Conceptions of Care vs. Fear-Based Safety and Preferences for Violence Intervention Approaches

Brief report for the 2026 American Public Health Association

Published

October 17, 2026

Abstract

Community safety can be promoted through education, legislation, and policing, among other approaches. Scholar activist Zach Norris 2021 theorizes community safety approaches can be classified into one of two safety paradigms: care-based safety or fear-based safety, conceptualizing care-based approaches by their aim to increase health access, economic resources and social cohesion, whereas, fear-based approaches rely on an in-group fearing a “dangerous” out-group. However, the field has yet to empirically determine whether these conceptions of safety are held by the public. In this mixed-method study of 84 Broome County, NY residents, we analyzed: 1) qualitative responses to “what does safety mean to you?” to examine lay conceptions of community safety using NVivo 14, and 2) quantitative ratings of various care-based (e.g., education) vs. fear-based approaches (e.g., policing) using R. Content analysis of participant responses confirmed a two-category safety framework: care-based safety (n=68) and fear-based safety (n=49), and an exploratory factor analysis of the ratings likewise separated care-based and fear-based approaches. Care-based approaches received a higher level of support (M = 4.88) than fear-based approaches (M = 2.46). Quantitatively, a Wilcoxon signed-rank test showed a statistically significant difference between support for care-based and fear-based approaches (V = 3176, p < .001). Additionally, a multiple linear regression controlling for racialized identity and gender revealed growth mindset (β = .46, p < .001) as the most important explanatory variable of supporting a care-based safety approach. The study suggests that the care vs. fear-based safety distinction could be a theoretical framework of community safety and a future direction for academic violence prevention research.

1 Background

Community violence is violence that occurs between unrelated individuals outside of the home.[1]

  • Norris (2021) has conceptualized community safety as:
    • Care-Based Safety: Approaches that promote accessibility of physical and mental health resources, increase economic security, and hold individuals accountable for harm to the community[2]
    • Fear-Based Safety: Approaches that rely on an in-group fearing an out-group because the out-group is viewed as “dangerous”[2]
  • Psychological Beliefs about People:
    • Growth/Incremental Mindset: Belief that people are capable of internal change[3]
    • Fixed Mindset: Belief that individuals are born with certain characteristics and traits that cannot or will not change[3]
  • The study examines care-based vs. fear-based safety by:
    • assessing Binghamton & Broome Co. community’s views of effective violence prevention/intervention (VPI)
    • assessing if growth/fixed mindsets influence VPI

2 Methods

## IMPORT
library(readxl)    # read in xlsx files

## TIDY + TRANSFORM + VISUALIZE
library(tidyverse) # loads tidyr, dplyr, ggplot2, and more
library(psych)     # composite scores, reliability, descriptive stats, and factor analysis
library(systemfonts) # register the Calibri font for plots

## COMMUNICATE
library(knitr)     # tables
# Calibri comes with Microsoft Office. On a Mac it lives inside the Word app,
# where R cannot see it, so register it by file path. On Windows, Calibri is
# already installed and this step is skipped.
calibri_folder <- "/Applications/Microsoft Word.app/Contents/Resources/DFonts"

if (file.exists(file.path(calibri_folder, "Calibri.ttf"))) {
  register_font(name       = "Calibri",
                plain      = file.path(calibri_folder, "Calibri.ttf"),
                bold       = file.path(calibri_folder, "Calibrib.ttf"),
                italic     = file.path(calibri_folder, "Calibrii.ttf"),
                bolditalic = file.path(calibri_folder, "Calibriz.ttf"))
}

match_fonts("Calibri")$path      # should end in Calibri.ttf
[1] "/Applications/Microsoft Word.app/Contents/Resources/DFonts/Calibri.ttf"
#source: systemfonts package (Pedersen et al.) https://systemfonts.r-lib.org/; Quarto brand.yml https://quarto.org/docs/authoring/brand.html
#explanation: _brand.yml sets Calibri for the report text, but plots are drawn by R, so R also needs to find Calibri; every plot below uses theme_bw(base_family = "Calibri"); if Calibri cannot be found, R falls back to its default sans-serif font

2.1 Data preparation

2.1.1 Import data once

# 1. alldata: the file exactly as exported. Never change it.
alldata <- read_excel("11.06.2025.data.team1.clean.xlsx")

nrow(alldata)   # how many people?
[1] 210
ncol(alldata)   # how many variables?
[1] 88
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/import-once.html
#explanation: alldata is the Qualtrics export cleaned by the team on 11.06.2025; every later chunk builds on a copy of it

2.1.2 Recode don’t know and prefer not to say to missing

# 2. cleandata: the copy you clean and build on. Every later chunk adds to it.
cleandata <- alldata

sum(cleandata == -99, na.rm = TRUE)   # prefer not to say: how many answers before cleaning?
[1] 18
sum(cleandata == -50, na.rm = TRUE)   # don't know: how many answers before cleaning?
[1] 0
cleandata[cleandata == -99] <- NA     # prefer not to say -> missing
cleandata[cleandata == -50] <- NA     # don't know -> missing

sum(cleandata == -99, na.rm = TRUE)   # should now be 0
[1] 0
sum(cleandata == -50, na.rm = TRUE)   # should now be 0
[1] 0
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/import-once.html
#explanation: -99 (prefer not to say) and -50 (don't know) are Qualtrics codes, not real answers, so they become NA before any other step; the person stays in the data and only that one answer is treated as missing

2.1.3 POC: White vs. People of Color

“What is your racial/ethnic identity?” allowed more than one answer, so RACIALIZED is stored as text: one code (for example “7”) or several codes joined by commas (for example “4,7”). Code 7 is White.

cleandata <- cleandata |>
  mutate(POC = case_when(
    RACIALIZED == "7"  ~ 0,     # White only
    !is.na(RACIALIZED) ~ 1      # any other identity, or two or more identities
  ))

# ALWAYS check a transformation: every RACIALIZED answer lands in the right group
cleandata |> count(RACIALIZED, POC)
# A tibble: 10 × 3
   RACIALIZED   POC     n
   <chr>      <dbl> <int>
 1 1,2,3,4,6      1     1
 2 2              1    22
 3 2,7            1     3
 4 3              1     2
 5 4              1     4
 6 4,5            1     1
 7 4,7            1     3
 8 5              1     1
 9 7              0    36
10 <NA>          NA   137
cleandata |> count(POC)
# A tibble: 3 × 2
    POC     n
  <dbl> <int>
1     0    36
2     1    37
3    NA   137
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/dplyr.html
#explanation: 0 = racialized as White (White was the only identity chosen), 1 = racialized as a Person of Color (every other identity, including people who chose White plus another identity); -99 (prefer not to say) was already set to NA, so it stays NA

2.1.4 WOMAN: man vs. woman

GENDER is coded 0 = woman, 1 = man, 2 = another gender identity (nonbinary, genderfluid, or genderqueer). Prefer not to say (-99) is already NA.

cleandata <- cleandata |>
  mutate(WOMAN = case_when(
    GENDER == 1 ~ 0,     # man
    GENDER == 0 ~ 1,     # woman
    GENDER == 2 ~ NA     # another gender identity is excluded (too few to analyze)
  ))

# ALWAYS check a transformation: every GENDER answer lands in the right group
cleandata |> count(GENDER, WOMAN)
# A tibble: 4 × 3
  GENDER WOMAN     n
   <dbl> <dbl> <int>
1      0     1    48
2      1     0    24
3      2    NA     2
4     NA    NA   136
cleandata |> count(WOMAN)
# A tibble: 3 × 2
  WOMAN     n
  <dbl> <int>
1     0    24
2     1    48
3    NA   138
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/dplyr.html
#explanation: 0 = man, 1 = woman; the two people who chose another gender identity are set to NA because two people are too few to form a group; note the original GENDER codes run the other way (0 = woman), so the check table confirms the flip

2.1.5 GROWTH: growth mindset composite

The eight FIXEDPERSON items are coded 1 = strongly agree to 6 = strongly disagree with a fixed-mindset statement, so disagreeing means a growth mindset. The four items ending in _R are worded the growth way and are reversed.

# Create keys for scoring: higher GROWTH = stronger belief that people can change
growth_keys <- list(
  GROWTH = c("FIXEDPERSON1_BASIC", "FIXEDPERSON2_DIFF", "-FIXEDPERSON3_CHANGE_R", "FIXEDPERSON4_OLD",
             "-FIXEDPERSON_ALL_R", "-FIXEDPERSON_ALWAYS_R", "FIXEDPERSON_CERTAIN", "-FIXEDPERSON_MATTER_R")
)

#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/composite.html
#explanation: growth_keys lists the eight mindset items; the minus sign reverses the four growth-worded _R items so that every item points the same way before averaging
growth_scores <- scoreItems(growth_keys, cleandata, impute = "none")
growth_scores                                    # alpha is in the first table
Call: scoreItems(keys = growth_keys, items = cleandata, impute = "none")

(Standardized) Alpha:
      GROWTH
alpha   0.87

Standard errors of unstandardized Alpha:
      GROWTH
ASE    0.026

Standardized Alpha of observed scales:
     GROWTH
[1,]   0.87

Average item correlation:
          GROWTH
average.r   0.46

Median item correlation:
GROWTH 
  0.43 

 Guttman 6* reliability: 
         GROWTH
Lambda.6   0.89

Signal/Noise based upon av.r : 
             GROWTH
Signal/Noise    6.9

Scale intercorrelations corrected for attenuation 
 raw correlations below the diagonal, alpha on the diagonal 
 corrected correlations above the diagonal:

Note that these are the correlations of the complete scales based on the correlation matrix,
 not the observed scales based on the raw items.
       GROWTH
GROWTH   0.87

 Average adjusted correlations within and between scales (MIMS)
[1] 0.46

 Average adjusted item x scale correlations within and between scales (MIMT)
[1] 0.73

 In order to see the item by scale loadings and frequency counts of the data
 print with the short option = FALSE
cleandata$GROWTH <- growth_scores$scores[, "GROWTH"]

summary(cleandata$GROWTH)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max.    NA's 
  2.000   3.625   4.125   4.185   4.670   6.000     135 
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/composite.html; psych package (Revelle)
#explanation: scoreItems() averages each person's available mindset items (impute = "none") and reports Cronbach's alpha; GROWTH runs from 1 (strong fixed mindset) to 6 (strong growth mindset)
cleandata <- cleandata |>
  mutate(GROWTH_cent = GROWTH - mean(GROWTH, na.rm = TRUE))

summary(cleandata$GROWTH_cent)                      # mean is now 0
    Min.  1st Qu.   Median     Mean  3rd Qu.     Max.     NA's 
-2.18487 -0.55987 -0.05987  0.00000  0.48477  1.81513      135 
#source: Aiken & West (1991) Multiple regression: Testing and interpreting interactions
#explanation: GROWTH_cent is GROWTH minus its average, so 0 means an average growth mindset; the regression intercept is then the predicted rating for a White respondent with an average growth mindset instead of at a growth score of 0, which nobody can have on a 1 to 6 scale

2.2 Effectiveness items: exploratory factor analysis

Research question: Do people’s beliefs about the effectiveness of violence prevention and intervention approaches organize into separate care-based and fear-based dimensions?

The EFFECTIVENESS_RATE block asked how much each person agrees that each of nine violence prevention and intervention approaches is effective (1 = strongly disagree to 6 = strongly agree). Five approaches were written as care-based (EFFECT_CARE_) and four as fear-based (EFFECT_FEAR_). This section tests whether beliefs about effectiveness actually organize into those two dimensions.

2.2.1 Select the items

effectdata <- cleandata |>
  select(EFFECT_CARE_COMM, EFFECT_CARE_EDUCATION, EFFECT_CARE_HVIP,
         EFFECT_CARE_MHRESOURCES, EFFECT_CARE_WELFARE,
         EFFECT_FEAR_LEG, EFFECT_FEAR_POLICE, EFFECT_FEAR_DEPORT, EFFECT_FEAR_ARMY)

names(effectdata)                    # the nine EFFECTIVENESS_RATE items
[1] "EFFECT_CARE_COMM"        "EFFECT_CARE_EDUCATION"  
[3] "EFFECT_CARE_HVIP"        "EFFECT_CARE_MHRESOURCES"
[5] "EFFECT_CARE_WELFARE"     "EFFECT_FEAR_LEG"        
[7] "EFFECT_FEAR_POLICE"      "EFFECT_FEAR_DEPORT"     
[9] "EFFECT_FEAR_ARMY"       
colSums(!is.na(effectdata))          # how many people answered each item
       EFFECT_CARE_COMM   EFFECT_CARE_EDUCATION        EFFECT_CARE_HVIP 
                     80                      82                      73 
EFFECT_CARE_MHRESOURCES     EFFECT_CARE_WELFARE         EFFECT_FEAR_LEG 
                     81                      78                      80 
     EFFECT_FEAR_POLICE      EFFECT_FEAR_DEPORT        EFFECT_FEAR_ARMY 
                     83                      81                      80 
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/dplyr.html
#explanation: effectdata holds only the nine effectiveness items so the correlation table and factor analysis use the same columns

2.2.2 Describe the items

effect_describe <- describe(effectdata)
kable(effect_describe[, c("n", "mean", "sd", "median", "min", "max", "skew")], digits = 2,
      caption = "Descriptive statistics for the nine effectiveness items")
Descriptive statistics for the nine effectiveness items
n mean sd median min max skew
EFFECT_CARE_COMM 80 4.79 1.11 5 1 6 -1.17
EFFECT_CARE_EDUCATION 82 4.78 1.19 5 1 6 -1.24
EFFECT_CARE_HVIP 73 4.01 0.99 4 2 6 -0.28
EFFECT_CARE_MHRESOURCES 81 5.14 0.97 5 1 6 -1.40
EFFECT_CARE_WELFARE 78 4.86 1.07 5 1 6 -1.25
EFFECT_FEAR_LEG 80 4.19 1.26 4 1 6 -0.42
EFFECT_FEAR_POLICE 83 3.29 1.44 4 1 6 -0.12
EFFECT_FEAR_DEPORT 81 1.83 1.16 1 1 6 1.57
EFFECT_FEAR_ARMY 80 2.12 1.33 2 1 6 1.19
#source: psych package (Revelle) https://personality-project.org/r/psych/
#explanation: describe() gives n, mean, sd, median, min, max, and skew for every item; the fear-based deportation and military items are strongly skewed toward disagree

2.2.3 Correlation table

options(knitr.kable.NA = "")              # print empty cells instead of NA in every table below

effect_cor <- corr.test(effectdata, use = "pairwise", adjust = "none")

effect_r <- round(effect_cor$r, 2)
effect_r[upper.tri(effect_r)] <- NA      # show the lower triangle only

