalogue of group deficits. It is a clearer account of the barriers for which education and public institutions can be held responsible, accompanied by evidence that the response reaches the children previously least visible in the aggregate.

A disaggregated household estimate identifies a distribution requiring explanation. The policy evidence record prevents a descriptive association from being converted directly into a causal claim or an intervention. It sets out proposed mechanisms, evidence for and against them, responsible institutions and the further test required.

The record should be completed jointly by statistical, education-sector and relevant social-policy specialists. Community evidence may reveal barriers not represented by survey variables. Participation by affected groups is a source of evidence and accountability, not a substitute for a representative estimate.

Household economic position may relate to direct fees, uniforms, books, transport, meals, foregone labour, seasonal work, distance, school availability, language, disability accommodation, safety, perceived quality, documentation, discrimination or household decision-making. These mechanisms operate at different institutional levels.

The survey may contain proxies for some mechanisms, but a proxy is not the mechanism itself. Rural residence does not prove distance; asset position does not measure school cost; parental education does not establish preference. The evidence record preserves this distinction.

Evidence for causation requires that the proposed cause precede the education event. A cross-sectional survey often records household conditions and school participation at one time. Some conditions may have been changed by the event: a household may reduce assets after sustained education expenditure, or a child’s non-participation may alter work and residence.

The mechanism assessment therefore asks when the circumstance began, whether the education event could affect it and whether longitudinal or historical information is available. Where temporal order is uncertain, the report retains associational language.

Each proposed mechanism is tested against credible alternatives. An economic gradient may be partly geographic if poorer households live where schools are sparse. A gender difference may vary by age because safety concerns or work expectations intensify at transition. A language difference may reflect instructional practice, local supply or administrative exclusion rather than language alone.

Alternative explanations are not listed to delay action indefinitely. They direct the choice of response and protect against interventions that address the wrong barrier. Where several mechanisms are supported, a combined response may be necessary.

No single hierarchy suits every policy question. A representative household survey is strong for population distribution when coverage is adequate. Administrative records can show school supply, attendance histories and operational reach but omit non-enrolled children. Qualitative inquiry can establish process and lived barriers but not national prevalence. An evaluation can estimate the effect of a defined intervention under its design conditions.

The evidence record selects sources by the claim required. Convergence across sources with different limitations strengthens an explanation. Disagreement should be investigated through concepts, populations, dates and measurement before one source is declared correct.

Where lawful and technically reliable, household and school information may be linked through protected identifiers or geographic aggregation. The linkage report states the match rate, false-match risk, unmatched characteristics and confidentiality controls. Low match rates among disadvantaged children can reproduce the original exclusion.

Aggregate linkage may compare survey-estimated eligible populations with school places, teacher deployment or facility conditions by area. It supports mechanism assessment without identifying households, but geographic scale and boundary consistency must be checked.

Interviews, observation and community inquiry should use purposive coverage of the groups and locations implicated by the statistical finding. The protocol records recruitment, who may be absent, language, safeguarding and analytic method. Quotations should not be selected solely for force; recurring and divergent accounts both matter.

Children’s accounts require age-appropriate consent and protection. Public reporting should remove identifying detail and should not expose a child to retaliation or stigma. The purpose is to understand institutional and household processes while preserving dignity.

An intervention should state the barrier, affected population, responsible actor, input, service change, expected education response and possible adverse effect. A transfer may reduce cost but fail where no school place exists. New facilities may improve proximity but not participation where instruction is unsafe or inaccessible.

The proposed response should therefore connect directly to evidenced mechanisms. Broad exhortations to households are not a quality-improvement strategy where the binding constraint lies in public provision.

Programme monitoring should measure whether the intended disadvantaged group receives the service. An overall increase in beneficiaries can coexist with lower reach among remote households or children lacking documents. Reach is assessed by eligible population, not simply by the composition of enrolled beneficiaries.

Delivery indicators may include application success, benefit receipt, travel time, accessible facilities, teacher presence, language provision and complaint resolution. Education outcomes follow after service reach and should not replace it.

