Getting started with cvpa
cvpa gives you weighted party support and policy-issue attitudes by
demographic group, over time, from Canadian public-opinion surveys going
back to 1945. As of version 0.2.0, that covers several hundred policy
questions across seven issue domains – economy and taxation, energy and
environment, health/education/social programs, immigration/diversity/
federalism, foreign policy and identity, social issues/rights/justice,
and democracy and institutions – in addition to vote choice and
intention. From cvpa’s point of view, vote is just one more question:
it’s decoded and analyzed the same way as any policy question.
Two ways to use cvpa
There are two separate paths through the package, and it’s worth
being clear about which one you’re on:
- The live data path
(
update_cvpa_data(), question_catalog(),
question_data(), wtd_response(),
response_gap_analysis()) reads Parquet files hosted
alongside the policy-attitudes
web app, covering every question in the catalog. This is almost
certainly what you want for new analysis.
- The legacy vote path (
vote_data,
wtd_vote(), gap_analysis()) is the original
vote-choice/intention dataset and functions the package shipped with
before 0.2.0, kept as-is for reproducing prior work (in particular Lucas, Romualdi,
Armstrong, and Goodyear-Grant (2025) on the Canadian Vote Intention
Dataset, Canadian Journal of Political Science 58 (3): 666-80).
wtd_vote()/gap_analysis() are, under the hood,
thin wrappers around the same general engine that powers the live data
path, so their behavior hasn’t changed – see the last section of this
vignette.
This vignette focuses on the live data path, since that’s what’s
new.
Setting up the local cache
cvpa doesn’t bundle policy-attitude data in the package itself –
there’s too much of it, and it’s updated periodically as new survey
waves are added. Instead, update_cvpa_data() downloads what
you ask for into a local cache (see cvpa_cache_dir()) the
first time you need it, and every other function reads only from that
cache. Nothing in the package other than update_cvpa_data()
ever touches the network, so you’re always in control of when a download
happens.
The question catalog and the demographic/party code-labels table are
small and are always refreshed by update_cvpa_data().
Respondent-level data for a specific question is only fetched when you
name it (there are several hundred questions, so there’s deliberately no
“fetch everything” option – the full dataset is available as
dataset.parquet from the cvpa_dataset
repository):
update_cvpa_data(questions = "bri_econ_013")
#> Fetching question catalog (vbl_data.parquet)...
#> Fetching code labels (code_labels.parquet)...
#> Fetching 1 question file(s)...
#> [1/1] bri_econ_013
Finding a question
question_catalog() lists every available question, one
row per question, deduplicated across the several hundred survey waves
each question has accumulated over the decades:
head(question_catalog(domain = "econ"), 10)
#> question domain domain_label issue_label
#> 1 env_econ_061 econ Economy and Taxation Agree free trade Can-US-Mexico
#> 2 env_econ_074 econ Economy and Taxation Appr cutting federal taxes
#> 3 env_econ_056 econ Economy and Taxation Appr fed govt: deficit reduction
#> 4 env_econ_057 econ Economy and Taxation Appr fed govt: taxation
#> 5 env_econ_060 econ Economy and Taxation Appr prov govt: agriculture
#> 6 env_econ_075 econ Economy and Taxation Appr prov govt: deficit reduction
#> 7 env_econ_049 econ Economy and Taxation Appr prov govt: econ development
#> 8 env_econ_046 econ Economy and Taxation Appr prov govt: privatization
#> 9 env_econ_073 econ Economy and Taxation Appr prov govt: taxation
#> 10 env_econ_058 econ Economy and Taxation Approve GST
#> question_wording
#> 1 Do you agree or disagree that there should be free trade between Canada, the U.S. and Mexico?
#> 2 Do you approve or disapprove of the federal government making a tax cut?
#> 3 Do you approve or disapprove of the way the current federal government is handling deficit reduction?
#> 4 Generally speaking, do you approve or disapprove of the way the current federal government is handling taxation?
#> 5 Do you approve or disapprove of the way your provincial government is handling agriculture and farming?
#> 6 Generally speaking, do you approve or disapprove of the way your current provincial government is handling deficit reduction?
#> 7 Generally speaking, do you approve or disapprove of the way your current provincial government is handling economic development?
#> 8 Do you approve or disapprove of the way your current provincial government is handling privatization?
