# Methods — Novels and short stories among U.S. book readers, 2017–2022

This document describes the survey definitions, calculations and limitations of the study. Observed absence of fiction does not establish commercial opportunity or purchase intent.

## Distinct research question
Among U.S. adults reporting reading books in the preceding 12 months, what proportion also report reading novels or short stories, and how did this association vary by age and education between July 2017 and July 2022? This is annual retrospective self/proxy-reported survey participation, not ATUS diary minutes or reading-day frequency.

## Observed results
Complete-case, supplement-weighted respondent intersections:

| Year | Book readers answering fiction yes/no (n) | Both yes (n) | Fiction among book readers | 95% replicate-weight CI |
|---|---:|---:|---:|---:|
| 2017 | 4,813 | 3,561 | 72.46% | 70.67–74.25% |
| 2022 | 3,336 | 2,372 | 69.33% | 67.14–71.52% |

Difference: −3.13 percentage points (95% CI −5.96 to −0.30), assuming independent cross-sections. This is an unadjusted single overall contrast, not a causal effect. At the same time, 7.92% of complete-case non-book readers reported fiction in 2017 and 7.65% in 2022. The fiction set is demonstrably NOT nested inside the book set. Dividing published marginal fiction percentages by published book percentages is invalid.

The 24 age and education cells across the two waves, including estimates, sample sizes, standard errors, and intervals, are in `results.csv`; cross-wave contrasts in `changes.csv`. Small groups are imprecise. No multiple-comparison adjustment has been applied; subgroup differences should remain exploratory. In particular, age 18–24 has only 183 book-reader complete cases in 2022 and a wide change interval. Do not relabel age groups as generations.

Direct standardization separately to the 2017 complete-case book-reader age composition gives 69.36% for 2022; to its education composition, 68.89%. These are separate descriptive adjustments, not joint age-and-education regression. The reference distribution is treated as fixed for the reported standard errors in `standardized.csv`; these are not unconditional uncertainty estimates for an estimated-reference contrast and should not support an adjusted significance claim.

## Compatible units, questions, and denominator
Both variables belong to Core 1 in both waves. `PEC1Q14A` asks about books read in the previous 12 months excluding books required for work or school; the 2022 questionnaire expressly adds print/electronic books, whereas the 2017 questionnaire uses a religious-text clarification. This is a wording difference, not identical measurement. `PEC1Q15A` separately asks about novels/short stories and expressly includes books, magazines, newspapers, paper or online. The fiction question itself does not repeat the book question's work/school exclusion; do not silently claim that it does. Neither question is a romance question or a purchasing question, nor is fiction necessarily read in book form. Specific listening questions are not included in our estimand; respondents may nevertheless have understood “reading” broadly.

We restrict to age >=18 and positive supplement weight. For the main conditional estimate, Q14A must equal 1 and Q15A must equal 1 or 2. Numerator is Q14A=1 AND Q15A=1; denominator includes the same persons' valid responses. This is P(fiction=yes | book=yes, fiction response observed), not a ratio of separate marginal prevalence. Respondents with missing fiction are not reclassified as non-readers. Core 2 structural skips are not treated as No. Other denominator variants (both valid, nonbook readers) are explicitly named in outputs.

All 2017 and 2022 valid intersections are observed directly. The analysis script generates the mutually exclusive response cross-tab including negative codes as `audit.json`; this generated diagnostic is not a listed public download. `-1,-1` structural-universe exclusions dominate missing cells; do not conflate them with refusal. Questionnaire layouts and actual respondent responses support these universe distinctions.

## Missingness sensitivity
2017 has 4,848 observed book readers, of whom 35 lack valid fiction; 2022 has 3,359, of whom 23 lack valid fiction. Their weighted shares are 1.38% and 0.97%. Assigning all missing fiction first No then Yes gives 2017 bounds 71.46–72.84%; 2022 bounds 68.66–69.62%. These are worst-case item-response bounds only, not confidence intervals, and do not repair survey nonresponse. Complete-case marginal book prevalence (52.44% in 2017) is not the NEA published book prevalence (52.7%) because we require valid fiction as well. This intentional denominator difference must stay visible.

## Important proxy and wording sensitivity
The 2017 interviewer instruction allows proxies under restricted callback/closeout/irritation circumstances; the 2022 instruction permits proxies whenever the respondent is unavailable. Among complete-case book readers, the weighted proxy-response share rose from 19.60% to 28.22%. Three 2022 book-reader complete cases have an unknown proxy flag. The main estimate includes them.