Responses can create exclusion through complex application, public labelling, unsafe travel, loss of income, displacement of unpaid work to another child or pressure to attend a poor-quality setting. Monitoring should include these foreseeable effects and a means for affected households to report them safely.

Where a targeted programme creates stigma or boundary errors, a universal service with proportionate support may better serve the public interest. Targeting accuracy is not the sole criterion; dignity, administrative feasibility and missed eligibility matter.

The evaluation design follows the decision. A phased introduction, comparison group, interrupted series or matched design may be available. The choice should account for selection into the programme, spillovers, concurrent reforms and ethical constraints. Baseline group disparities should be measured with the same definitions planned for follow-up.

An outcome change among beneficiaries is not an effect estimate without a credible counterfactual. Where causal evaluation is not feasible, process, reach and before-after evidence may still support operational judgement if limitations are explicit.

The move from disparity to action is disciplined when each claim is no stronger than its evidence and each response addresses an identified barrier. Uncertainty may justify staged action and better measurement; it does not justify describing unequal participation as a household failing.

The final policy statement should identify what is known, what remains uncertain, what the responsible institution will change and how distributional effect will be assessed. This is a more accountable use of disaggregated evidence than either unsupported certainty or indefinite qualification.

Education disaggregation can expose sensitive information about poverty, disability, migration, ethnicity, work and school participation. The statistical value of a detailed table must be balanced against the risk that a child, household or small community can be identified directly or by combining published information.

Protection applies to tables, maps, narrative cases, downloadable files and research access. Removal of names is not sufficient where rare characteristics, precise geography or household composition permits recognition.

The disclosure review considers cell size, dominance, rarity, geographic concentration, availability of external lists, sensitivity of the attribute and harm from identification. A cell containing several children may still be identifying if all belong to one known institution or settlement.

The review should distinguish identity disclosure, where a person is recognised, from attribute disclosure, where new sensitive information can be inferred. It should also consider group disclosure: a publication can stigmatise a small community even without naming individuals.

Primary suppression removes cells that fail a confidentiality rule. Secondary suppression prevents reconstruction of a primary cell from totals. Category aggregation may reduce risk but should preserve analytical meaning. Rounding and perturbation can add protection if applied consistently and explained.

Suppression should not be signalled as a zero and should not be used to conceal an adverse result. The table distinguishes confidentiality suppression from statistical unreliability and unavailable data. These reasons require separate symbols.

A safe table can become unsafe when compared with another release. Subtracting overlapping age groups, areas or years may reveal a protected cell. The disclosure officer should review the planned release against previously published tables and foreseeable combinations.

Corrections also require review: publishing both original and revised detailed cells may permit differencing. Version transparency can be maintained while protecting the underlying small count through aggregated explanation.

Fine geographic maps can identify communities and imply greater accuracy than the survey supports. Areas with small samples should be aggregated, suppressed or shown as not estimable. Point locations of households or services associated with sensitive groups should not be published.

Boundary files and labels should match the observation date. Where area names themselves reveal a protected institution or settlement, a broader geography is required.

Case descriptions should remove or alter combinations of age, location, household structure, occupation and school that enable recognition, without changing the substantive evidence. Composite accounts should be clearly described as such and should not be presented as a verbatim individual history.

Direct quotations require consent compatible with the intended publication and should avoid details that expose the speaker. Protection is heightened where a child discusses work, disability, violence, migration status or exclusion by a named service.

A public-use file should remove direct identifiers and review indirect identifiers, rare categories, dates and fine geography. Top- or bottom-coding, category aggregation and controlled perturbation may be required. The accompanying documentation records changes that affect analysis.

Weights may themselves reveal rare population structures or locations and require review. Removing all design information, however, prevents valid estimation. A balance may provide broad strata and protected cluster identifiers sufficient for approximate design-based analysis.

