Replication of ‘Does poverty promote a different and harmful way of thinking? The links between economic scarcity, concrete construal level and risk behaviors’ experiment 2 by Caballero et al. (2021, Current Psychology)

Author

Morgan Tompkins

Published

December 1, 2025

Introduction

Justification

I chose Caballero et al. (2023) Study 2 because it experimentally manipulates perceived economic scarcity to test how it shifts people’s construal level, or their tendency to think in concrete versus abstract terms. This focus on cognitive style and self-control under scarcity directly connects to my interests in the psychology of poverty, dignity, and institutional precarity. I am interested in how experiences of resource insecurity shape short-term versus long-term decision-making, and how interventions might target these different mindsets to help break individuals out of poverty cycles.

Caballero et al. (2023) build on Construal Level Theory (Trope & Liberman, 2003), which proposes that psychological distance influences whether people think concretely or abstractly. In their Study 2, participants were randomly assigned to a “scarcity” or “non-scarcity” condition within a fictional society called Bimboola, and the researchers measured how this manipulation changed participants’ thinking styles using the Behavioral Identification Form (BIF). They found that those in the scarcity condition thought more concretely about actions and goals, confirming that perceived financial limitation causally narrows one’s cognitive focus to the immediate present.

Original Authors:Pilar Carrera (pilar.carrera@uam.es); Amparo Caballero (amparo.caballero@uam.es); Itziar Fernández (ifernandez@psi.uned.es); Pilar Aguilar (mpaguilar@uloyola.es); Dolores Muñoz (lola.munnoz@uam.es)

Stimuli & Procedures

This replication will follow Caballero et al. (2023) Study 2. The original sample included 120 undergraduate participants (102 women) who were randomly assigned to a scarcity or non-scarcity condition. All data were collected individually in lab cubicles via Qualtrics.

The procedure consisted of: 1. BIF (Pre-test): 12 items from the Behavioral Identification Form, where each action (e.g., “locking a door”) is described at two levels: 1. a concrete “how” description (“putting the key in the lock”) and 2. an abstract “why” description (“securing the house”). Participants select which feels more natural.

  1. Economic-scarcity manipulation (fictional society): Participants imagined living in a fictional society (Bimboola) organized by five income groups.

    • Scarcity condition: assigned to the lowest income tier (≤ 400 Bimboolean per month) and required to choose items such as housing, transportation, and leisure from unattractive, low-resource options while viewing the better alternatives available to wealthier groups.
    • Non-scarcity condition: assigned to a comfortable middle-income tier (1,201–3,000 Bimboolean) and made similar choices among adequate options.

  2. BIF (Post-test): A second 12-item subset of the BIF (non-overlapping with the pre-test) measured change in abstraction.

  3. Manipulation checks: “My group is poor” and “My group is rich” on 7-point scales, plus an estimated group-income question.

  4. Basic demographic collection & debrief.

Challenges:
One challenge in replicating this study include adapting the materials to an English-speaking Prolific population rather than the original Spanish-speaking undergraduate sample. Another challenge involves maintaining participant attention across both the pre and post-manipulation Behavioral Identification Form (BIF) scales, which may induce some fatigue, especially in the second half. Finally, ensuring that the scarcity manipulation feels psychologically real in an online environment without the presence of an experimenter (as was the case in the original study) will be important for preserving the strength of the original effect that was does individually in cubicles. —

Methods

Power Analysis

This study relied on an effect size consistent with prior work on economic scarcity and unhealthy behavior (medium effect, f = 0.30; Bratanova et al., 2016; Laran & Salerno, 2013). Power analyses conducted with G*Power indicated that, to detect this effect with 80% power at α = .05, the required sample sizes were N = 90 for designs with two conditions.

Recruitment followed the practical constraints of the lab schedule: once the target sample sizes identified by the power analysis were reached, data collection continued through the full two-week period during which lab space had been allocated. The Study 3 sample reflected voluntary participation by students enrolled in a psychology course as part of a larger survey effort. All measures, manipulations, and exclusions were fully disclosed.