#> 9 Generally speaking, do you approve or disapprove of the way your current provincial government is handling taxation?
#> 10 As you may know, the federal government has a tax on most goods and services, known as the Goods and Services Tax or GST. Do you approve or disapprove of this tax?
#> n_waves type value_labels
#> 1 9 b Disagree, Agree
#> 2 9 b Disapprove, Approve
#> 3 10 b Disapprove, Approve
#> 4 10 b Disapprove, Approve
#> 5 25 b Disapprove, Approve
#> 6 1 b Disapprove, Approve
#> 7 17 b Disapprove, Approve
#> 8 5 b Disapprove, Approve
#> 9 35 b Disapprove, Approve
#> 10 9 b Disapprove, Approve
domain and type ("b" for
binary questions, "lsm" for 3-point less/same/more-style
questions, "nsa" for 3-point never/sometimes/always- style
questions, or "party" for vote choice/intention) both
filter the catalog if you want to narrow it down. With no arguments,
question_catalog() returns everything, vote included:
nrow(question_catalog())
#> [1] 630
table(question_catalog()$domain)
#>
#> dpin econ ener hwps idim intd soim vote
#> 34 139 57 85 118 72 123 2
Once you’ve picked a question,
question_response_labels() shows you what its response
categories actually are and how they’re coded:
question_response_labels("bri_econ_013")
#> code label sort_order
#> 1 -1 Less 1
#> 2 0 Same 2
#> 3 1 More 3
Fetching and inspecting a question’s data
question_data() reads the (now cached) respondent-level
data for one question and, by default, decodes both the response and
every demographic column into labeled factors:
d <- question_data("bri_econ_013")
head(d)
#> # A data frame: 6 × 18
#> response_id survey_id response weight year vote_intention vote_choice region
#> * <chr> <chr> <fct> <dbl> <dbl> <fct> <fct> <fct>
#> 1 CES2015_Com… 2015.dta More 0.835 2015 Undecided PC/Conserv… Ontar…
#> 2 2019+Canadi… 2019.dta Less 1.35 2019 PC/Conservati… <NA> Prair…
#> 3 2021 Canadi… 2021.dta More 0.546 2021 Liberal Bloc Quebe… Quebec
#> 4 ces2004-6-8… 2008.dta Same 1.31 2008 <NA> Green Prair…
#> 5 2021 Canadi… 2021.dta Same 0.631 2021 PC/Conservati… PC/Conserv… Prair…
#> 6 2021 Canadi… 2021.dta Less 1.94 2021 PC/Conservati… PC/Conserv… Prair…
#> # ℹ 10 more variables: province <fct>, com_500 <fct>, com_100 <fct>,
#> # gender <fct>, age_cats <fct>, religion <fct>, degree <fct>, language <fct>,
#> # union_household <fct>, occupation <fct>
Every question’s data has the same shape: response,
weight, year, plus the demographic and vote
columns listed in cvpa_grouping_vars – see
demographic_vars for what each demographic variable
measures and what categories it takes:
cvpa_grouping_vars
#> [1] "vote_intention" "vote_choice" "region" "province"
#> [5] "com_500" "com_100" "gender" "age_cats"
#> [9] "religion" "degree" "language" "union_household"
#> [13] "occupation"
demographic_vars
#> variable
#> 1 region
#> 2 province
#> 3 com_500
#> 4 com_100
#> 5 gender
#> 6 age_cats
#> 7 religion
#> 8 degree
#> 9 language
#> 10 union_household
#> 11 occupation
#> description
#> 1 Region of residence
#> 2 Province or territory of residence
#> 3 Community size, split at 500,000 population
#> 4 Community size, split at 100,000 population
#> 5 Respondent gender. Binary Man/Woman only -- a non-binary/other response option exists in some surveys from 2019 onward but is too sparse to include here; see the Canadian Opinion Research Archive (https://www.queensu.ca/cora/) for the full data.
