Recipes: the language divide in the Liberal vote, step by step

Version française

This page downloads 11 studies (about 16 MB) the first time it runs. options(qesR.cache = "disk") keeps them on disk for later sessions.

Since the first provincial election studies, the Quebec vote has been read through language and the national question (Hamilton and Pinard 1976; Pinard and Hamilton 1978; Bélanger et al. 2018). The Parti libéral du Québec (PLQ) wins most of the anglophone and allophone vote, and its fortunes among francophones decide elections. A common reading holds that this language divide is the national question under another name: non-francophones vote Liberal because they are federalists. If that reading is right, the divide should have narrowed after 2012, as the Coalition avenir Québec (CAQ), a nationalist party that sets independence aside, courted francophone federalists and the sovereignty question was said to lose ground. That decline is itself less certain than it seems: in federal elections, the weight of support for independence in the vote held remarkably steady from 2006 to 2021 (Daoust and Gareau-Paquette 2024).

Overall, the Quebec Election Studies give this reading only limited support. The gap in the Liberal vote between non-francophones and francophones was about 38 percentage points in 2007 and 50 in 2012; it widened in 2014 and 2018, and it was still 48 points in 2022. Support for independence accounts for part of it, but only part: in the four studies from 2012 to 2022, once support for independence is held constant, the gap falls from 54 to 42 points. What did change is the Liberal electorate itself. Francophones left it (20% of them voted PLQ in 2007, 6% in 2022), so that non-francophones went from 31% of Liberal voters to 71% (58% had their share of the voters in each sample stayed at its 2007 level).

The recipes below build that answer step by step; each runs as is.

library(qesR)
library(survey)
library(ggplot2)

1. One vote-choice variable for every study

vote_choice pools three questions into one column: the vote reported after the election, the intention with the undecided pushed toward a party, and the plain intention. vote_choice__type says which question each row answers.

studies <- c("qes1998", "qes2007", "qes2007_panel", "qes2008", "qes_crop_2007_2010",
             "qes2012", "qes2012_panel", "qes2014", "qes2018", "qes2018_panel", "qes2022")
h <- qes_harmonize(studies = studies, targets = c("vote_choice", "lang_mother"),
                   layout = "long", quiet = TRUE)
table(h$study, h$vote_choice__type)
#>                     
#>                      recall intention_push intention
#>   qes_crop_2007_2010      0          22848      1179
#>   qes1998              1483           1483         0
#>   qes2007              2175              0         0
#>   qes2007_panel        2054           1997        53
#>   qes2008              1151              0         0
#>   qes2012              1505              0         0
#>   qes2012_panel         844            844         0
#>   qes2014              1517              0         0
#>   qes2018              3072              0         0
#>   qes2018_panel         842           1250         0
#>   qes2022              1220           1521         0

In this long format, a study with a wave before and a wave after the election has one row per wave, so the same respondent appears once with an intention and once with a reported vote. The CROP polls, fielded between elections, only have intentions.

2. Keep the reported vote only

Setting an intention in one study against a reported vote in another would compare two different questions. For the language divide we keep the reported vote, which every Quebec Election Study asked after the election, and read each study in the default format, one row per respondent. types keeps only some of the questions a pooled variable brings together:

r <- qes_harmonize(studies = studies, targets = c("vote_choice", "lang_mother", "sov_support"),
                   types = list(vote_choice = "recall"), quiet = TRUE)
table(r$study, r$vote_choice__type, useNA = "ifany")
#>                     
#>                      recall intention_push intention  <NA>
#>   qes_crop_2007_2010      0              0         0 24027
#>   qes1998              1483              0         0     0
#>   qes2007              2175              0         0     0
#>   qes2007_panel        2054              0         0   388
#>   qes2008              1151              0         0     0
#>   qes2012              1505              0         0     0
#>   qes2012_panel         844              0         0     0
#>   qes2014              1517              0         0     0
#>   qes2018              3072              0         0     0
#>   qes2018_panel         842              0         0   408
#>   qes2022              1220              0         0   301

vote_choice is now missing for the CROP polls, and for respondents not reached after the vote. When a study asked several of these questions and no type is set, the reported vote comes first, then the pushed intention, then the plain one (qes_spec("pooled") lists that order).