Planned Sample

Participants will be recruited to reach the planned sample size of N = 72, based on a power analysis targeting 80% power to detect the original effect size (f = 0.33) reported by Caballero et al. (2021). The study uses a questionnaire format and corresponds to Experiment 2 in the original paper. Data collection will end once the planned sample size is reached. There are no preregistered materials or open data associated with the original study. The original sample consisted of N = 120 participants. The sampling frame for the current replication follows the same general structure, without any additional preselection criteria or known demographic quotas. Estimated participation time is 15 minutes, with compensation set at $2.00 per person (equivalent to $8.00 per hour), for a total projected cost of $191.52.

Materials

Economic Scarcity Manipulation (Bimboola Procedure)

Economic scarcity was manipulated an adaptation of the Bimboola paradigm (Jetten et al., 2015) to manipulate the experience of economic scarcity. Participants were assigned to one of two economic conditions:

Scarcity: participants were placed in the lowest income group, described as living on the poverty line and having difficulty living comfortably.

Nonscarcity (Economic Sufficiency): participants were placed in the third income group, described as having enough money to live comfortably.

In both conditions, participants imagined themselves living in Bimboola and selected a house, vehicle, phone, and leisure activity from sets of three items associated with each income group. All groups’ options were visible during the choice process.

Behavioral Identification Form (BIF; Vallacher & Wegner, 1989)

Construal level was measured using the BIF, one of the most widely used tools for assessing abstraction. The original BIF includes 25 items; Study 2 used two sets of 12 items each. Participants first completed 12 randomly selected items before the economic manipulation, and then completed 12 additional items afterward (only one item from the original scale was not used at any point).

Each item presents an action with two descriptions: one abstract (scored 1) and one concrete (scored 0). For example, “locking a door” could be described as “securing the house” (abstract) or “putting the key in the lock” (concrete). The total number of abstract choices indexes higher-level construal.

Procedure

“Participants responded to the online survey in separate cubicles. First, they completed twelve items randomly chosen from the BIF (Vallacher & Wegner, 1989). The original version of the BIF (Vallacher & Wegner, 1989) includes 25 items. In this scale, participants are presented with different actions and must choose between two options for each action. One option describes the action in concrete terms, while the other option describes the action in abstract terms. For example, participants must choose whether “locking a door” is best described as “securing the house” (abstract level; scored as 1) or “putting the key in the lock” (concrete level; scored as 0). The number of abstract descriptions selected serves as a measure of abstraction: higher scores indicate higher abstraction. Then, participants imagined hemselves living in Bimboola. The same manipulation used in Study 1 was adopted, with a slight variation in the amounts of money: the wealthiest group earned more than 100,000 Bimboolean € per month; the second group earned between 3001 and 100,000 Bimboolean € per month; the third group earned between 1201 and 3000 Bimboolean € per month; the fourth group earned between 400 and 1200 Bimboolean € per month; and the fifth group earned less than 400 Bimboolean € per month (on the poverty line). The procedure was identical to that in Study 1; participants had to choose a house, a vehicle, a phone, and a leisure activity from a group of three items associated with each economic group. Participants observed the items of all groups when making the selection. The participants in the scarcity condition were assigned to the fifth group (on the poverty line); the participants in the nonscarcity condition were assigned to the third group (enough money to live comfortably). After the Bimboola manipulation, participants completed the second part of the BIF scale, with twelve new items (only one randomly selected item from the original scale was not used). Finally, they provided demographic information and responded to the three manipulation check items used in Study 1.”

The procedures of this study were followed with the following exceptions:

  1. Income groups were re-adjusted to match a U.S. rather than Spanish context. The new ranges approximate U.S. household income quintiles from Household Income data. They were then converted to monthly figures and reflect meaningful distinctions in purchasing power and living standards parallel to the Spanish brackets. Group 1 corresponds to income levels below the U.S. poverty threshold; Groups 2 and 3 map onto lower- and middle-income households; Group 4 reflects upper-middle-income earners; and Group 5 captures high-income households without inflating the top category unrealistically. This update keeps the relative spacing of the original categories while adapting to US cultural contexts. The revised income brackets are as follows:

Group 1 earns less than 1,500 $ Bimbolianos (DB) per month (they are below the poverty line), Group 2 earns between 1,500 and 4,000 DB per month, Group 3 earns between 4,001 and 8,000 DB per month, Group 4 earns between 8,001 and 20,000 DB per month, and, Group 5 earns more than 20,000 DB per month.