# abbreviate each item to everything after the second _ (EFFECT_CARE_COMM -> COMM)
rownames(effect_r) <- sub("^EFFECT_(CARE|FEAR)_", "", rownames(effect_r))
colnames(effect_r) <- sub("^EFFECT_(CARE|FEAR)_", "", colnames(effect_r))

effect_r_print <- ifelse(is.na(effect_r), "", sprintf("%.2f", effect_r))   # two decimals in every cell
dimnames(effect_r_print) <- dimnames(effect_r)

kable(effect_r_print, caption = "Pearson correlations among the effectiveness items (pairwise deletion). Care: COMM, EDUCATION, HVIP, MHRESOURCES, WELFARE; fear: LEG, POLICE, DEPORT, ARMY.")
Pearson correlations among the effectiveness items (pairwise deletion). Care: COMM, EDUCATION, HVIP, MHRESOURCES, WELFARE; fear: LEG, POLICE, DEPORT, ARMY.
COMM EDUCATION HVIP MHRESOURCES WELFARE LEG POLICE DEPORT ARMY
COMM 1.00
EDUCATION 0.45 1.00
HVIP 0.25 0.24 1.00
MHRESOURCES 0.31 0.33 0.37 1.00
WELFARE 0.36 0.15 0.31 0.49 1.00
LEG -0.04 0.08 -0.04 -0.15 -0.10 1.00
POLICE 0.00 -0.10 0.05 -0.14 -0.16 0.23 1.00
DEPORT -0.24 -0.24 0.13 -0.26 -0.20 0.17 0.40 1.00
ARMY -0.11 -0.18 0.29 -0.03 -0.21 -0.01 0.40 0.70 1.00
#source: psych package (Revelle) https://personality-project.org/r/psych/; The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/correlation.html
#explanation: corr.test() computes every pairwise Pearson correlation; sub() drops the EFFECT_CARE_ or EFFECT_FEAR_ prefix so the table fits on the page (COMM = community-based interventions, EDUCATION = educational programs, HVIP = hospital-based violence intervention, MHRESOURCES = mental health resources, WELFARE = welfare and economic support, LEG = legislation, POLICE = policing, DEPORT = deportation, ARMY = military or National Guard deployment); care items correlate positively with each other, fear items correlate positively with each other, and the two sets are near zero or negative with each other
effect_p <- round(effect_cor$p, 3)
effect_p[upper.tri(effect_p)] <- NA

rownames(effect_p) <- sub("^EFFECT_(CARE|FEAR)_", "", rownames(effect_p))
colnames(effect_p) <- sub("^EFFECT_(CARE|FEAR)_", "", colnames(effect_p))

effect_p_print <- ifelse(is.na(effect_p), "", sprintf("%.3f", abs(effect_p)))   # abs() removes rounding signs like -0.000   # three decimals in every cell
dimnames(effect_p_print) <- dimnames(effect_p)

kable(effect_p_print, caption = "p-values for the correlations above (unadjusted)")
p-values for the correlations above (unadjusted)
COMM EDUCATION HVIP MHRESOURCES WELFARE LEG POLICE DEPORT ARMY
COMM 0.000
EDUCATION 0.000 0.000
HVIP 0.037 0.038 0.000
MHRESOURCES 0.007 0.003 0.001 0.000
WELFARE 0.002 0.207 0.010 0.000 0.000
LEG 0.726 0.466 0.752 0.197 0.377 0.000
POLICE 0.998 0.397 0.693 0.216 0.158 0.038 0.000
DEPORT 0.034 0.032 0.270 0.022 0.079 0.139 0.000 0.000
ARMY 0.326 0.107 0.014 0.775 0.077 0.957 0.000 0.000 0.000
#source: psych package (Revelle) https://personality-project.org/r/psych/
#explanation: adjust = "none" keeps the raw p-values; with about 80 people a correlation of about .22 or larger is significant at p < .05
effect_n <- effect_cor$n
effect_n[upper.tri(effect_n)] <- NA

rownames(effect_n) <- sub("^EFFECT_(CARE|FEAR)_", "", rownames(effect_n))
colnames(effect_n) <- sub("^EFFECT_(CARE|FEAR)_", "", colnames(effect_n))

kable(effect_n, caption = "Number of people behind each correlation (pairwise n)")
Number of people behind each correlation (pairwise n)
COMM EDUCATION HVIP MHRESOURCES WELFARE LEG POLICE DEPORT ARMY
COMM 80
EDUCATION 79 82
HVIP 71 73 73
MHRESOURCES 77 80 72 81
WELFARE 75 77 70 77 78
LEG 77 79 71 78 76 80
POLICE 80 81 73 80 78 79 83
DEPORT 78 79 72 78 76 78 81 81
ARMY 77 78 71 77 75 77 80 79 80
#source: psych package (Revelle) https://personality-project.org/r/psych/
#explanation: pairwise deletion uses every person who answered both items, so n varies from 70 to 83 across pairs

2.2.4 Check suitability for factor analysis

effectdata_complete <- effectdata |>
  drop_na()                          # factor analysis needs complete cases

nrow(effectdata_complete)            # people who answered all nine items
[1] 66
effect_kmo <- KMO(effectdata_complete)
effect_kmo                           # overall MSA and one MSA per item
Kaiser-Meyer-Olkin factor adequacy
Call: KMO(r = effectdata_complete)
Overall MSA =  0.62
MSA for each item = 
       EFFECT_CARE_COMM   EFFECT_CARE_EDUCATION        EFFECT_CARE_HVIP 
                   0.71                    0.59                    0.78 
EFFECT_CARE_MHRESOURCES     EFFECT_CARE_WELFARE         EFFECT_FEAR_LEG 
                   0.63                    0.57                    0.40 
     EFFECT_FEAR_POLICE      EFFECT_FEAR_DEPORT        EFFECT_FEAR_ARMY 
                   0.79                    0.60                    0.54 
effect_bartlett <- cortest.bartlett(cor(effectdata_complete), n = nrow(effectdata_complete))
effect_bartlett                      # chi-square, df, p
$chisq
[1] 157.5386

$p.value
[1] 3.81639e-17

$df
[1] 36
#source: psych package (Revelle) https://personality-project.org/r/psych/; Kaiser (1974) https://doi.org/10.1007/BF02291575
#explanation: KMO above .60 and a significant Bartlett test mean the correlations are strong enough to factor; legislation has the lowest item MSA, a first sign it does not belong with the other items

2.2.5 How many factors?

par(family = "Calibri")                 # base-graphics plots also use Calibri
set.seed(1)
effect_parallel <- fa.parallel(effectdata_complete, fm = "minres", fa = "fa", n.iter = 1000)

Parallel analysis: factors are kept while the observed eigenvalue (blue) is above the eigenvalue from random data (red).
Parallel analysis suggests that the number of factors =  2  and the number of components =  NA 
effect_parallel$nfact                 # number of factors suggested
[1] 2
round(effect_parallel$fa.values, 2)   # observed eigenvalues
[1]  2.12  1.09  0.31  0.03 -0.02 -0.17 -0.29 -0.39 -0.56
#source: psych package (Revelle) https://personality-project.org/r/psych/; Horn (1965) https://doi.org/10.1007/BF02289447
#explanation: parallel analysis compares the eigenvalues from our data with eigenvalues from 1000 random datasets of the same size; two eigenvalues are larger than chance, so two factors are suggested
efa1 <- fa(effectdata_complete, nfactors = 1, fm = "minres")
print(efa1$loadings, cutoff = 0.30)

Loadings:
                        MR1   
EFFECT_CARE_COMM         0.535
EFFECT_CARE_EDUCATION    0.452
EFFECT_CARE_HVIP         0.333
EFFECT_CARE_MHRESOURCES  0.740
EFFECT_CARE_WELFARE      0.589
EFFECT_FEAR_LEG               
EFFECT_FEAR_POLICE      -0.411
EFFECT_FEAR_DEPORT      -0.552
EFFECT_FEAR_ARMY        -0.348

                 MR1
SS loadings    2.118
Proportion Var 0.235
#source: psych package (Revelle) https://personality-project.org/r/psych/
#explanation: in a one-factor solution care items load positively and fear items load negatively, forcing care and fear onto opposite ends of a single line; the fit statistics below show this model fits poorly
efa2 <- fa(effectdata_complete, nfactors = 2, fm = "minres", rotate = "oblimin")
print(efa2$loadings, cutoff = 0.30)

Loadings:
                        MR1    MR2   
EFFECT_CARE_COMM         0.602       
EFFECT_CARE_EDUCATION    0.473       
EFFECT_CARE_HVIP         0.623       
EFFECT_CARE_MHRESOURCES  0.756       
EFFECT_CARE_WELFARE      0.549       
EFFECT_FEAR_LEG                      
EFFECT_FEAR_POLICE              0.535
EFFECT_FEAR_DEPORT              0.775
EFFECT_FEAR_ARMY                0.864

                 MR1   MR2
SS loadings    1.924 1.757
Proportion Var 0.214 0.195
Cumulative Var 0.214 0.409
round(efa2$Phi, 2)                    # correlation between the two factors
      MR1   MR2
MR1  1.00 -0.17
MR2 -0.17  1.00
#source: psych package (Revelle) https://personality-project.org/r/psych/
#explanation: oblimin rotation lets the two factors correlate; one factor holds the five care items and the other holds police, deportation, and military; the factors correlate only about -.17
efa3 <- fa(effectdata_complete, nfactors = 3, fm = "minres", rotate = "oblimin")
print(efa3$loadings, cutoff = 0.30)

Loadings:
                        MR1    MR2    MR3   
EFFECT_CARE_COMM                0.454       
EFFECT_CARE_EDUCATION                  1.000
EFFECT_CARE_HVIP         0.318  0.474       
EFFECT_CARE_MHRESOURCES         0.780       
EFFECT_CARE_WELFARE             0.699       
EFFECT_FEAR_LEG                -0.307       
EFFECT_FEAR_POLICE       0.526              
EFFECT_FEAR_DEPORT       0.750              
EFFECT_FEAR_ARMY         0.876              

                 MR1   MR2   MR3
SS loadings    1.718 1.674 1.169
Proportion Var 0.191 0.186 0.130
Cumulative Var 0.191 0.377 0.507
#source: psych package (Revelle) https://personality-project.org/r/psych/
#explanation: the third factor is a single item (education, loading 1.00, a Heywood case), so it does not describe a real third dimension
efa_fit <- data.frame(
  factors = c(1, 2, 3),
  chisq   = round(c(efa1$STATISTIC, efa2$STATISTIC, efa3$STATISTIC), 1),
  df      = c(efa1$dof, efa2$dof, efa3$dof),
  p       = signif(c(efa1$PVAL, efa2$PVAL, efa3$PVAL), 2),
  TLI     = round(c(efa1$TLI, efa2$TLI, efa3$TLI), 2),
  RMSEA   = round(c(efa1$RMSEA[1], efa2$RMSEA[1], efa3$RMSEA[1]), 3),
  RMSR    = round(c(efa1$rms, efa2$rms, efa3$rms), 3),
  BIC     = round(c(efa1$BIC, efa2$BIC, efa3$BIC), 1)
)
kable(efa_fit, caption = "Fit of the one-, two-, and three-factor solutions")
Fit of the one-, two-, and three-factor solutions
factors chisq df p TLI RMSEA RMSR BIC
1 86.6 27 0.000 0.34 0.182 0.160 -26.6
2 27.4 19 0.096 0.87 0.080 0.065 -52.2
3 14.8 12 0.250 0.93 0.058 0.038 -35.4
#source: psych package (Revelle) https://personality-project.org/r/psych/; Hu & Bentler (1999) https://doi.org/10.1080/10705519909540118
#explanation: TLI near .90 or above and RMSEA near .08 or below are acceptable; the one-factor model misses both, the two-factor model is acceptable, and the three-factor model adds little while producing a Heywood case, so two factors are retained

2.2.6 Two-factor solution

efa2_loadings <- unclass(efa2$loadings)
care_factor <- names(which.max(efa2_loadings["EFFECT_CARE_MHRESOURCES", ]))   # the factor the care items load on
fear_factor <- names(which.max(efa2_loadings["EFFECT_FEAR_ARMY", ]))          # the factor the fear items load on

efa2_table <- data.frame(
  item        = rownames(efa2_loadings),
  care_factor = round(efa2_loadings[, care_factor], 2),
  fear_factor = round(efa2_loadings[, fear_factor], 2),
  communality = round(efa2$communality, 2)
)
kable(efa2_table, row.names = FALSE,
      caption = "Two-factor EFA: oblimin pattern loadings and communalities")
Two-factor EFA: oblimin pattern loadings and communalities
item care_factor fear_factor communality
EFFECT_CARE_COMM 0.60 -0.01 0.36
EFFECT_CARE_EDUCATION 0.47 -0.05 0.23
EFFECT_CARE_HVIP 0.62 0.29 0.41
EFFECT_CARE_MHRESOURCES 0.76 -0.11 0.61
EFFECT_CARE_WELFARE 0.55 -0.14 0.35
EFFECT_FEAR_LEG -0.15 0.03 0.03
EFFECT_FEAR_POLICE -0.11 0.54 0.32
EFFECT_FEAR_DEPORT -0.17 0.77 0.67
EFFECT_FEAR_ARMY 0.11 0.86 0.73
#source: psych package (Revelle) https://personality-project.org/r/psych/
#explanation: a loading of .30 or more means the item belongs to that factor; EFFECT_FEAR_LEG loads below .20 on both factors and has a communality of .03, so legislation is not measured by either dimension; EFFECT_CARE_HVIP loads mainly on care with a secondary loading on fear, so it is kept out of the care composite
par(family = "Calibri")                 # base-graphics plots also use Calibri
fa.diagram(efa2, cut = 0.30, simple = FALSE, main = "Effectiveness items: two factors")

Two-factor structure of the effectiveness items (loadings of .30 or more shown).
#source: psych package (Revelle) https://personality-project.org/r/psych/
#explanation: the diagram draws each item next to the factor it loads on; legislation has no arrow because no loading reaches .30

2.2.7 Check the solution with polychoric correlations

efa2_poly <- fa(effectdata_complete, nfactors = 2, fm = "minres", rotate = "oblimin", cor = "poly")
print(efa2_poly$loadings, cutoff = 0.30)

Loadings:
                        MR2    MR1   
EFFECT_CARE_COMM         0.647       
EFFECT_CARE_EDUCATION    0.563       
EFFECT_CARE_HVIP         0.648  0.329
EFFECT_CARE_MHRESOURCES  0.867       
EFFECT_CARE_WELFARE      0.548       
EFFECT_FEAR_LEG                      
EFFECT_FEAR_POLICE              0.581
EFFECT_FEAR_DEPORT              0.733
EFFECT_FEAR_ARMY                0.905

                 MR2   MR1
SS loadings    2.353 1.831
Proportion Var 0.261 0.203
Cumulative Var 0.261 0.465
round(efa2_poly$Phi, 2)
      MR2   MR1
MR2  1.00 -0.22
MR1 -0.22  1.00
#source: psych package (Revelle) https://personality-project.org/r/psych/
#explanation: polychoric correlations treat the 1-6 ratings as ordered categories instead of continuous numbers; the same two factors appear with the same items, so the result does not depend on treating the ratings as continuous

2.3 Composites and reliability

The EFA supports two composites. Legislation (loaded on neither factor) and HVIP (loaded on both) are kept out of both and reported as neutral single items.

