Structural Drivers of Financial Nihilism — Initial Analysis
Cleaning → FAI construction → cohort-by-FAI logit
Olivia Shipley
2026-03-25
0. Setup
library(ipumsr)
library(dplyr)
library(data.table)
library(R.utils)
library(ggplot2)
library(lmtest)
library(sandwich)
1. Load data
ddi <- read_ipums_ddi("~/Desktop/DA6833/FinancialNihilism/usa_00003.xml")
data <- fread("~/Desktop/DA6833/FinancialNihilism/usa_00003.csv.gz")
setDT(data)
stopifnot(all(unique(data$YEAR) %in% c(1970, 1980, 1990, 2000, 2010, 2015, 2019, 2022, 2024)))
2. Sample construction
# drop children
data <- data[AGE >= 18]
# Gen by birth year
data[, BIRTHYEAR := YEAR - AGE]
data[, GENERATION := fcase(
BIRTHYEAR >= 1928 & BIRTHYEAR <= 1945, "Silent",
BIRTHYEAR >= 1946 & BIRTHYEAR <= 1964, "Boomer",
BIRTHYEAR >= 1965 & BIRTHYEAR <= 1980, "GenX",
BIRTHYEAR >= 1981 & BIRTHYEAR <= 1996, "Millennial",
BIRTHYEAR >= 1997 & BIRTHYEAR <= 2012, "GenZ",
default = "Other"
)]
data <- data[GENERATION != "Other"]
# Dropping 1970 & silent gen from the sample
# Reasoning:
# (1) MORTAMT1 is not collected in the 1970 IPUMS sample (per usa_00003.cbk),
# so owner housing-cost burden cannot be computed for 1970. This silently
# dropped 1970 owners from the FAI sample, leaving 1970 represented almost
# entirely by renters, which produced a degenerate reference cell in the
# earlier model fit (intercept ~ -16 on the logit scale).
# (2) Silent generation (born 1928-1945) appears in the 25-35 window almost
# exclusively in 1970 in this extract. Dropping 1970 effectively drops
# Silent regardless.
# Boomer becomes the natural reference cohort: cleanly observed from 1980 onward
# across multiple sample years with full mortgage data.
data <- data[YEAR != 1970]
data <- data[GENERATION != "Silent"]
data[, GENERATION := factor(GENERATION,
levels = c("Boomer", "GenX", "Millennial", "GenZ"))]
# prime household-formation window
dt_prime <- data[AGE >= 25 & AGE <= 35]
n_before_cpi <- nrow(dt_prime)
n_before_cpi
## [1] 3367742
3. Inflation adjustment
cpi_lookup <- data.table(
YEAR = c(1980, 1990, 2000, 2010, 2015, 2019, 2022, 2024),
cpi = c(82.4, 130.7, 172.2, 218.1, 237.0, 255.7, 292.7, 313.5)
)
stopifnot(length(setdiff(unique(dt_prime$YEAR), cpi_lookup$YEAR)) == 0)
# Recode IPUMS missing sentinels
dt_prime[VALUEH == 9999999, VALUEH := NA]
dt_prime[HHINCOME == 9999999, HHINCOME := NA]
dt_prime[INCTOT == 9999999, INCTOT := NA]
dt_prime[MORTAMT1 == 99999, MORTAMT1 := NA]
dt_prime[COSTELEC %in% c(99993, 99994, 99995, 99999), COSTELEC := NA]
dt_prime[COSTGAS %in% c(99993, 99994, 99995, 99999), COSTGAS := NA]
dt_prime[COSTWATR %in% c(99993, 99994, 99995, 99999), COSTWATR := NA]
dt_prime[COSTFUEL %in% c(99993, 99994, 99995, 99999), COSTFUEL := NA]
# merge cpi on year
dt_prime <- cpi_lookup[dt_prime, on = "YEAR"]
stopifnot(nrow(dt_prime) == n_before_cpi)
# inflation-adjust to 2024
cpi_2024 <- 313.5
dt_prime[, valueh_real := VALUEH * (cpi_2024 / cpi)]
dt_prime[, hhinc_real := HHINCOME * (cpi_2024 / cpi)]
dt_prime[, inc_real := INCTOT * (cpi_2024 / cpi)]
dt_prime[, rentgrs_real := RENTGRS * (cpi_2024 / cpi)]
dt_prime[, mortamt1_real := MORTAMT1 * (cpi_2024 / cpi)]
dt_prime[, costelec_real := COSTELEC * (cpi_2024 / cpi)]
dt_prime[, costgas_real := COSTGAS * (cpi_2024 / cpi)]
dt_prime[, costwatr_real := COSTWATR * (cpi_2024 / cpi)]
dt_prime[, costfuel_real := COSTFUEL * (cpi_2024 / cpi)]
# Restrict to own/rent universe (single application; was duplicated in original)
dt_prime <- dt_prime[OWNERSHP %in% c(1, 2)]
4. FAI components
4a. Housing cost burden ratio (universal — defined for owners and renters)