Images of housing, phone, transportation, and vacation options were also adjusted to reflect a U.S. context, borrowing images from Zillow, CarMax, and Google where applicable.

Analysis Plan

All analyses will follow procedures as closely aligned with the original article as possible. Data will first be inspected for incomplete responses, with participants excluded only if they did not finish the BIF measures or the manipulation checks, consistent with the original study’s practice of analyzing all voluntary participants who completed the survey. No additional exclusion rules or covariates were used in the original paper, and none will be introduced. Response text for BIF items will be cleaned by lowercasing and trimming whitespace to ensure accurate matching with the scoring dictionary.

A standardized dictionary will be constructed for all 25 Behavioral Identification Form items, detailing each item’s concrete and abstract response options. Participants’ BIF responses will then be reshaped into long format and coded as abstract (1) or concrete (0). For each participant, pre-test abstraction scores will be calculated using items 1–11, and post-test abstraction scores will be calculated using items 12–24. Condition-level means and standard deviations will be summarized in tables, accompanied by visualizations illustrating changes in abstraction from pre- to post-manipulation across scarcity and nonscarcity conditions. An additional item-level table will report the percentage of abstract selections for each BIF item by condition to mirror the interpretive granularity of the original study.

The manipulation checks described in the original results will be reproduced through two between-subjects ANOVAs testing whether participants in the scarcity condition rated their assigned group as poorer and less rich than those in the nonscarcity condition. These tests verify that the Bimboola manipulation induced the intended subjective perceptions of scarcity.

The key analysis of interest is a mixed-design ANOVA with time (pre-test vs. post-test BIF abstraction score) as a within-subject factor and condition (scarcity vs. nonscarcity) as a between-subject factor. The condition × time interaction is the critical test of the hypothesis, as the original study reported a significant interaction showing that scarcity lowered abstraction while nonscarcity increased it. Main effects of time and condition will be examined but are not expected to differ based on the original findings. To verify baseline comparability, pre-test abstraction scores will be compared across conditions to confirm the absence of initial group differences prior to the manipulation.

This analytic strategy replicates the data handling, scoring procedures, exclusion criteria, manipulation checks, and inferential tests reported in the original article.

Differences from Original Study

The primary differences between the original study and the planned replication involve the sample, setting, and mode of data collection. Whereas the original research was conducted with undergraduate students in Spain who completed the survey in laboratory cubicles, the replication will use a U.S.-based online sample recruited through Prolific. The study will therefore be administered entirely online rather than in person. No substantive changes will be made to the materials or procedures beyond formatting adjustments required for online delivery. Based on the nature of the manipulation and the measures, these differences are not expected to meaningfully alter the results. The original article does not identify cultural or in-person contextual factors as necessary conditions for observing the effect, and the BIF measure and Bimboola manipulation have been validated in diverse settings. Accordingly, the differences between the samples and recruitment platforms are not anticipated to influence the pattern of findings.

Results

Data preparation

library(foreign) # for reading spss formatted data
library(tidyr)
library(dplyr)

Attaching package: 'dplyr'
The following objects are masked from 'package:stats':

    filter, lag
The following objects are masked from 'package:base':

    intersect, setdiff, setequal, union
library(stringr) # useful for some string manipulation
library(ggplot2)
library(kableExtra)

Attaching package: 'kableExtra'
The following object is masked from 'package:dplyr':

    group_rows
library(lubridate)

Attaching package: 'lubridate'
The following objects are masked from 'package:base':

    date, intersect, setdiff, union
library(ez)
#install.packages("showtext")
library(showtext)
Loading required package: sysfonts
Loading required package: showtextdb
font_add_google("Mulish", "Mulish")
showtext_auto()

Import data

df = read.csv("PSYCH251_December 1, 2025_11.34.csv")
#head(df)

Data exclusion / filtering

valid_id <- "692ccf9ea85f7e5cca050dca"

df_select <- df %>%
  filter(grepl("^\\d{4}-\\d{2}-\\d{2}", StartDate)) %>%
  mutate(StartDate = ymd_hms(StartDate)) %>%
  filter(as_date(StartDate) == Sys.Date() - 1) %>% # keep only yesterday
  filter(STUDY_ID == valid_id)

Prepare data for analysis - create columns etc.