# Create keys for scoring: care-based and fear-based effectiveness
effect_keys <- list(
  CARE_EFFECT = c("EFFECT_CARE_COMM", "EFFECT_CARE_EDUCATION",
                  "EFFECT_CARE_MHRESOURCES", "EFFECT_CARE_WELFARE"),
  FEAR_EFFECT = c("EFFECT_FEAR_POLICE", "EFFECT_FEAR_DEPORT", "EFFECT_FEAR_ARMY")
)

#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/composite.html
#explanation: effect_keys tells R which items belong to the care-based composite and which belong to the fear-based composite; EFFECT_FEAR_LEG is left out of both because the EFA showed it loads on neither factor, and EFFECT_CARE_HVIP is left out because it loads on both (.62 care, .29 fear)
library(psych)
# Using scoreItems() - recommended
effect_scores <- scoreItems(effect_keys, cleandata, impute = "none")
effect_scores                                  # alpha for each composite is in the first table
Call: scoreItems(keys = effect_keys, items = cleandata, impute = "none")

(Standardized) Alpha:
      CARE_EFFECT FEAR_EFFECT
alpha        0.68        0.74

Standard errors of unstandardized Alpha:
      CARE_EFFECT FEAR_EFFECT
ASE         0.061       0.068

Standardized Alpha of observed scales:
     CARE_EFFECT FEAR_EFFECT
[1,]        0.67        0.74

Average item correlation:
          CARE_EFFECT FEAR_EFFECT
average.r        0.34        0.49

Median item correlation:
CARE_EFFECT FEAR_EFFECT 
       0.34        0.40 

 Guttman 6* reliability: 
         CARE_EFFECT FEAR_EFFECT
Lambda.6        0.67        0.71

Signal/Noise based upon av.r : 
             CARE_EFFECT FEAR_EFFECT
Signal/Noise         2.1         2.8

Scale intercorrelations corrected for attenuation 
 raw correlations below the diagonal, alpha on the diagonal 
 corrected correlations above the diagonal:

Note that these are the correlations of the complete scales based on the correlation matrix,
 not the observed scales based on the raw items.
            CARE_EFFECT FEAR_EFFECT
CARE_EFFECT        0.68       -0.37
FEAR_EFFECT       -0.26        0.74

 Average adjusted correlations within and between scales (MIMS)
            CARE_ FEAR_
CARE_EFFECT  0.34      
FEAR_EFFECT -0.21  0.49

 Average adjusted item x scale correlations within and between scales (MIMT)
            CARE_ FEAR_
CARE_EFFECT  0.71      
FEAR_EFFECT -0.22  0.82

 In order to see the item by scale loadings and frequency counts of the data
 print with the short option = FALSE
# Add to your dataframe
cleandata$CARE_EFFECT <- effect_scores$scores[, "CARE_EFFECT"]
cleandata$FEAR_EFFECT <- effect_scores$scores[, "FEAR_EFFECT"]

# Legislation and HVIP stay on their own as neutral approaches
cleandata$LEGISLATION_EFFECT <- cleandata$EFFECT_FEAR_LEG
cleandata$HVIP_EFFECT        <- cleandata$EFFECT_CARE_HVIP

summary(cleandata$CARE_EFFECT)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max.    NA's 
   2.00    4.50    5.00    4.88    5.50    6.00     126 
summary(cleandata$FEAR_EFFECT)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max.    NA's 
  1.000   1.667   2.333   2.456   3.000   6.000     127 
summary(cleandata$LEGISLATION_EFFECT)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max.    NA's 
  1.000   3.750   4.000   4.188   5.000   6.000     130 
summary(cleandata$HVIP_EFFECT)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max.    NA's 
  2.000   3.000   4.000   4.014   5.000   6.000     137 
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/composite.html; psych package (Revelle)
#explanation: scoreItems() averages each person's available items (impute = "none") and reports Cronbach's alpha; CARE_EFFECT is the mean of four care items (community, education, mental health, welfare), FEAR_EFFECT is the mean of three fear items, and LEGISLATION_EFFECT and HVIP_EFFECT are single items kept apart as neutral approaches
effect_reliability <- data.frame(
  composite    = c("Care-based effectiveness (CARE_EFFECT)", "Fear-based effectiveness (FEAR_EFFECT)"),
  items        = c("Community, education, mental health, welfare", "Police, deportation, military"),
  n_items      = c(4, 3),
  alpha        = round(c(effect_scores$alpha[, "CARE_EFFECT"], effect_scores$alpha[, "FEAR_EFFECT"]), 2),
  mean_item_r  = round(c(effect_scores$av.r[, "CARE_EFFECT"], effect_scores$av.r[, "FEAR_EFFECT"]), 2),
  n            = c(sum(!is.na(cleandata$CARE_EFFECT)), sum(!is.na(cleandata$FEAR_EFFECT))),
  mean         = round(c(mean(cleandata$CARE_EFFECT, na.rm = TRUE), mean(cleandata$FEAR_EFFECT, na.rm = TRUE)), 2),
  sd           = round(c(sd(cleandata$CARE_EFFECT, na.rm = TRUE), sd(cleandata$FEAR_EFFECT, na.rm = TRUE)), 2)
)
kable(effect_reliability, row.names = FALSE,
      col.names = c("Composite", "Items", "k", "Alpha", "Mean r", "n", "M", "SD"),
      caption = "Reliability and descriptive statistics for the effectiveness composites. Legislation and HVIP are reported as single neutral items.")
Reliability and descriptive statistics for the effectiveness composites. Legislation and HVIP are reported as single neutral items.
Composite Items k Alpha Mean r n M SD
Care-based effectiveness (CARE_EFFECT) Community, education, mental health, welfare 4 0.68 0.34 84 4.88 0.81
Fear-based effectiveness (FEAR_EFFECT) Police, deportation, military 3 0.74 0.49 83 2.46 1.12
#source: psych package (Revelle) https://personality-project.org/r/psych/; The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/composite.html
#explanation: effect_scores$alpha holds Cronbach's alpha and effect_scores$av.r holds the average correlation between items in each composite; alpha around .70 is acceptable for a short scale, and the mean inter-item r shows how strongly the items hang together regardless of how many there are
cor.test(cleandata$CARE_EFFECT, cleandata$FEAR_EFFECT)

    Pearson's product-moment correlation

data:  cleandata$CARE_EFFECT and cleandata$FEAR_EFFECT
t = -1.9058, df = 81, p-value = 0.06023
alternative hypothesis: true correlation is not equal to 0
95 percent confidence interval:
 -0.404763230  0.008928639
sample estimates:
       cor 
-0.2071596 
cor.test(cleandata$CARE_EFFECT, cleandata$LEGISLATION_EFFECT)

    Pearson's product-moment correlation

data:  cleandata$CARE_EFFECT and cleandata$LEGISLATION_EFFECT
t = -0.6036, df = 78, p-value = 0.5479
alternative hypothesis: true correlation is not equal to 0
95 percent confidence interval:
 -0.2836528  0.1538361
sample estimates:
        cor 
-0.06818549 
cor.test(cleandata$FEAR_EFFECT, cleandata$LEGISLATION_EFFECT)

    Pearson's product-moment correlation

data:  cleandata$FEAR_EFFECT and cleandata$LEGISLATION_EFFECT
t = 1.4578, df = 77, p-value = 0.149
alternative hypothesis: true correlation is not equal to 0
95 percent confidence interval:
 -0.05937813  0.37153135
sample estimates:
      cor 
0.1638839 
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/correlation.html
#explanation: care and fear composites are close to independent, and legislation is only weakly related to either, which supports treating it as a neutral approach rather than folding it into a composite

2.4 Sample characteristics

The analytic sample is everyone who rated the effectiveness items (that is, anyone with a care-based or fear-based composite score).

sampledata <- cleandata |>
  filter(!is.na(CARE_EFFECT) | !is.na(FEAR_EFFECT)) |>
  mutate(AGE = as.numeric(AGE))                    # AGE was stored as text

nrow(sampledata)                                   # analytic sample size
[1] 84
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/dplyr.html
#explanation: sampledata keeps the people who answered the effectiveness questions, so Tables 1a and 1b describe the same people the rest of the report analyzes; AGE is converted from text to a number so it can be averaged
sampledata <- sampledata |>
  mutate(
    GENDER_3CAT = factor(
      case_when(
        GENDER == 0     ~ "Woman",
        GENDER == 1     ~ "Man",
        GENDER == 2     ~ "Another gender identity",
        is.na(GENDER)   ~ "Not reported"
      ),
      levels = c("Woman", "Man", "Another gender identity", "Not reported")),
    RACIALIZED_6CAT = factor(
      case_when(
        is.na(RACIALIZED)                ~ "Not reported",
        str_detect(RACIALIZED, ",")      ~ "Two or more identities",
        RACIALIZED == "7"                ~ "White",
        RACIALIZED == "2"                ~ "Asian",
        RACIALIZED == "3"                ~ "Black",
        !is.na(RACIALIZED)               ~ "Another identity"      # codes 1, 4, 5, 6, 8: too few to stand alone
      ),
      levels = c("White", "Asian", "Black", "Two or more identities", "Another identity", "Not reported")),
    POLITICAL_7CAT = factor(
      case_when(
        POLITICAL_BELIEFS == 1     ~ "Far left/leftist",
        POLITICAL_BELIEFS == 2     ~ "Very liberal",
        POLITICAL_BELIEFS == 3     ~ "Liberal",
        POLITICAL_BELIEFS == 4     ~ "Moderate",
        POLITICAL_BELIEFS == 5     ~ "Conservative",
        POLITICAL_BELIEFS == 6     ~ "Very conservative",
        POLITICAL_BELIEFS == 7     ~ "Far-right/alt-right",
        POLITICAL_BELIEFS == 8     ~ "Other",
        is.na(POLITICAL_BELIEFS)   ~ "Not reported"
      ),
      levels = c("Far left/leftist", "Very liberal", "Liberal", "Moderate", "Conservative",
                 "Very conservative", "Far-right/alt-right", "Other", "Not reported"))
  )

# ALWAYS check a transformation
sampledata |> count(GENDER, GENDER_3CAT)
# A tibble: 4 × 3
  GENDER GENDER_3CAT                 n
   <dbl> <fct>                   <int>
1      0 Woman                      48
2      1 Man                        24
3      2 Another gender identity     2
4     NA Not reported               10
sampledata |> count(RACIALIZED, RACIALIZED_6CAT)
# A tibble: 10 × 3
   RACIALIZED RACIALIZED_6CAT            n
   <chr>      <fct>                  <int>
 1 1,2,3,4,6  Two or more identities     1
 2 2          Asian                     22
 3 2,7        Two or more identities     3
 4 3          Black                      2
 5 4          Another identity           4
 6 4,5        Two or more identities     1
 7 4,7        Two or more identities     3
 8 5          Another identity           1
 9 7          White                     36
10 <NA>       Not reported              11
sampledata |> count(POLITICAL_BELIEFS, POLITICAL_7CAT)
# A tibble: 8 × 3
  POLITICAL_BELIEFS POLITICAL_7CAT          n
              <dbl> <fct>               <int>
1                 1 Far left/leftist        8
2                 2 Very liberal           13
3                 3 Liberal                17
4                 4 Moderate               18
5                 5 Conservative            1
6                 6 Very conservative       1
7                 7 Far-right/alt-right     1
8                NA Not reported           25
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/dplyr.html
#explanation: each demographic code becomes a labeled group for the table; people who checked two or more racialized identities form one group, the four smallest single identities are combined, and anyone who skipped a question or chose prefer not to say is counted as Not reported so every column adds to the full sample
n_sample <- nrow(sampledata)

gender_rows <- sampledata |>
  count(GENDER_3CAT, .drop = FALSE) |>
  mutate(characteristic = paste0("   ", GENDER_3CAT),
         percent = sprintf("%.1f", 100 * n / n_sample)) |>
  select(characteristic, n, percent)

race_rows <- sampledata |>
  count(RACIALIZED_6CAT, .drop = FALSE) |>
  mutate(characteristic = paste0("   ", RACIALIZED_6CAT),
         percent = sprintf("%.1f", 100 * n / n_sample)) |>
  select(characteristic, n, percent)

political_rows <- sampledata |>
  count(POLITICAL_7CAT, .drop = FALSE) |>
  mutate(characteristic = paste0("   ", POLITICAL_7CAT),
         percent = sprintf("%.1f", 100 * n / n_sample)) |>
  select(characteristic, n, percent)

categorical_table <- bind_rows(
  data.frame(characteristic = "Gender identity", n = NA, percent = ""),
  gender_rows,
  data.frame(characteristic = "Racialized identity", n = NA, percent = ""),
  race_rows,
  data.frame(characteristic = "Political beliefs", n = NA, percent = ""),
  political_rows
)

kable(categorical_table, row.names = FALSE, align = c("l", "r", "r"),
      col.names = c("Characteristic", "n", "%"),
      caption = paste0("Table 1a. Categorical sociodemographic characteristics of the analytic sample (N = ", n_sample, ")"))
Table 1a. Categorical sociodemographic characteristics of the analytic sample (N = 84)
Characteristic n %
Gender identity
   Woman 48 57.1
   Man 24 28.6
   Another gender identity 2 2.4
   Not reported 10 11.9
Racialized identity
   White 36 42.9
   Asian 22 26.2
   Black 2 2.4
   Two or more identities 8 9.5
   Another identity 5 6.0
   Not reported 11 13.1
Political beliefs
   Far left/leftist 8 9.5
   Very liberal 13 15.5
   Liberal 17 20.2
   Moderate 18 21.4
   Conservative 1 1.2
   Very conservative 1 1.2
   Far-right/alt-right 1 1.2
   Other 0 0.0
   Not reported 25 29.8
#source: APA Publication Manual (7th ed.), tables of sample characteristics; The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/describe.html
#explanation: each block counts one categorical demographic with count(); n and percent are separate columns, and percents use the full analytic sample as the denominator so the Not reported rows show how much is missing
continuous_table <- data.frame(
  variable     = c("Age (years)", "Subjective social status (ladder, 1-10)"),
  n            = c(sum(!is.na(sampledata$AGE)), sum(!is.na(sampledata$SOCIALSTATUS))),
  mean         = round(c(mean(sampledata$AGE, na.rm = TRUE), mean(sampledata$SOCIALSTATUS, na.rm = TRUE)), 1),
  sd           = round(c(sd(sampledata$AGE, na.rm = TRUE),   sd(sampledata$SOCIALSTATUS, na.rm = TRUE)), 1),
  minimum      = c(min(sampledata$AGE, na.rm = TRUE), min(sampledata$SOCIALSTATUS, na.rm = TRUE)),
  maximum      = c(max(sampledata$AGE, na.rm = TRUE), max(sampledata$SOCIALSTATUS, na.rm = TRUE)),
  not_reported = c(sprintf("%d (%.1f%%)", sum(is.na(sampledata$AGE)),          100 * sum(is.na(sampledata$AGE)) / n_sample),
                   sprintf("%d (%.1f%%)", sum(is.na(sampledata$SOCIALSTATUS)), 100 * sum(is.na(sampledata$SOCIALSTATUS)) / n_sample))
)

kable(continuous_table, row.names = FALSE, align = c("l", "r", "r", "r", "r", "r", "r"),
      col.names = c("Characteristic", "n", "M", "SD", "Min", "Max", "Not reported, n (%)"),
      caption = paste0("Table 1b. Continuous sociodemographic characteristics of the analytic sample (N = ", n_sample, ")"))
Table 1b. Continuous sociodemographic characteristics of the analytic sample (N = 84)
Characteristic n M SD Min Max Not reported, n (%)
Age (years) 73 21.6 9.0 18 73 11 (13.1%)
Subjective social status (ladder, 1-10) 73 5.9 1.8 2 10 11 (13.1%)
#source: APA Publication Manual (7th ed.), tables of sample characteristics; The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/describe.html
#explanation: age and subjective social status are continuous, so each gets its own row with the number of people who answered, the mean, standard deviation, lowest and highest value, and how many did not report it

2.5 Correlations between sociodemographics and perceived effectiveness

Research question: Are age, gender, racialized identity, political beliefs, and subjective social status related to how effective people believe each approach is?