Where detailed data have substantial public value but cannot be released openly, a restricted-access arrangement may permit approved research under legal, physical and technical controls. Applications should specify purpose, variables, outputs and destruction or retention arrangements. Output checking remains necessary.

Restricted access is not a substitute for a useful public release. Core definitions, national findings, limitations and non-disclosive tables should remain available to the public.

The release authority records who conducted the disclosure review, which rule was applied and how statistical usefulness was considered. A clear review trail supports consistent treatment across politically favourable and unfavourable results.

Confidentiality protects participation in public statistics and the safety of children and households. It should be implemented as a professional statistical control, not invoked broadly to avoid reporting inequalities that can be disclosed safely.

This contract defines the minimum agreement among the human-readable table, structured report object and downloadable comma-separated file. It permits a public reader, an analyst and an archival system to reach the same value, label and limitation from the same release.

The contract does not prescribe a software platform. It specifies content and behaviour that should survive transfer between systems and remain intelligible without proprietary tools.

Each report has a document number, title, publication date, evidence cut-off and version. Each table has a stable identifier unique within the report. A corrected version retains the document identity while recording the new version date and the nature of the correction.

Table identifiers should not be reused for a different concept. If a replacement changes the population, outcome or grouping materially, the version note must make that change visible.

Every table consists of a title, column definitions, ordered rows, notes, source statement and a link to its downloadable representation. Column definitions include a stable key and a public label. Cells contain text or numeric values without hidden formatting as meaning.

Spanning headers may improve presentation but should not be the only location of a variable definition. A flat downloadable header must remain intelligible. Blank, zero, not applicable, unavailable and suppressed are distinct states and require controlled symbols or explicit null values.

The analytic value is retained at sufficient precision for reproduction. The visible value follows an established rounding rule. Percentages should identify whether they are stored as 0–1 proportions or 0–100 values; the same report should not mix conventions invisibly.

Thousands separators, decimal marks and percent signs belong to presentation and should not corrupt the downloadable numeric field. If a cell contains an interval, separate lower and upper numeric fields are preferable to a single textual string in the structured object.

Row and column labels should carry the measured concept, population unit and period where these are not unambiguous from the title. Abbreviations are defined in notes. Symbols for precision and suppression are not reused for unrelated meanings.

Notes follow a consistent order: population and denominator; outcome definition; grouping construction; weighting and precision; missing and excluded data; suppression; rounding; source. A table should not depend on a note located only elsewhere in the report for a condition that changes interpretation.

The comma-separated file is encoded in UTF-8 and contains one header row followed by one row for each visible data row. Fields containing commas, quotation marks or line breaks are enclosed and escaped according to the file convention. The order of columns and rows matches the public table.

Machine-oriented keys may be included in a separate structured object, while the CSV header uses comprehensible labels. If keys are included in the CSV, documentation maps them to the visible labels. Formula cells are exported as values, not executable expressions.

The reading presentation should preserve header association and allow horizontal movement where a table is wider than the display. It should not shrink text below a readable size or convert the table into an image. A concise caption and notes remain adjacent to the table.

Where a very wide table cannot be understood on a narrow display, it may be divided into substantively coherent panels. Each panel repeats the necessary group identifier. The downloadable file retains the full table and the relationship among panels.

The source statement identifies the data source and analysis responsibility without implying endorsement by the source institution. Constructed examples are labelled as such and are not included in a table that could be mistaken for observed national data.

Preliminary, revised and final status should be visible. A final label means that release controls were completed; it does not mean that the statistic is free of the stated uncertainty and limitations.

The parity test compares title, headers, row order, cell values, null states, notes and source across the visible table, structured object and CSV. Comparison is cell-level and should occur after generation of the final publication files.

A correction records the affected table and cells, original and replacement values where disclosure permits, reason, date and effect on the findings. The corrected CSV and structured object are regenerated from the same authoritative source. Replacing only the visible text creates inconsistent evidence.

The prior release remains archived and clearly superseded. Users who obtained the earlier downloadable file should be able to identify that a correction exists through the document number and version record.