# NOTE: in IPUMS, MORTAMT1 is the MONTHLY mortgage payment, but
# COSTELEC/COSTGAS/COSTWATR/COSTFUEL are ANNUAL dollar amounts.
# RENTGRS already includes utilities and is reported monthly
dt_prime[, burden_ratio := fifelse(
OWNERSHP == 2, # renter
rentgrs_real / (hhinc_real / 12),
# owner:
(mortamt1_real +
(costelec_real + costgas_real + costwatr_real + costfuel_real) / 12) /
(hhinc_real / 12)
)]
# trm infinities and extreme values
dt_prime[is.infinite(burden_ratio) | burden_ratio > 10 | burden_ratio < 0,
burden_ratio := NA_real_]
4b. Price-to-income ratio (with year-level imputation for renters)
# owner P/I: standard
dt_prime[, price_to_income_hh := valueh_real / hhinc_real]
year_median_value <- dt_prime[OWNERSHP == 1 & !is.na(valueh_real),
.(median_valueh_real = median(valueh_real,
na.rm = TRUE)),
by = YEAR]
dt_prime <- year_median_value[dt_prime, on = "YEAR"]
# Renter P/I = year-median owner home value / household income
dt_prime[OWNERSHP == 2,
price_to_income_hh := median_valueh_real / hhinc_real]
dt_prime[is.infinite(price_to_income_hh) | price_to_income_hh < 0,
price_to_income_hh := NA_real_]
dt_prime[price_to_income_hh > 50, price_to_income_hh := NA_real_]
dt_prime[, .(
n = .N,
burden_med = median(burden_ratio, na.rm = TRUE),
p2i_med = median(price_to_income_hh, na.rm = TRUE)
), by = OWNERSHP]
## OWNERSHP n burden_med p2i_med
## <int> <int> <num> <num>
## 1: 1 1883587 0.1916667 2.169197
## 2: 2 1365678 0.2263830 3.498834
5. FAI via PCA
fai_input <- dt_prime[!is.na(burden_ratio) & !is.na(price_to_income_hh),
.(burden_ratio, price_to_income_hh)]
pca_fit <- prcomp(fai_input, center = TRUE, scale. = TRUE)
print(pca_fit$rotation)
## PC1 PC2
## burden_ratio -0.7071068 0.7071068
## price_to_income_hh -0.7071068 -0.7071068
print(summary(pca_fit))
## Importance of components:
## PC1 PC2
## Standard deviation 1.2588 0.6445
## Proportion of Variance 0.7923 0.2077
## Cumulative Proportion 0.7923 1.0000
dt_prime[!is.na(burden_ratio) & !is.na(price_to_income_hh),
FAI := pca_fit$x[, 1]]
sign_check <- cor(dt_prime$FAI, dt_prime$price_to_income_hh, use = "complete.obs")
if (sign_check < 0) {
dt_prime[, FAI := -FAI]
}
# distribution of FAI by generation
dt_prime[!is.na(FAI), .(
n = .N,
fai_mean = mean(FAI),
fai_median = median(FAI)
), by = GENERATION]