# abbreviated names (see readme.qmd) keep Table 2 narrow enough for the page
demographic_vars <- sampledata |>
  select(AGE, WOMAN, POC,
         RIGHT = POLITICAL_BELIEFS,      # 1 = far left ... 7 = far right
         SSS   = SOCIALSTATUS)           # subjective social status ladder, 1 to 10

effectiveness_vars <- sampledata |>
  select(COMM        = EFFECT_CARE_COMM,
         EDUCATION   = EFFECT_CARE_EDUCATION,
         MHRESOURCES = EFFECT_CARE_MHRESOURCES,
         WELFARE     = EFFECT_CARE_WELFARE,
         HVIP        = EFFECT_CARE_HVIP,
         LEG         = EFFECT_FEAR_LEG,
         POLICE      = EFFECT_FEAR_POLICE,
         DEPORT      = EFFECT_FEAR_DEPORT,
         ARMY        = EFFECT_FEAR_ARMY,
         CARE_EFFECT, FEAR_EFFECT)

names(demographic_vars)
[1] "AGE"   "WOMAN" "POC"   "RIGHT" "SSS"  
names(effectiveness_vars)
 [1] "COMM"        "EDUCATION"   "MHRESOURCES" "WELFARE"     "HVIP"       
 [6] "LEG"         "POLICE"      "DEPORT"      "ARMY"        "CARE_EFFECT"
[11] "FEAR_EFFECT"
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/dplyr.html
#explanation: select() keeps the variables for Table 2 and renames them in the same step; effectiveness items are named after the part of the variable name after the second _ (EFFECT_CARE_COMM -> COMM); RIGHT is political beliefs (higher = more conservative) and SSS is subjective social status; WOMAN and POC are binary, so their correlations are point-biserial
demo_effect_cor <- corr.test(effectiveness_vars, demographic_vars, use = "pairwise", adjust = "none")

demo_effect_r <- demo_effect_cor$r
demo_effect_p <- demo_effect_cor$p
demo_effect_stars <- ifelse(demo_effect_p < .001, "***",
                     ifelse(demo_effect_p < .01,  "**",
                     ifelse(demo_effect_p < .05,  "*", "")))

correlation_table <- data.frame(
  measure = rownames(demo_effect_r),
  type    = c("Care", "Care", "Care", "Care", "Neutral", "Neutral",
              "Fear", "Fear", "Fear", "Care", "Fear"),
  AGE     = sprintf("%.2f%s", demo_effect_r[, "AGE"],   demo_effect_stars[, "AGE"]),
  WOMAN   = sprintf("%.2f%s", demo_effect_r[, "WOMAN"], demo_effect_stars[, "WOMAN"]),
  POC     = sprintf("%.2f%s", demo_effect_r[, "POC"],   demo_effect_stars[, "POC"]),
  RIGHT   = sprintf("%.2f%s", demo_effect_r[, "RIGHT"], demo_effect_stars[, "RIGHT"]),
  SSS     = sprintf("%.2f%s", demo_effect_r[, "SSS"],   demo_effect_stars[, "SSS"])
)

kable(correlation_table, row.names = FALSE, align = c("l", "l", "r", "r", "r", "r", "r"),
      col.names = c("Measure", "Type", "AGE", "WOMAN", "POC", "RIGHT", "SSS"),
      caption = "Table 2. Correlations between sociodemographic characteristics and perceived effectiveness of each approach")
Table 2. Correlations between sociodemographic characteristics and perceived effectiveness of each approach
Measure Type AGE WOMAN POC RIGHT SSS
COMM Care 0.14 -0.00 -0.21 -0.18 -0.09
EDUCATION Care 0.02 -0.08 -0.10 -0.23 -0.34**
MHRESOURCES Care -0.44*** 0.14 0.07 -0.35** -0.12
WELFARE Care 0.17 0.25* 0.09 -0.42** 0.02
HVIP Neutral -0.26* 0.02 0.11 -0.18 -0.03
LEG Neutral -0.13 -0.12 0.05 -0.02 -0.01
POLICE Fear -0.15 0.02 0.07 0.48*** 0.16
DEPORT Fear -0.13 -0.14 -0.01 0.50*** 0.19
ARMY Fear -0.24 -0.05 0.07 0.46*** 0.07
CARE_EFFECT Care -0.01 0.07 -0.09 -0.41** -0.19
FEAR_EFFECT Fear -0.21 -0.07 0.05 0.57*** 0.16
range(demo_effect_cor$n)                           # smallest and largest pairwise n
[1] 53 73
#source: psych package (Revelle) https://personality-project.org/r/psych/; The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/correlation.html
#explanation: corr.test() computes the Pearson correlation of every effectiveness measure (rows) with every sociodemographic variable (columns) using everyone who answered both; stars mark p < .05 (*), p < .01 (**), and p < .001 (***), unadjusted

Note. Pearson correlations with pairwise deletion (n = 53 to 73). Measures: COMM = community-based interventions; EDUCATION = educational programs; MHRESOURCES = mental health resources; WELFARE = welfare and economic support; HVIP = hospital-based violence intervention; LEG = legislation; POLICE = policing; DEPORT = deportation; ARMY = military or National Guard deployment; CARE_EFFECT and FEAR_EFFECT = care-based and fear-based composites. Variables: AGE = age in years; WOMAN = 1 woman, 0 man; POC = 1 Person of Color, 0 White; RIGHT = political beliefs, 1 (far left) to 7 (far right); SSS = subjective social status, MacArthur ladder 1 to 10. WOMAN and POC correlations are point-biserial. * p < .05, ** p < .01, *** p < .001.

2.6 Figure 1. Perceived effectiveness of each approach

Research question: Do people rate care-based approaches as more effective than fear-based approaches for achieving community safety and preventing violence?

2.6.1 Make the data long and label each approach

effectdata_long <- cleandata |>
  select(EFFECT_CARE_COMM, EFFECT_CARE_EDUCATION, EFFECT_CARE_HVIP,
         EFFECT_CARE_MHRESOURCES, EFFECT_CARE_WELFARE,
         EFFECT_FEAR_LEG, EFFECT_FEAR_POLICE, EFFECT_FEAR_DEPORT, EFFECT_FEAR_ARMY) |>
  pivot_longer(cols = everything(), names_to = "VARIABLE", values_to = "EFFECTIVENESS") |>
  filter(!is.na(EFFECTIVENESS))

effectdata_long <- effectdata_long |>
  mutate(
    APPROACH = case_when(
      VARIABLE == "EFFECT_CARE_COMM"        ~ "Community",
      VARIABLE == "EFFECT_CARE_EDUCATION"   ~ "Education",
      VARIABLE == "EFFECT_CARE_HVIP"        ~ "HVIP",
      VARIABLE == "EFFECT_CARE_MHRESOURCES" ~ "Mental health",
      VARIABLE == "EFFECT_CARE_WELFARE"     ~ "Welfare",
      VARIABLE == "EFFECT_FEAR_LEG"         ~ "Legislation",
      VARIABLE == "EFFECT_FEAR_POLICE"      ~ "Police",
      VARIABLE == "EFFECT_FEAR_DEPORT"      ~ "Deportation",
      VARIABLE == "EFFECT_FEAR_ARMY"        ~ "Military"
    ),
    TYPE = case_when(
      VARIABLE == "EFFECT_FEAR_LEG"         ~ "Mixed or neutral",   # loaded on neither factor
      VARIABLE == "EFFECT_CARE_HVIP"        ~ "Mixed or neutral",   # loaded on both factors
      str_detect(VARIABLE, "EFFECT_CARE_")  ~ "Care-based",
      str_detect(VARIABLE, "EFFECT_FEAR_")  ~ "Fear-based"
    ),
    TYPE = factor(TYPE, levels = c("Care-based", "Mixed or neutral", "Fear-based"))
  )

# ALWAYS check a transformation
effectdata_long |> count(APPROACH, TYPE)
# A tibble: 9 × 3
  APPROACH      TYPE                 n
  <chr>         <fct>            <int>
1 Community     Care-based          80
2 Deportation   Fear-based          81
3 Education     Care-based          82
4 HVIP          Mixed or neutral    73
5 Legislation   Mixed or neutral    80
6 Mental health Care-based          81
7 Military      Fear-based          80
8 Police        Fear-based          83
9 Welfare       Care-based          78
#source: R for Data Science (2e), Ch. 5 Data tidying (Wickham et al., 2023) https://r4ds.hadley.nz/data-tidy.html; Ch. 3 Data transformation https://r4ds.hadley.nz/data-transform.html
#explanation: pivot_longer() stacks the nine items into one EFFECTIVENESS column so they can share one plot; APPROACH is a short label for the axis and TYPE colors the bars, with legislation (loaded on neither factor) and HVIP (loaded on both) marked Mixed or neutral so they are grey

2.6.2 Summarise each approach

effectiveness_means <- effectdata_long |>
  group_by(APPROACH, TYPE) |>
  summarise(n = n(),
            mean_effectiveness = mean(EFFECTIVENESS),
            sd_effectiveness = sd(EFFECTIVENESS),
            se_effectiveness = sd(EFFECTIVENESS) / sqrt(n()),
            ci_lower = mean(EFFECTIVENESS) - 1.96 * sd(EFFECTIVENESS) / sqrt(n()),
            ci_upper = mean(EFFECTIVENESS) + 1.96 * sd(EFFECTIVENESS) / sqrt(n()),
            .groups = "drop") |>
  arrange(desc(mean_effectiveness))

kable(effectiveness_means, digits = 2,
      caption = "Mean perceived effectiveness of each approach, in descending order")
Mean perceived effectiveness of each approach, in descending order
APPROACH TYPE n mean_effectiveness sd_effectiveness se_effectiveness ci_lower ci_upper
Mental health Care-based 81 5.14 0.97 0.11 4.92 5.35
Welfare Care-based 78 4.86 1.07 0.12 4.62 5.10
Community Care-based 80 4.79 1.11 0.12 4.54 5.03
Education Care-based 82 4.78 1.19 0.13 4.52 5.04
Legislation Mixed or neutral 80 4.19 1.26 0.14 3.91 4.46
HVIP Mixed or neutral 73 4.01 0.99 0.12 3.79 4.24
Police Fear-based 83 3.29 1.44 0.16 2.98 3.60
Military Fear-based 80 2.12 1.33 0.15 1.83 2.42
Deportation Fear-based 81 1.83 1.16 0.13 1.57 2.08
#source: R for Data Science (2e), Ch. 3 Data transformation (Wickham et al., 2023) https://r4ds.hadley.nz/data-transform.html
#explanation: one row per approach with n, mean, sd, standard error, and 95 percent confidence interval; arrange(desc()) puts the most effective approach first, and this order is reused for the plot

2.6.3 Compare care-based with fear-based effectiveness

# every person rated both sets, so the comparison is paired
effectdata_paired <- cleandata |>
  filter(!is.na(CARE_EFFECT), !is.na(FEAR_EFFECT))

nrow(effectdata_paired)                                   # people with both composites
[1] 83
mean(effectdata_paired$CARE_EFFECT)                       # care-based: average of the four blue items
[1] 4.884538
mean(effectdata_paired$FEAR_EFFECT)                       # fear-based: average of the three purple items
[1] 2.455823
mean(effectdata_paired$CARE_EFFECT - effectdata_paired$FEAR_EFFECT)   # mean paired difference
[1] 2.428715
wilcoxon_care_fear <- wilcox.test(effectdata_paired$CARE_EFFECT, effectdata_paired$FEAR_EFFECT, paired = TRUE)
wilcoxon_care_fear

    Wilcoxon signed rank test with continuity correction

data:  effectdata_paired$CARE_EFFECT and effectdata_paired$FEAR_EFFECT
V = 3176, p-value = 8.547e-14
alternative hypothesis: true location shift is not equal to 0
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/compare2groups.html; R documentation ?wilcox.test
#explanation: the same people rated both care-based and fear-based approaches, so a paired Wilcoxon signed-rank test compares each person's two composites; V is the signed-rank statistic and the p-value shows care-based approaches are rated far more effective than fear-based approaches

2.6.4 Plot the means

# Put the approaches in descending order of their mean
effectiveness_means <- effectiveness_means |>
  mutate(APPROACH = factor(APPROACH, levels = APPROACH))
effectdata_long <- effectdata_long |>
  mutate(APPROACH = factor(APPROACH, levels = levels(effectiveness_means$APPROACH)))

plot1_effectiveness_barplot <- ggplot(effectiveness_means,
                                      aes(x = APPROACH, y = mean_effectiveness, fill = TYPE)) +
  geom_col(color = "black") +
  geom_errorbar(aes(ymin = ci_lower, ymax = ci_upper), width = 0.25, linewidth = 0.8) +
  geom_text(aes(label = sprintf("%.2f", mean_effectiveness), y = ci_upper), vjust = -0.6, size = 3.8, family = "Calibri") +
  scale_fill_manual(values = c("Care-based" = "#8297ce",    # care-based indicators
                               "Mixed or neutral" = "#bfbfbf",   # legislation and HVIP
                               "Fear-based" = "#a29ccc")) + # fear-based indicators: #453a98 lightened 50% toward white
  scale_y_continuous(breaks = 1:6, limits = c(0, 6.6),
                     labels = c("1: Strongly Disagree", "2: Disagree", "3: Slightly Disagree",
                                "4: Slightly Agree", "5: Agree", "6: Strongly Agree")) +
  labs(x = "Violence prevention and intervention approach",
       y = "Agreement that the approach is effective at\nachieving community safety and preventing violence",
       fill = NULL) +
  theme_bw(base_family = "Calibri") +
  theme(legend.position = "inside",
        # left edge midway between HVIP and Police (bar 6.5 of 9), centered at 6: Strongly Agree;
        # positions are fractions of the panel: x = (6.5 - 0.4) / 9.2, y = (6 + 0.33) / 7.26
        legend.position.inside = c(0.663, 0.872),
        legend.justification = c(0, 0.5),
        legend.direction = "horizontal",
        legend.background = element_rect(fill = "white", color = "grey70"),
        axis.text.x = element_text(angle = 40, hjust = 1))

plot1_effectiveness_barplot

Average perceived effectiveness of each community violence prevention approach (bars = means, whiskers = 95% CIs). Care-based effectiveness averages the four blue items and fear-based averages the three purple items; legislation and HVIP (grey) are excluded because they loaded on neither or both factors. Care-based approaches were rated more effective than fear-based approaches (paired Wilcoxon signed-rank test, V = 3176, p < .001, n = 83).
# poster quality: 600 dpi PNG plus an SVG vector file that stays sharp at any size
ggsave("plotsTEAM/plot1_effectiveness_barplot.png", plot = plot1_effectiveness_barplot,
       width = 14.4, height = 8, dpi = 600)
ggsave("plotsTEAM/plot1_effectiveness_barplot.svg", plot = plot1_effectiveness_barplot,
       width = 14.4, height = 8)

#source: ggplot2 documentation https://ggplot2.tidyverse.org/reference/geom_bar.html; The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/visualize-comparison.html
#explanation: geom_col() draws one bar per approach at its mean and geom_errorbar() adds the 95 percent confidence interval; care-based approaches are rated most effective, legislation sits in the middle, and the three fear-based approaches are rated least effective

2.6.5 Figure 1, light blue and purple version

plot1b_effectiveness_barplot_light <- ggplot(effectiveness_means,
                                             aes(x = APPROACH, y = mean_effectiveness, fill = TYPE)) +
  geom_col(color = "black") +
  geom_errorbar(aes(ymin = ci_lower, ymax = ci_upper), width = 0.25, linewidth = 0.8) +
  geom_text(aes(label = sprintf("%.2f", mean_effectiveness), y = ci_upper), vjust = -0.6, size = 3.8, family = "Calibri") +
  scale_fill_manual(values = c("Care-based"       = "#ccdaf6",    # light blue
                               "Mixed or neutral" = "#e4e4e4",    # light grey (brand neutral-light)
                               "Fear-based"       = "#d8d2e8")) + # light purple
  scale_y_continuous(breaks = 1:6, limits = c(0, 6.6),
                     labels = c("1: Strongly Disagree", "2: Disagree", "3: Slightly Disagree",
                                "4: Slightly Agree", "5: Agree", "6: Strongly Agree")) +
  labs(x = "Violence prevention and intervention approach",
       y = "Agreement that the approach is effective at\nachieving community safety and preventing violence",
       fill = NULL) +
  theme_bw(base_family = "Calibri") +
  theme(legend.position = "inside",
        # left edge midway between HVIP and Police (bar 6.5 of 9), centered at 6: Strongly Agree;
        # positions are fractions of the panel: x = (6.5 - 0.4) / 9.2, y = (6 + 0.33) / 7.26
        legend.position.inside = c(0.663, 0.872),
        legend.justification = c(0, 0.5),
        legend.direction = "horizontal",
        legend.background = element_rect(fill = "white", color = "grey70"),
        axis.text.x = element_text(angle = 40, hjust = 1))

plot1b_effectiveness_barplot_light

Average perceived effectiveness of each community violence prevention approach (bars = means, whiskers = 95% CIs). Care-based effectiveness averages the four blue items and fear-based averages the three purple items; legislation and HVIP (grey) are excluded because they loaded on neither or both factors. Care-based approaches were rated more effective than fear-based approaches (paired Wilcoxon signed-rank test, V = 3176, p < .001, n = 83).
# poster quality: 600 dpi PNG plus an SVG vector file that stays sharp at any size
ggsave("plotsTEAM/plot1b_effectiveness_barplot_light.png", plot = plot1b_effectiveness_barplot_light,
       width = 14.4, height = 8, dpi = 600)
ggsave("plotsTEAM/plot1b_effectiveness_barplot_light.svg", plot = plot1b_effectiveness_barplot_light,
       width = 14.4, height = 8)

#source: ggplot2 documentation https://ggplot2.tidyverse.org/reference/geom_bar.html
#explanation: the same bar plot as Figure 1 with lighter fills: light blue (#ccdaf6) for care-based, light purple (#d8d2e8) for fear-based, and light grey (#e4e4e4) for legislation and HVIP; the black outlines keep the bars distinct because the two light fills are close in tone

3 Explaining Care-based Support

Research question: Does growth mindset explain support for care-based approaches beyond racialized identity and gender, and does its association differ by racialized identity or gender?

