DATA 606 Final Project: The Hemline Theory

1 Abstract

This project evaluates the long-standing “hemline theory,” which claims that skirt lengths move with the economy - shorter hemlines during economic booms and longer hemlines during downturns. To test whether this idea has any validity, I combined Google Trends search data for “mini skirt,” “midi skirt,” and “maxi skirt” with U.S. quarterly economic indicators, including real GDP growth and retail clothing sales from 2004–2025. All datasets were aggregated to the quarterly level and merged, with COVID-disrupted periods flagged to prevent distortion.

Through exploratory data analysis, correlation heatmaps, and two regression models (simple and multiple regression), I examined whether interest in different skirt lengths shows any relationship with economic performance. Across all approaches, results consistently showed no meaningful relationship between hemline trends and GDP growth. Neither the simple regression nor the multiple regression (which controlled for retail clothing sales) produced significant effects, and both models explained virtually none of the variance in GDP growth. Diagnostic plots confirmed that the models themselves behaved well; the problem is simply that skirt-length interest does not predict economic outcomes. Overall, this analysis provides no evidence in support of the hemline theory.

2 Introduction

The hemline theory is the theory that women’s skirt hemlines rise and fall with the economy (i.e. short skirts being more popular during a boom, and longer skirts being more popular in a downturn). I’ve see miscellaneous references to this older (about 100ish years old) theory resurfacing recently in social media, questioning if trending maxi skirts are a recession indicator.

The below image visually describes the Hemline Theory, and was referenced in a LinkedIn post this past year (2025).

The Hemline Theory. Source: Mages Studio Private Limited

TikTok has a whole slew of posts on it casually discussing the hemline theory. Blogs have tried to tackle the index theory.

I’m skeptical, so my goal is to analyze if the hemline theory is true or not.

2.1 Research Question

Do changes in skirt-length fashion interest (measured by Google searches for mini, midi or maxi skirts) correspond with changes in U.S. economic conditions? Specifically, do these fashion-trend signals relate to quarterly GDP growth and to consumer spending on clothing and apparel?

3 Reproducability Requirements

• To reproduce this code, you will need to get a FRED API key and store it using REnviron, referred to as “FRED_API_KEY”.

4 Data Acquisition

For this analysis, I have gathered data from the following sources:

• Google Trends: gtrendsR package (https://cran.r-project.org/package=gtrendsR)
• Census.gov: Monthly Retail Clothing Trade (US) data is available for download as a csv.(https://www.census.gov/econ/currentdata/)
• FRED real GDP: fredr package (https://fred.stlouisfed.org/series/GDPC1)

# File Path
if (interactive()) {
  projPath <- dirname(file.path(getSourceEditorContext()$path))   # getting where .Rmd is located
} else {
  projPath <- "."  # default for knitting
}

# FRED API Key
fred_api_key <- Sys.getenv("FRED_API_KEY")
fredr_set_key(fred_api_key)

4.0.1 Google Trends Skirt + Clothing Data: API

From Google Trends, I am obtaining data from 2004 to start of 2025 on how often “mini skirt”, “midi skirt” and “maxi skirt” trended in searches. I’m doing this relative to overall searches for “clothing” in order to properly assess skirt length sentiment while controlling for waxing/waning overall clothing interest.

# set google searches of interest
keywords <- c("mini skirt", "midi skirt", "maxi skirt", "clothing")

# google api call 
# added logic here to pull the data from csv if knitting. This is due to a gtrends api errror where sometimes it errors out on a new session, and I don't want to risk knitting failures.
if (interactive()) {
  # Live API call
  gtr <- gtrends(
    keyword = keywords,
    geo = "US",
    time = "2004-01-01 2025-01-01",
    gprop = "web",
    hl = "en-US"
  )
  
  trend_df <- gtr$interest_over_time %>% 
    select(date, keyword, hits)

  write_csv(trend_df, "trend_data_with_clothing_2004.csv")
  
} else {
  # KNITTING → read cached version
  # trend_df <- read_csv("file.path(projPath, "trend_data_with_clothing_2004.csv")) # reading from local saved from interactive session
  trend_df <- read_csv("https://raw.githubusercontent.com/cdube89128/DATA-607/refs/heads/main/Final_Project/trend_data_with_clothing_2004.csv") # reading from github
} 

## Rows: 1012 Columns: 3
## ── Column specification ────────────────────────────────────────────────────────
## Delimiter: ","
## chr  (2): keyword, hits
## dttm (1): date
## 
## ℹ Use `spec()` to retrieve the full column specification for this data.
## ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.