3. Where each value comes from

prov <- qes_provenance(r, level = "pooled")
prov[prov$pooled == "vote_choice", c("study", "type", "wave", "grade", "n_value")]
#>                 study   type wave       grade n_value
#> 1             qes1998 recall post  comparable    1126
#> 7             qes2007 recall post  comparable    1727
#> 13      qes2007_panel recall post approximate    1494
#> 19            qes2008 recall post  comparable     898
#> 25 qes_crop_2007_2010 recall <NA>        <NA>       0
#> 31            qes2012 recall post   identical    1274
#> 37      qes2012_panel recall post approximate     633
#> 43            qes2014 recall post  comparable    1283
#> 49            qes2018 recall post  comparable    2016
#> 55      qes2018_panel recall post  comparable     704
#> 61            qes2022 recall  pes  comparable    1101

Each study’s question has a grade. It is identical when the question, the options and the format are the same, comparable when the differences (the order of the options, whether “don’t know” is offered) should not move the shares, and approximate when a format or a filter may move them. Only the 2007 panel and the 2012 panel are approximate here. The charts below therefore leave the panels aside and keep one Quebec Election Study per election.

d <- qes_design(r[r$study == "qes2018", ], weight = "weight_post")
d <- subset(d, !is.na(vote_choice))
caq18 <- svyciprop(~I(vote_choice == "CAQ"), d, method = "logit")
caq18
#>                                2.5% 97.5%
#> I(vote_choice == "CAQ") 0.358 0.336 0.381

The CAQ received 37.4% of the valid votes in 2018, and 35.8% of respondents to the 2018 study reported voting for it (95% confidence interval, 33.6 to 38.1). The official result thus falls inside that interval.

5. By group

plq18 <- svyby(~I(vote_choice == "PLQ"), ~lang_mother, subset(d, !is.na(lang_mother)),
               svyciprop, method = "logit", vartype = "ci")
plq18
#>         lang_mother I(vote_choice == "PLQ")      ci_l      ci_u
#> French       French               0.1223020 0.1070069 0.1394418
#> English     English               0.7418131 0.6814276 0.7942093
#> Other         Other               0.6416905 0.5267904 0.7423376

The output gives the Liberal vote of each mother-tongue group in 2018, with its interval; the realignment page follows these shares over every election. The allophone estimate rests on 84 respondents, hence its wider interval. The next recipes group anglophones and allophones as non-francophones, since the allophone samples are too small to follow on their own from one election to the next.

6. Every study, one at a time

Studies differ in population, mode, wording and weighting, so each estimate is made within one study before any comparison. The 1998 polls surveyed francophones only, so they have no non-francophones to compare, and the series starts in 2007. The 2008 study has no recommended weight (its weights are calibrated on the reported vote; see the weights table): its estimates are unweighted, and the charts draw them as hollow points.

qes <- c("qes2007", "qes2008", "qes2012", "qes2014", "qes2018", "qes2022")
r$francophone <- factor(ifelse(r$lang_mother == "French", "francophone", "other"))

# one design per study; a study without a post-election weight (2008) gets
# a unit weight, so its estimates are unweighted
is_weighted <- function(s, data = r) !all(is.na(data$weight_post[data$study == s]))
study_design <- function(s, data = r) {
  x <- data[data$study == s, ]
  if (!is_weighted(s, data)) x$weight_post <- 1
  qes_design(x, weight = "weight_post")
}

