A useful average can still conceal its components
Among currently employed U.S. adults aged sixteen to sixty-five in the 2017 PIAAC household survey, a single indicator of reading books at least weekly in either work or everyday life combines three different groups. Their pooled assessed-literacy mean is 281.5 points. That number is a valid description of the combined population, not an erroneous average. Yet it does not show that the work-only component has a mean of 261.5, while the everyday-only and both-context components have means of 284.7 and 284.6. These are historical, unadjusted group summaries, not individual assessments or effects of reading.
The distinction is not simply that a national average hides variation. The pooling rule is explicit: weekly book reading in at least one of two separately recorded contexts becomes a single positive category. Within that category, everyday-only readers account for 65.1%, work-only readers for 13.6%, and readers reporting both contexts for 21.3%. Consequently, the pooled result is numerically much closer to the two larger, higher-mean components than to the smaller work-only component. Calling everyone in that pool a frequent book reader preserves frequency in either context but discards which context supplied it.
This study measures that descriptive coarsening with a fixed set of comparisons. It retains all four joint reading states, reports their composition, and examines a mean and one lower-tail proficiency share. It does not estimate a complete proficiency distribution or invent a general information-loss index. Nor does it argue that researchers should never combine contexts. The question is what a particular combined statistic no longer tells a reader, and how much uncertainty accompanies the component descriptions that restore that information.
Two book questions, one historical worker population
The source is the U.S. 2017 household public-use file from the Programme for the International Assessment of Adult Competencies, or PIAAC. The present analysis selects people currently employed and aged sixteen to sixty-five, with valid answers to both book-frequency questions and a positive finite final survey weight. The resulting reading-response domain contains 2,395 records. It is not a description of all U.S. adults, current conditions, schoolchildren, unemployed people, or everyone who has held a job. The employment definition also reaches qualifying self-employment and family-business work, not employees alone.
The work item, G_Q01E, asks how often respondents read books in their job. Official questionnaire routing links the selected current workers to their current rather than a previous job. The everyday item, H_Q01E, concerns reading books, fiction or nonfiction, outside work. For respondents enrolled in education, the instructions explicitly include reading for studies. Everyday therefore does not mean pleasure-only, reading physically at home, or reading unrelated to education. Both contexts can include electronic displays; neither item identifies a particular book, genre, language or amount of text.
Both items offer five frequency categories. This study fixes at least weekly as the two highest responses and below weekly as the other three valid responses. Below weekly includes never, but is not synonymous with never. Missing, refused, unknown and routed responses are not turned into nonreading. Another field, H_Q01E_T, is a collapsed trend recode of the everyday item, not a second independent report. Treating that recode as a separate instrument would manufacture a comparison that the underlying observations do not provide.
The four states and the composition of the combined group
In the selected worker population, 53.3% report book reading below weekly in both contexts, 30.4% at least weekly only outside work, 6.4% at least weekly only at work, and 10.0% at least weekly in both. These percentages are weighted population descriptions, not raw sample percentages. Rounding means the displayed components need not sum to exactly one hundred. The four groups describe a joint frequency classification; none identifies lifelong readers, permanent nonreaders, reading quality, or a fixed personal identity.
Conditioning on weekly reading in either context changes the denominator. Within that combined group, the estimated component shares are 65.1% everyday-only, with an approximate 95% interval of 61.5% to 68.4%; 13.6% work-only, with an interval of 11.4% to 16.2%; and 21.3% both, with an interval of 18.4% to 24.6%. These conditional shares explain the mixture represented by the pooled proficiency statistic. They should not be confused with the corresponding shares among all eligible workers.
The unweighted group counts are 1,234 below weekly in both contexts, 765 everyday-only, 157 work-only and 239 both, giving 1,161 records in the weekly-in-either pool. Each group satisfies the study's predetermined minimum of sixty-two observations, Kish effective support of at least thirty, and positive denominators in every active replicate. The Kish measure describes dispersion of survey weights, not a guarantee of design-adjusted precision. Meeting these support checks permits estimation; it does not make every component equally precise or remove the need to show intervals.