The outcome is CARE_EFFECT, the mean of four care-based items (community-based interventions, educational programs, mental health resources, welfare) rated from 1 (strongly disagree) to 6 (strongly agree) that each approach is effective. Predictors enter in three blocks (hierarchical regression): Block 1 adds racialized identity and gender, Block 2 adds growth mindset, and Block 3 adds the interactions of growth mindset with racialized identity and gender. Each block is compared with the one before it, so the change in R² shows what that block explains beyond the earlier ones.

3.1 Model sample

# the same people in every block, so the R-squared changes can be compared
caredata_model <- cleandata |>
  filter(!is.na(CARE_EFFECT), !is.na(POC), !is.na(WOMAN), !is.na(GROWTH_cent)) |>
  mutate(POC_LABEL   = factor(POC,   levels = c(0, 1), labels = c("White", "People of Color")),
         WOMAN_LABEL = factor(WOMAN, levels = c(0, 1), labels = c("Man", "Woman")))

nrow(caredata_model)
[1] 71
# ALWAYS check a transformation
caredata_model |> count(POC, POC_LABEL, WOMAN, WOMAN_LABEL)
# A tibble: 4 × 5
    POC POC_LABEL       WOMAN WOMAN_LABEL     n
  <dbl> <fct>           <dbl> <fct>       <int>
1     0 White               0 Man            12
2     0 White               1 Woman          22
3     1 People of Color     0 Man            12
4     1 People of Color     1 Woman          25
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html
#explanation: hierarchical regression compares models fit to the same people, so caredata_model keeps only people with every variable used in any block; POC_LABEL and WOMAN_LABEL turn the 0/1 codes into words for the plots

3.2 Block 1: Racialized identity and gender

Do racialized identity and gender explain care-based support on their own?

# regmodel1: Block 1, CARE_EFFECT predicted by POC and WOMAN
regmodel1 <- lm(CARE_EFFECT ~ POC + WOMAN, data = caredata_model)

summary(regmodel1)

Call:
lm(formula = CARE_EFFECT ~ POC + WOMAN, data = caredata_model)

Residuals:
    Min      1Q  Median      3Q     Max 
-2.8059 -0.4334  0.1941  0.5862  1.1941 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept)   4.8109     0.1954  24.620   <2e-16 ***
POC          -0.1079     0.1966  -0.549    0.585    
WOMAN         0.1029     0.2076   0.495    0.622    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.8273 on 68 degrees of freedom
Multiple R-squared:  0.007741,  Adjusted R-squared:  -0.02144 
F-statistic: 0.2653 on 2 and 68 DF,  p-value: 0.7678
nobs(regmodel1)
[1] 71
# caption for Figure 2, built from the model so the numbers stay correct
figure2_caption <- sprintf("Block 1. Mean care-based effectiveness by racialized identity and by gender (points = means, whiskers = 95%% confidence intervals; n = %d). Together, racialized identity and gender explained %.1f%% of the variance (F(%d, %d) = %.2f, p = %.2f).", nobs(regmodel1), 100 * summary(regmodel1)$r.squared, summary(regmodel1)$fstatistic[2], summary(regmodel1)$fstatistic[3], summary(regmodel1)$fstatistic[1], pf(summary(regmodel1)$fstatistic[1], summary(regmodel1)$fstatistic[2], summary(regmodel1)$fstatistic[3], lower.tail = FALSE))

#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html
#explanation: Block 1 enters only the sociodemographic covariates; POC is the difference between People of Color and White respondents and WOMAN the difference between women and men, each holding the other constant

3.2.1 Figure 2. Care-based support by racialized identity and gender

care_group_means <- bind_rows(
  caredata_model |>
    group_by(GROUP = as.character(POC_LABEL)) |>
    summarise(n = n(), mean = mean(CARE_EFFECT), se = sd(CARE_EFFECT) / sqrt(n()), .groups = "drop") |>
    mutate(VARIABLE = "Racialized identity"),
  caredata_model |>
    group_by(GROUP = as.character(WOMAN_LABEL)) |>
    summarise(n = n(), mean = mean(CARE_EFFECT), se = sd(CARE_EFFECT) / sqrt(n()), .groups = "drop") |>
    mutate(VARIABLE = "Gender")
) |>
  mutate(GROUP    = factor(GROUP, levels = c("White", "People of Color", "Man", "Woman")),
         VARIABLE = factor(VARIABLE, levels = c("Racialized identity", "Gender")))

care_group_means                                  # n, mean, and standard error for each group
# A tibble: 4 × 5
  GROUP               n  mean    se VARIABLE           
  <fct>           <int> <dbl> <dbl> <fct>              
1 People of Color    37  4.77 0.146 Racialized identity
2 White              34  4.88 0.128 Racialized identity
3 Man                24  4.76 0.162 Gender             
4 Woman              47  4.86 0.122 Gender             
plot2_care_block1_groups <- ggplot(care_group_means, aes(x = GROUP, y = mean, color = GROUP)) +
  geom_errorbar(aes(ymin = mean - 1.96 * se, ymax = mean + 1.96 * se), width = 0.15, linewidth = 0.8) +
  geom_point(size = 3.5) +
  geom_text(aes(label = sprintf("n = %d", n), y = 1.2), color = "grey30", size = 3.5, family = "Calibri") +
  facet_wrap(~ VARIABLE, scales = "free_x") +
  scale_color_manual(values = c("White" = "#E69F00", "People of Color" = "#009E73",
                                "Man" = "#D55E00", "Woman" = "#CC79A7")) +
  scale_y_continuous(breaks = 1:6, limits = c(1, 6),
                     labels = c("1: Strongly Disagree", "2: Disagree", "3: Slightly Disagree",
                                "4: Slightly Agree", "5: Agree", "6: Strongly Agree")) +
  labs(x = NULL, y = "Care-based effectiveness") +
  theme_bw(base_family = "Calibri") +
  theme(legend.position = "none")

plot2_care_block1_groups

Block 1. Mean care-based effectiveness by racialized identity and by gender (points = means, whiskers = 95% confidence intervals; n = 71). Together, racialized identity and gender explained 0.8% of the variance (F(2, 68) = 0.27, p = 0.77).
ggsave("plotsTEAM/plot2_care_block1_groups.png", plot = plot2_care_block1_groups,
       width = 8, height = 5, dpi = 300)

#source: ggplot2 documentation https://ggplot2.tidyverse.org/reference/geom_linerange.html
#explanation: each point is a group mean with a 95 percent confidence interval; overlapping intervals at nearly the same height show that racialized identity and gender make little difference to care-based support

Block 1 explained 0.8% of the variance in care-based support (F(2, 68) = 0.27, p = 0.768).

3.3 Block 2: Growth mindset

Does growth mindset explain care-based support beyond racialized identity and gender?

# regmodel2: Block 2, regmodel1 + GROWTH_cent
regmodel2 <- lm(CARE_EFFECT ~ POC + WOMAN + GROWTH_cent, data = caredata_model)

summary(regmodel2)

Call:
lm(formula = CARE_EFFECT ~ POC + WOMAN + GROWTH_cent, data = caredata_model)

Residuals:
    Min      1Q  Median      3Q     Max 
-2.5238 -0.4748  0.1848  0.5123  1.2093 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept)  4.86799    0.17681  27.533  < 2e-16 ***
POC          0.02454    0.18031   0.136 0.892173    
WOMAN       -0.08229    0.19274  -0.427 0.670779    
GROWTH_cent  0.41824    0.10271   4.072 0.000126 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.7462 on 67 degrees of freedom
Multiple R-squared:  0.2046,    Adjusted R-squared:  0.169 
F-statistic: 5.744 on 3 and 67 DF,  p-value: 0.001471
confint(regmodel2)
                 2.5 %    97.5 %
(Intercept)  4.5150836 5.2208985
POC         -0.3353646 0.3844347
WOMAN       -0.4669930 0.3024127
GROWTH_cent  0.2132267 0.6232456
care_change_12 <- anova(regmodel1, regmodel2)       # F change from Block 1 to Block 2
care_change_12
Analysis of Variance Table

Model 1: CARE_EFFECT ~ POC + WOMAN
Model 2: CARE_EFFECT ~ POC + WOMAN + GROWTH_cent
  Res.Df    RSS Df Sum of Sq      F    Pr(>F)    
1     68 46.539                                  
2     67 37.306  1    9.2326 16.581 0.0001255 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
# caption for Figure 3, built from the model so the numbers stay correct
figure3_caption <- sprintf("As growth mindset increases, people report higher perceived effectiveness of care-based approaches, controlling for racialized identity and gender (b = %s, p %s, ΔR² = %s; n = %d). The line is the least-squares fit with a 95%% confidence band.", sub("^0", "", sprintf("%.2f", coef(regmodel2)["GROWTH_cent"])), ifelse(care_change_12[["Pr(>F)"]][2] < .001, "< .001", sub("= 0", "= ", sprintf("= %.3f", care_change_12[["Pr(>F)"]][2]))), sub("^0", "", sprintf("%.2f", summary(regmodel2)$r.squared - summary(regmodel1)$r.squared)), nobs(regmodel2))

#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html; R documentation ?anova.lm
#explanation: anova() compares Block 2 with Block 1; its F test asks whether adding growth mindset explains significantly more variance; GROWTH_cent is the change in care-based effectiveness for each one-point increase in growth mindset, holding racialized identity and gender constant

3.3.1 Figure 3. Growth mindset and care-based support

plot3_care_block2_growth <- ggplot(caredata_model, aes(x = GROWTH, y = CARE_EFFECT)) +
  geom_point(position = position_jitter(width = 0.03, height = 0.03, seed = 1),
             size = 2, alpha = 0.7, color = "#8297ce") +
  geom_smooth(method = "lm", formula = y ~ x, se = TRUE, color = "#8297ce", fill = "#8297ce", alpha = 0.25) +
  scale_y_continuous(breaks = 1:6, limits = c(1, 6.2),
                     labels = c("1: Strongly Disagree", "2: Disagree", "3: Slightly Disagree",
                                "4: Slightly Agree", "5: Agree", "6: Strongly Agree")) +
  scale_x_continuous(breaks = 2:6,
                     labels = c("Disagree", "Slightly Disagree", "Slightly Agree", "Agree", "Strongly Agree")) +
  coord_cartesian(xlim = c(1.9, 6.1)) +
  labs(x = "Agreement with Growth Mindset Statements",
       y = "Care-based effectiveness") +
  theme_bw(base_family = "Calibri") +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))

plot3_care_block2_growth

As growth mindset increases, people report higher perceived effectiveness of care-based approaches, controlling for racialized identity and gender (b = .42, p < .001, ΔR² = .20; n = 71). The line is the least-squares fit with a 95% confidence band.
# poster quality: 600 dpi PNG plus an SVG vector file that stays sharp at any size
ggsave("plotsTEAM/plot3_care_block2_growth.png", plot = plot3_care_block2_growth,
       width = 8, height = 6, dpi = 600)
ggsave("plotsTEAM/plot3_care_block2_growth.svg", plot = plot3_care_block2_growth,
       width = 8, height = 6)

#source: ggplot2 documentation https://ggplot2.tidyverse.org/reference/geom_smooth.html
#explanation: each dot is one person and the line is the overall line of best fit; blue marks care-based support, as in the brand guide

3.3.2 Figure 3, dark blue version

plot3b_care_block2_growth_darkblue <- ggplot(caredata_model, aes(x = GROWTH, y = CARE_EFFECT)) +
  geom_point(position = position_jitter(width = 0.03, height = 0.03, seed = 1),
             size = 2, alpha = 0.7, color = "#1f3762") +
  geom_smooth(method = "lm", formula = y ~ x, se = TRUE, color = "#1f3762", fill = "#1f3762", alpha = 0.2) +
  scale_y_continuous(breaks = 1:6, limits = c(1, 6.2),
                     labels = c("1: Strongly Disagree", "2: Disagree", "3: Slightly Disagree",
                                "4: Slightly Agree", "5: Agree", "6: Strongly Agree")) +
  scale_x_continuous(breaks = 2:6,
                     labels = c("Disagree", "Slightly Disagree", "Slightly Agree", "Agree", "Strongly Agree")) +
  coord_cartesian(xlim = c(1.9, 6.1)) +
  labs(x = "Agreement with Growth Mindset Statements",
       y = "Care-based effectiveness") +
  theme_bw(base_family = "Calibri") +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))

plot3b_care_block2_growth_darkblue

As growth mindset increases, people report higher perceived effectiveness of care-based approaches, controlling for racialized identity and gender (b = .42, p < .001, ΔR² = .20; n = 71). The line is the least-squares fit with a 95% confidence band.
# poster quality: 600 dpi PNG plus an SVG vector file that stays sharp at any size
ggsave("plotsTEAM/plot3b_care_block2_growth_darkblue.png", plot = plot3b_care_block2_growth_darkblue,
       width = 8, height = 6, dpi = 600)
ggsave("plotsTEAM/plot3b_care_block2_growth_darkblue.svg", plot = plot3b_care_block2_growth_darkblue,
       width = 8, height = 6)

#source: ggplot2 documentation https://ggplot2.tidyverse.org/reference/geom_smooth.html
#explanation: the same plot as Figure 3 with points, line, and confidence band in dark blue (#1f3762) for use on the poster; the data, model, and caption are unchanged

Adding growth mindset raised R² from 0.008 to 0.205 (ΔR² = 0.197, F change(1, 67) = 16.58, p < .001). Each one-point increase in growth mindset went with a 0.42-point increase in care-based effectiveness.

3.4 Block 3: Growth mindset by racialized identity and gender

Does the association between growth mindset and care-based support differ by racialized identity or gender?