The release package contains source metadata, the report text, structured reader object, references, one CSV for each table, the automated integrity audit and the professional review record. Analytic source files may reside in a protected archive but should be linked through an internal provenance identifier.

Long-term intelligibility requires controlled definitions as well as files. The archive therefore retains code lists, questionnaire wording, welfare-index construction, weight method, variance method and suppression rules. A table without these materials may remain technically readable while losing its substantive meaning.

Release authority is granted only when statistical calculation, evidence dates, language, table parity, confidentiality and population claims have all been reviewed. A passed file test cannot compensate for an unsupported policy conclusion, and a strong narrative cannot compensate for conflicting tables.

The contract treats publication as the last stage of research quality. It makes the public record reproducible, protects against silent alteration and ensures that disaggregated findings remain usable for scrutiny and improvement after the immediate policy debate has passed.

The release should enable a reader to move from a principal statement to the relevant table, definition, source and limitation without reconstructing the analysis independently. This route begins with a cited finding in the narrative, identifies the table and population, and leads to the methodological note and downloadable values. Cross-references should use stable section and table identifiers rather than page numbers alone, because pagination may change across reading formats.

Questions of interpretation should be answerable from the publication package: whether the estimate concerns enrolment or attendance; which children enter the denominator; how household position was constructed; which population was excluded; and how uncertainty was assessed. Where an answer depends on restricted material, the public note should identify the nature of that material and the reason for restriction.

The reader route is tested by a reviewer who did not prepare the analysis. The reviewer selects each headline finding and independently locates its value, denominator, comparison and evidence qualification. Failure to complete this route indicates a publication defect even if the underlying calculation is correct.

When files move between an editorial repository, a publication service and an archive, transfer validation should compare byte-level checks or another controlled integrity record. The objective is to detect truncation, character conversion, replacement of a tested table or loss of notes. The validation concerns the released content, not the internal design of the receiving system.

Special attention is required for quotation marks, dashes, mathematical symbols, non-English names and line breaks within CSV fields. A file that opens in one desktop application may still parse incorrectly under a different regional setting. The acceptance test should therefore use a standards-conforming parser and should verify every row, not rely on visual inspection of the first lines.

If the receiving environment transforms Markdown into another presentation form, headings, lists, formulas, captions and table notes should remain in their declared order. Links to references and downloads should resolve to the version included in the release package. A transformation that changes a numeric value, drops a qualification or converts a null to zero fails acceptance.

Education structures, geographic boundaries, household concepts and welfare measures can change after publication. The archive should preserve the definitions current at the evidence cut-off rather than silently update historical text to later terminology. A later explanatory note may be attached, but it should be dated and kept distinct from the original finding.

Machine-readable fields should likewise preserve historical category labels and code lists. If an archive maps them to a later classification, both the original and mapped values should be retained with the correspondence rule. The mapping must not imply that the two categories are identical where boundaries or eligibility changed.

Persistent interpretation also requires preservation of negative information: categories not collected, populations outside the frame, cells suppressed and estimates judged unfit for release. Removing these states from an archive can make later users believe the evidence was complete.

The final sign-off identifies the accountable research lead, statistical reviewer, evidence reviewer, editorial reviewer and disclosure reviewer. Each confirms a defined aspect of the release rather than offering a general approval. Unresolved reservations are recorded with the decision taken and the person responsible for that decision.

The statistical reviewer confirms population, estimator, weights, precision and cell reproduction. The evidence reviewer confirms that cited sources were available by the cut-off and support the claim made. The editorial reviewer confirms that findings, qualifications and public responsibilities are expressed accurately and without prejudicial group language. The disclosure reviewer confirms that the combined release does not expose protected persons or communities.

Sign-off is repeated after a material correction or replacement of a table. Minor typographic correction may follow a proportionate route, provided it cannot alter meaning. The dated record completes the chain from household observation to public claim and establishes who was responsible for the integrity of the released evidence.