4.0.2 Census.gov Monthly Retail Clothing Trade (US) Data: CSV

From census.gov, I am obtaining data from 1992 to 2025 on retail clothing sales in the US.

• Source: https://www.census.gov/econ/currentdata/
• Downloaded as 2 csvs and stored on Github

# URLS
url_nsa <- "https://raw.githubusercontent.com/cdube89128/DATA-607/refs/heads/main/Final_Project/Not%20Seasonally%20Adjusted%20Clothing%20Sales.csv"
url_sa  <- "https://raw.githubusercontent.com/cdube89128/DATA-607/refs/heads/main/Final_Project/Seasonally%20Adjusted%20Clothing%20Sales.csv"

# Read in Retail Clothing Trade data
clothing_nsa <- read_csv(url_nsa, skip = 7, show_col_types = FALSE)
clothing_sa  <- read_csv(url_sa,  skip = 7, show_col_types = FALSE) 

4.0.3 Economic data from FRED: API

From the FRED API, I am getting Gross Domestic Product data from 2004 to start of 2025.

# Using real GDP (series “GDPC1”) = Real Gross Domestic Product, 
gdp <- fredr(
  series_id = "GDPC1",
  observation_start = as.Date("2004-01-01"),
  observation_end = as.Date("2025-01-01"),
  frequency = "q"  # quarterly
)

# get quarter over quarter growth rate
gdp <- gdp %>%
  arrange(date) %>%
  mutate(
    gdp_growth = (value / lag(value) - 1) * 100
  ) %>%
  filter(!is.na(gdp_growth))

5 Data Wrangling and Transformations

• Enriching, grouping, cleaning up date formatting to allow all data to be incorporated together
• GDP data is quarterly, so all of the data will need to be changed to the same quarterly time periods

5.0.1 Cleaning and Transformation: Retail Data

• Clean data formatting to read dates
• Convert from monthly to quarterly

# Retail data
# Convert monthly to quarterly
# Clean data formatting

# Convert NSA dataset
clothing_nsa_clean <- clothing_nsa %>%
  mutate( 
    date = parse_date_time(Period, orders = c("b-y", "b-Y")),
    sales_nsa = as.numeric(Value)
  ) %>%
  select(date, sales_nsa)

# Convert SA dataset
clothing_sa_clean <- clothing_sa %>%
  mutate(
    date = parse_date_time(Period, orders = c("b-y", "b-Y")),
    sales_sa = as.numeric(Value)
  ) %>%
  select(date, sales_sa)

# Aggregate nsa to quarterly
clothing_nsa_quarterly <- clothing_nsa_clean %>%
  mutate(
    year = year(date),
    quarter = quarter(date)
  ) %>%
  group_by(year, quarter) %>%
  summarise(
    clothing_sales_nsa = mean(sales_nsa, na.rm = TRUE),
    .groups = "drop"
  )

# Aggregate sa to quarterly
clothing_sa_quarterly <- clothing_sa_clean %>%
  mutate(
    year = year(date),
    quarter = quarter(date)
  ) %>%
  group_by(year, quarter) %>%
  summarise(
    clothing_sales_sa = mean(sales_sa, na.rm = TRUE),
    .groups = "drop"
  )

5.0.2 Cleaning and Tranformation: Google Trends Data

• Treat Google Trends data where search hits are “<1” to be handled numerically
• Convert from monthly to quarterly
• Create important clothing-normalized variables to represent search interest for skirts and skirt lengths relative to overall clothing interest
• Create important variable hemline_index as a numeric value ranging 1:3.
  • Closer to 1 = preferred mini skirts in Google data
  • Closer to 3 = preferred maxi skirts in Google data

trend_df <- trend_df %>%
  mutate(
    hits = case_when(
      hits == "<1" ~ 0.49,
      .default  = as.numeric(hits)
    )
  )

## Warning: There was 1 warning in `mutate()`.
## ℹ In argument: `hits = case_when(hits == "<1" ~ 0.49, .default =
##   as.numeric(hits))`.
## Caused by warning in `vec_case_when()`:
## ! NAs introduced by coercion