What the assessed-literacy means describe
The everyday-only mean is 284.7 points, with an approximate pointwise 95% interval from 280.4 to 289.0. The work-only mean is 261.5, with an interval from 252.3 to 270.6. For weekly reading in both contexts, the mean is 284.6, with an interval from 277.2 to 292.0. The pooled weekly-in-either mean is 281.5, with an interval from 277.8 to 285.3. The below-weekly-in-both reference group has a mean of 269.4, with an interval from 265.9 to 272.9.
These numbers summarize assessed literacy in English, not intelligence, personal worth, literary taste or the quality of books people choose. PIAAC measures adults' ability to understand and use written information through a structured assessment. The context questions are self-reported frequencies, not experiments that assign books or reading conditions. Differences may reflect educational histories, language experiences, selection into jobs, job tasks and other characteristics. A lower group mean cannot establish that reading at work reduces proficiency, just as a higher one cannot establish a benefit caused by leisure reading.
The mean interval widths also differ substantially: 8.7 score points for everyday-only, 18.3 for work-only, and 14.8 for both contexts. The work-only component is the smallest of the three and has the widest mean interval. The pooled interval is narrower, at 7.5 points, but that precision concerns the pooled target, not every component within it. Replacing each component with the pooled mean would supply a more precise answer to a different descriptive question, rather than resolve uncertainty about the component itself.
One lower-tail summary adds a different perspective
The second proficiency summary is the share below a score of 226, corresponding to Level 1 or below under the official continuous boundary. The threshold is applied separately to each plausible value before population estimates are pooled. It is not applied to an individual's average plausible value, and no person is labelled as having a diagnosed level. This single lower-tail measure complements the mean, but the pair still does not describe an entire distribution or show how every part of the scale differs.
The estimated share at Level 1 or below is 11.8% for everyday-only, with an approximate interval of 9.2% to 15.2%; 23.3% for work-only, with an interval of 15.0% to 34.3%; and 11.3% for both, with an interval of 7.1% to 17.5%. The pooled frequent group has a share of 13.3%, with an interval of 10.6% to 16.5%. The below-weekly-in-both reference group has a share of 19.1%, with an interval of 16.3% to 22.3%.
The work-only lower-tail interval is appreciably wider than those for the other frequent-context components. That uncertainty remains visible even though its point estimate differs from the pool. The descriptive component range is 11.3% to 23.3%; it is not a confidence interval, a ranking procedure, or an independently selected endpoint. All specified component results are retained. There is no search over additional cutpoints, age subgroups, occupations or survey years to find a stronger contrast after inspecting these results.
Fixed contrasts include differences compatible with zero
The prespecified everyday-only minus work-only comparison is 23.3 score points, with an approximate interval from 13.8 to 32.7. For the Level 1-or-below share, the corresponding difference is −11.5 percentage points, with an interval from −21.2 to −1.8. These are descriptive contrasts within the selected worker population. Their intervals account for the survey replication and plausible-value uncertainty, but are not adjusted for multiple comparisons and do not support a family-wide confirmatory claim or a causal interpretation.
Comparing each component with the pooled frequent group answers the coarsening question more directly. Everyday-only is 3.2 score points above the pool, with an interval from 0.6 to 5.8, and its lower-tail share is 1.4 percentage points lower, with an interval from −3.5 to 0.6. Work-only is 20.1 score points below the pool, with an interval from −27.9 to −12.2, and its lower-tail share is 10.0 points higher, with an interval from 1.9 to 18.1. The latter component is not well represented by simply quoting the pooled point estimates as though they described every context.
The both-context component is 3.1 score points above the pool, with an interval from −3.4 to 9.5, and 2.0 percentage points lower on the lower-tail measure, with an interval from −6.3 to 2.3. These intervals include zero, as does the everyday-only lower-tail contrast. They are reported rather than discarded as inconvenient. An interval including zero is not proof of equivalence, and a nonzero interval is not proof of a mechanism. The complete pattern is more informative than a headline selecting only comparisons that look especially different.
Why pooling is not itself a statistical error
For each proficiency summary, the combined frequent-group estimate equals the component estimates weighted by their shares in the assessed frequent population. This identity holds before rounding. The calculation checks it for all ten plausible values and every full-weight and replicate calculation, rather than verifying only the final average. The identity explains why a legitimate pool can lie close to some components and far from another. It does not make the pooled result biased, and it does not designate any component as the correct population for every research purpose.