# regmodel3: Block 3, regmodel2 + POC:GROWTH_cent + WOMAN:GROWTH_cent
regmodel3 <- lm(CARE_EFFECT ~ POC + WOMAN + GROWTH_cent + POC:GROWTH_cent + WOMAN:GROWTH_cent, data = caredata_model)

summary(regmodel3)

Call:
lm(formula = CARE_EFFECT ~ POC + WOMAN + GROWTH_cent + POC:GROWTH_cent + 
    WOMAN:GROWTH_cent, data = caredata_model)

Residuals:
    Min      1Q  Median      3Q     Max 
-2.4991 -0.4440  0.1902  0.5285  1.2327 

Coefficients:
                  Estimate Std. Error t value Pr(>|t|)    
(Intercept)        4.83019    0.18729  25.789   <2e-16 ***
POC                0.03862    0.18336   0.211    0.834    
WOMAN             -0.06443    0.20098  -0.321    0.750    
GROWTH_cent        0.34012    0.23173   1.468    0.147    
POC:GROWTH_cent   -0.04869    0.20879  -0.233    0.816    
WOMAN:GROWTH_cent  0.15425    0.22629   0.682    0.498    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.7545 on 65 degrees of freedom
Multiple R-squared:  0.211, Adjusted R-squared:  0.1503 
F-statistic: 3.476 on 5 and 65 DF,  p-value: 0.007601
confint(regmodel3)
                       2.5 %    97.5 %
(Intercept)        4.4561385 5.2042456
POC               -0.3275790 0.4048184
WOMAN             -0.4658253 0.3369612
GROWTH_cent       -0.1226838 0.8029262
POC:GROWTH_cent   -0.4656733 0.3682930
WOMAN:GROWTH_cent -0.2976903 0.6061939
care_change_23 <- anova(regmodel2, regmodel3)       # F change from Block 2 to Block 3
care_change_23
Analysis of Variance Table

Model 1: CARE_EFFECT ~ POC + WOMAN + GROWTH_cent
Model 2: CARE_EFFECT ~ POC + WOMAN + GROWTH_cent + POC:GROWTH_cent + WOMAN:GROWTH_cent
  Res.Df    RSS Df Sum of Sq      F Pr(>F)
1     67 37.306                           
2     65 37.006  2   0.29982 0.2633 0.7693
# captions for Figures 4 and 5, built from the model so the numbers stay correct
figure4_caption <- sprintf("Block 3. Growth mindset and care-based effectiveness for White respondents and People of Color (n = %d). Lines are least-squares fits within each group with 95%% confidence bands. The growth slope did not differ by racialized identity (POC x growth mindset: b = %.2f, p = %.2f).", nobs(regmodel3), coef(regmodel3)["POC:GROWTH_cent"], summary(regmodel3)$coefficients["POC:GROWTH_cent", "Pr(>|t|)"])
figure5_caption <- sprintf("Block 3. Growth mindset and care-based effectiveness for men and women (n = %d). Lines are least-squares fits within each group with 95%% confidence bands. The growth slope did not differ by gender (WOMAN x growth mindset: b = %.2f, p = %.2f).", nobs(regmodel3), coef(regmodel3)["WOMAN:GROWTH_cent"], summary(regmodel3)$coefficients["WOMAN:GROWTH_cent", "Pr(>|t|)"])

#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html; Aiken & West (1991) Multiple regression: Testing and interpreting interactions
#explanation: Block 3 adds both interactions at once; anova() tests whether they explain more variance together; GROWTH_cent is now the growth slope for White men, and each interaction is how much steeper or flatter the slope is for People of Color or for women

3.4.1 Figure 4. Growth mindset and care-based support by racialized identity

plot4_care_block3_poc <- ggplot(caredata_model, aes(x = GROWTH, y = CARE_EFFECT, color = POC_LABEL, fill = POC_LABEL)) +
  geom_point(position = position_jitter(width = 0.03, height = 0.03, seed = 1), size = 2, alpha = 0.7) +
  geom_smooth(method = "lm", formula = y ~ x, se = TRUE, alpha = 0.15) +
  scale_color_manual(values = c("White" = "#E69F00", "People of Color" = "#009E73")) +
  scale_fill_manual(values  = c("White" = "#E69F00", "People of Color" = "#009E73")) +
  scale_y_continuous(breaks = 1:6, limits = c(1, 6.2),
                     labels = c("1: Strongly Disagree", "2: Disagree", "3: Slightly Disagree",
                                "4: Slightly Agree", "5: Agree", "6: Strongly Agree")) +
  scale_x_continuous(breaks = 2:6,
                     labels = c("Disagree", "Slightly Disagree", "Slightly Agree", "Agree", "Strongly Agree")) +
  coord_cartesian(xlim = c(1.9, 6.1)) +
  labs(x = "Agreement with Growth Mindset Statements",
       y = "Care-based effectiveness",
       color = NULL, fill = NULL) +
  theme_bw(base_family = "Calibri") +
  theme(legend.position = "top",
        axis.text.x = element_text(angle = 45, hjust = 1))

plot4_care_block3_poc

Block 3. Growth mindset and care-based effectiveness for White respondents and People of Color (n = 71). Lines are least-squares fits within each group with 95% confidence bands. The growth slope did not differ by racialized identity (POC x growth mindset: b = -0.05, p = 0.82).
ggsave("plotsTEAM/plot4_care_block3_poc.png", plot = plot4_care_block3_poc,
       width = 8, height = 6, dpi = 300)

#source: ggplot2 documentation https://ggplot2.tidyverse.org/reference/geom_smooth.html
#explanation: a separate line of best fit for White respondents and People of Color; similar slopes match a non-significant POC:GROWTH_cent interaction; orange and green mark racialized identity so that blue and purple keep their care and fear meaning

3.4.2 Figure 5. Growth mindset and care-based support by gender

plot5_care_block3_woman <- ggplot(caredata_model, aes(x = GROWTH, y = CARE_EFFECT, color = WOMAN_LABEL, fill = WOMAN_LABEL)) +
  geom_point(position = position_jitter(width = 0.03, height = 0.03, seed = 1), size = 2, alpha = 0.7) +
  geom_smooth(method = "lm", formula = y ~ x, se = TRUE, alpha = 0.15) +
  scale_color_manual(values = c("Man" = "#D55E00", "Woman" = "#CC79A7")) +
  scale_fill_manual(values  = c("Man" = "#D55E00", "Woman" = "#CC79A7")) +
  scale_y_continuous(breaks = 1:6, limits = c(1, 6.2),
                     labels = c("1: Strongly Disagree", "2: Disagree", "3: Slightly Disagree",
                                "4: Slightly Agree", "5: Agree", "6: Strongly Agree")) +
  scale_x_continuous(breaks = 2:6,
                     labels = c("Disagree", "Slightly Disagree", "Slightly Agree", "Agree", "Strongly Agree")) +
  coord_cartesian(xlim = c(1.9, 6.1)) +
  labs(x = "Agreement with Growth Mindset Statements",
       y = "Care-based effectiveness",
       color = NULL, fill = NULL) +
  theme_bw(base_family = "Calibri") +
  theme(legend.position = "top",
        axis.text.x = element_text(angle = 45, hjust = 1))

plot5_care_block3_woman

Block 3. Growth mindset and care-based effectiveness for men and women (n = 71). Lines are least-squares fits within each group with 95% confidence bands. The growth slope did not differ by gender (WOMAN x growth mindset: b = 0.15, p = 0.50).
ggsave("plotsTEAM/plot5_care_block3_woman.png", plot = plot5_care_block3_woman,
       width = 8, height = 6, dpi = 300)

#source: ggplot2 documentation https://ggplot2.tidyverse.org/reference/geom_smooth.html
#explanation: a separate line of best fit for men and women; similar slopes match a non-significant WOMAN:GROWTH_cent interaction; vermillion and pink mark gender so that blue and purple keep their care and fear meaning

Adding the two interactions changed R² by only 0.006 (F change(2, 65) = 0.26, p = 0.769), so the growth mindset association with care-based support was similar across racialized identity and gender.

3.5 Table 3. Hierarchical regression summary

care_b1 <- summary(regmodel1)$coefficients
care_b2 <- summary(regmodel2)$coefficients
care_b3 <- summary(regmodel3)$coefficients

care_stars1 <- ifelse(care_b1[, "Pr(>|t|)"] < .001, "***", ifelse(care_b1[, "Pr(>|t|)"] < .01, "**", ifelse(care_b1[, "Pr(>|t|)"] < .05, "*", "")))
care_stars2 <- ifelse(care_b2[, "Pr(>|t|)"] < .001, "***", ifelse(care_b2[, "Pr(>|t|)"] < .01, "**", ifelse(care_b2[, "Pr(>|t|)"] < .05, "*", "")))
care_stars3 <- ifelse(care_b3[, "Pr(>|t|)"] < .001, "***", ifelse(care_b3[, "Pr(>|t|)"] < .01, "**", ifelse(care_b3[, "Pr(>|t|)"] < .05, "*", "")))

care_hierarchical_table <- data.frame(
  predictor = c("Intercept", "Person of Color (vs. White)", "Woman (vs. man)", "Growth mindset",
                "Person of Color x Growth mindset", "Woman x Growth mindset",
                "R²", "ΔR²", "F change (df)", "p (F change)"),
  block1 = c(sprintf("%.2f (%.2f)%s", care_b1[, "Estimate"], care_b1[, "Std. Error"], care_stars1),
             "", "", "",
             sprintf("%.3f", summary(regmodel1)$r.squared),
             sprintf("%.3f", summary(regmodel1)$r.squared),
             sprintf("%.2f (%d, %d)", summary(regmodel1)$fstatistic[1], summary(regmodel1)$fstatistic[2], summary(regmodel1)$fstatistic[3]),
             sprintf("%.3f", pf(summary(regmodel1)$fstatistic[1], summary(regmodel1)$fstatistic[2], summary(regmodel1)$fstatistic[3], lower.tail = FALSE))),
  block2 = c(sprintf("%.2f (%.2f)%s", care_b2[, "Estimate"], care_b2[, "Std. Error"], care_stars2),
             "", "",
             sprintf("%.3f", summary(regmodel2)$r.squared),
             sprintf("%.3f", summary(regmodel2)$r.squared - summary(regmodel1)$r.squared),
             sprintf("%.2f (%d, %d)", care_change_12$F[2], care_change_12$Df[2], care_change_12$Res.Df[2]),
             ifelse(care_change_12[["Pr(>F)"]][2] < .001, "< .001", sprintf("%.3f", care_change_12[["Pr(>F)"]][2]))),
  block3 = c(sprintf("%.2f (%.2f)%s", care_b3[, "Estimate"], care_b3[, "Std. Error"], care_stars3),
             sprintf("%.3f", summary(regmodel3)$r.squared),
             sprintf("%.3f", summary(regmodel3)$r.squared - summary(regmodel2)$r.squared),
             sprintf("%.2f (%d, %d)", care_change_23$F[2], care_change_23$Df[2], care_change_23$Res.Df[2]),
             ifelse(care_change_23[["Pr(>F)"]][2] < .001, "< .001", sprintf("%.3f", care_change_23[["Pr(>F)"]][2])))
)

kable(care_hierarchical_table, row.names = FALSE, align = c("l", "r", "r", "r"),
      col.names = c("Predictor", "Block 1", "Block 2", "Block 3"),
      caption = "Table 3. Hierarchical linear regression explaining care-based support")
Table 3. Hierarchical linear regression explaining care-based support
Predictor Block 1 Block 2 Block 3
Intercept 4.81 (0.20)*** 4.87 (0.18)*** 4.83 (0.19)***
Person of Color (vs. White) -0.11 (0.20) 0.02 (0.18) 0.04 (0.18)
Woman (vs. man) 0.10 (0.21) -0.08 (0.19) -0.06 (0.20)
Growth mindset 0.42 (0.10)*** 0.34 (0.23)
Person of Color x Growth mindset -0.05 (0.21)
Woman x Growth mindset 0.15 (0.23)
R² 0.008 0.205 0.211
ΔR² 0.008 0.197 0.006
F change (df) 0.27 (2, 68) 16.58 (1, 67) 0.26 (2, 65)
p (F change) 0.768 < .001 0.769
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html; APA Publication Manual (7th ed.), hierarchical regression tables
#explanation: each column is one block; cells are unstandardized b with its standard error in parentheses and stars for p < .05 (*), .01 (**), and .001 (***); R² is the variance explained by the block's model, ΔR² is what that block adds beyond the one before, and F change tests whether the addition is significant

Note. n = 71. Cells show b (SE). Growth mindset is centered at its sample mean (GROWTH_cent). In Block 3, the growth mindset coefficient is the slope for White men. * p < .05, ** p < .01, *** p < .001.

3.6 Model checks

shapiro.test(resid(regmodel2))

    Shapiro-Wilk normality test

data:  resid(regmodel2)
W = 0.93882, p-value = 0.001801
round(max(cooks.distance(regmodel2)), 2)       # largest influence of any one person
[1] 0.13
round(max(hatvalues(regmodel2)), 2)              # largest leverage of any one person
[1] 0.15
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html; Cook (1977) https://doi.org/10.1080/00401706.1977.10489493
#explanation: checks for the Block 2 model, the one interpreted because Block 3 adds little; shapiro.test() checks whether the residuals are roughly normal, and a Cook's distance above 1 or leverage near 1 would mean one person sets the results

4 Explaining Fear-based Support

Research question: Does growth mindset explain support for fear-based approaches beyond racialized identity and gender, and does its association differ by racialized identity or gender?

The outcome is FEAR_EFFECT, the mean of three fear-based items (policing, deportation, military or National Guard deployment) rated from 1 (strongly disagree) to 6 (strongly agree) that each approach is effective. Predictors enter in three blocks (hierarchical regression): Block 1 adds racialized identity and gender, Block 2 adds growth mindset, and Block 3 adds the interactions of growth mindset with racialized identity and gender. Each block is compared with the one before it, so the change in R² shows what that block explains beyond the earlier ones.

4.1 Model sample

# the same people in every block, so the R-squared changes can be compared
feardata_model <- cleandata |>
  filter(!is.na(FEAR_EFFECT), !is.na(POC), !is.na(WOMAN), !is.na(GROWTH_cent)) |>
  mutate(POC_LABEL   = factor(POC,   levels = c(0, 1), labels = c("White", "People of Color")),
         WOMAN_LABEL = factor(WOMAN, levels = c(0, 1), labels = c("Man", "Woman")))

nrow(feardata_model)
[1] 70
# ALWAYS check a transformation
feardata_model |> count(POC, POC_LABEL, WOMAN, WOMAN_LABEL)
# A tibble: 4 × 5
    POC POC_LABEL       WOMAN WOMAN_LABEL     n
  <dbl> <fct>           <dbl> <fct>       <int>
1     0 White               0 Man            12
2     0 White               1 Woman          22
3     1 People of Color     0 Man            12
4     1 People of Color     1 Woman          24
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html
#explanation: hierarchical regression compares models fit to the same people, so feardata_model keeps only people with every variable used in any block; POC_LABEL and WOMAN_LABEL turn the 0/1 codes into words for the plots

4.2 Block 1: Racialized identity and gender

Do racialized identity and gender explain fear-based support on their own?