# Make the google trends data wider 
trend_wide <- trend_df %>%
  mutate(
    keyword = case_when(
      keyword == "mini skirt" ~ "mini",
      keyword == "midi skirt" ~ "midi",
      keyword == "maxi skirt" ~ "maxi",
      keyword == "clothing"   ~ "clothing",
      TRUE ~ keyword
    )
  ) %>%
  pivot_wider(names_from = keyword, values_from = hits) %>%
  mutate(
    total_skirt = mini + midi + maxi,

    # clothing normalization
    mini_to_clothing = mini / clothing,
    midi_to_clothing = midi / clothing,
    maxi_to_clothing = maxi / clothing,
    skirt_to_clothing = total_skirt / clothing
  )

# Quarterly Google Trends Dataset
# because gdp is quarterly

trend_quarterly <- trend_wide %>%
  mutate(
    year = year(date),
    quarter = quarter(date)
  ) %>%
  group_by(year, quarter) %>%
  summarise(
    
    # clothing-normalized variables
    mini_to_clothing = mean(mini_to_clothing, na.rm = TRUE),
    midi_to_clothing = mean(midi_to_clothing, na.rm = TRUE),
    maxi_to_clothing = mean(maxi_to_clothing, na.rm = TRUE),
    skirt_to_clothing = mean(skirt_to_clothing, na.rm = TRUE),
    
    # pure hits data
    mini     = mean(mini, na.rm = TRUE),
    midi     = mean(midi, na.rm = TRUE),
    maxi     = mean(maxi, na.rm = TRUE),
    clothing = mean(clothing, na.rm = TRUE),
    
    # hemline index
    hemline_index = (1*mini + 2*midi + 3*maxi) / (mini + midi + maxi),
    
    .groups = "drop"
  ) %>%
  mutate(
    date = as_date(paste(year, (quarter - 1) * 3 + 1, "01", sep = "-"))
  )

5.0.3 Merging Data

• Merge the 3 datasets into 1 clean data set

# Merging All Quarterly Datasets

# GDP is already quarterly, just selecting the columns I want
gdp_quarterly <- gdp %>%
  mutate(
    gdp_value  = value, 
    year = year(date),
    quarter = quarter(date)
  ) %>%
  select(date, gdp_value, gdp_growth, year, quarter)

# Merge everything
df_full <- gdp_quarterly %>%
  inner_join(trend_quarterly, by = c("year", "quarter")) %>%
  left_join(clothing_sa_quarterly,  by = c("year", "quarter")) %>%
  left_join(clothing_nsa_quarterly, by = c("year", "quarter"))

# Clean the duplicate date column
df_full <- df_full %>%
  rename(date = date.x) %>%
  select(-date.y)

df_full %>%
  head(10) %>%
  kable(format = "html", caption = "Preview of Merged Quarterly Dataset (df_full)") %>%
  kable_styling(full_width = FALSE, bootstrap_options = c("striped", "hover"))

Preview of Merged Quarterly Dataset (df_full)

date	gdp_value	gdp_growth	year	quarter	mini_to_clothing	midi_to_clothing	maxi_to_clothing	skirt_to_clothing	mini	midi	maxi	clothing	hemline_index	clothing_sales_sa	clothing_sales_nsa
2004-04-01	15366.85	0.7749523	2004	2	0.0126455	0	0	0.0126455	1.00	0	0	79.33333	1	11011.67	10559.67
2004-07-01	15512.62	0.9485939	2004	3	0.0121570	0	0	0.0121570	1.00	0	0	82.33333	1	11113.33	10684.33
2004-10-01	15670.88	1.0202081	2004	4	0.0107054	0	0	0.0107054	1.00	0	0	93.66667	1	11419.00	14182.00
2005-01-01	15844.73	1.1093634	2005	1	0.0122210	0	0	0.0122210	1.00	0	0	82.00000	1	11603.67	9836.00
2005-04-01	15922.78	0.4926245	2005	2	0.0121210	0	0	0.0121210	1.00	0	0	82.66667	1	11822.67	11262.00
2005-07-01	16047.59	0.7838140	2005	3	0.0121133	0	0	0.0121133	1.00	0	0	82.66667	1	11748.00	11276.00
2005-10-01	16136.73	0.5555165	2005	4	0.0092750	0	0	0.0092750	0.83	0	0	89.66667	1	12173.33	15116.67
2006-01-01	16353.83	1.3453838	2006	1	0.0089527	0	0	0.0089527	0.66	0	0	75.33333	1	12233.00	10258.67
2006-04-01	16396.15	0.2587528	2006	2	0.0115985	0	0	0.0115985	0.83	0	0	72.00000	1	12486.67	12041.67
2006-07-01	16420.74	0.1499559	2006	3	0.0133446	0	0	0.0133446	1.00	0	0	75.00000	1	12689.00	12242.00