plq <- do.call(rbind, lapply(qes, function(s) {
  ds <- subset(study_design(s), !is.na(vote_choice) & !is.na(francophone))
  est <- svyby(~I(vote_choice == "PLQ"), ~francophone, ds, svyciprop, method = "logit", vartype = "ci")
  data.frame(study = s, year = as.integer(substr(ds$variables$year[1], 1, 4)),
             weighted = is_weighted(s), group = est$francophone,
             pct = 100 * est[[2]], lo = 100 * est$ci_l, hi = 100 * est$ci_u)
}))
print(plq, digits = 3, row.names = FALSE)
#>    study year weighted       group   pct    lo    hi
#>  qes2007 2007     TRUE francophone 20.14 17.91 22.59
#>  qes2007 2007     TRUE       other 58.44 47.99 68.18
#>  qes2008 2008    FALSE francophone 32.68 29.46 36.08
#>  qes2008 2008    FALSE       other 78.74 70.68 85.05
#>  qes2012 2012     TRUE francophone 15.49 13.23 18.05
#>  qes2012 2012     TRUE       other 65.46 57.20 72.87
#>  qes2014 2014     TRUE francophone 24.18 21.25 27.37
#>  qes2014 2014     TRUE       other 86.01 77.91 91.47
#>  qes2018 2018     TRUE francophone 12.23 10.70 13.94
#>  qes2018 2018     TRUE       other 71.83 66.55 76.58
#>  qes2022 2022     TRUE francophone  6.13  4.42  8.44
#>  qes2022 2022     TRUE       other 54.01 40.32 67.11

Line chart of the share of the reported vote for the PLQ at each Quebec election from 2007 to 2022, among francophones and among non-francophones. Among francophones it goes from 20% to 6%; among non-francophones it stays between 54% and 86%. Values in the table view. — At every election since 2007, non-francophones were 38 to 62 points more likely than francophones to vote LiberalShare of the reported vote for the PLQ among francophones and non-francophones (mother tongue), one Quebec Election Study per election, with 95% confidence intervals

Election	Study	Francophones, % PLQ [95% CI]	Non-francophones, % PLQ [95% CI]	Gap [95% CI]	n, francophones	n, non-francophones	Weighting
2007	QES 2007	20.1 [17.9, 22.6]	58.4 [48.0, 68.2]	+38.3 pts [27.9, 48.7]	1525	163	weighted
2008	QES 2008	32.7 [29.5, 36.1]	78.7 [70.7, 85.1]	+46.1 pts [38.2, 53.9]	771	127	unweighted: no validated weight
2012	QES 2012	15.5 [13.2, 18.1]	65.5 [57.2, 72.9]	+50.0 pts [41.8, 58.2]	1094	180	weighted
2014	QES 2014	24.2 [21.3, 27.4]	86.0 [77.9, 91.5]	+61.8 pts [54.5, 69.1]	981	201	weighted
2018	QES 2018	12.2 [10.7, 13.9]	71.8 [66.5, 76.6]	+59.6 pts [54.3, 64.9]	1656	359	weighted
2022	QES 2022	6.1 [4.4, 8.4]	54.0 [40.3, 67.1]	+47.9 pts [34.1, 61.6]	883	121	weighted

Two results emerge. First, the divide is large at every election, from 38 points in 2007 to 62 points in 2014, and it did not clearly narrow after 2012. Relative to 2012 (about 50 points), the gap widened in 2014 and 2018 and was back to about 48 points in 2022, an estimate whose interval (34 to 62) includes the 2012 gap. Second, the two groups did not move in step: the francophone Liberal vote fell from one election to the next (the realignment page follows it), while the non-francophone vote moved with no clear trend, and in 2022 it was about 4 points below its 2007 level. The narrower gap of 2022 comes from the non-francophone side: from 2018 to 2022, the Liberal vote fell 18 points among non-francophones and 6 among francophones. The non-francophone estimate of 2022, however, rests on only 121 respondents, the smallest sample of the series (the largest has 359), and its interval is wide. This is also why the page does not split non-francophones into anglophones and allophones.

7. Who votes Liberal

The same data can be read from the other side: not the share of each group that votes Liberal, but the share of Liberal voters who belong to each group. Read this way, the change is plain.

comp <- do.call(rbind, lapply(qes, function(s) {
  ds <- subset(study_design(s), vote_choice %in% "PLQ" & !is.na(francophone))
  est <- svyciprop(~I(francophone == "other"), ds, method = "logit")
  data.frame(study = s, year = plq$year[plq$study == s][1], weighted = is_weighted(s),
             pct = 100 * as.numeric(est), lo = 100 * attr(est, "ci")[[1]], hi = 100 * attr(est, "ci")[[2]])
}))
print(comp, digits = 3, row.names = FALSE)
#>    study year weighted  pct   lo   hi
#>  qes2007 2007     TRUE 30.9 25.3 37.2
#>  qes2008 2008    FALSE 28.4 23.9 33.4
#>  qes2012 2012     TRUE 49.5 42.9 56.2
#>  qes2014 2014     TRUE 34.8 29.5 40.4
#>  qes2018 2018     TRUE 57.2 52.5 61.8
#>  qes2022 2022     TRUE 71.2 60.0 80.3

