# Sources — Political Efficacy (External)
## "Americans Have Stopped Believing Officials Listen"
### Supports: Chapter 2, Observation 7 — Political Efficacy (Diagnostic). STUB as of August 2026; not yet in the draft. Issue number TBD — assign before uploading to tsb-assets.
### Last updated: August 20, 2026 (chart rebuilt — see Corrections Log)

> **Numbering note.** This chart was built while political efficacy was Observation 8. Observations 7 and 8 were subsequently swapped (efficacy moved to 7; EO structural weight moved to 8). Every reference below now reads "Obs 7." If you find "Obs 8" anywhere in this folder's assets, it is stale and refers to this observation, not to `obs08-eo-structural-weight`.

---

## What This Chart Shows

Two independent measures of **external political efficacy** — whether Americans believe the political system hears them — plotted together from 1952 to 2024. Both fall. Both are built from ANES data, by different hands, on different rules, and they track each other for seventy years.

**Navy line (primary): the item this chapter quotes.** ANES's VCF0609 — "Public officials don't care much what people like me think" — showing the share who **disagree**, as a share of Agree + Disagree only. That is, of the Americans who took a side either way, how many said officials *do* care. It runs from 75.0% in 1960 down to 16.4% in 2024.

**Lighter dashed line (secondary): ANES's own index.** VCF0648, the external efficacy index, 0–100, higher = more efficacious. It averages VCF0609 with a second item (VCF0613, "people like me don't have any say about what government does") and scores "neither" at the midpoint of 50. It runs from 73.5 in 1960 to 28.5 in 2020, and **stops there** — ANES has not scored it for 2024.

### Why the primary line uses a side-takers denominator

This is the one methodological choice in the chart that is ours rather than ANES's, so it is worth stating plainly.

"Neither agree nor disagree" was not offered as a response before 1988. From 1988 on it absorbs between 9.8% and 23.3% of respondents. ANES's published share (agree ÷ agree+disagree+neither) is the right way to report *agreement*, but it makes the *disagreement* share unusable as a trend, because the middle option eats into it at a single point in time for reasons that have nothing to do with opinion. The seam is visible in the raw numbers:

| | 1986 | 1988 | change |
|---|---|---|---|
| Disagree, ANES denominator | 45.0% | 37.6% | **−7.4** |
| Disagree, side-takers only | 45.0% | 42.5% | −2.6 |
| VCF0648 index | 45.0 | 48.7 | +3.6 |

Nothing happened to American opinion between 1986 and 1988; ANES added a response box. A 7.4-point drop that points the same direction as the trend being argued is exactly the artifact a hostile reader should find, so the chart does not use that series. Restricting the denominator to respondents who took a side gives a consistent denominator in every year of the series, at the cost of dropping the middle category entirely. The residual −2.6 is real movement plus ordinary sampling noise.

The secondary line exists as the check on that choice: VCF0648 handles "neither" the opposite way — keeping it, scored at 50 — and lands on the same trajectory. Two different treatments of the awkward category, one conclusion.

**All three constructions are in `data.csv`.** Anyone who prefers ANES's published denominator can plot `pct_disagree_vcf0609` from the same file and see what it does.

### The shape of the decline

The fall is not smooth, and the chart does not smooth it. Efficacy drops steeply through the 1960s and 70s, crosses the halfway line around 1972, and after 1988 swings hard from cycle to cycle — 1990, 1992, and 1994 alternate by 15 points or more. Both lines swing together, which is evidence the swings are real rather than an artifact of either construction.

The single genuine reversal is **2002**: 59.2% on the primary line against 36.7% in 2000, and 60.6 against 46.3 on the index. This is the post-9/11 rally effect, and it reverts completely by 2004. See the correction below — an earlier read of this session suspected it was a questionnaire artifact, and the data does not support that.

As the Obs 7 stub notes, this is a distinct construct from Approval (Obs 6): Approval grades current officials' performance; this asks whether people believe the system hears them at all, independent of who's in office. That's why it's read as closer to a legitimacy measure than a performance measure.