6 Exploratory Data Analysis

• Exploring the data to gauge what form the final analysis should take.

6.1 Time-Series Plots

Beginning EDA with time-series plots of a number of variables of interest.

# GDP Growth over Time
ggplot(df_full, aes(x = date, y = gdp_growth)) +
  geom_line(color = "steelblue", linewidth = 1) +
  labs(
    title = "Quarterly U.S. GDP Growth (2015–2025)",
    x = "Quarter",
    y = "GDP Growth (%)"
  ) +
  theme_minimal()

# GDP Value over Time
ggplot(df_full, aes(x = date, y = gdp_value)) +
  geom_line(color = "darkblue", linewidth = 1) +
  labs(
    title = "Quarterly U.S. GDP Value (2015–2025)",
    x = "Quarter",
    y = "GDP Value"
  ) +
  theme_minimal()

# Retail Sales over Time (adjusted)
ggplot(df_full, aes(x = date, y = clothing_sales_sa)) +
  geom_line(color = "darkgreen", linewidth = 1) +
  labs(
    title = "Quarterly U.S. Clothing Retail Sales (Seasonally Adjusted)",
    x = "Quarter",
    y = "Sales (Millions USD)"
  ) +
  theme_minimal()

# Clothing Sales & Google Search Interest over Time 
ggplot(df_full, aes(x = date)) +
  # For legend purposes: mapping each line to color group
  geom_line(aes(y = clothing_sales_nsa, color = "Clothing Sales (NSA)"), linewidth = 1) +
  geom_line(aes(y = clothing * 200, color = "Clothing Search Interest"), linewidth = 1) +
  # legend
  scale_color_manual(values = c(
    "Clothing Sales (NSA)" = "darkgreen",
    "Clothing Search Interest" = "violet"
  )) +
  scale_y_continuous(
    name = "Clothing Sales (Millions USD)",
    sec.axis = sec_axis(~ . / 200,
                        name = "Clothing Search Interest")
  ) +
  labs(
    title = "Clothing Sales vs Google Search Interest",
    x = "Quarter",
    color = "Legend"
  ) +
  theme_minimal()

# Skirt to Clothing Ratio of Interest over Time
ggplot(df_full, aes(x = date, y = skirt_to_clothing)) +
  geom_line(color = "purple", linewidth = 1) +
  labs(
    title = "Relative Search Interest: Skirt Searches vs. Clothing Searches",
    x = "Quarter",
    y = "Skirt-to-Clothing Ratio"
  ) +
  theme_minimal()

# Skirt Length Interest Relative to Clothing Over Time
df_long_norm <- df_full %>%
  select(date, mini_to_clothing, midi_to_clothing, maxi_to_clothing) %>%
  pivot_longer(-date, names_to = "variable", values_to = "value")

ggplot(df_long_norm, aes(x = date, y = value, color = variable)) +
  geom_line(linewidth = 1) +
  scale_color_manual(values = c("mini_to_clothing" = "hotpink",
                                "midi_to_clothing" = "forestgreen",
                                "maxi_to_clothing" = "navy")) +
  labs(
    title = "Mini vs Midi vs Maxi Skirt Search Interest (Relative to Clothing)",
    x = "Quarter",
    y = "Normalized Search Interest",
    color = "Trend"
  ) +
  theme_minimal()

# Search Interst (Raw values) Over Time
df_long_raw <- df_full %>%
  select(date, mini, midi, maxi) %>%
  pivot_longer(-date, names_to = "variable", values_to = "value")

ggplot(df_long_raw, aes(x = date, y = value, color = variable)) +
  geom_line(linewidth = 1) +
  scale_color_manual(
    values = c(
      mini = "hotpink",
      midi = "forestgreen",
      maxi = "navy"
    )
  ) +
  labs(
    title = "Raw Google Search Interest: Clothing vs Skirt Types",
    x = "Quarter",
    y = "Search Index",
    color = "Search Term"
  ) +
  theme_minimal()