## GENERATION n fai_mean fai_median
## <fctr> <int> <num> <num>
## 1: Boomer 324426 -0.13227404 -0.4698059
## 2: GenX 304186 -0.05050935 -0.4257556
## 3: Millennial 702982 0.07155292 -0.3052493
## 4: GenZ 67257 0.11860417 -0.2399871
6. Model preparation
# IMPORTANT: filter on `price_to_income_hh`, NOT on the PCA-based `FAI`.
model_dt <- dt_prime[!is.na(price_to_income_hh) & !is.na(MARST) &
!is.na(RACE) & !is.na(HISPAN) & !is.na(EDUC)]
# outcome: 1 = own, 0 = rent
model_dt[, own := as.integer(OWNERSHP == 1)]
# Collapse RACE to 4 categories (matching Census/JCHS convention)
# IPUMS RACE codes: 1=White, 2=Black, 3=AI/AN, 4=Chinese, 5=Japanese,
# 6=Other Asian/PI, 7=Other, 8=2 races, 9=3+ races
model_dt[, RACE_cat := fcase(
RACE == 1, "White",
RACE == 2, "Black",
RACE %in% c(4, 5, 6), "Asian/PI",
default = "Other"
)]
model_dt[, RACE_cat := factor(RACE_cat,
levels = c("White", "Black", "Asian/PI", "Other"))]
# Collapse EDUC to 4 categories (uses the coarse 0-11 EDUC variable)
model_dt[, EDUC_cat := fcase(
EDUC <= 5, "<HS",
EDUC == 6, "HS",
EDUC %in% c(7, 8, 9), "Some college",
EDUC >= 10, "BA+",
default = NA_character_
)]
model_dt[, EDUC_cat := factor(EDUC_cat,
levels = c("<HS", "HS", "Some college", "BA+"))]
# MARST and HISPAN as factors
model_dt[, MARST := factor(MARST)]
model_dt[, HISPAN := factor(HISPAN)]
model_dt <- model_dt[!is.na(EDUC_cat)]
7. Headline logit: OWNERSHP ~ FAI × GENERATION + controls
The model uses person-weights (PERWT) for point estimation and
year-clustered robust standard errors (via sandwich::vcovCL) for
inference.
# Normalize PERWT to have mean 1.
model_dt[, w := PERWT / mean(PERWT)]
# Note: m1 uses the PCA-based FAI which requires both burden_ratio and
# price_to_income_hh to be non-NA.
m1_dt <- model_dt[!is.na(FAI)]
m1 <- glm(
own ~ FAI * GENERATION + MARST + RACE_cat + HISPAN + EDUC_cat,
data = m1_dt,
weights = w,
family = binomial(link = "logit")
)
vcov_cluster <- vcovCL(m1, cluster = ~YEAR, type = "HC1")
coef_table <- coeftest(m1, vcov. = vcov_cluster)
print(coef_table)
##
## z test of coefficients:
##
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -2.8283604 0.6982524 -4.0506 5.108e-05 ***