Create clean dataset

df_clean <- df_select %>%
  mutate(participant_id = row_number()) %>%
  rename(rating_poor = group_construal_1,
         rating_rich = group_construal_2) %>%
  mutate(rating_poor = str_extract(rating_poor, "\\d+") %>% as.numeric(),
         rating_rich = str_extract(rating_rich, "\\d+") %>% as.numeric()) %>%
  mutate(manip_pass = case_when(
      condition == "scarcity" & bimboola_earn == "Less than 1,500 DB per month" ~ TRUE,
      condition == "sufficiency" & bimboola_earn == "Between 4,001 and 8,000 DB per month" ~ TRUE,
      TRUE ~ FALSE)) %>% 
  select(participant_id, starts_with("BIF"), manip_pass, rating_poor, rating_rich, age, gender, condition) 

write.csv(df_clean, "pilotB_anonymized.csv", row.names = FALSE)

Read in clean data (open access would start from here?)

df_clean <- read.csv("pilotB_anonymized.csv")
#head(df_clean)

Create BIF dictionary

# The standard 25 BIF items. 
bif_item_descriptions <- data.frame(
  BIF_Item = paste0("BIF_", 1:25), 
  
  Item_Topic = c(
    "Organization", "Knowledge Acquisition", "Civic Duty/Defense", "Hygiene", "Nutrition/Consumption",
    "Resource Acquisition", "Home Improvement", "Housekeeping", "Painting", "Housing Maintenance",
    "Gardening", "Security", "Civic Duty", "Recreation", "Self-Reflection",
    "Dental Health", "Assessment", "Social Interaction", "Self-Control", "Eating/Physiology",
    "Gardening/Exercise", "Transportation", "Health Appointment", "Child Interaction", "Social Contact"
  ),
  
  # The exact text of the CONCRETE (0) choice based on user's list
  Concrete_Choice = c(
    "Writing things down",                   # BIF 1: Writing things down (b)
    "Following lines of print",             # BIF 2: Following lines of print (a)
    "Signing up",                           # BIF 3: Signing up (b)
    "Putting clothes into the machine",     # BIF 4: Putting clothes into the machine (b)
    "Pulling an apple off a branch",        # BIF 5: Pulling an apple off a branch (b)
    "Wielding an axe",                      # BIF 6: Wielding an axe (a)
    "Using a yardstick",                    # BIF 7: Using a yardstick (b)
    "Vacuuming the floor",                  # BIF 8: Vacuuming the floor (b)
    "Applying brush strokes",               # BIF 9: Applying brush strokes (a)
    "Writing a check",                      # BIF 10: Writing a check (b)
    "Watering plants",                      # BIF 11: Watering plants (a)
    "Putting a key in the lock",            # BIF 12: Putting a key in the lock (a)
    "Marking a ballot",                     # BIF 13: Marking a ballot (b)
    "Holding on to branches",               # BIF 14: Holding on to branches (b)
    "Answering questions",                  # BIF 15: Answering questions (a)
    "Moving a brush around in one's mouth", # BIF 16: Moving a brush around in one's mouth (b)
    "Answering questions",                  # BIF 17: Answering questions (a)
    "Saying hellow",                         # BIF 18: Saying hello (a)
    "Saying \"no\"",                        # BIF 19: Saying "no" (a)
    "Chewing and swallowing",               # BIF 20: Chewing and swallowing (b)
    "Planting seeds",                       # BIF 21: Planting seeds (a)
    "Following a map",                      # BIF 22: Following a map (a)
    "Going to the dentist",                 # BIF 23: Going to the dentist (b)
    "Using simple words",                   # BIF 24: Using simple words (b)
    "Moving a finger"                       # BIF 25: Moving a finger (a)
  ),
  