# Hemline Index over Time
ggplot(df_full, aes(x = date, y = hemline_index)) +
  geom_line(color = "pink", linewidth = 1) +
  labs(
    title = "Hemline Index over Time",
    subtitle = "(1:3 where 1 is the most mini and 3 is the most maxi)",
    x = "Quarter",
    y = "Hemline Index "
  ) +
  theme_minimal()

So far, I can see that:

• GDP value, GDP growth and retail sales were heavily impacted by COVID.
• Among Google searches for skirts, mini skirts dominated in 2005, starting losing out to maxi skirts around 2013, then start taking back popularity share in 2023.
• In general, searches for skirts had local peaks around 2014, and going into 2025 (and potentially still climbing)

6.1.1 COVID flag

Based on the EDA so far, adding in this flag to indicate where COVID happenings heavily impacted the data. This way, COVID impacted data can be filtered out.

df_full <- df_full %>%
  mutate(
    covid_disrupted = case_when(
      # COVID disruption period
      (year == 2020 & quarter %in% c(1, 2, 3, 4)) |
      (year == 2021 & quarter %in% c(1, 2)) ~ 1,
      
      TRUE ~ 0
    )
  )

6.2 Scatter Plots

Continuing EDA with a number of scatterplots of key relationships. If the Hemline Theory is true, some relationships should be visible here between hemline and economic markers.

# Helper function for consistent formatting
plot_scatter <- function(df, xvar, yvar, xlab, ylab, title) {
  ggplot(df, aes(x = .data[[xvar]], y = .data[[yvar]])) +
    geom_point(color = "darkgray", alpha = 0.8, size = 2) +
    geom_smooth(method = "loess", color = "steelblue", linewidth = 1.1, se = FALSE) +
    theme_minimal() +
    labs(title = title, x = xlab, y = ylab)
}


# 1. Skirt-to-Clothing vs GDP Growth
plot_scatter(
  df_full %>% filter(covid_disrupted == 0),
  "skirt_to_clothing", "gdp_growth",
  "Skirt-to-Clothing Ratio",
  "GDP Growth (%)",
  "GDP Growth vs Skirt-to-Clothing Search Interest"
)

## `geom_smooth()` using formula = 'y ~ x'

# 2. Mini-to-Clothing vs GDP Growth
plot_scatter(
  df_full %>% filter(covid_disrupted == 0),
  "mini_to_clothing", "gdp_growth",
  "Mini-to-Clothing Search Ratio",
  "GDP Growth (%)",
  "GDP Growth vs Mini Skirt Search Interest"
)

## `geom_smooth()` using formula = 'y ~ x'

# 3. Maxi-to-Clothing vs GDP Growth
plot_scatter(
  df_full %>% filter(covid_disrupted == 0),
  "maxi_to_clothing", "gdp_growth",
  "Maxi-to-Clothing Search Ratio",
  "GDP Growth (%)",
  "GDP Growth vs Maxi Skirt Search Interest"
)

## `geom_smooth()` using formula = 'y ~ x'

# 4. Clothing Sales vs GDP Growth
plot_scatter(
  df_full %>% filter(covid_disrupted == 0),
  "clothing_sales_sa", "gdp_growth",
  "Clothing Retail Sales (SA, millions USD)",
  "GDP Growth (%)",
  "GDP Growth vs Clothing Retail Sales"
)

## `geom_smooth()` using formula = 'y ~ x'

# 5. Clothing Sales vs Skirt-to-Clothing
plot_scatter(
  df_full %>% filter(covid_disrupted == 0),
  "clothing_sales_nsa", "skirt_to_clothing",
  "Clothing Retail Sales (SA, millions USD)",
  "Skirt-to-Clothing Ratio",
  "Skirt Interest vs Clothing Retail Sales"
)

## `geom_smooth()` using formula = 'y ~ x'

# 6. Hemline Index vs GDP Growth
plot_scatter(
  df_full %>% filter(covid_disrupted == 0),
  "hemline_index", "gdp_growth",
  "Hemline Index",
  "GDP Growth (%)",
  "GDP Growth vs hemline Index"
)