#> 6 Age category
#> 7 Religious affiliation
#> 8 University degree attainment
#> 9 Language spoken
#> 10 Union membership in the household
#> 11 Respondent's occupation
#> categories
#> 1 Atlantic; Quebec; Ontario; Prairies; British Columbia; North
#> 2 NL; PEI; NS; NB; QC; ON; MB; SK; AB; BC; YT; NT; NU
#> 3 Population <= 500k; Population > 500k
#> 4 Population <= 100k; Population > 100k
#> 5 Man; Woman
#> 6 18-29; 30-49; 50+
#> 7 Catholic; Protestant; Jewish; Other; None
#> 8 No University Degree; University Degree
#> 9 French; English; French and English; Other
#> 10 No Union Member in Household; Union Member in Household
#> 11 Managers, Executives; Professional; Routine Non-manual; Working Class; Armed Forces; Farmer/Fiserman; Retired; Unemployed; Student; Other
Notice that vote_intention and vote_choice
are in that list alongside the demographics – every question’s data
carries how each respondent voted, so you can group any policy question
by vote just as easily as by gender or region.
Weighted response shares over time
wtd_response() computes each response category’s
weighted share of the total, for every year (and, optionally,
grouping-variable combination):
wtd_response(d, grouping_vars = "gender")
#> # A tibble: 54 × 5
#> year response gender count response_share
#> <dbl> <fct> <fct> <dbl> <dbl>
#> 1 2000 Less Man 118. 0.0715
#> 2 2000 Less Woman 63.5 0.0374
#> 3 2000 Same Man 315. 0.191
#> 4 2000 Same Woman 222. 0.131
#> 5 2000 More Man 1218. 0.738
#> 6 2000 More Woman 1411. 0.832
#> 7 2004 Less Man 134. 0.0675
#> 8 2004 Less Woman 63.5 0.0309
#> 9 2004 Same Man 372. 0.187
#> 10 2004 Same Woman 276. 0.134
#> # ℹ 44 more rows
A quick look at how the “More” category (support for reducing the
rich-poor gap) has moved over time, by gender:
library(ggplot2)
more_by_year <- subset(wtd_response(d, grouping_vars = "gender"), response == "More")
ggplot(more_by_year, aes(x = year, y = response_share, color = gender)) +
geom_line() +
geom_point() +
scale_color_manual(values = c(Man = "steelblue", Woman = "firebrick")) +
labs(
x = "Year", y = "Share answering \"More\"",
title = "Support for reducing the rich-poor gap, by gender",
color = NULL
) +
theme_minimal()
plot of chunk plot-wtd-response
Testing whether a gap is significant
response_gap_analysis() generalizes
gap_analysis() to any question: a difference-of-proportions
test between two levels of a grouping variable, for each response
category, for each year:
response_gap_analysis(d, grouping_var = "gender")
#> # A tibble: 27 × 13
#> # Groups: year [9]
#> year response cnt_Man cnt_Woman n_Man n_Woman flag estimate1 estimate2
#> <dbl> <fct> <dbl> <dbl> <dbl> <dbl> <chr> <dbl> <dbl>
#> 1 2000 Less 118. 63.5 1651. 1697. p-val rel… 0.0715 0.0374
#> 2 2000 Same 315. 222. 1651. 1697. p-val rel… 0.191 0.131
#> 3 2000 More 1218. 1411. 1651. 1697. p-val rel… 0.738 0.832
#> 4 2004 Less 134. 63.5 1991. 2054. p-val rel… 0.0675 0.0309
#> 5 2004 Same 372. 276. 1991. 2054. p-val rel… 0.187 0.134
#> 6 2004 More 1485. 1714. 1991. 2054. p-val rel… 0.745 0.835
#> 7 2006 Less 62.5 19.6 956. 914. p-val rel… 0.0653 0.0215
#> 8 2006 Same 177. 107. 956. 914. p-val rel… 0.185 0.117
#> 9 2006 More 716. 788. 956. 914. p-val rel… 0.749 0.862
#> 10 2008 Less 94.0 61.8 1609. 1665. p-val rel… 0.0584 0.0371
#> # ℹ 17 more rows
#> # ℹ 4 more variables: p.value <dbl>, conf.low <dbl>, conf.high <dbl>,
#> # diff <dbl>
The flag column tells you whether every cell going into
that year’s test had an expected count above 5
("p-val reliable") or not ("p-val unreliable")
– see make_flag().
Vote is just another question
Because vote choice and intention live in the same catalog and the
same resp/question=<code>/data_0.parquet layout as
every policy question, everything above works identically for
"vote_choice"/"vote_intention":
update_cvpa_data(questions = "vote_choice")
#> Fetching question catalog (vbl_data.parquet)...