An independent raw-file recomputation using PUNXTPR3=1 gives self-response-only estimates 73.48% (2017; n=4,124) and 71.45% (2022; n=2,660). Their change is −2.03 pp, 95% CI −5.06 to +1.01: it is NOT distinguishable from zero at 95% confidence. This sensitivity is not a superior population estimate (self respondents are a selected subgroup), but materially limits an unqualified interpretation of the all-response change as behavioral decline. Keep this caveat alongside any headline year comparison, not only in footnotes. The separate self-response sensitivity script and its output are not included in the listed public downloads; `analyze.py` reproduces the main all-response estimates, not this proxy sensitivity. The altered print/electronic clarification is an additional cross-wave measurement limitation.

## Computation and uncertainty
Download `download.py`, `analyze.py`, and `source-manifest.json` into a working directory. Keep a separate copy of the supplied manifest before running the downloader: `download.py` writes a new manifest and reuses any source files already present without checking their expected hashes. Install pandas and NumPy in your Python 3 environment, then run from the directory containing the scripts:

```sh
python download.py
python analyze.py
```

The downloader retrieves official source files into its own directory. Check its manifest for download errors and compare the source byte lengths and SHA-256 hashes with the saved supplied manifest before analysis. Raw source files are not bundled. The analysis requires `jul17pub.zip`, `jul17rep.zip`, `jul22pub.csv`, and `jul22rep.csv` beside `analyze.py`; it leaves those raw files unmodified. Script paths are resolved relative to each script, so an explicit path to the script also works from another directory.

The analysis writes or overwrites `results.csv`, `changes.csv`, `missingness_bounds.csv`, `standardized.csv`, and the generated diagnostic `audit.json` beside the script. The nine listed public downloads are this `README.md`, `download.py`, `analyze.py`, `source-manifest.json`, `claimset.json`, and the four result CSVs; `claimset.json` is a supplied summary, not an output regenerated by the analysis. For 2017, record-layout positions are 1-indexed: QSTNUM 815–819, OCCURNUM 820–821, age 122–123, education 137–138, book 1045–1046, fiction 1050–1051, supplement weight 1484–1493. 2022 uses Census CSV columns. Note fiction position changes between waves; never reuse the 2017 fixed-width fiction offset for 2022.

Replicate weights join uniquely on QSTNUM/OCCURNUM and are asserted present for all positive-weight adults. The script prints the unique `PWSUPWGT / replicate0` ratios; equality was verified for the analyzed files. Raw weight values have four implied decimals; normalization cancels that common factor for every reported ratio. Core 1 population counts would require the documented factor 2 and weight division by 10,000; we deliberately report no population counts. The factor 2 also cancels ratios. Raw weight sums are diagnostics in implied-decimal units, not population counts. No ordinary CPS labor-force weight is substituted.

For each ratio, recalculate numerator and denominator with each of the 160 supplement replicate weights. Variance is `(4/160)*sum((replicate_estimate-full_estimate)^2)`; intervals are normal 1.96 SE intervals. This matches NEA Technical Appendix's BRR Fay(0.5), mse specification. Cross-wave SE is sqrt(SE2017²+SE2022²), independent repeated samples five years apart, not a matched-person analysis. Script reports percentage-point units throughout.

## Sources, versions, limitations, and rights
Sources are official Census microdata and documentation, official NEA report/appendix; each downloaded file has URL, byte length and SHA-256 in `source-manifest.json`. The scripts use direct public Census files, not the Census API. ICPSR 38936 v2 metadata (2026-02-02) was inspected but its interactive download was NOT used; v2 notes describe documentation changes for other modules. We analyze actual Census files, not an assumed ICPSR download.

Official sources:
- https://www.census.gov/data/datasets/2017/demo/cps/cps-arts.html
- https://www.census.gov/data/datasets/2022/demo/cps/cps-arts.html
- https://www.icpsr.umich.edu/web/NADAC/studies/38936
- https://www.arts.gov/impact/research/publications/arts-participation-2022-technical-summary-report

Data are U.S. federal public-use statistical files. Do not imply NEA/Census endorsement. Preserve source attribution and confidentiality protections; no reidentification, individual outreach, or record linkage to private people was attempted. Distribution rights of any third-party material in reports are separate from use of the public-use data.

Survey population: civilian noninstitutionalized household adults 18+, not children or institutional residents. Supplement participation, household selection, spouse/partner proxies, recall and social desirability, and differential nonresponse may affect estimates. Weighting/replicates address design and some nonresponse uncertainty, not all bias. Year differences cannot establish a pandemic effect, marketing opportunity, preference change among the same individuals, or a causal education effect. 2022 is the latest SPPA wave located and used here; this is not a claim about 2026 reading prevalence.

Source retrieval and original calculations were completed 2026-10-30 UTC.