## `geom_smooth()` using formula = 'y ~ x'

# 7. Hemline Index vs Clothing Sales (not seasonally normalized)
plot_scatter(
  df_full %>% filter(covid_disrupted == 0),
  "hemline_index", "clothing_sales_nsa",
  "Hemline Index",
  "Clothing Retail Sales (SA, millions USD)",
  "Hemline Index vs Clothing Retail Sales"
)

## `geom_smooth()` using formula = 'y ~ x'

So far, this is not looking good for the Hemline Theory. There are no strong visible trends between hemline and any economic markers in any of the above scatter plots. Skirt Interest vs Clothing retail sales is the only scatter plot which shows a clear trend - which is moreso a dummy check and looks exactly as expected.

6.3 Correlation Heatmap

Final aspect of EDA here are correlation matrices.

# Adding in some lagged data first
df_full <- df_full %>%
  mutate(
    gdp_growth_lead1 = lead(gdp_growth, 1),
    gdp_value_lead1  = lead(gdp_value, 1)
  )

# Select only continuous numeric variables for correlation
corr_vars <- df_full %>%
  select(
    gdp_growth,
    gdp_growth_lead1,
    gdp_value,
    gdp_value_lead1,
    clothing_sales_sa,
    clothing_sales_nsa,
    mini_to_clothing,
    midi_to_clothing,
    maxi_to_clothing,
    skirt_to_clothing,
    mini,
    midi,
    maxi,
    clothing,
    hemline_index
  )

# Correlation matrix
corr_matrix <- cor(corr_vars, use = "pairwise.complete.obs")

# Convert to long format for ggplot heatmap shtuffs
corr_long <- as.data.frame(as.table(corr_matrix))
colnames(corr_long) <- c("Var1", "Var2", "Correlation")

# Heatmap visual for correlation plot
ggplot(corr_long, aes(x = Var1, y = Var2, fill = Correlation)) +
  geom_tile(color = "white") +
  geom_text(aes(label = sprintf("%.2f", Correlation)), size = 3) +
  scale_fill_gradient2(
    low = "navy",
    mid = "white",
    high = "darkred",
    midpoint = 0,
    limits = c(-1, 1)
  ) +
  theme_minimal() +
  theme(
    axis.text.x = element_text(angle = 45, hjust = 1),
    panel.grid = element_blank()
  ) +
  labs(
    title = "Correlation Heatmap of Key Variables",
    fill = "Correlation"
  )

## Same ish thing but without the outlier covid data 

# Filter out COVID-disrupted quarters
df_no_covid <- df_full %>% 
  filter(covid_disrupted == 0)

# Select numeric variables
corr_vars_nc <- df_no_covid %>%
  select(
    gdp_growth,
    gdp_growth_lead1,
    gdp_value,
    gdp_value_lead1,
    clothing_sales_sa,
    clothing_sales_nsa,
    mini_to_clothing,
    midi_to_clothing,
    maxi_to_clothing,
    skirt_to_clothing,
    mini,
    midi,
    maxi,
    clothing,
    hemline_index
  )

# correlation matrix
corr_matrix_nc <- cor(corr_vars_nc, use = "pairwise.complete.obs")

# Convert to long format for ggplot
corr_long_nc <- as.data.frame(as.table(corr_matrix_nc))
colnames(corr_long_nc) <- c("Var1", "Var2", "Correlation")

# Heatmap visual
ggplot(corr_long_nc, aes(x = Var1, y = Var2, fill = Correlation)) +
  geom_tile(color = "white") +
  geom_text(aes(label = sprintf("%.2f", Correlation)), size = 3) +
  scale_fill_gradient2(
    low = "navy", mid = "white", high = "darkred",
    midpoint = 0, limits = c(-1, 1)
  ) +
  labs(
    title = "Correlation Heatmap (COVID-Free Quarters Only)",
    fill = "Correlation"
  ) +
  theme_minimal() +
  theme(
    axis.text.x = element_text(angle = 45, hjust = 1),
    panel.grid = element_blank()
  )

The correlation matrices double down on this not looking good for the hemline theory. When performed both with and without COVID era data included, gdp_growth has no significant correlations with any other variables. The measure of gdp_value does show positive correlation with other variables, but based on what has been seen so far, this is more likely due to the nature of gdp, as something that will drift up in value over time, and many other measures such as the number of searches for skirts have also drifted up over time. A notable contrast to this is the clothing variable, which actually did drift down over time (as seen earlier in the EDA section).