---

## Getting the Data: Why There Is a Script in This Folder

The ANES Time Series Cumulative Data File is the authoritative source for this variable, and it is **not in this repo** — the CSV release is roughly 163 MB and the raw `.dat` is roughly 145 MB, both far past GitHub's 100 MB hard file limit (and past the 50 MB warning threshold). Committing it is not an option, and it also could not be moved through the Claude session's device bridge, which timed out on it repeatedly.

The workaround, and the reproducible path for anyone re-deriving these numbers:

1. **Download the CDF yourself** from the ANES data center (free account required):
   **https://electionstudies.org/data-center/anes-time-series-cumulative-data-file/**
   Take the CSV release. This folder's two codebook HTML files (`...VARIABLE CODEBOOK.html`, `...APPENDICES.html`) are committed here because they are small and are what the methodology decisions below are argued from.
2. **Run `anes_efficacy_extract.py`** (in this folder) against it:
   ```
   python3 anes_efficacy_extract.py anes_timeseries_cdf_csv_20260205.csv obs07_efficacy.csv
   ```
   It streams only five columns (VCF0004, VCF0009z, VCF0609, VCF0613, VCF0648), so memory stays flat regardless of file size, and writes a ~1.4 KB result.
3. **`obs07_efficacy.csv`** — the script's output — *is* committed here. It is the full weighted derivation behind `data.csv`, including the columns `data.csv` doesn't carry (disagree, neither, both denominators, and the composite index).

So the repo holds the recipe and the result, not the ingredient. Re-running step 2 against a future ANES release is the intended way to extend this chart.

---

## Data and Sourcing

### VCF0609, all years — % agree "officials don't care" (weighted)
- **Formula:** For each study year, (sum of VCF0009z weights where VCF0609 = 1 "Agree") ÷ (sum of VCF0009z weights where VCF0609 ∈ {1, 2, 3} "Agree/Disagree/Neither"). Don't-know/refused (code 9) and not-applicable (code 0, blank) are excluded from the denominator. **"Neither" IS in the denominator** — see the correction below.
- **Source:** ANES Time Series Cumulative Data File, 1948–2024 (release dated 2026-02-05). Variable VCF0609, "R Feels Public Officials Don't Care" / "Government Officials Care What People Like Respondent Think." **https://electionstudies.org/data-center/anes-time-series-cumulative-data-file/**
- **Weight:** VCF0009z. Per the CDF codebook's Appendices document (WEIGHTS section): weight family (0009 vs. 0010 vs. 0011) depends on which "type" of 1970 data a variable uses — VCF0609 is not listed among the Type 1/Type 2 exception variables, so it uses the default Type 0 → the "0009" family. Suffix (x/y/z) depends on sample: "z" is the combined face-to-face + web sample, the right choice for a single continuous trend spanning the pre- and post-2012 mode change.
- **Confidence level:** High. Primary source, full cumulative microdata, ANES's own harmonized coding and cross-year weight.
- **Sample sizes:** Range from ~1,000 (1986, 2008) to ~7,400 (2020) valid responses per year; see `data.csv` and `obs07_efficacy.csv` for exact per-year N.
- **Which years have data:** ANES did not field this item every study year. Years with no data (not plotted): 1948, 1954, 1958, 1962, and even-year-only cycles where the item wasn't included (2006, 2010, 2014, 2018, 2022). All years the item *was* fielded are plotted — nothing was cherry-picked or dropped for shape.
- **Known measurement notes:**
  - The response scale format changed over the item's history (fewer categories pre-1988, a 5-point scale 1988+), but ANES's own VCF0609 harmonization already collapses "agree strongly"/"agree somewhat" into one Agree code across every year.
  - **"Neither agree nor disagree" did not exist as a code before 1988.** It is 0.0% in every year 1952–1986 and runs 9.8%–23.3% from 1988 on. Because it sits in the denominator, some of the apparent post-1988 movement in *disagree* is the middle category absorbing share rather than genuine opinion change. This matters enormously for any inverted version of this chart — see "Denominator" below.
  - 2008 and 2012 fielded this item to only a random half-sample of respondents (per CDF notes); the weight variable accounts for this, but per-year N is correspondingly smaller those years.