#> Fetching code labels (code_labels.parquet)...
#> Fetching 1 question file(s)...
#> [1/1] vote_choice
vc <- question_data("vote_choice", years = 2000:2021)
wtd_response(vc, grouping_vars = "gender")
#> # A tibble: 120 × 5
#> year response gender count response_share
#> <dbl> <fct> <fct> <dbl> <dbl>
#> 1 2000 Liberal Man 1218. 0.415
#> 2 2000 Liberal Woman 1365. 0.456
#> 3 2000 PC/Conservative Man 408. 0.139
#> 4 2000 PC/Conservative Woman 414. 0.138
#> 5 2000 CCF/NDP Man 265. 0.0904
#> 6 2000 CCF/NDP Woman 339. 0.113
#> 7 2000 Bloc Quebecois Man 304. 0.104
#> 8 2000 Bloc Quebecois Woman 298. 0.0997
#> 9 2000 Reform/Alliance Man 670. 0.228
#> 10 2000 Reform/Alliance Woman 520. 0.174
#> # ℹ 110 more rows
The legacy vote path
If you’re working with the original vote_data dataset –
for instance to reproduce results from the paper it was built for (Lucas, Romualdi,
Armstrong, and Goodyear-Grant (2025)) – wtd_vote() and
gap_analysis() are unchanged:
data(vote_data)
wtd_vote(vote_data, "intention", years = 2000:2022, grouping_vars = "gender")
#> The following years are not available: 2012, 2013, 2014, 2016, 2017, 2018
#> # A tibble: 200 × 5
#> year vote gender count party_support
#> <dbl> <fct> <fct> <dbl> <dbl>
#> 1 2000 Liberal Man 5357. 0.451
#> 2 2000 Liberal Woman 5318. 0.483
#> 3 2000 Conservative Man 1263. 0.106
#> 4 2000 Conservative Woman 1247. 0.113
#> 5 2000 NDP Man 1045. 0.0880
#> 6 2000 NDP Woman 1146. 0.104
#> 7 2000 Social Credit Man 80.9 0.00681
#> 8 2000 Social Credit Woman 59.4 0.00539
#> 9 2000 Bloc Quebecois Man 1041. 0.0876
#> 10 2000 Bloc Quebecois Woman 942. 0.0855
#> # ℹ 190 more rows
gap_analysis(vote_data, "intention", years = 2000:2022, grouping_var = "gender")
#> The following years are not available: 2012, 2013, 2014, 2016, 2017, 2018
#> # A tibble: 93 × 13
#> # Groups: year [16]
#> year vote cnt_Man cnt_Woman n_Man n_Woman flag estimate1 estimate2
#> <dbl> <fct> <dbl> <dbl> <dbl> <dbl> <chr> <dbl> <dbl>
#> 1 2000 Liberal 5357. 5318. 11883. 11012. p-va… 0.451 0.483
#> 2 2000 Conservative 1263. 1247. 11883. 11012. p-va… 0.106 0.113
#> 3 2000 NDP 1045. 1146. 11883. 11012. p-va… 0.0880 0.104
#> 4 2000 Social Cred… 80.9 59.4 11883. 11012. p-va… 0.00681 0.00539
#> 5 2000 Bloc Quebec… 1041. 942. 11883. 11012. p-va… 0.0876 0.0855
#> 6 2000 Reform/Alli… 2742. 2008. 11883. 11012. p-va… 0.231 0.182
#> 7 2000 Other 353. 292. 11883. 11012. p-va… 0.0297 0.0265
#> 8 2001 Liberal 1426. 1598. 3315. 3361. p-va… 0.430 0.476
#> 9 2001 Conservative 581. 525. 3315. 3361. p-va… 0.175 0.156
#> 10 2001 NDP 332. 377. 3315. 3361. p-va… 0.100 0.112
#> # ℹ 83 more rows
#> # ℹ 4 more variables: p.value <dbl>, conf.low <dbl>, conf.high <dbl>,
#> # diff <dbl>
Note that vote_data predates the cp3 Parquet pipeline
and has a single community_size column rather than the live
data path’s com_100/com_500 split – see
demographic_vars for the live path’s version of each
demographic variable.