# Meta-analysis of the Indigenous survey questions catalogue *Built 2026-08-04 and revised 2026-08-05 for the 465-question catalogue. Question wordings, response distributions and source URLs live in the catalogue itself (`Indigenous_Survey_Questions_Dataset.xlsx`); the coding decisions live in `analysis_coding_sheet.csv`; every figure below traces back to a named primary source through `Indigenous_Survey_Harmonized_Supportive.csv`. Polarity and supportive-side codings were made with Claude Code (Anthropic) under conventions Ron Levi accepted on 2026-08-04; 37 of the 167 pro-coded items carry `polarity_confidence=medium` and can be re-filtered out by anyone who disagrees with those calls.* *Revision note, 2026-08-05: 21 items were added from three sources that are not polling firms and carry an asterisk in the catalogue (see methods.html s.6): the 2023 Residential School Denialism Survey, 19 items, and one police item each from GSS-35 and the crowdsourced COVID discrimination file. Sixteen of the denialism survey's items are pro-coded and enter the figures below; its sample excludes Indigenous respondents by design, so those rows sit in the general-population column alongside the other non-Indigenous-only samples already there. The two police items carry no support axis, and the COVID file is flagged DO NOT POOL and appears in no pooled statistic. Every number below was regenerated; the one substantive change is noted in the remedy-gradient section.* ## What was measured Of the 465 catalogued questions, 167 ask for a position on an Indigenous claim, policy or acknowledgment and were coded with a supportive direction; the other 298 measure knowledge, perceptions, experiences, evaluations of governments and churches, or open-ended meanings, and carry no support axis. For 165 of the 167 a supportive share could be harmonized (226 measurements once waves, population splits and battery targets are counted); the two others have no public data. The harmonized share means: the percent of respondents on the side that `analysis_coding_sheet.csv` names as supportive for that item, formed from published nets where the source printed one, from summed scale points, from a named forced-choice option, or (for five Ipsos releases that printed only one side of a binary) as 100 minus the published share, a basis flagged `residual` because an unreported don't-know share would shrink it. The 23 CES figures are our own weighted computations from public microdata, not published toplines. The 26 CRB items enter as index points (percent of scale range, rescaled from published means), kept out of every pooled percent statistic. Across the 121 deduplicated general-population percent measurements, the supportive share averages 53 with a standard deviation of 16, and the spread is the finding: items sit anywhere from 18 (a veto for Indigenous governments over BC land use, ARI 2024) to 87 (federal help searching burial sites, Ipsos 2021). ## 1. Within-instrument trends Fifteen series repeat the same instrument (or a chain the pollster itself treats as one trend item) in at least two years. Most move little: of the thirteen without wording or scale changes, nine change by 5 points or fewer from first to last measurement. The exceptions run in both directions. Agreement that the federal government "must act now" to raise Indigenous quality of life, on the same Ipsos wording each time, rose from 62 (2010) and 63 (2013) to 75 (Feb 2020, during the Wet'suwet'en blockades) and 81 (June 2021, after the Kamloops announcement). The ARI forced choice on unique status moved from 47 choosing the unique-status side in 2018 to 55 in 2023 and 55 in 2025. In the other direction, the view that governments have "not gone far enough" on reconciliation (COT) fell from a 46 peak in 2022 to 37 in 2025, and the share saying Indigenous Peoples "often" face discrimination (Race Relations survey) rose from 39 (2019) to 49 (2021) before falling to 31 (2024), below its 2019 level. Endpoint arithmetic would call that series a decline of 8; the path is a spike and a retreat. Two chains are flagged rather than trusted: the cultural-genocide item (70 agree in 2015, 67 in 2025) switched from a binary to a five-point scale with a not-sure option, and the residential-schools "harm continues" pair drifts in wording across 2018, 2022 and 2023. ## 2. Pooled year effects (a modest mood index) Averaging each multi-wave series' deviation from its own mean, by year, gives a common-movement index (`Meta_MoodIndex.csv`). Years 2020 through 2023 sit above series means (+1.6 to +5.7), 2018, 2019 and 2024 sit below (-2.3 to -3.1), and 2025 is near zero (+0.8). Read with care: the index rests on at most 7 series in any year, its composition changes yearly (2022 is one series; 2010 and 2013 are the Ipsos federal-action item alone), and no smoothing or dyad-ratio weighting was applied. A consistent sign pattern does survive the caveats: every series measured in 2021 sat above its own mean that year, and every series measured in 2024 sat at or below