Line chart of the share of non-francophones among respondents who reported a PLQ vote, at each Quebec election from 2007 to 2022, with a dashed line at 50%. It goes from 31% in 2007 to 71% in 2022. Values in the table view. — Non-francophones went from 31% of Liberal voters in 2007 to 71% in 2022Share of respondents whose mother tongue is not French among those who reported a vote for the PLQ, one Quebec Election Study per election, with 95% confidence intervals

Election	Study	Non-francophones among PLQ voters, % [95% CI]	n, PLQ voters	Non-francophones among all voters, %	Non-francophones among PLQ voters, at their 2007 share of all voters, %	Weighting
2007	QES 2007	30.9 [25.3, 37.2]	424	13.4	30.9	weighted
2008	QES 2008	28.4 [23.9, 33.4]	352	14.1	27.1	unweighted: no validated weight
2012	QES 2012	49.5 [42.9, 56.2]	278	18.8	39.5	weighted
2014	QES 2014	34.8 [29.5, 40.4]	417	13.0	35.4	weighted
2018	QES 2018	57.2 [52.5, 61.8]	490	18.5	47.5	weighted
2022	QES 2022	71.2 [60.0, 80.3]	114	21.9	57.6	weighted

Non-francophones made up 31% of PLQ voters in 2007 and 71% in 2022. The path is not smooth, and each point rests on a few hundred Liberal voters at most. The share also depends on something other than the Liberal vote of each group: the weighted share of non-francophones among the voters of each study, which swings more than any change in the population would. That share is about 13% in 2007 and 2014, and 19 to 22% in 2012, 2018 and 2022; the dip of 2014 lines up with that year’s low share. Held at its 2007 level, the share of non-francophones among Liberal voters would still reach 58% in 2022. The rise holds, then, but its size depends on the composition of each sample. In other words, the PLQ went from a party whose voters were mostly francophone to one whose voters mostly are not. This is not because non-francophones rallied to it: francophones left.

8. Sovereignty: one variable, several wordings

sov_support pools the referendum questions. The wording changed over time, and the pooled column records it in sov_support__type: shares should be compared within one wording only.

table(r$study, r$sov_support__type)
#>                     
#>                      independence sovereign_country partnership_1995_push
#>   qes_crop_2007_2010            0                 0                     0
#>   qes1998                       0                 0                  1483
#>   qes2007                       0                 0                  2175
#>   qes2007_panel                 0                 0                  2050
#>   qes2008                       0                 0                  1151
#>   qes2012                    1505                 0                     0
#>   qes2012_panel                 0               844                     0
#>   qes2014                    1517                 0                     0
#>   qes2018                    3072                 0                     0
#>   qes2018_panel                 0                 0                     0
#>   qes2022                    1521                 0                     0
#>                     
#>                      partnership_1995 favour
#>   qes_crop_2007_2010                0      0
#>   qes1998                           0      0
#>   qes2007                           0      0
#>   qes2007_panel                     0      0
#>   qes2008                           0      0
#>   qes2012                           0      0
#>   qes2012_panel                     0      0
#>   qes2014                           0      0
#>   qes2018                           0      0
#>   qes2018_panel                     0    842
#>   qes2022                           0      0

The four Quebec Election Studies from 2012 to 2022 asked the same question on independence, which the next recipe uses. The earlier studies asked the partnership question of the 1995 referendum, and the 2012 and 2018 panels used wordings of their own.