# regmodel4: Block 1, FEAR_EFFECT predicted by POC and WOMAN
regmodel4 <- lm(FEAR_EFFECT ~ POC + WOMAN, data = feardata_model)

summary(regmodel4)

Call:
lm(formula = FEAR_EFFECT ~ POC + WOMAN, data = feardata_model)

Residuals:
    Min      1Q  Median      3Q     Max 
-1.5396 -0.8215 -0.0988  0.5959  3.1785 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept)  2.48818    0.25447   9.778 1.58e-14 ***
POC          0.05141    0.25728   0.200    0.842    
WOMAN       -0.16356    0.27091  -0.604    0.548    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 1.076 on 67 degrees of freedom
Multiple R-squared:  0.005929,  Adjusted R-squared:  -0.02374 
F-statistic: 0.1998 on 2 and 67 DF,  p-value: 0.8194
nobs(regmodel4)
[1] 70
# caption for Figure 6, built from the model so the numbers stay correct
figure6_caption <- sprintf("Block 1. Mean fear-based effectiveness by racialized identity and by gender (points = means, whiskers = 95%% confidence intervals; n = %d). Together, racialized identity and gender explained %.1f%% of the variance (F(%d, %d) = %.2f, p = %.2f).", nobs(regmodel4), 100 * summary(regmodel4)$r.squared, summary(regmodel4)$fstatistic[2], summary(regmodel4)$fstatistic[3], summary(regmodel4)$fstatistic[1], pf(summary(regmodel4)$fstatistic[1], summary(regmodel4)$fstatistic[2], summary(regmodel4)$fstatistic[3], lower.tail = FALSE))

#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html
#explanation: Block 1 enters only the sociodemographic covariates; POC is the difference between People of Color and White respondents and WOMAN the difference between women and men, each holding the other constant

4.2.1 Figure 6. Fear-based support by racialized identity and gender

fear_group_means <- bind_rows(
  feardata_model |>
    group_by(GROUP = as.character(POC_LABEL)) |>
    summarise(n = n(), mean = mean(FEAR_EFFECT), se = sd(FEAR_EFFECT) / sqrt(n()), .groups = "drop") |>
    mutate(VARIABLE = "Racialized identity"),
  feardata_model |>
    group_by(GROUP = as.character(WOMAN_LABEL)) |>
    summarise(n = n(), mean = mean(FEAR_EFFECT), se = sd(FEAR_EFFECT) / sqrt(n()), .groups = "drop") |>
    mutate(VARIABLE = "Gender")
) |>
  mutate(GROUP    = factor(GROUP, levels = c("White", "People of Color", "Man", "Woman")),
         VARIABLE = factor(VARIABLE, levels = c("Racialized identity", "Gender")))

fear_group_means                                  # n, mean, and standard error for each group
# A tibble: 4 × 5
  GROUP               n  mean    se VARIABLE           
  <fct>           <int> <dbl> <dbl> <fct>              
1 People of Color    36  2.43 0.170 Racialized identity
2 White              34  2.38 0.193 Racialized identity
3 Man                24  2.51 0.247 Gender             
4 Woman              46  2.35 0.146 Gender             
plot6_fear_block1_groups <- ggplot(fear_group_means, aes(x = GROUP, y = mean, color = GROUP)) +
  geom_errorbar(aes(ymin = mean - 1.96 * se, ymax = mean + 1.96 * se), width = 0.15, linewidth = 0.8) +
  geom_point(size = 3.5) +
  geom_text(aes(label = sprintf("n = %d", n), y = 1.2), color = "grey30", size = 3.5, family = "Calibri") +
  facet_wrap(~ VARIABLE, scales = "free_x") +
  scale_color_manual(values = c("White" = "#E69F00", "People of Color" = "#009E73",
                                "Man" = "#D55E00", "Woman" = "#CC79A7")) +
  scale_y_continuous(breaks = 1:6, limits = c(1, 6),
                     labels = c("1: Strongly Disagree", "2: Disagree", "3: Slightly Disagree",
                                "4: Slightly Agree", "5: Agree", "6: Strongly Agree")) +
  labs(x = NULL, y = "Fear-based effectiveness") +
  theme_bw(base_family = "Calibri") +
  theme(legend.position = "none")

plot6_fear_block1_groups

Block 1. Mean fear-based effectiveness by racialized identity and by gender (points = means, whiskers = 95% confidence intervals; n = 70). Together, racialized identity and gender explained 0.6% of the variance (F(2, 67) = 0.20, p = 0.82).
ggsave("plotsTEAM/plot6_fear_block1_groups.png", plot = plot6_fear_block1_groups,
       width = 8, height = 5, dpi = 300)

#source: ggplot2 documentation https://ggplot2.tidyverse.org/reference/geom_linerange.html
#explanation: each point is a group mean with a 95 percent confidence interval; overlapping intervals at nearly the same height show that racialized identity and gender make little difference to fear-based support

Block 1 explained 0.6% of the variance in fear-based support (F(2, 67) = 0.20, p = 0.819).

4.3 Block 2: Growth mindset

Does growth mindset explain fear-based support beyond racialized identity and gender?

# regmodel5: Block 2, regmodel4 + GROWTH_cent
regmodel5 <- lm(FEAR_EFFECT ~ POC + WOMAN + GROWTH_cent, data = feardata_model)

summary(regmodel5)

Call:
lm(formula = FEAR_EFFECT ~ POC + WOMAN + GROWTH_cent, data = feardata_model)

Residuals:
    Min      1Q  Median      3Q     Max 
-1.5317 -0.7664 -0.1150  0.5332  3.0791 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept)  2.45481    0.25088   9.785 1.79e-14 ***
POC         -0.03897    0.25780  -0.151   0.8803    
WOMAN       -0.05023    0.27356  -0.184   0.8549    
GROWTH_cent -0.26647    0.14655  -1.818   0.0736 .  
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 1.058 on 66 degrees of freedom
Multiple R-squared:  0.05335,   Adjusted R-squared:  0.01032 
F-statistic:  1.24 on 3 and 66 DF,  p-value: 0.3024
confint(regmodel5)
                 2.5 %     97.5 %
(Intercept)  1.9539182 2.95569540
POC         -0.5536951 0.47575291
WOMAN       -0.5964019 0.49594300
GROWTH_cent -0.5590744 0.02612935
fear_change_12 <- anova(regmodel4, regmodel5)       # F change from Block 1 to Block 2
fear_change_12
Analysis of Variance Table

Model 1: FEAR_EFFECT ~ POC + WOMAN
Model 2: FEAR_EFFECT ~ POC + WOMAN + GROWTH_cent
  Res.Df    RSS Df Sum of Sq      F  Pr(>F)  
1     67 77.517                              
2     66 73.820  1    3.6978 3.3061 0.07356 .
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
# caption for Figure 7, built from the model so the numbers stay correct
figure7_caption <- sprintf("Block 2. Growth mindset and fear-based effectiveness (n = %d). The line is the least-squares fit with a 95%% confidence band. Controlling for racialized identity and gender, growth mindset explained an additional %.1f%% of the variance (b = %.2f, F change(%d, %d) = %.2f, p %s).", nobs(regmodel5), 100 * (summary(regmodel5)$r.squared - summary(regmodel4)$r.squared), coef(regmodel5)["GROWTH_cent"], fear_change_12$Df[2], fear_change_12$Res.Df[2], fear_change_12$F[2], ifelse(fear_change_12[["Pr(>F)"]][2] < .001, "< .001", sprintf("= %.3f", fear_change_12[["Pr(>F)"]][2])))

#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html; R documentation ?anova.lm
#explanation: anova() compares Block 2 with Block 1; its F test asks whether adding growth mindset explains significantly more variance; GROWTH_cent is the change in fear-based effectiveness for each one-point increase in growth mindset, holding racialized identity and gender constant

4.3.1 Figure 7. Growth mindset and fear-based support

plot7_fear_block2_growth <- ggplot(feardata_model, aes(x = GROWTH, y = FEAR_EFFECT)) +
  geom_point(position = position_jitter(width = 0.03, height = 0.03, seed = 1),
             size = 2, alpha = 0.7, color = "#a29ccc") +
  geom_smooth(method = "lm", formula = y ~ x, se = TRUE, color = "#453a98", fill = "#a29ccc", alpha = 0.25) +
  scale_y_continuous(breaks = 1:6, limits = c(0.8, 6.2),
                     labels = c("1: Strongly Disagree", "2: Disagree", "3: Slightly Disagree",
                                "4: Slightly Agree", "5: Agree", "6: Strongly Agree")) +
  scale_x_continuous(breaks = 2:6,
                     labels = c("Disagree", "Slightly Disagree", "Slightly Agree", "Agree", "Strongly Agree")) +
  coord_cartesian(xlim = c(1.9, 6.1)) +
  labs(x = "Agreement with Growth Mindset Statements",
       y = "Fear-based effectiveness") +
  theme_bw(base_family = "Calibri") +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))

plot7_fear_block2_growth

Block 2. Growth mindset and fear-based effectiveness (n = 70). The line is the least-squares fit with a 95% confidence band. Controlling for racialized identity and gender, growth mindset explained an additional 4.7% of the variance (b = -0.27, F change(1, 66) = 3.31, p = 0.074).
ggsave("plotsTEAM/plot7_fear_block2_growth.png", plot = plot7_fear_block2_growth,
       width = 8, height = 6, dpi = 300)

#source: ggplot2 documentation https://ggplot2.tidyverse.org/reference/geom_smooth.html
#explanation: each dot is one person and the line is the overall line of best fit; purple marks fear-based support, as in the brand guide

Adding growth mindset raised R² from 0.006 to 0.053 (ΔR² = 0.047, F change(1, 66) = 3.31, p = 0.074). Each one-point increase in growth mindset went with a 0.27-point decrease in fear-based effectiveness.

4.4 Block 3: Growth mindset by racialized identity and gender

Does the association between growth mindset and fear-based support differ by racialized identity or gender?

# regmodel6: Block 3, regmodel5 + POC:GROWTH_cent + WOMAN:GROWTH_cent
regmodel6 <- lm(FEAR_EFFECT ~ POC + WOMAN + GROWTH_cent + POC:GROWTH_cent + WOMAN:GROWTH_cent, data = feardata_model)

summary(regmodel6)

Call:
lm(formula = FEAR_EFFECT ~ POC + WOMAN + GROWTH_cent + POC:GROWTH_cent + 
    WOMAN:GROWTH_cent, data = feardata_model)

Residuals:
    Min      1Q  Median      3Q     Max 
-1.5466 -0.7597 -0.1092  0.5168  3.0862 

Coefficients:
                  Estimate Std. Error t value Pr(>|t|)    
(Intercept)        2.46883    0.26691   9.250 2.08e-13 ***
POC               -0.04077    0.26306  -0.155    0.877    
WOMAN             -0.07610    0.28677  -0.265    0.792    
GROWTH_cent       -0.16295    0.33019  -0.494    0.623    
POC:GROWTH_cent   -0.10965    0.29899  -0.367    0.715    
WOMAN:GROWTH_cent -0.05206    0.32265  -0.161    0.872    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 1.073 on 64 degrees of freedom
Multiple R-squared:  0.05566,   Adjusted R-squared:  -0.01811 
F-statistic: 0.7545 on 5 and 64 DF,  p-value: 0.5859
confint(regmodel6)
                       2.5 %    97.5 %
(Intercept)        1.9356239 3.0020455
POC               -0.5663028 0.4847562
WOMAN             -0.6489818 0.4967873
GROWTH_cent       -0.8225843 0.4966830
POC:GROWTH_cent   -0.7069546 0.4876645
WOMAN:GROWTH_cent -0.6966401 0.5925117
fear_change_23 <- anova(regmodel5, regmodel6)       # F change from Block 2 to Block 3
fear_change_23
Analysis of Variance Table

Model 1: FEAR_EFFECT ~ POC + WOMAN + GROWTH_cent
Model 2: FEAR_EFFECT ~ POC + WOMAN + GROWTH_cent + POC:GROWTH_cent + WOMAN:GROWTH_cent
  Res.Df    RSS Df Sum of Sq      F Pr(>F)
1     66 73.820                           
2     64 73.639  2   0.18051 0.0784 0.9246
# captions for Figures 8 and 9, built from the model so the numbers stay correct
figure8_caption <- sprintf("Block 3. Growth mindset and fear-based effectiveness for White respondents and People of Color (n = %d). Lines are least-squares fits within each group with 95%% confidence bands. The growth slope did not differ by racialized identity (POC x growth mindset: b = %.2f, p = %.2f).", nobs(regmodel6), coef(regmodel6)["POC:GROWTH_cent"], summary(regmodel6)$coefficients["POC:GROWTH_cent", "Pr(>|t|)"])
figure9_caption <- sprintf("Block 3. Growth mindset and fear-based effectiveness for men and women (n = %d). Lines are least-squares fits within each group with 95%% confidence bands. The growth slope did not differ by gender (WOMAN x growth mindset: b = %.2f, p = %.2f).", nobs(regmodel6), coef(regmodel6)["WOMAN:GROWTH_cent"], summary(regmodel6)$coefficients["WOMAN:GROWTH_cent", "Pr(>|t|)"])

#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html; Aiken & West (1991) Multiple regression: Testing and interpreting interactions
#explanation: Block 3 adds both interactions at once; anova() tests whether they explain more variance together; GROWTH_cent is now the growth slope for White men, and each interaction is how much steeper or flatter the slope is for People of Color or for women

4.4.1 Figure 8. Growth mindset and fear-based support by racialized identity

plot8_fear_block3_poc <- ggplot(feardata_model, aes(x = GROWTH, y = FEAR_EFFECT, color = POC_LABEL, fill = POC_LABEL)) +
  geom_point(position = position_jitter(width = 0.03, height = 0.03, seed = 1), size = 2, alpha = 0.7) +
  geom_smooth(method = "lm", formula = y ~ x, se = TRUE, alpha = 0.15) +
  scale_color_manual(values = c("White" = "#E69F00", "People of Color" = "#009E73")) +
  scale_fill_manual(values  = c("White" = "#E69F00", "People of Color" = "#009E73")) +
  scale_y_continuous(breaks = 1:6, limits = c(0.8, 6.2),
                     labels = c("1: Strongly Disagree", "2: Disagree", "3: Slightly Disagree",
                                "4: Slightly Agree", "5: Agree", "6: Strongly Agree")) +
  scale_x_continuous(breaks = 2:6,
                     labels = c("Disagree", "Slightly Disagree", "Slightly Agree", "Agree", "Strongly Agree")) +
  coord_cartesian(xlim = c(1.9, 6.1)) +
  labs(x = "Agreement with Growth Mindset Statements",
       y = "Fear-based effectiveness",
       color = NULL, fill = NULL) +
  theme_bw(base_family = "Calibri") +
  theme(legend.position = "top",
        axis.text.x = element_text(angle = 45, hjust = 1))

plot8_fear_block3_poc

Block 3. Growth mindset and fear-based effectiveness for White respondents and People of Color (n = 70). Lines are least-squares fits within each group with 95% confidence bands. The growth slope did not differ by racialized identity (POC x growth mindset: b = -0.11, p = 0.72).
ggsave("plotsTEAM/plot8_fear_block3_poc.png", plot = plot8_fear_block3_poc,
       width = 8, height = 6, dpi = 300)

#source: ggplot2 documentation https://ggplot2.tidyverse.org/reference/geom_smooth.html
#explanation: a separate line of best fit for White respondents and People of Color; similar slopes match a non-significant POC:GROWTH_cent interaction; orange and green mark racialized identity so that blue and purple keep their care and fear meaning

4.4.2 Figure 9. Growth mindset and fear-based support by gender

plot9_fear_block3_woman <- ggplot(feardata_model, aes(x = GROWTH, y = FEAR_EFFECT, color = WOMAN_LABEL, fill = WOMAN_LABEL)) +
  geom_point(position = position_jitter(width = 0.03, height = 0.03, seed = 1), size = 2, alpha = 0.7) +
  geom_smooth(method = "lm", formula = y ~ x, se = TRUE, alpha = 0.15) +
  scale_color_manual(values = c("Man" = "#D55E00", "Woman" = "#CC79A7")) +
  scale_fill_manual(values  = c("Man" = "#D55E00", "Woman" = "#CC79A7")) +
  scale_y_continuous(breaks = 1:6, limits = c(0.8, 6.2),
                     labels = c("1: Strongly Disagree", "2: Disagree", "3: Slightly Disagree",
                                "4: Slightly Agree", "5: Agree", "6: Strongly Agree")) +
  scale_x_continuous(breaks = 2:6,
                     labels = c("Disagree", "Slightly Disagree", "Slightly Agree", "Agree", "Strongly Agree")) +
  coord_cartesian(xlim = c(1.9, 6.1)) +
  labs(x = "Agreement with Growth Mindset Statements",
       y = "Fear-based effectiveness",
       color = NULL, fill = NULL) +
  theme_bw(base_family = "Calibri") +
  theme(legend.position = "top",
        axis.text.x = element_text(angle = 45, hjust = 1))

plot9_fear_block3_woman

Block 3. Growth mindset and fear-based effectiveness for men and women (n = 70). Lines are least-squares fits within each group with 95% confidence bands. The growth slope did not differ by gender (WOMAN x growth mindset: b = -0.05, p = 0.87).
ggsave("plotsTEAM/plot9_fear_block3_woman.png", plot = plot9_fear_block3_woman,
       width = 8, height = 6, dpi = 300)

#source: ggplot2 documentation https://ggplot2.tidyverse.org/reference/geom_smooth.html
#explanation: a separate line of best fit for men and women; similar slopes match a non-significant WOMAN:GROWTH_cent interaction; vermillion and pink mark gender so that blue and purple keep their care and fear meaning

Adding the two interactions changed R² by only 0.002 (F change(2, 64) = 0.08, p = 0.925), so the growth mindset association with fear-based support was similar across racialized identity and gender.