The denominator matters to that identity. Reading-response shares are defined among eligible respondents with both frequency answers. Proficiency summaries and mixture weights are defined among those also having all ten finite literacy plausible values. In this selected sample, both domains contain the same 2,395 records, so there is no additional assessment loss overall or within any context. That observed equality does not justify conflating the definitions. In another release or population, using broader background shares to reconstruct assessed-domain means could fail.
A component and its pool share observations, so their estimation errors are dependent. The analysis calculates each component-minus-pool contrast within every replicate and plausible value, preserving that covariance before pooling uncertainty. Adding two independently treated standard errors would not represent the actual comparison. This is especially important when a large component strongly influences the pool. The goal is a transparent description of the chosen recode, not a claim that finer classifications are always better, or that every possible difference deserves a separate study.
A planned composition adjustment was not supported
Before examining proficiency results, the design specified a descriptive standardization across three age groups and three education groups. The age bands were sixteen to thirty-four, thirty-five to fifty-four and fifty-five to sixty-five. Education groups followed the source's ISCED-based EDCAT6 codes, rather than assumed U.S. diploma equivalences. The proposed target was the common age-by-education distribution of assessed frequent readers with valid education. This would have compared the three contexts under the same observed composition, without making the comparison causal.
The plan required every one of the twenty-seven context-by-cell combinations to have at least twenty observations, Kish effective support of at least ten and positive replicate denominators. Ten cells failed that fixed gate. Therefore the entire standardized sensitivity analysis was omitted. Cells were not merged, a regression was not substituted, and an age-only comparison was not introduced after seeing the failure. The internal cell counts were used to decide whether the planned estimator was supported, not to publish small-cell proficiency results.
Consequently, the reported contrasts remain unadjusted for age and education. The study cannot determine how much of the context pattern those characteristics explain, let alone isolate a direct relationship with book-reading context. Reporting this omission is part of the result's interpretation, not an administrative footnote. The unsupported sensitivity does not erase the supported unadjusted description, but it sharply limits claims that components would remain different under a common composition. No adjusted or composition-independent result is implied.
Replicate weights and plausible values serve different purposes
The file contains eighty replicate-weight fields, but only forty-five are active for this U.S. design. The remaining thirty-five exactly repeat the full-sample weight. The paired-jackknife variance uses the sum of squared deviations of active replicate estimates from the full-weight estimate, not their average and not an eighty-replicate bootstrap formula. As a source check, separate reproductions match published employment and book-frequency marginal percentages and their standard errors. These benchmarks validate the basic weighting implementation before the new joint-context summaries are interpreted.
Literacy proficiency additionally uses ten plausible values. Each is suitable for population estimation; none should be treated as a definitive individual test score. For each value, all estimates and contrasts are recalculated with the full weight and forty-five replicates. Total variance combines the average within-value sampling variance with the between-value sample variance multiplied by 1.1. Proportion intervals use a delta-method logit transformation, while differences of proportions remain on their difference scale. Means and mean contrasts use approximate symmetric intervals in score points.
The displayed estimates, standard errors and interval endpoints are rounded only after calculation. The study's minimum counts, Kish thresholds and precision flags are declared analytical rules, not represented as a universal official guarantee. Positive proportions receive a caution if relative standard error exceeds thirty percent; very wide proportion intervals receive an imprecision flag. None of these rules suppresses a near-zero signed contrast merely because a relative error is unhelpful there. Boundary or degenerate intervals would be marked unavailable rather than presented as exact certainty.
An extension of existing research, with a reproducible boundary
Researchers already combine and distinguish work and everyday reading for different purposes. Reder, Gauly and Lechner's 2020 longitudinal work uses the higher task frequency across contexts in engagement measures. Tighe and colleagues' 2024 study combines work and home indices in profiles of U.S. adults with lower literacy and explicitly discusses further context-specific analysis. Kyröläinen and Kuperman's 2021 international analysis includes separate work and home reading indices. This study does not discover context separation, refute those designs, or claim the first connection between reading practices and proficiency.