it. The measured window of elevated support coincides with the 2020 blockades coverage and the 2021 unmarked-graves announcements, and had receded by 2024, though with these few series the timing cannot be separated from which surveys happened to field. ## 3. Indigenous versus non-Indigenous respondents Where one instrument reports both populations in the same wave, the Indigenous share on the supportive side is higher in 27 of 34 pairs. Among the 8 percent-unit pairs the mean gap is +11 points, ranging from -2 (YRB youth on "inequality must fall first") to +23 (COT 2025, governments have not gone far enough: 59 among Indigenous respondents, 36 among non-Indigenous). Among the 26 CRB index pairs (2022) the mean gap is +6 index points; the largest sit on the residential-schools harm battery (+10 to +15.5), while the nine representation items ("Indigenous Peoples are decision makers in...") show gaps near zero because both populations lean toward disagreement. Sampling notes matter here: CRB's Indigenous sample is an online panel (n=1,034), YRB's is n=682 youth, and NJS publishes only subgroup breakdowns, so the White comparator stands in for non-Indigenous in that one pair. ## 4. Question wording Fourteen wording families contain two or more distinct instruments measuring the same construct. Within those families the family accounts for 64 percent of the variance in supportive shares (eta-squared, n=52 measurements), leaving 36 percent moving within family across wordings, formats and years. The benchmark for "years alone" is the one long same-wording series available: 19 points across 2010-2021 on the Ipsos federal-action item. Several within-family spreads match or exceed that benchmark inside a much shorter window, which suggests, though on few cases, that format choices move answers about as much as a decade of events: - **Consent over resource development, 51-point family range.** An agree-disagree statement ("no one should be allowed to develop... unless the community agrees") drew 66-69 agreement (COT 2020-21); a forced choice with a "depends" option drew 35 for holding off until consent (COT 2019); a four-option authority menu drew 18-25 for the strongest option (ARI 2024-25); Ipsos's "should not proceed without Indigenous involvement" drew 52 (2022). Population and year differ across these, but the ordering tracks the response format: offered a middle route, respondents take it. - **Statues and renaming, same month.** In June 2021 Ipsos found 52 for removing statues and 56 for renaming buildings, while CES (computed) found 49 for renaming; ARI's case-by-case version in 2023 reached 61 once its middle option is counted as supportive. The 2021 cluster spans 11 points on nearly simultaneous fieldwork. - **The word "genocide", same survey.** In ARI's July 2025 survey, 67 agreed with "cultural genocide" for residential schools and 52 with "genocide" for the broader history, a 15-point difference produced inside one questionnaire. - **Stable families exist too.** Treaties-ignored (60 to 62 across 2018-2025), the MMIW inquiry (80, 79 across 2015-16), and the taxpayer-support resentment items (37, 36 across 2012-13; the CES 2021 cousin lands at 38) barely move across wordings or years. ### Robustness check: remedy tier versus response format The gradient claim (commemoration and acknowledgment at the top, spending and status in the middle, jurisdiction at the bottom) could be an instrument artifact, because jurisdiction items are disproportionately asked as multi-option menus whose middle options absorb support. To check, every general-population item was hand-classified by remedy tier (memory / money-status / power, with diffuse act-now items and non-remedy items set aside) and by format (rate-one-proposition scale / binary forced choice / menu of three or more options); waves were averaged so trend items count once. `Meta_GradientRobustness.csv` holds the cells; `robustness_gradient.py` holds the item-level assignments. Two parts of the gradient survive, one does not. - **The power penalty survives every format.** Power items sit about 6 points below memory and money items among scales (57 versus 63, n=10 versus 38), about 15 points below among forced choices (36 versus 51, n=4 versus 13), and lowest among menus (30 versus 36-43). Within the scale cell the spread still tracks the strength of the claim, from 67-73 for consent-as-principle and involvement wordings down to 40-45 for making UNDRIP law and self-administered justice. - **The format ladder is larger than any tier difference.** Holding tier composition to the three gradient tiers, scales average 62, forced choices 48, menus 35, a step of roughly 14 points per format. Offering a counter-statement costs support; offering a middle route costs more. - **The memory-versus-money step does not survive.