4.5 Table 4. Hierarchical regression summary

fear_b1 <- summary(regmodel4)$coefficients
fear_b2 <- summary(regmodel5)$coefficients
fear_b3 <- summary(regmodel6)$coefficients

fear_stars1 <- ifelse(fear_b1[, "Pr(>|t|)"] < .001, "***", ifelse(fear_b1[, "Pr(>|t|)"] < .01, "**", ifelse(fear_b1[, "Pr(>|t|)"] < .05, "*", "")))
fear_stars2 <- ifelse(fear_b2[, "Pr(>|t|)"] < .001, "***", ifelse(fear_b2[, "Pr(>|t|)"] < .01, "**", ifelse(fear_b2[, "Pr(>|t|)"] < .05, "*", "")))
fear_stars3 <- ifelse(fear_b3[, "Pr(>|t|)"] < .001, "***", ifelse(fear_b3[, "Pr(>|t|)"] < .01, "**", ifelse(fear_b3[, "Pr(>|t|)"] < .05, "*", "")))

fear_hierarchical_table <- data.frame(
  predictor = c("Intercept", "Person of Color (vs. White)", "Woman (vs. man)", "Growth mindset",
                "Person of Color x Growth mindset", "Woman x Growth mindset",
                "R²", "ΔR²", "F change (df)", "p (F change)"),
  block1 = c(sprintf("%.2f (%.2f)%s", fear_b1[, "Estimate"], fear_b1[, "Std. Error"], fear_stars1),
             "", "", "",
             sprintf("%.3f", summary(regmodel4)$r.squared),
             sprintf("%.3f", summary(regmodel4)$r.squared),
             sprintf("%.2f (%d, %d)", summary(regmodel4)$fstatistic[1], summary(regmodel4)$fstatistic[2], summary(regmodel4)$fstatistic[3]),
             sprintf("%.3f", pf(summary(regmodel4)$fstatistic[1], summary(regmodel4)$fstatistic[2], summary(regmodel4)$fstatistic[3], lower.tail = FALSE))),
  block2 = c(sprintf("%.2f (%.2f)%s", fear_b2[, "Estimate"], fear_b2[, "Std. Error"], fear_stars2),
             "", "",
             sprintf("%.3f", summary(regmodel5)$r.squared),
             sprintf("%.3f", summary(regmodel5)$r.squared - summary(regmodel4)$r.squared),
             sprintf("%.2f (%d, %d)", fear_change_12$F[2], fear_change_12$Df[2], fear_change_12$Res.Df[2]),
             ifelse(fear_change_12[["Pr(>F)"]][2] < .001, "< .001", sprintf("%.3f", fear_change_12[["Pr(>F)"]][2]))),
  block3 = c(sprintf("%.2f (%.2f)%s", fear_b3[, "Estimate"], fear_b3[, "Std. Error"], fear_stars3),
             sprintf("%.3f", summary(regmodel6)$r.squared),
             sprintf("%.3f", summary(regmodel6)$r.squared - summary(regmodel5)$r.squared),
             sprintf("%.2f (%d, %d)", fear_change_23$F[2], fear_change_23$Df[2], fear_change_23$Res.Df[2]),
             ifelse(fear_change_23[["Pr(>F)"]][2] < .001, "< .001", sprintf("%.3f", fear_change_23[["Pr(>F)"]][2])))
)

kable(fear_hierarchical_table, row.names = FALSE, align = c("l", "r", "r", "r"),
      col.names = c("Predictor", "Block 1", "Block 2", "Block 3"),
      caption = "Table 4. Hierarchical linear regression explaining fear-based support")
Table 4. Hierarchical linear regression explaining fear-based support
Predictor Block 1 Block 2 Block 3
Intercept 2.49 (0.25)*** 2.45 (0.25)*** 2.47 (0.27)***
Person of Color (vs. White) 0.05 (0.26) -0.04 (0.26) -0.04 (0.26)
Woman (vs. man) -0.16 (0.27) -0.05 (0.27) -0.08 (0.29)
Growth mindset -0.27 (0.15) -0.16 (0.33)
Person of Color x Growth mindset -0.11 (0.30)
Woman x Growth mindset -0.05 (0.32)
R² 0.006 0.053 0.056
ΔR² 0.006 0.047 0.002
F change (df) 0.20 (2, 67) 3.31 (1, 66) 0.08 (2, 64)
p (F change) 0.819 0.074 0.925
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html; APA Publication Manual (7th ed.), hierarchical regression tables
#explanation: each column is one block; cells are unstandardized b with its standard error in parentheses and stars for p < .05 (*), .01 (**), and .001 (***); R² is the variance explained by the block's model, ΔR² is what that block adds beyond the one before, and F change tests whether the addition is significant

Note. n = 70. Cells show b (SE). Growth mindset is centered at its sample mean (GROWTH_cent). In Block 3, the growth mindset coefficient is the slope for White men. * p < .05, ** p < .01, *** p < .001.

4.6 Model checks

shapiro.test(resid(regmodel5))

    Shapiro-Wilk normality test

data:  resid(regmodel5)
W = 0.92828, p-value = 0.0006209
round(max(cooks.distance(regmodel5)), 2)       # largest influence of any one person
[1] 0.2
round(max(hatvalues(regmodel5)), 2)              # largest leverage of any one person
[1] 0.16
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html; Cook (1977) https://doi.org/10.1080/00401706.1977.10489493
#explanation: checks for the Block 2 model, the one interpreted because Block 3 adds little; shapiro.test() checks whether the residuals are roughly normal, and a Cook's distance above 1 or leverage near 1 would mean one person sets the results

5 Comparing Care-based and Fear-based Support

Research question: How do the explanatory variables compare for care-based and fear-based support, and do the growth mindset results hold without the covariates?

5.1 Sensitivity analysis: growth mindset alone

Adding racialized identity and gender drops anyone missing them. These models keep everyone with a growth mindset score and an effectiveness score, to check that the growth mindset results do not depend on the covariates or on who was dropped.

# regmodel7: CARE_EFFECT predicted by GROWTH_cent alone (no covariates)
regmodel7 <- lm(CARE_EFFECT ~ GROWTH_cent, data = cleandata)
summary(regmodel7)

Call:
lm(formula = CARE_EFFECT ~ GROWTH_cent, data = cleandata)

Residuals:
    Min      1Q  Median      3Q     Max 
-2.5752 -0.4332  0.1950  0.4632  1.2177 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept)  4.85889    0.08458  57.450  < 2e-16 ***
GROWTH_cent  0.41421    0.09606   4.312 4.99e-05 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.7325 on 73 degrees of freedom
  (135 observations deleted due to missingness)
Multiple R-squared:  0.203, Adjusted R-squared:  0.1921 
F-statistic: 18.59 on 1 and 73 DF,  p-value: 4.988e-05
# regmodel8: FEAR_EFFECT predicted by GROWTH_cent alone (no covariates)
regmodel8 <- lm(FEAR_EFFECT ~ GROWTH_cent, data = cleandata)
summary(regmodel8)

Call:
lm(formula = FEAR_EFFECT ~ GROWTH_cent, data = cleandata)

Residuals:
     Min       1Q   Median       3Q      Max 
-1.50387 -0.74411 -0.09696  0.57216  3.10074 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept)   2.3959     0.1201  19.950   <2e-16 ***
GROWTH_cent  -0.2482     0.1363  -1.822   0.0727 .  
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 1.033 on 72 degrees of freedom
  (136 observations deleted due to missingness)
Multiple R-squared:  0.04406,   Adjusted R-squared:  0.03078 
F-statistic: 3.318 on 1 and 72 DF,  p-value: 0.07266
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html
#explanation: simple regressions with growth mindset as the only predictor; they use everyone with both scores, so their n is larger than the hierarchical models'
sensitivity_table <- data.frame(
  outcome = c("Care-based support", "Care-based support", "Fear-based support", "Fear-based support"),
  model   = c("Block 2: growth mindset + racialized identity + gender", "Growth mindset alone",
              "Block 2: growth mindset + racialized identity + gender", "Growth mindset alone"),
  n       = c(nobs(regmodel2), nobs(regmodel7), nobs(regmodel5), nobs(regmodel8)),
  b       = round(c(coef(regmodel2)["GROWTH_cent"], coef(regmodel7)["GROWTH_cent"],
                    coef(regmodel5)["GROWTH_cent"], coef(regmodel8)["GROWTH_cent"]), 2),
  CI_95   = c(sprintf("[%.2f, %.2f]", confint(regmodel2)["GROWTH_cent", 1], confint(regmodel2)["GROWTH_cent", 2]),
              sprintf("[%.2f, %.2f]", confint(regmodel7)["GROWTH_cent", 1], confint(regmodel7)["GROWTH_cent", 2]),
              sprintf("[%.2f, %.2f]", confint(regmodel5)["GROWTH_cent", 1], confint(regmodel5)["GROWTH_cent", 2]),
              sprintf("[%.2f, %.2f]", confint(regmodel8)["GROWTH_cent", 1], confint(regmodel8)["GROWTH_cent", 2])),
  p       = c(summary(regmodel2)$coefficients["GROWTH_cent", "Pr(>|t|)"], summary(regmodel7)$coefficients["GROWTH_cent", "Pr(>|t|)"],
              summary(regmodel5)$coefficients["GROWTH_cent", "Pr(>|t|)"], summary(regmodel8)$coefficients["GROWTH_cent", "Pr(>|t|)"])
)
sensitivity_table$p <- ifelse(sensitivity_table$p < .001, "< .001", sprintf("%.3f", sensitivity_table$p))

kable(sensitivity_table, row.names = FALSE, align = c("l", "l", "r", "r", "r", "r"),
      col.names = c("Outcome", "Model", "n", "Growth mindset b", "95% CI", "p"),
      caption = "Table 5. Sensitivity analysis: the growth mindset coefficient with and without covariates")
Table 5. Sensitivity analysis: the growth mindset coefficient with and without covariates
Outcome Model n Growth mindset b 95% CI p
Care-based support Block 2: growth mindset + racialized identity + gender 71 0.42 [0.21, 0.62] < .001
Care-based support Growth mindset alone 75 0.41 [0.22, 0.61] < .001
Fear-based support Block 2: growth mindset + racialized identity + gender 70 -0.27 [-0.56, 0.03] 0.074
Fear-based support Growth mindset alone 74 -0.25 [-0.52, 0.02] 0.073
#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html
#explanation: if the growth mindset coefficient and its p-value barely change when the covariates are removed and more people are included, the result does not depend on the covariates or on which people were dropped

Note. Most of the sample loss happens before any model is fit: of the 210 rows in the Qualtrics export, 84 people rated the effectiveness items, and 75 of them also completed the growth mindset items. Removing racialized identity and gender adds back only 4 people to the care-based model.

5.2 Figure 10. Standardized regression coefficients

Which explanatory variables matter, and for which kind of support? Standardized coefficients (β) put every predictor on the same scale: the change in support, in standard deviations, for a one standard deviation change in the predictor.

# regmodel9 and regmodel10: the Block 2 models with every variable standardized (mean 0, SD 1)
regmodel9  <- lm(scale(CARE_EFFECT) ~ scale(POC) + scale(WOMAN) + scale(GROWTH_cent), data = caredata_model)
regmodel10 <- lm(scale(FEAR_EFFECT) ~ scale(POC) + scale(WOMAN) + scale(GROWTH_cent), data = feardata_model)

standardized_coefficients <- bind_rows(
  data.frame(outcome = "Care-based support",
             term    = c("Person of Color (vs. White)", "Woman (vs. man)", "Growth mindset"),
             beta    = coef(regmodel9)[-1],
             lower   = confint(regmodel9)[-1, 1],
             upper   = confint(regmodel9)[-1, 2]),
  data.frame(outcome = "Fear-based support",
             term    = c("Person of Color (vs. White)", "Woman (vs. man)", "Growth mindset"),
             beta    = coef(regmodel10)[-1],
             lower   = confint(regmodel10)[-1, 1],
             upper   = confint(regmodel10)[-1, 2])
) |>
  mutate(term    = factor(term, levels = c("Woman (vs. man)", "Person of Color (vs. White)", "Growth mindset")),
         outcome = factor(outcome, levels = c("Care-based support", "Fear-based support")))

kable(standardized_coefficients, digits = 2, row.names = FALSE,
      col.names = c("Outcome", "Predictor", "β", "95% CI lower", "95% CI upper"))
Outcome Predictor β 95% CI lower 95% CI upper
Care-based support Person of Color (vs. White) 0.02 -0.21 0.24
Care-based support Woman (vs. man) -0.05 -0.27 0.18
Care-based support Growth mindset 0.46 0.24 0.69
Fear-based support Person of Color (vs. White) -0.02 -0.26 0.23
Fear-based support Woman (vs. man) -0.02 -0.27 0.22
Fear-based support Growth mindset -0.23 -0.48 0.02
# caption for Figure 10, built from the models so the numbers stay correct
figure10_caption <- sprintf("Standardized regression coefficients (β) with 95%% confidence intervals from the Block 2 models (care-based n = %d, fear-based n = %d). Intervals that cross the dashed line at zero are not statistically significant. Growth mindset is the only predictor whose interval excludes zero, and only for care-based support (β = %.2f).", nobs(regmodel9), nobs(regmodel10), coef(regmodel9)["scale(GROWTH_cent)"])

#source: The Quantitative Playbook for Public Health Research in R (McCarty, 2026) https://shanemccarty.github.io/FRIplaybook/regression.html; R documentation ?scale
#explanation: scale() turns every variable into standard deviation units before fitting, so the coefficients are standardized betas that can be compared across predictors and across the two outcomes; the binary predictors are standardized too, which is the usual convention for reporting beta
plot10_standardized_coefficients <- ggplot(standardized_coefficients,
                                           aes(x = beta, y = term, color = outcome)) +
  geom_vline(xintercept = 0, linetype = "dashed", color = "grey50") +
  geom_errorbar(aes(xmin = lower, xmax = upper), width = 0.2, linewidth = 0.9, orientation = "y",
                position = position_dodge(width = 0.5)) +
  geom_point(size = 3.5, position = position_dodge(width = 0.5)) +
  scale_color_manual(values = c("Care-based support" = "#8297ce",
                                "Fear-based support" = "#453a98")) +
  labs(x = "Standardized coefficient (β) with 95% confidence interval",
       y = NULL, color = NULL) +
  theme_bw(base_family = "Calibri") +
  theme(legend.position = "top")

plot10_standardized_coefficients

Standardized regression coefficients (β) with 95% confidence intervals from the Block 2 models (care-based n = 71, fear-based n = 70). Intervals that cross the dashed line at zero are not statistically significant. Growth mindset is the only predictor whose interval excludes zero, and only for care-based support (β = 0.46).
ggsave("plotsTEAM/plot10_standardized_coefficients.png", plot = plot10_standardized_coefficients,
       width = 8, height = 4.5, dpi = 300)

#source: ggplot2 documentation https://ggplot2.tidyverse.org/reference/geom_linerange.html
#explanation: each point is a standardized coefficient and each bar its 95 percent confidence interval, blue for care-based support and purple for fear-based support; only growth mindset for care-based support sits clearly away from zero

References

Norris, Z. (2021). Defund fear: Safety without policing, prisons, and punishment. Beacon Press.