7 Data (Statistical) Analysis

I am opting to try two models here, one linear regression, and one multiple regression, to see if hemline interest is able to predict any trend in the economy.

Null Hypothesis (H₀):There is no relationship between hemline_index and GDP growth.

Alternative Hypothesis (Hₐ): There is a relationship between hemline_index and GDP growth.

7.1 Model 1: Linear Regression

First, I am fitting a simple linear regression to evaluate the direct relationship between hemline interest and GDP growth. This establishes a baseline test of the core hemline theory.

Formally: the dependent variable is quarterly gdp_growth, and the independent variable is the hemline_index.

# Filter to non-COVID periods
df_nc <- df_full %>% filter(covid_disrupted == 0)

# Fit simple regression
model1 <- lm(gdp_growth ~ hemline_index, data = df_nc)

# Print regression summary
summary(model1)

## 
## Call:
## lm(formula = gdp_growth ~ hemline_index, data = df_nc)
## 
## Residuals:
##      Min       1Q   Median       3Q      Max 
## -2.70421 -0.22266  0.08326  0.30084  1.18031 
## 
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)
## (Intercept)    0.43331    0.28593   1.515    0.134
## hemline_index  0.05862    0.16455   0.356    0.723
## 
## Residual standard error: 0.5666 on 76 degrees of freedom
## Multiple R-squared:  0.001667,   Adjusted R-squared:  -0.01147 
## F-statistic: 0.1269 on 1 and 76 DF,  p-value: 0.7227

# Visualization: Regression line + residual cloud
ggplot(df_nc, aes(x = hemline_index, y = gdp_growth)) +
  geom_point(alpha = 0.7, color = "gray40") +
  geom_smooth(method = "lm", color = "darkred", se = TRUE, linewidth = 1.1) +
  theme_minimal() +
  labs(
    title = "Model 1: GDP Growth vs Hemline Index",
    x = "Hemline Index",
    y = "GDP Growth (%)"
  )

## `geom_smooth()` using formula = 'y ~ x'

# Diagnostic Visual
par(mfrow=c(2,2))
plot(model1) # found out I can just plot the model like this instad of manually making each diagnostic plot 

Model 1: GDP Growth ~ Hemline Index

Looking at this first model:

• p = 0.72 (not significant)
• R² = 0 (0% of variance explained)
• Slope = 0.06 (When hemline index increases by 1 (e.g., shift toward maxi skirts), GDP growth increases by only0.06 percentage points.)

Looking at the Diagnostic Plots:

• The residuals versus fitted values plot appears like a random cloud; it’s good that we have constant variance.
• The Q-Q plot shows the points falling along the line, with only a bit of a tail at the top and bottom, also good and indicating constant variance of the residuals.
• The scale-location plot looks pretty much like a random cloud, again an indicator of constant variance of residuals.
• The residuals versus leverage plot doesn’t show anything with high values, so there is no data point that is overly influencing the model.
• The fitted range is pretty narrow [0.45, 0.57]. This means that predicted values barely change (they stay between 0.45 and 0.57). The model is only predicting a tiny difference in GDP growth

Thoughts:

While the diagnostic plots come back pretty clean , this model barely predicts any variance. Good diagnostic plots don’t save a model that cannot predict anything. I fail to reject the null hypothesis that there is no relationship between hemline_index and GDP growth.

7.2 Model 2: Multiple Regression

Second, I’m fitting multiple regression to assess whether hemline interest predicts GDP growth after accounting for broader consumer spending patterns. This allows me to test whether any hemline effect is present once the key economic control variable of clothing sales is included.

Formally: the dependent variable is quarterly gdp_growth, and the independent variables are the hemline_index and clothing_sales_nsa.