### VCF0648 — ANES's external efficacy index (0–100)
- **What it is:** ANES's own composite, built from VCF0609 **and** VCF0613 ("people like me don't have any say about what government does"). Each component is recoded 1=0, 2=100, 3=50, the recoded values are totaled and divided by the number of valid responses, and the result is rounded. Missing code 999 = "not scored in both VCF0609 and VCF0613."
- **Role in the chart:** plotted as the secondary dashed line. It is **already polarity-correct** — higher means officials are seen as *more* responsive — and it handles "Neither" honestly by scoring it at the midpoint rather than letting it distort a two-way share.
- **Range in this series:** 73.5 (1960, highest) down to 28.5 (2020, lowest).
- **⚠️ Not available for 2024.** Every 2024 row is coded 999, because VCF0613 was not scored for 2024 in the 2026-02-05 release even though VCF0609 was. Any chart built on VCF0648 currently ends at 2020 and cannot show the most recent reading. Worth re-checking on the next CDF release.

---

## Corrections to the Prior Version of This File

### 1. The denominator question is settled: "Neither" IS included

The previous `chart.html` reading note said *"'Neither' (1988+) and don't-know/refused responses are excluded from the denominator."* That was **wrong**, and it contradicted the formula stated in this file.

Verified directly: the `pct_agree` column of `obs07_efficacy.csv`, computed with denominator {1,2,3}, reproduces every published figure in `data.csv` to the decimal (1952 = 35.78, 1960 = 25.04, 2002 = 31.50, 2024 = 70.98). The alternative denominator {1,2} produces materially different numbers from 1988 on (e.g. 2024 = 83.59, not 70.98). This file's formula was correct; the chart's reading note was the error and has been fixed.

**Consequence for any inverted chart:** `100 − agree` is *not* "Americans who believe officials care." It is disagree + neither. The honest inversions are either `pct_disagree`, `pct_agree_excl_neither` inverted, or VCF0648 — not `100 − pct_agree`.

### 2. The 2002 dip is real — retracting the "questionnaire artifact" hypothesis

The 2002 questionnaire did read the middle option aloud (*"Do you agree, neither agree nor disagree, or disagree with this statement?"* — codebook, 2000 TELEPHONE/2002 note), and 2002's "Neither" share is 22.8%, well above the 10.5% of 2000 and 15.4% of 2004. It was reasonable to suspect the dip was mostly a middle-option artifact.

**The data does not support that.** Normalizing the format out entirely — agree as a share of agree+disagree only — 2002 still falls from 63.4% (2000) to 40.8%, a 22.6-point drop, against a 25.2-point drop on the published measure. The question wording accounts for roughly 2.6 points of a 25-point move. ANES's own composite index tells the same story: 60.6 in 2002 against 46.3 in 2000 and 46.6 in 2004 — a genuine one-cycle spike in efficacy that reverts.

**The existing callout — "31.5% — 2002, a one-year drop, the post-9/11 rally effect" — stands as written.** A footnote that the 2002 item also offered an explicit middle option would be honest, but the attribution itself is sound.

### 3. The stub's "1984 peak" claim is supported after all — on the composite index

The previous version of this file concluded that the stub's claim ("one analysis found external efficacy declined by nearly half from its 1984 peak") was *not supported*, and recommended dropping it. That conclusion was based on VCF0609 alone, where 1984 is a local dip (42.6% agreement, against 54.5% in 1980 and 55.0% in 1986). That much remains true.