## FAI 0.1208920 0.0043321 27.9060 < 2.2e-16 ***
## GENERATIONGenX 0.0555485 0.6743806 0.0824 0.9343527
## GENERATIONMillennial -0.8699802 0.6801624 -1.2791 0.2008699
## GENERATIONGenZ -1.2657461 0.6693103 -1.8911 0.0586083 .
## MARST2 -0.1359686 0.1452338 -0.9362 0.3491675
## MARST3 -0.9564324 0.0979400 -9.7655 < 2.2e-16 ***
## MARST4 -0.8481850 0.0933000 -9.0909 < 2.2e-16 ***
## MARST5 -0.3544176 0.0940607 -3.7680 0.0001646 ***
## MARST6 -0.4528536 0.0795644 -5.6917 1.258e-08 ***
## RACE_catBlack -0.6929560 0.0987096 -7.0201 2.216e-12 ***
## RACE_catAsian/PI -0.4706524 0.0478319 -9.8397 < 2.2e-16 ***
## RACE_catOther -0.0358675 0.0605424 -0.5924 0.5535587
## HISPAN1 -0.8895947 0.0824093 -10.7948 < 2.2e-16 ***
## HISPAN2 -0.5884271 0.0947675 -6.2092 5.327e-10 ***
## HISPAN3 -1.1087214 0.1494737 -7.4175 1.193e-13 ***
## HISPAN4 -0.6116646 0.0628639 -9.7300 < 2.2e-16 ***
## EDUC_catHS 0.6503610 0.0542101 11.9970 < 2.2e-16 ***
## EDUC_catSome college 0.7465290 0.1013269 7.3675 1.738e-13 ***
## EDUC_catBA+ 0.7622140 0.1170553 6.5116 7.437e-11 ***
## FAI:GENERATIONGenX 0.0229925 0.0271469 0.8470 0.3970128
## FAI:GENERATIONMillennial -0.0952060 0.0495867 -1.9200 0.0548588 .
## FAI:GENERATIONGenZ -0.3536628 0.0813378 -4.3481 1.373e-05 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
# Pseudo-R^2 (McFadden) and N
ll_full <- logLik(m1)
m0 <- glm(own ~ 1, data = m1_dt, weights = w,
family = binomial(link = "logit"))
ll_null <- logLik(m0)
mcfadden_r2 <- as.numeric(1 - ll_full / ll_null)
cat("N (observations) :", nrow(m1_dt), "\n")
## N (observations) : 1398851
cat("McFadden's R^2 :", round(mcfadden_r2, 4), "\n")
## McFadden's R^2 : 0.056
7b. Robustness: P/I-only FAI (load-bearing spec)
The headline m1 uses an FAI built from both burden_ratio and
price_to_income_hh. Because burden_ratio is computed with different
formulas for owners and renters, it is mechanically related to tenure
status, which is why m1 produced a McFadden R² of 0.75 This chunk fits an alternative model m1_pi using a P/I-only FAI, where P/I is identically defined for owners (own home) and renters (year-median
home value) and is therefore not mechanically tied to tenure.
# P/I-only FAI: standardized price_to_income_hh (no PCA needed for one variable;
model_dt[, FAI_pi := scale(price_to_income_hh)[, 1]]
m1_pi <- glm(
own ~ FAI_pi * GENERATION + MARST + RACE_cat + HISPAN + EDUC_cat,
data = model_dt[!is.na(FAI_pi)],
weights = w,
family = binomial(link = "logit")
)
vcov_cluster_pi <- vcovCL(m1_pi, cluster = ~YEAR, type = "HC1")
coef_table_pi <- coeftest(m1_pi, vcov. = vcov_cluster_pi)
print(coef_table_pi)
##
## z test of coefficients:
##
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) 0.107520 0.071291 1.5082 0.1315102