  # The exact text of the ABSTRACT (1) choice based on user's list
  Abstract_Choice = c(
    "Getting organized",                    # BIF 1: Getting organized (a)
    "Gaining knowledge",                    # BIF 2: Gaining knowledge (b)
    "Helping the Nation's defense",         # BIF 3: Helping the Nation's defense (a)
    "Removing odor from the clothes",          # BIF 4: Removing odors from clothes (a)
    "Getting something to eat",             # BIF 5: Getting something to eat (a)
    "Getting firewood",                     # BIF 6: Getting firewood (b)
    "Getting ready to remodel",             # BIF 7: Getting ready to remodel (a)
    "Showing one's cleanliness",            # BIF 8: Showing one's cleanliness (a)
    "Making the room look fresh",           # BIF 9: Making the room look fresh (b)
    "Maintaining a place to live",          # BIF 10: Maintaining a place to live (a)
    "Making the room look fresh",            # BIF 11: Making the room look nice (b)
    "Securing the house",                   # BIF 12: Securing the house (b)
    "Influencing the election",             # BIF 13: Influencing the election (a)
    "Getting a good view",                  # BIF 14: Getting a good view (a)
    "Revealing what you're like",           # BIF 15: Revealing what you're like (b)
    "Preventing tooth decay",               # BIF 16: Preventing tooth decay (a)
    "Showing one's knowledge",              # BIF 17: Showing one's knowledge (b)
    "Showing friendliness",                 # BIF 18: Showing friendliness (b)
    "Showing moral courage",                # BIF 19: Showing moral courage (b)
    "Getting nutrition",                    # BIF 20: Getting nutrition (a)
    "Getting fresh vegetables",             # BIF 21: Getting fresh vegetables (b)
    "Seeing countryside",                   # BIF 22: Seeing countryside (b)
    "Protecting your teeth",                # BIF 23: Protecting your teeth (a)
    "Teaching a child something",           # BIF 24: Teaching a child something (a)
    "Seeing if someone's home"              # BIF 25: Seeing if someone's home (b)
  ),
  stringsAsFactors = FALSE
) %>%
  mutate(
    Concrete_Choice = str_trim(tolower(Concrete_Choice)),
    Abstract_Choice = str_trim(tolower(Abstract_Choice))
  )

Confirmatory analysis

Step 0: Remove participants who didn’t pass attention check

df_manip_pass <- df_clean %>%
  filter(manip_pass == "TRUE")

# How many failed? 
nrow(df_clean) - nrow(df_manip_pass)
[1] 1

Step 1: Check manipulation checks

# Do scarcity participants view themselves at poorer in bimboola society?
#t.test(rating_poor ~ condition, data = df_manip_pass)

# Do sufficiency participants view themselves as well off in bimboola society? 
#t.test(rating_rich ~ condition, data = df_manip_pass)

Step 2: Score participant’s BIF responses for variables BIF 1-24 (abstract vs. concrete)

# Select relevant data
bif_counts_df <- df_manip_pass %>%
  select(starts_with("BIF"), condition, participant_id) %>%
  
  pivot_longer(
    cols = starts_with("BIF"),
    names_to = "BIF_Item",
    values_to = "Choice" # 1 = Abstract, 0 = Concrete
  ) %>%
  mutate(
    Choice = str_trim(tolower(Choice)), 
    condition = str_trim(tolower(condition))
  ) %>%
  
  left_join(bif_item_descriptions, by = "BIF_Item") %>%
  
  mutate(
    abstract = case_when(
     Choice == Abstract_Choice ~ 1, 
     Choice == Concrete_Choice ~ 0,
     TRUE ~ NA_real_ 
    )
  )