9. Pooling studies in one model

Pooling studies only makes sense in a model that keeps them apart, with study fixed effects, so that each association is estimated within studies. pool = "equal" gives every study the same total weight among the rows it receives, so we select the rows of the model first; by default a study counts in proportion to its number of respondents. We use a linear probability model, so a coefficient times 100 reads in percentage points. The first model has only the study fixed effects; the second adds support for independence; the third lets the language gap differ from one study to the next.

four <- c("qes2012", "qes2014", "qes2018", "qes2022")
x <- r[r$study %in% four & !is.na(r$vote_choice) & !is.na(r$francophone) & !is.na(r$sov_support), ]
dp <- qes_design(x, weight = "weight_post", pool = "equal")
m1 <- svyglm(I(vote_choice == "PLQ") ~ francophone + study, dp)
m2 <- svyglm(I(vote_choice == "PLQ") ~ francophone + I(sov_support == "Yes") + study, dp)
m3 <- svyglm(I(vote_choice == "PLQ") ~ francophone * study + I(sov_support == "Yes"), dp)
100 * rbind(study_effects = coef(m1)[2], plus_independence = coef(m2)[2])
#>                   francophoneother
#> study_effects             54.07327
#> plus_independence         42.37054
regTermTest(m3, ~ francophone:study)
#> Wald test for francophone:study
#>  in svyglm(formula = I(vote_choice == "PLQ") ~ francophone * study + 
#>     I(sov_support == "Yes"), design = dp)
#> F =  2.150366  on  3  and  4851  df: p= 0.091757

Across the four elections, and with study fixed effects only, non-francophones are about 54 points more likely than francophones to vote Liberal (95% confidence interval, 49 to 59). Among voters who give the same answer to the independence question, the gap is still 42 points. Support for independence thus accounts for about 12 of the 54 points of the language divide in the Liberal vote. In other words, most of the divide is not the national question under another name, at least as one question on independence measures it. Finally, the evidence that the gap changed between 2012 and 2022 is weak (a test of equal gaps across the four studies gives p = 0.09).

10. The ADQ and the CAQ as one series, against the vote count

The harmonized columns keep the Action démocratique du Québec (ADQ) and the CAQ apart. qes_party_lineage() adds a column that joins them, for time series; its rows fielded before the merger of 2012 are graded approximate.

l <- qes_party_lineage(r, cols = "vote_choice")
table(l$study, l$vote_choice_lineage)[, c("PLQ", "PQ", "ADQ/CAQ", "QS")]
#>                     
#>                      PLQ  PQ ADQ/CAQ  QS
#>   qes_crop_2007_2010   0   0       0   0
#>   qes1998            384 517     204   0
#>   qes2007            439 521     583  75
#>   qes2007_panel      428 460     480  57
#>   qes2008            352 336     144  38
#>   qes2012            278 509     324  96
#>   qes2012_panel      167 243     151  45
#>   qes2014            489 344     277 131
#>   qes2018            490 392     700 342
#>   qes2018_panel      200 108     263  87
#>   qes2022            146 206     364 209

Weighted within each study and set against the official result, that column gives a first check on the reported vote. p is a plain ggplot2 chart to adapt (print it with p).

lineage <- do.call(rbind, lapply(qes, function(s) {
  ds <- subset(study_design(s, l), !is.na(vote_choice_lineage))
  est <- svyciprop(~I(vote_choice_lineage == "ADQ/CAQ"), ds, method = "logit")
  ci <- attr(est, "ci")
  data.frame(study = s, year = plq$year[plq$study == s][1], weighted = is_weighted(s, l),
             pct = 100 * as.numeric(est), lo = 100 * ci[[1]], hi = 100 * ci[[2]])
}))
# the official share of valid votes of the ADQ (2007, 2008) and of the CAQ
# (2012 on), from Élections Québec
official <- data.frame(year = c(2007, 2008, 2012, 2014, 2018, 2022),
                       pct = c(30.8, 16.4, 27.1, 23.1, 37.4, 41.0))
# the chart's labels, in English or French
labels_of <- function(lang = "en") {
  if (lang == "fr") {
    list(series = c(lineage = "ADQ (2007, 2008), puis CAQ : vote déclaré", official = "Résultat officiel"),
         weight = c(`TRUE` = "pondéré", `FALSE` = "non pondéré"), y = "% du vote déclaré")
  } else {
    list(series = c(lineage = "ADQ (2007, 2008), then CAQ: reported vote", official = "Official result"),
         weight = c(`TRUE` = "weighted", `FALSE` = "unweighted"), y = "% of the reported vote")
  }
}
lab <- labels_of(lang = "en")