Its narrower contribution is a fully specified audit of one book-frequency recode within one historical worker sample: which components enter the pool, how two assessed-literacy summaries differ from that pool, and how much uncertainty accompanies the fixed comparisons. The everyday and work questions do not measure reading quality, language or the contents of a book. Assessment is in English, and selection into both work and reading practices remains unaddressed. No recommendation that changing a reading context will produce a particular proficiency gain follows from these cross-sectional data.
The public reproduction program accepts the original permitted NCES archive and produces screened, rounded aggregates only. Source fingerprints, methods and attribution accompany the code; respondent records, weights and individual values are not redistributed. The practical conclusion is that a weekly-in-either-context average can be useful while leaving out substantively different components. A careful reader should ask what the indicator combines, retain the valid pooled target, inspect the fixed component comparisons and their intervals, and remember that the planned composition adjustment was not supported by these data.
Book-reading contexts and assessed literacy
United States · PIAAC 2017 · Currently employed ages 16–65
Full chart and source
Scroll horizontally to inspect the chart. ↔
Chart and values
Scroll horizontally to see all columns.
| Context | Estimate (score points) | SE | Lower 95 | Upper 95 | Interval width |
|---|---|---|---|---|---|
| Below weekly in both contexts | 269.4 | 1.8 | 265.9 | 272.9 | 6.9 |
| Weekly everyday only | 284.7 | 2.2 | 280.4 | 289.0 | 8.7 |
| Weekly at work only | 261.5 | 4.7 | 252.3 | 270.6 | 18.3 |
| Weekly in both contexts | 284.6 | 3.8 | 277.2 | 292.0 | 14.8 |
| Weekly in either context (pooled) | 281.5 | 1.9 | 277.8 | 285.3 | 7.5 |
A assessed domain. Ten plausible values and 45 active JK2 replicates; approximate pointwise 95% intervals, not multiplicity-adjusted. No individual diagnosis or causal inference. Below 226 applied separately per plausible value for the lower-tail share. Interval widths are calculated before display rounding; subtracting rounded endpoints may not reproduce the displayed width.
Book reading: four fixed mean contrasts
United States · PIAAC 2017 · Currently employed ages 16–65
Full chart and source
Scroll horizontally to inspect the chart. ↔
Chart and values
Scroll horizontally to see all columns.
| Comparison | Difference (score points) | SE | Lower 95 | Upper 95 | Interval width |
|---|---|---|---|---|---|
| Everyday-only minus work-only | 23.3 | 4.8 | 13.8 | 32.7 | 18.9 |
| Everyday-only minus pool | 3.2 | 1.3 | 0.6 | 5.8 | 5.1 |
| Work-only minus pool | −20.1 | 4.0 | −27.9 | −12.2 | 15.7 |
| Both-context minus pool | 3.1 | 3.3 | −3.4 | 9.5 | 12.9 |
Joint contrasts within each replicate and plausible value preserve covariance, including overlapping component/pool groups. Approximate pointwise 95% intervals; null-compatible results retained. Difference scale, not independent-interval subtraction. Interval widths are calculated before display rounding; subtracting rounded endpoints may not reproduce the displayed width.
Data & downloads
Scroll horizontally to see all columns.
| Reading context | Records | Share (%) | SE (pp) | Lower 95 (%) | Upper 95 (%) |
|---|---|---|---|---|---|
| Below weekly in both contexts | 1,234 | 53.3 | 1.1 | 51.2 | 55.4 |
| Weekly everyday only | 765 | 30.4 | 1.0 | 28.5 | 32.4 |
| Weekly at work only | 157 | 6.4 | 0.6 | 5.3 | 7.5 |
| Weekly in both contexts | 239 | 10.0 | 0.8 | 8.4 | 11.7 |
B: 2,395 current workers ages 16–65 with both valid frequency responses. Below weekly is not never. Weighted percentages and approximate pointwise 95% intervals; displayed values may not sum exactly after rounding. Counts are unweighted.