bif_counts_df  
# A tibble: 72 × 8
   condition participant_id BIF_Item Choice           Item_Topic Concrete_Choice
   <chr>              <int> <chr>    <chr>            <chr>      <chr>          
 1 scarcity               1 BIF_1    getting organiz… Organizat… writing things…
 2 scarcity               1 BIF_2    gaining knowled… Knowledge… following line…
 3 scarcity               1 BIF_3    helping the nat… Civic Dut… signing up     
 4 scarcity               1 BIF_4    removing odor f… Hygiene    putting clothe…
 5 scarcity               1 BIF_5    getting somethi… Nutrition… pulling an app…
 6 scarcity               1 BIF_6    getting firewood Resource … wielding an axe
 7 scarcity               1 BIF_7    getting ready t… Home Impr… using a yardst…
 8 scarcity               1 BIF_8    showing one's c… Housekeep… vacuuming the …
 9 scarcity               1 BIF_9    making the room… Painting   applying brush…
10 scarcity               1 BIF_10   maintaining a p… Housing M… writing a check
# ℹ 62 more rows
# ℹ 2 more variables: Abstract_Choice <chr>, abstract <dbl>

Step 3: Compute simple pre-test BIF score (BIF 1-12) & post-test BIF score (BIF 13-24)

bif_pre_post_df <- bif_counts_df %>%
  mutate(BIF_Index = as.numeric(str_replace(BIF_Item, "BIF_", ""))) %>%
  group_by(participant_id, condition) %>%
  summarise(
    pre_test_score = sum(abstract[BIF_Index < 12], na.rm = TRUE),
    post_test_score = sum(abstract[BIF_Index >= 12 & BIF_Index <= 24], na.rm = TRUE),
    .groups = 'drop') 
  
pre_post_summary_table <- bif_pre_post_df %>%
  group_by(condition) %>%
  summarise(
  respondent_count = n(),  
  pre_score_avg = round(mean(pre_test_score, na.rm = TRUE),1),
  pre_score_sd = round(sd(pre_test_score, na.rm = TRUE),1),
  post_score_avg = round(mean(post_test_score, na.rm = TRUE),1),
  post_score_sd = round(sd(post_test_score, na.rm = TRUE),1),
  .groups = 'drop'
  )

bif_summary_table_final <- pre_post_summary_table %>%
  # 1. Start with kable (mandatory for kableExtra)
  kable(
    caption = "Item-Level Comparison of Abstract Choice Percentage by Condition",
    align = "lcccrr",
    digits = 1
  ) %>%
  # 2. Add the styling (striped effect comes from 'bootstrap_options')
  kable_styling(
    bootstrap_options = c("striped", "hover", "condensed", "responsive"),
    full_width = FALSE
  )

Step 4: Run the 2×2 ANOVA (within: time, between: condition)

# Reshape to long 
bif_long <- bif_pre_post_df %>%
  pivot_longer(
    cols = c(pre_test_score, post_test_score),
    names_to = "time",
    values_to = "BIF_score"
  ) %>%
  mutate(
    time        = factor(time,
                         levels = c("pre_test_score", "post_test_score"),
                         labels = c("pre", "post")),
    condition   = factor(condition),
    participant_id = factor(participant_id)
  )

# Mixed ANOVA
anova_results <- ezANOVA(
  data    = bif_long,
  dv      = .(BIF_score),
  wid     = .(participant_id),
  within  = .(time),
  between = .(condition),
  type    = 3,
  detailed = TRUE
)
Warning: Data is unbalanced (unequal N per group). Make sure you specified a
well-considered value for the type argument to ezANOVA().
anova_results
$ANOVA
          Effect DFn DFd SSn SSd     F         p p<.05       ges
1    (Intercept)   1   1 300  25 12.00 0.1789124       0.9230769
2      condition   1   1  12  25  0.48 0.6142777       0.3243243
3           time   1   1   3   0   Inf 0.0000000     * 0.1071429
4 condition:time   1   1   3   0   Inf 0.0000000     * 0.1071429
Step 4.1 Effect of condition at post
#t.test(post_test_score ~ condition, data = bif_pre_post_df)
Step 4.2 Effect of time within each condition
# scarcity_df <- bif_pre_post_df %>% filter(condition == "scarcity")
# 
# t.test(scarcity_df$pre_test_score, scarcity_df$post_test_score, paired = TRUE)

Step 5: ANCOVA

ancova_model <- lm(post_test_score ~ condition + pre_test_score,
                   data = bif_pre_post_df)

anova(ancova_model)
Warning in anova.lm(ancova_model): ANOVA F-tests on an essentially perfect fit
are unreliable
Analysis of Variance Table

Response: post_test_score
               Df Sum Sq Mean Sq F value Pr(>F)
condition       1   13.5    13.5     NaN    NaN
pre_test_score  1   12.5    12.5     NaN    NaN
Residuals       0    0.0     NaN               
summary(ancova_model)

Call:
lm(formula = post_test_score ~ condition + pre_test_score, data = bif_pre_post_df)

Residuals:
ALL 3 residuals are 0: no residual degrees of freedom!