But this file also speculated that the original analysis may have used a composite index rather than the single item. **It did, or something very like it.** On VCF0648:

| Year | Index | |
|---|---|---|
| 1982 | 52.2 | |
| **1984** | **62.7** | **local peak** |
| 1986 | 45.0 | |
| 2016 | 34.9 | −44% from 1984 |
| 2020 | 28.5 | −55% from 1984 |

1984 *is* a local peak on the composite, and the decline from it to 2016 is 44% — "nearly half," almost exactly the stub's phrasing.

**Recommended handling in the draft:** the claim can be used, with one qualification. 1984 is a *local* peak — the highest reading since 1980 — not the series peak, which is 1960 (73.5). Write it as "its 1984 local peak" or "the highest level since the early 1980s," and cite VCF0648 rather than VCF0609. Do not write "its 1984 peak" unqualified.

---

## Known Limitations

- **The primary line's denominator is our choice, not ANES's.** Restricting to side-takers is defensible and documented above, but it is a derivation. Anyone reproducing the chart from ANES's published tables will get the `pct_disagree_vcf0609` column instead, which is lower from 1988 on. Both are in `data.csv`.
- **Dropping "neither" discards real respondents.** Between 9.8% and 23.3% of post-1988 respondents are excluded from the primary line. If ambivalence is itself a symptom of low efficacy — a defensible reading — the primary line understates the decline; the index line, which scores them at 50, is the more conservative of the two.
- **The two lines diverge after 1988,** by up to about 13 points (2012). This is expected (midpoint scoring pulls the index toward 50) and is explained on the chart, but it does mean the "two measures agree" claim is about *trajectory*, not *level*.
- **VCF0648 cannot currently be extended to 2024** (see above), so the secondary line stops at 2020 while the primary line runs to 2024. Worth re-checking on the next CDF release.
- **The 1984 divergence is real and instructive**: the single item shows a dip, the composite shows a local peak. They are measuring overlapping but not identical things. See the correction below.
- **Year-to-year swings are real but noisy** at the sample sizes involved (roughly 1,000–2,200 valid responses in most years, ~3,000–7,400 in 2012/2016/2020/2024) — treat single-cycle jumps (e.g., 1988→1990, −15.7 points on the primary line) with appropriate caution; a couple of points of any given year's estimate is within normal sampling variation.

---

## Corrections Log

**2026-07-20:** Replaced the initial 5-point, partly secondary-sourced version with the full VCF0609 series (29 verified years) computed directly from the ANES Time Series Cumulative Data File, weighted with VCF0009z.

**2026-08-20 (b):** **Rebuilt the chart** at editorial request, to remove the "higher is worse" reading cost. Was "Agreement That Officials Don't Care Has Roughly Doubled Since the 1950s" — a single rising line of *agreement*. Now "Americans Have Stopped Believing Officials Listen" — two falling lines of *efficacy*: the side-takers disagree share through 2024, plus VCF0648 as an independent check through 2020. Note that `100 − agree` was **not** used, and is not a valid construction here: on ANES's denominator it equals disagree + neither, which is not "believe officials care." Callout stats reworked from 25%/31.5%/71.0% (agreement) to 75%/59%/16% (efficacy). `data.csv` rebuilt to carry all three constructions plus the raw components.

**2026-08-20 (a):** Re-derived from the CDF with `anes_efficacy_extract.py`, which also pulls disagree, neither, both denominators, and VCF0648. Four changes: (a) renumbered all "Obs 8" references to "Obs 7" after the observation swap; (b) added the ANES data-center download URL and documented why the source file is not in the repo and how to reproduce the extract; (c) confirmed "Neither" is in the denominator and corrected the contradictory reading note in `chart.html`; (d) **reversed two prior conclusions** — the 2002 dip is a genuine rally effect rather than a questionnaire artifact, and the stub's "1984 peak" claim is supported on the composite index (with the "local peak" qualification) rather than unsupported.

## Contact
For corrections or questions: Jason, jason@tracerfitness.com