Scroll horizontally to see all columns.
| Frequent context | B share (%) | B lower95 (%) | B upper95 (%) | A mixture weight (%) |
|---|---|---|---|---|
| Weekly everyday only | 65.1 | 61.5 | 68.4 | 65.1 |
| Weekly at work only | 13.6 | 11.4 | 16.2 | 13.6 |
| Weekly in both contexts | 21.3 | 18.4 | 24.6 | 21.3 |
Conditional shares among frequent readers. B background and A assessed domains both have 1,161 records here; their definitions remain distinct. Assessed-domain shares weight proficiency mixture identities.
Scroll horizontally to see all columns.
| Context | Estimate (score points) | SE | Lower 95 | Upper 95 | Interval width |
|---|---|---|---|---|---|
| Below weekly in both contexts | 269.4 | 1.8 | 265.9 | 272.9 | 6.9 |
| Weekly everyday only | 284.7 | 2.2 | 280.4 | 289.0 | 8.7 |
| Weekly at work only | 261.5 | 4.7 | 252.3 | 270.6 | 18.3 |
| Weekly in both contexts | 284.6 | 3.8 | 277.2 | 292.0 | 14.8 |
| Weekly in either context (pooled) | 281.5 | 1.9 | 277.8 | 285.3 | 7.5 |
A assessed domain. Ten plausible values and 45 active JK2 replicates; approximate pointwise 95% intervals, not multiplicity-adjusted. No individual diagnosis or causal inference. Below 226 applied separately per plausible value for the lower-tail share. Interval widths are calculated before display rounding; subtracting rounded endpoints may not reproduce the displayed width.
Scroll horizontally to see all columns.
| Context | Estimate (percent) | SE | Lower 95 | Upper 95 | Interval width |
|---|---|---|---|---|---|
| Below weekly in both contexts | 19.1 | 1.5 | 16.3 | 22.3 | 6.0 |
| Weekly everyday only | 11.8 | 1.5 | 9.2 | 15.2 | 6.0 |
| Weekly at work only | 23.3 | 4.9 | 15.0 | 34.3 | 19.3 |
| Weekly in both contexts | 11.3 | 2.6 | 7.1 | 17.5 | 10.4 |
| Weekly in either context (pooled) | 13.3 | 1.5 | 10.6 | 16.5 | 5.9 |
A assessed domain. Ten plausible values and 45 active JK2 replicates; approximate pointwise 95% intervals, not multiplicity-adjusted. No individual diagnosis or causal inference. Below 226 applied separately per plausible value for the lower-tail share. Interval widths are calculated before display rounding; subtracting rounded endpoints may not reproduce the displayed width.
Scroll horizontally to see all columns.
| Comparison | Difference (score points) | SE | Lower 95 | Upper 95 | Interval width |
|---|---|---|---|---|---|
| Everyday-only minus work-only | 23.3 | 4.8 | 13.8 | 32.7 | 18.9 |
| Everyday-only minus pool | 3.2 | 1.3 | 0.6 | 5.8 | 5.1 |
| Work-only minus pool | −20.1 | 4.0 | −27.9 | −12.2 | 15.7 |
| Both-context minus pool | 3.1 | 3.3 | −3.4 | 9.5 | 12.9 |
Joint contrasts within each replicate and plausible value preserve covariance, including overlapping component/pool groups. Approximate pointwise 95% intervals; null-compatible results retained. Difference scale, not independent-interval subtraction. Interval widths are calculated before display rounding; subtracting rounded endpoints may not reproduce the displayed width.
Scroll horizontally to see all columns.
| Comparison | Difference (percentage points) | SE | Lower 95 | Upper 95 | Interval width |
|---|---|---|---|---|---|
| Everyday-only minus work-only | −11.5 | 5.0 | −21.2 | −1.8 | 19.4 |
| Everyday-only minus pool | −1.4 | 1.1 | −3.5 | 0.6 | 4.1 |
| Work-only minus pool | 10.0 | 4.1 | 1.9 | 18.1 | 16.2 |
| Both-context minus pool | −2.0 | 2.2 | −6.3 | 2.3 | 8.7 |
Joint contrasts within each replicate and plausible value preserve covariance, including overlapping component/pool groups. Approximate pointwise 95% intervals; null-compatible results retained. Difference scale, not independent-interval subtraction. Interval widths are calculated before display rounding; subtracting rounded endpoints may not reproduce the displayed width.