## FAI_pi -0.895731 0.163472 -5.4794 4.268e-08 ***
## GENERATIONGenX 0.137791 0.109311 1.2605 0.2074717
## GENERATIONMillennial 0.208291 0.104417 1.9948 0.0460639 *
## GENERATIONGenZ 0.165904 0.075999 2.1830 0.0290386 *
## MARST2 -0.295277 0.076904 -3.8395 0.0001233 ***
## MARST3 -0.733821 0.056679 -12.9469 < 2.2e-16 ***
## MARST4 -0.652322 0.071613 -9.1089 < 2.2e-16 ***
## MARST5 -0.114145 0.041455 -2.7535 0.0058961 **
## MARST6 -0.622398 0.055160 -11.2835 < 2.2e-16 ***
## RACE_catBlack -0.640652 0.013057 -49.0640 < 2.2e-16 ***
## RACE_catAsian/PI -0.408098 0.031961 -12.7686 < 2.2e-16 ***
## RACE_catOther -0.150248 0.040425 -3.7167 0.0002018 ***
## HISPAN1 -0.334442 0.045407 -7.3654 1.766e-13 ***
## HISPAN2 -0.646181 0.087119 -7.4172 1.196e-13 ***
## HISPAN3 -0.222617 0.053045 -4.1967 2.708e-05 ***
## HISPAN4 -0.641774 0.027173 -23.6183 < 2.2e-16 ***
## EDUC_catHS 0.449925 0.016369 27.4863 < 2.2e-16 ***
## EDUC_catSome college 0.466115 0.025605 18.2039 < 2.2e-16 ***
## EDUC_catBA+ 0.386295 0.037756 10.2313 < 2.2e-16 ***
## FAI_pi:GENERATIONGenX 0.143127 0.193051 0.7414 0.4584540
## FAI_pi:GENERATIONMillennial 0.367830 0.173869 2.1156 0.0343822 *
## FAI_pi:GENERATIONGenZ 0.302454 0.163888 1.8455 0.0649665 .
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
ll_full_pi <- logLik(m1_pi)
m0_pi <- glm(own ~ 1, data = model_dt[!is.na(FAI_pi)], weights = w,
family = binomial(link = "logit"))
ll_null_pi <- logLik(m0_pi)
mcfadden_r2_pi <- as.numeric(1 - ll_full_pi / ll_null_pi)
cat("N (P/I-only model) :", nrow(model_dt[!is.na(FAI_pi)]), "\n")
## N (P/I-only model) : 3150831
cat("McFadden's R^2 (P/I-only) :", round(mcfadden_r2_pi, 4), "\n")
## McFadden's R^2 (P/I-only) : 0.0843
cat("McFadden's R^2 (full FAI) :", round(mcfadden_r2, 4), "\n")
## McFadden's R^2 (full FAI) : 0.056
# cohort slope comparison
fai_main <- coef(m1)["FAI"]
fai_main_pi <- coef(m1_pi)["FAI_pi"]
cat("FAI slope by cohort (full FAI / P/I-only FAI):\n")
## FAI slope by cohort (full FAI / P/I-only FAI):
cat(sprintf(" Boomer : %6.2f / %6.2f\n",
fai_main,
fai_main_pi))
## Boomer : 0.12 / -0.90
for (g in c("GenX", "Millennial", "GenZ")) {
i_full <- coef(m1)[paste0("FAI:GENERATION", g)]
i_pi <- coef(m1_pi)[paste0("FAI_pi:GENERATION", g)]
cat(sprintf(" %-10s : %6.2f / %6.2f\n",
g,
fai_main + i_full,
fai_main_pi + i_pi))
}
## GenX : 0.14 / -0.75
## Millennial : 0.03 / -0.53
## GenZ : -0.23 / -0.59
7c. Parallel outcome: cost burden
This model swaps the outcome from own to cost_burdened (= 1 if a
household pays more than 30% of household income on housing, the JCHS
threshold), keeping the same right-hand side. The question is
different: not “who owns?” but “who is squeezed?”. Because burden_ratio
is now part of the outcome rather than the predictor, the mechanical
relationship that inflated m1’s R² is no longer an issue!!
# binary cost-burdened outcome
model_dt[, cost_burdened := as.integer(burden_ratio > 0.30)]
cat("N with cost_burdened defined:", sum(!is.na(model_dt$cost_burdened)), "\n")
## N with cost_burdened defined: 1398851
cat("Share cost-burdened (raw) :",
round(mean(model_dt$cost_burdened, na.rm = TRUE), 3), "\n")
## Share cost-burdened (raw) : 0.32
model_dt[!is.na(cost_burdened),
.(share_burdened = round(weighted.mean(cost_burdened, w), 3),
n = .N),
by = .(GENERATION, OWNERSHP)][order(GENERATION, OWNERSHP)]
## GENERATION OWNERSHP share_burdened n
## <fctr> <int> <num> <int>
## 1: Boomer 1 0.227 22521
## 2: Boomer 2 0.268 301905
## 3: GenX 1 0.300 23299
## 4: GenX 2 0.320 280887
## 5: Millennial 1 0.288 25921
## 6: Millennial 2 0.370 677061
## 7: GenZ 1 0.259 1593
## 8: GenZ 2 0.371 65664
m2 <- glm(
cost_burdened ~ FAI_pi * GENERATION + MARST + RACE_cat + HISPAN + EDUC_cat,
data = model_dt[!is.na(cost_burdened) & !is.na(FAI_pi)],
weights = w,
family = binomial(link = "logit")
)
vcov_cluster_m2 <- vcovCL(m2, cluster = ~YEAR, type = "HC1")
coef_table_m2 <- coeftest(m2, vcov. = vcov_cluster_m2)
print(coef_table_m2)
##
## z test of coefficients:
##
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -1.370632 0.254062 -5.3949 6.857e-08 ***