Coefficients:
                     Estimate Std. Error t value Pr(>|t|)
(Intercept)                 3        NaN     NaN      NaN
conditionsufficiency       -3        NaN     NaN      NaN
pre_test_score              1        NaN     NaN      NaN

Residual standard error: NaN on 0 degrees of freedom
Multiple R-squared:      1, Adjusted R-squared:    NaN 
F-statistic:   NaN on 2 and 0 DF,  p-value: NA

Visual

# Shared dodge so bars, error bars, and labels align
pd <- position_dodge(width = 0.7)

pre_post_long_for_plot <- bif_pre_post_df %>%
  # wide → long
  pivot_longer(
    cols = c(pre_test_score, post_test_score),
    names_to = "Test_Phase",
    values_to = "Score"
  ) %>%
  mutate(
    Test_Phase = case_when(
      Test_Phase == "pre_test_score"  ~ "Pre-Test",
      Test_Phase == "post_test_score" ~ "Post-Test",
      TRUE ~ Test_Phase
    ),
    Test_Phase = factor(Test_Phase, levels = c("Pre-Test", "Post-Test")),
    condition  = stringr::str_to_title(condition)
  ) %>%
  group_by(condition, Test_Phase) %>%
  summarise(
    Mean_Score = mean(Score, na.rm = TRUE),
    SE_Score   = sd(Score, na.rm = TRUE) / sqrt(n()),
    .groups    = "drop"
  ) %>%
  mutate(
    SE_Score_fixed = ifelse(is.na(SE_Score), 0, SE_Score),
    label_y        = Mean_Score + SE_Score_fixed + 0.3
  )

pre_post_long_for_plot %>%
  ggplot(aes(x = Test_Phase, y = Mean_Score, fill = condition)) +
  geom_col(
    position = pd,
    color    = "white",
    width    = 0.6
  ) +
  geom_errorbar(
    aes(
      ymin = Mean_Score - SE_Score,
      ymax = Mean_Score + SE_Score
    ),
    position = pd,
    width    = 0.15,
    linewidth = 0.8,
    na.rm    = TRUE
  ) +
  geom_text(
    aes(
      y     = label_y,
      label = round(Mean_Score, 1)
    ),
    position = pd,
    vjust    = -0.1,
    size     = 4.5,
    fontface = "bold",
    color    = "black"
  ) +
  scale_fill_manual(values = c(
    "Scarcity"    = "#2C6CF5",
    "Sufficiency" = "#B0B0B0"
  )) +
  labs(
    title    = "Abstract BIF Score by Condition and Test Phase",
    subtitle = "Mean Abstract BIF Score (±1 SE, Max 12)",
    y        = NULL,
    x        = "Test Phase",
    fill     = "Condition"
  ) +
  scale_y_continuous(
    limits = c(0, NA),
    expand = expansion(mult = c(0, 0.08))
  ) +
  coord_cartesian(clip = "off") +
  theme_minimal(base_size = 16) +
  theme(
    text = element_text(family = "Mulish"),
    axis.text.y        = element_blank(),
    axis.title.y       = element_blank(),
    axis.ticks.y       = element_blank(),
    panel.grid.major.y = element_blank(),
    panel.grid.minor.y = element_blank(),
    panel.grid.major.x = element_blank(),
    panel.grid.minor.x = element_blank(),
    axis.line.x        = element_line(color = "black", linewidth = 0.5),
    plot.title         = element_text(face = "bold", size = 20),
    plot.subtitle      = element_text(size = 13),
    legend.position    = "bottom",
    plot.margin        = margin(t = 20, r = 10, b = 10, l = 10)
  )