p <- ggplot(lineage, aes(year, pct)) +
  geom_line(data = official, aes(colour = "official")) +
  geom_point(data = official, aes(colour = "official"), shape = 45, size = 9) +
  geom_linerange(aes(ymin = lo, ymax = hi, colour = "lineage")) +
  geom_point(aes(colour = "lineage", fill = weighted), shape = 21, size = 3, stroke = 1) +
  scale_colour_manual(values = c(lineage = "#1d97b0", official = "grey20"), labels = lab$series) +
  scale_fill_manual(values = c(`TRUE` = "#1d97b0", `FALSE` = "white"), labels = lab$weight) +
  scale_x_continuous(breaks = c(2007, 2012, 2014, 2018, 2022)) +
  scale_y_continuous(limits = c(0, 50)) +
  labs(x = NULL, y = lab$y, colour = NULL, fill = NULL)
print(lineage, digits = 3, row.names = FALSE)
#>    study year weighted  pct   lo   hi
#>  qes2007 2007     TRUE 31.6 29.0 34.3
#>  qes2008 2008    FALSE 16.0 13.8 18.6
#>  qes2012 2012     TRUE 25.4 22.8 28.2
#>  qes2014 2014     TRUE 23.1 20.5 26.0
#>  qes2018 2018     TRUE 35.8 33.6 38.1
#>  qes2022 2022     TRUE 33.0 29.3 36.9

Dot chart with confidence intervals of the share of the reported vote for the ADQ in 2007 and 2008 and for the CAQ from 2012 to 2022, in the Quebec Election Studies, with the official result as a line with ticks. In 2022 the study has 33% against 41% officially. Values in the table view. — With the ADQ and the CAQ joined, the reported vote is within 2 points of the official result until 2018, and 8 points short in 2022Share of the reported vote for the ADQ (2007, 2008) and the CAQ (2012 on), one Quebec Election Study per election, with 95% confidence intervals and the official result

Election	Study	Party	Reported vote, % [95% CI]	Official, % of valid votes	Difference	Weighting
2007	QES 2007	ADQ	31.6 [29.0, 34.3]	30.8	+0.8 pts	weighted
2008	QES 2008	ADQ	16.0 [13.8, 18.6]	16.4	−0.4 pts	unweighted: no validated weight
2012	QES 2012	CAQ	25.4 [22.8, 28.2]	27.1	−1.7 pts	weighted
2014	QES 2014	CAQ	23.1 [20.5, 26.0]	23.1	0.0 pts	weighted
2018	QES 2018	CAQ	35.8 [33.6, 38.1]	37.4	−1.6 pts	weighted
2022	QES 2022	CAQ	33.0 [29.3, 36.9]	41.0	−8.0 pts	weighted

Up to 2018, the reported vote stays within 2 points of the official result. In 2022, contrary to the usual overreporting of the winner after an election, it falls 8 points short, while the same study puts the PLQ at 17% against 14% officially. The levels of 2022 call for some caution. The argument above rests on differences between groups within each study, which are less exposed than levels to an error of this kind, though not immune to it.

What the recipes show

In sum, the Quebec Election Studies give only limited support to the idea that the language divide is the national question under another name. Support for independence explains part of the gap in the Liberal vote, and the gap did not clearly narrow after 2012. What changed is who votes Liberal: as francophones left, the PLQ became a party whose voters are mostly non-francophones. Our analysis is not without limits. First, mother tongue is a coarse measure. Recent work also treats language as an attitude, the worry that French is under threat, and that worry varies among francophones themselves (Daoust and Gareau-Paquette 2026). Second, the estimates are associations within cross-sections, not causes, and the non-francophone samples are small, and their share of each sample varies from one study to the next. Of course, the divide does not hold for every party: support for the Parti conservateur du Québec (PCQ) cut across it in 2022 (Bélanger et al. 2025). The next question is where the francophone Liberals went, which the pages on realignment and on changing minds during the campaign take up.