# Fit regression with clothing sales as control
model2 <- lm(gdp_growth ~ hemline_index + clothing_sales_nsa, data = df_nc)

# Print summary
summary(model2)

## 
## Call:
## lm(formula = gdp_growth ~ hemline_index + clothing_sales_nsa, 
##     data = df_nc)
## 
## Residuals:
##      Min       1Q   Median       3Q      Max 
## -2.73932 -0.20511  0.08042  0.33918  0.97136 
## 
## Coefficients:
##                      Estimate Std. Error t value Pr(>|t|)  
## (Intercept)        -1.874e-02  3.918e-01  -0.048   0.9620  
## hemline_index      -3.444e-02  1.720e-01  -0.200   0.8418  
## clothing_sales_nsa  4.164e-05  2.499e-05   1.666   0.0998 .
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.5601 on 75 degrees of freedom
## Multiple R-squared:  0.03731,    Adjusted R-squared:  0.01164 
## F-statistic: 1.453 on 2 and 75 DF,  p-value: 0.2403

# Visualization: Partial effect of hemline_index (added-variable plot)
avPlots(model2, id = FALSE)

# Partial-regression-plot
df_nc <- df_full %>% filter(covid_disrupted == 0)

# Residuals of GDP growth after controlling for clothing_sales_nsa
resid_gdp <- resid(lm(gdp_growth ~ clothing_sales_nsa, data = df_nc))

# Residuals of hemline_index after controlling for clothing_sales_nsa
resid_hem <- resid(lm(hemline_index ~ clothing_sales_nsa, data = df_nc))

# dataframe for ggplot
partial_df <- tibble(
  resid_gdp = resid_gdp,
  resid_hem = resid_hem
)

# Plot
ggplot(partial_df, aes(x = resid_hem, y = resid_gdp)) +
  geom_point(alpha = 0.7, color = "gray40") +
  geom_smooth(method = "lm", se = TRUE, color = "darkred", linewidth = 1.1) +
  theme_minimal() +
  labs(
    title = "Partial Regression Plot: Hemline Index Effect\n(Controlling for Clothing Sales)",
    x = "Residual Hemline Index\n(after removing clothing_sales_sa effect)",
    y = "Residual GDP Growth (%)\n(after removing clothing_sales_sa effect)"
  )

## `geom_smooth()` using formula = 'y ~ x'

# Diagnostic Visual
par(mfrow=c(2,2))
plot(model2)

Model 2: GDP Growth ~ Hemline Index + Clothing Sales (NSA)

Looking at this first model:

• p = 0.84 (not significant)
• Adj. R² = -0.01 (-1% of variance explained)

Looking at the Diagnostic Plots (this is almost the same as the first model):

• The residuals versus fitted values plot appears like a random cloud; it’s good that we have constant variance.
• The Q-Q plot shows the points falling along the line, with only a bit of a tail at the bottom, also good and indicating constant variance of the residuals.
• The scale-location plot looks pretty much like a random cloud, again an indicator of constant variance of residuals.
• The residuals versus leverage plot doesn’t show anything with high values, so there is no data point that is overly influencing the model.
• The fitted range is pretty narrow [0.3, 0.8]. This means that predicted values barely change (they stay between 0.3 and 0.8). The model is only predicting a tiny difference in GDP growth

Thoughts:

Adding clothing sales doesn’t strengthen the model. The Hemline index still shows no predictive value, even after controlling for spending. I fail to reject the null hypothesis that there is no relationship between hemline_index and GDP growth.

8 Conclusions

After going through the full analysis of exploratory data analysis, correlation checks, and two regression models, it’s clear that the hemline theory does not hold up in this dataset. While skirt-length search trends do fluctuate over time and show meaningful shifts in popularity (mini vs. maxi cycles are very visible in the early and late 2010s), none of these fashion signals show any connection to actual economic performance. GDP growth has no measurable relationship with skirt interest, whether examined directly or after controlling for broader retail consumption patterns.

Both regression models produced extremely low R² values, nonsignificant coefficients, and very narrow fitted ranges, meaning the models barely changed their predictions regardless of what the hemline index is doing. The diagnostic plots looked clean, so the issue wasn’t with model assumptions. It’s simply that the predictors offer no explanatory power. Even the correlation heatmaps reinforced this, with GDP growth essentially uncorrelated with all skirt-related variables.

In short, based on 20 years of search trends and economic data, there is no evidence that skirt lengths move with the economy. Fashion may respond to cultural cycles, aesthetics, or social trends, but it does not appear to function as an economic indicator. The hemline theory, at least amidst the modern Google-Trends-era of assessing interest, does not hold up.