** Within forced choices the two tiers nearly tie (52 versus 51), and within scales money now sits slightly above memory (66.9 versus 62.7); the pooled difference, 7.5 points, arose because both bare spending items were asked as menus (28, 28, 39) while most memory items were asked as scales. Named-necessity funding polls with the acknowledgment items (clean water and housing 83, language preservation 66), so the middle of the original gradient was a format artifact plus a real distinction between abstract budget lines and named necessities. - **The denialism items moved the memory cell, 2026-08-05.** Adding the nine 2023 denialism items pulled the memory-scale mean from 67.0 to 62.7 (n from 29 to 33) and turned the memory-money tie among scales into a small money advantage. Four of them sit at the bottom of the cell, from 37 for whether the graves may not contain Indigenous people to 42 for whether radar can reliably locate them. Acknowledging documented harm in the abstract and rejecting a specific denialist proposition are not the same measurement, and the memory tier now mixes both. Read the memory figures as an average over two kinds of item, not one. Cells are small (money-status scales n=5, memory menus n=2) and the tier and format codes are one coder's judgment, laid out item by item in the script for anyone to recut. Diffuse act-now items, for the record, score highest of all among scales (69, n=6, range 62-81). Construct explains 14 percent of variance across all 137 measurements, year 22 percent, source organization 5 percent. These are one-way decompositions on overlapping factors, not a causal partition; family membership, construct and year are confounded with one another and with mode and house effects, which nothing in this design can separate. ## 5. The attitude space: an MCA of the CES 2021 microdata Aggregate toplines cannot say whether the corpus's many items tap one public disposition or several, because that structure lives inside respondents. The CES 2021 public microdata allows the test at the individual level. A weighted specific multiple correspondence analysis (run_ces_mca.py; plane and full coordinates in ces_mca.html) took as active the eight attitude items answered by the same 3,926 respondents, the four residential-schools action items, reconciliation spending, and the three resentment statements, with don't-know responses set passive so response style cannot pose as substance. One axis organizes nearly everything. The first dimension, carrying 80 percent of the adjusted inertia (22 percent raw; Benzecri's adjustment flatters, so both are reported), runs from opposing the actions, wanting less spending and endorsing resentment statements to strongly supporting every action, wanting more spending and rejecting resentment. Symbolic items (renaming) and material ones (spending, ceasing court actions) load on the same axis rather than separating, so in 2021 the people who supported one tended to support the others. The second dimension (17 percent adjusted) is an intensity contrast, emphatic answers at either pole against middle and neither responses, not a second substantive divide. The projected categories behave as a valid axis should: the Indigenous feeling thermometer and perceived-discrimination bands order monotonically along dimension 1, and party identification spreads from Conservative (-0.57) through Liberal (+0.15) to Green and NDP (+0.52, +0.53). Education barely moves (-0.10 to +0.10 across three bands), age and region little more; position on this axis reads far better from party and affect than from schooling. Read together with section 4, this suggests one reconciliation of the whole corpus: a single underlying disposition, with question wordings acting as thresholds of different heights on it, would produce both the aggregate spread (18 to 87) and the unidimensional individual structure. Limits, named: one election year, and 2021 sat at the mood peak, so the unidimensionality could be period-specific; a non-probability panel; the module-overlap subsample; our category collapses; and eight items cannot rule out dimensions that unmeasured items (jurisdiction, policing) might open. ## Files - `Indigenous_Survey_Harmonized_Supportive.csv` (210 rows) and `Harmonized_MatchReport.csv`: the outcome table and per-item accounting. - `Meta_Trends.csv`, `Meta_MoodIndex.csv`, `Meta_Gaps.csv`, `Meta_WordingFamilies.csv`: the four analyses. - `Meta_GradientRobustness.csv`: the tier-by-format robustness cells. - `build_harmonized_outcome.py`, `run_meta_analyses.py`, `robustness_gradient.py`: rebuild everything; no network needed. - `run_ces_mca.py`, `build_ces_mca_page.py`, `CES_MCA_Categories.csv`, `ces_mca.html`: the CES 2021 attitude-space analysis (needs the public CES microdata, DOI in the script). - `analysis_coding_sheet.csv`: the coding authority (hand-maintained; do not regenerate). Known limits, named: five residual-basis shares overstate support by any unreported don't-know share; CES figures depend on our analytic choices and are so labelled; CRB indexes are not shares of people; single-wave items dominate the corpus (13 usable series carry all trend claims); polarity judgments on 37 items are medium-confidence and listed as such in the coding sheet.