## FAI_pi 1.969094 0.012407 158.7098 < 2.2e-16 ***
## GENERATIONGenX 0.164747 0.256388 0.6426 0.520504
## GENERATIONMillennial 0.090600 0.238943 0.3792 0.704560
## GENERATIONGenZ -0.095727 0.222346 -0.4305 0.666809
## MARST2 0.072467 0.036790 1.9697 0.048868 *
## MARST3 0.121191 0.053492 2.2656 0.023476 *
## MARST4 0.021424 0.055073 0.3890 0.697266
## MARST5 0.166442 0.111362 1.4946 0.135017
## MARST6 -0.017376 0.022409 -0.7754 0.438100
## RACE_catBlack 0.026833 0.042047 0.6382 0.523371
## RACE_catAsian/PI 0.325594 0.034412 9.4617 < 2.2e-16 ***
## RACE_catOther 0.040716 0.023834 1.7083 0.087581 .
## HISPAN1 0.325819 0.024828 13.1228 < 2.2e-16 ***
## HISPAN2 0.293186 0.028812 10.1759 < 2.2e-16 ***
## HISPAN3 0.540179 0.052363 10.3160 < 2.2e-16 ***
## HISPAN4 0.555419 0.036645 15.1569 < 2.2e-16 ***
## EDUC_catHS 0.034019 0.042442 0.8015 0.422819
## EDUC_catSome college 0.162858 0.053036 3.0707 0.002136 **
## EDUC_catBA+ 0.150463 0.063760 2.3598 0.018284 *
## FAI_pi:GENERATIONGenX 0.242472 0.120175 2.0177 0.043627 *
## FAI_pi:GENERATIONMillennial -0.027295 0.079259 -0.3444 0.730562
## FAI_pi:GENERATIONGenZ -0.125320 0.031990 -3.9175 8.948e-05 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
ll_full_m2 <- logLik(m2)
m0_m2 <- glm(cost_burdened ~ 1,
data = model_dt[!is.na(cost_burdened) & !is.na(FAI_pi)],
weights = w, family = binomial(link = "logit"))
ll_null_m2 <- logLik(m0_m2)
mcfadden_r2_m2 <- as.numeric(1 - ll_full_m2 / ll_null_m2)
cat("N (cost-burden model) :",
sum(!is.na(model_dt$cost_burdened) & !is.na(model_dt$FAI_pi)), "\n")
## N (cost-burden model) : 1398851
cat("McFadden's R^2 (cost-burden) :", round(mcfadden_r2_m2, 4), "\n")
## McFadden's R^2 (cost-burden) : 0.2802
# Implied cohort effects on log-odds of being cost-burdened (at FAI_pi = 0,
# reference covariates) — direct read of the GENERATION main effects.
cat("Cohort effect on log-odds(cost-burdened) vs. Boomer reference, FAI=0:\n")
## Cohort effect on log-odds(cost-burdened) vs. Boomer reference, FAI=0:
for (g in c("GenX", "Millennial", "GenZ")) {
est <- coef(m2)[paste0("GENERATION", g)]
cat(sprintf(" %-10s : %+.3f\n", g, est))
}
## GenX : +0.165
## Millennial : +0.091
## GenZ : -0.096
8. Predicted probabilities by generation across FAI range
fai_grid <- seq(quantile(model_dt$FAI_pi, 0.05, na.rm = TRUE),
quantile(model_dt$FAI_pi, 0.95, na.rm = TRUE),
length.out = 50)
modal <- function(x) names(sort(table(x), decreasing = TRUE))[1]
pred_grid <- expand.grid(
FAI_pi = fai_grid,
GENERATION = levels(model_dt$GENERATION),
MARST = modal(model_dt$MARST),
RACE_cat = modal(model_dt$RACE_cat),
HISPAN = modal(model_dt$HISPAN),
EDUC_cat = modal(model_dt$EDUC_cat)
)
pred_grid$GENERATION <- factor(pred_grid$GENERATION,
levels = levels(model_dt$GENERATION))
pred_grid$MARST <- factor(pred_grid$MARST, levels = levels(model_dt$MARST))
pred_grid$RACE_cat <- factor(pred_grid$RACE_cat, levels = levels(model_dt$RACE_cat))
pred_grid$HISPAN <- factor(pred_grid$HISPAN, levels = levels(model_dt$HISPAN))
pred_grid$EDUC_cat <- factor(pred_grid$EDUC_cat, levels = levels(model_dt$EDUC_cat))
pred_grid$pred_own <- predict(m1_pi, newdata = pred_grid, type = "response")
ggplot(pred_grid, aes(x = FAI_pi, y = pred_own, color = GENERATION)) +
geom_line(size = 1) +
scale_y_continuous(labels = scales::percent, limits = c(0, 1)) +
labs(
title = "Predicted homeownership probability by affordability and generation",
subtitle = "Holding marital status, race, ethnicity, and education at modal values",
x = "Affordability index (z-scored price-to-income; higher = less attainable)",
y = "Predicted Pr(own home)",
color = "Generation"
) +
theme_minimal()
8. save cleaned data and fitted models
out_dir <- "~/Desktop/DA6833/FinancialNihilism"
saveRDS(model_dt, file.path(out_dir, "model_dt.rds"))
saveRDS(m1_pi, file.path(out_dir, "m1_pi.rds"))
saveRDS(m2, file.path(out_dir, "m2.rds"))
cat("Saved:\n")
## Saved:
cat(" model_dt.rds (rows = ", nrow(model_dt), ")\n", sep = "")
## model_dt.rds (rows = 3150831)
cat(" m1_pi.rds (homeownership model, P/I-only FAI)\n")
## m1_pi.rds (homeownership model, P/I-only FAI)
cat(" m2.rds (cost-burden model)\n")
## m2.rds (cost-burden model)