Evidence & method
How we calculated it
- Historical 2017 U.S. household PUF; current workers ages 16–65 with two valid book-frequency answers and positive full weight.
- B reading-response and A all-ten-PV-assessed domains separately defined; both 2,395 here, with no additional assessment loss.
- Weekly=responses 4–5; below weekly=1–3. Everyday outside work includes studies; current-job routing verified.
- All 45 active JK2 replicates, sum squared deviations; ten plausible values with mean within variance plus 1.1×between sample variance.
- Mean and PV<226 share for all four states and frequent pool; fixed everyday-only/work-only and component/pool contrasts calculated jointly.
- 920 mixture identities across summaries/PVs/replicates; fixed 27-cell standardization failed and entirely omitted.
- Study support, precision and disclosure gates applied; public outputs rounded only after calculation; no respondent redistribution.
What the data cannot tell us
- Historical selected workers, not all adults or current conditions. Frequencies self-reported; below weekly is not nonreading.
- Everyday is not pleasure-only or necessarily at home; no book, genre, material amount or reading-language identification.
- English proficiency assessment; plausible values support population estimation, not individual diagnosis.
- Unadjusted comparisons cannot distinguish age, education, job demands, selection, language experiences or other explanations.
- Pool is a valid combined target, not aggregation bias. Two summaries are not full distributions.
- Approximate pointwise intervals are not multiplicity-adjusted; null-compatible contrasts do not establish equivalence.
- Ten failed standardization cells; no adjusted results, cell merging or regression rescue. No causal reading-effect recommendation.
Original sources
- NCES: PIAAC 2017 original ASCII public-use archive
Original archive supplies DAT and SAS schema. Source and entry fingerprints checked; public reproduction requires the reader’s original permitted archive. No respondent data redistributed.
Accessed October 30, 2026
- NCES: U.S. PIAAC 2012/2014/2017 technical report
Chapter 8 and Appendix E establish 45 active JK2 replicates, placeholder weights, ten plausible values and variance pooling; continuous Level 1 boundary is below 226.
Accessed October 30, 2026
- NCES: 2017 household background questionnaire
Current-worker routing selects current job; everyday reading excludes work but includes study for enrolled respondents. Context is not pleasure-only or a physical location.
Accessed October 30, 2026
- NCES PIAAC data usage agreement
Statistical use only; no identification attempts or linkage to individually identifiable records. Public outputs contain screened aggregates, not respondent records.
Accessed October 30, 2026
- NCES: 2017 background compendium
Official weighted marginal estimates and standard errors provide separate benchmark checks for employment and both book-frequency items. They do not supply the new joint-context proficiency results.
Accessed October 30, 2026
- NCES: PIAAC data files
Official public-use source listing distinguishes accessible household files for statistical analysis from licensed restricted-use files.
Accessed October 30, 2026
- Reder, Gauly and Lechner (2020): Practice makes perfect
Prior longitudinal research combines task frequencies across work and outside-work settings. The present historical descriptive coarsening audit does not refute that choice or establish a causal practice effect.
Accessed October 30, 2026
- Tighe and colleagues (2024): Profiles of skills-use engagement
Prior U.S. low-literacy profile study combines work/home indices and discusses context- and item-specific research. Context separation is not claimed as a newly discovered research gap.
Accessed October 30, 2026
- Kyröläinen and Kuperman (2021): Predictors of literacy in adulthood
Prior international research includes separate work and home reading indices. This study focuses on one explicit book-item recode, not first evidence that reading contexts differ.
Accessed October 30, 2026
Cite & reuse this study
Alisson Bento. What a weekly book-reader average leaves out: work and everyday contexts. Secondary analysis of NCES PIAAC U.S. 2017 household public-use microdata. Sources accessed October 30, 2026. https://alissonbento.com/research/us-book-reading-context-coarsening-and-literacy-2017/
Source information: National Center for Education Statistics, PIAAC U.S. 2017 household public-use file. Statistical-use/no-identification conditions apply. Original selection, calculation and explanation: Alisson Bento. No NCES or OECD endorsement, agency insignia, respondent records or third-party article figures reproduced.
Corrections or editorial questions