Other: Interrupted TS, ARFIMA, Spectral Analysis

This section applies specialized time series techniques to address specific features of the data. Interrupted Time Series (ITS) analysis evaluates the impact of discrete interventions like the 2008 Financial Crisis on economic outcomes. ARFIMA (AutoRegressive Fractionally Integrated Moving Average) models handle long-memory processes where shocks decay slowly. Spectral analysis decomposes series into frequency components, revealing cyclical patterns. These advanced methods complement the core analyses by capturing structural breaks, persistence, and periodicities that standard models may miss.

Interrupted Time Series Analysis

Step 1: Variable Selection

We selected Self-Employment as our outcome variable for interrupted time series analysis, with the 2008 Financial Crisis (specifically September 2008, Lehman Brothers collapse) as the intervention point.

Rationale: Self-employment is an ideal candidate for ITS analysis because our exploratory analysis revealed a clear structural break around 2008-2010 where decades of steady growth abruptly stalled. The financial crisis represents a well-defined intervention with both immediate shocks (credit freeze, wealth destruction) and sustained effects (regulatory changes, risk aversion). This variable is central to our entrepreneurship pathway hypothesis - testing whether the crisis fundamentally altered entrepreneurial activity in the U.S.

Step 2: Time Series Plot with Intervention Line

Show Code
# Load self-employment data
self_employment <- read.csv('data/raw/LNS12032192.csv')

# Create date column and time series
self_emp_df <- self_employment %>%
  mutate(date = as.Date(paste0(observation_date, "-01"))) %>%
  rename(self_employed = LNS12032192) %>%
  select(date, self_employed)

# Define intervention date (September 2008 - Lehman Brothers collapse)
intervention_date <- as.Date("2008-09-01")

# Plot with intervention line
ggplot(self_emp_df, aes(x = date, y = self_employed)) +
  geom_line(color = "steelblue", linewidth = 0.8) +
  geom_vline(xintercept = as.numeric(intervention_date),
             color = "red", linetype = "dashed", linewidth = 1) +
  annotate("text", x = intervention_date, y = max(self_emp_df$self_employed) * 0.95,
           label = "2008 Financial Crisis\n(Sep 2008)", color = "red", hjust = -0.1) +
  labs(
    title = "U.S. Self-Employment Over Time",
    subtitle = "Red line indicates 2008 Financial Crisis intervention",
    x = "Year",
    y = "Self-Employed Persons (Thousands)",
    caption = "Source: Bureau of Labor Statistics (LNS12032192)"
  ) +
  theme_minimal() +
  theme(plot.title = element_text(face = "bold", size = 14))

The visual inspection reveals a clear change in trend around the intervention point. Pre-2008, self-employment showed consistent growth; post-2008, the series exhibits stagnation and increased volatility, supporting our hypothesis that the crisis fundamentally disrupted entrepreneurial pathways.

Step 3: Data Preparation

Show Code
# Create ITS variables
self_emp_its <- self_emp_df %>%
  mutate(
    # Xt: Sequential time variable (months since start)
    Xt = as.numeric(difftime(date, min(date), units = "days")) / 30.44,

    # Zt: Binary intervention indicator (0 = before crisis, 1 = after)
    Zt = ifelse(date >= intervention_date, 1, 0),

    # Pt: Time since intervention (0 before, months after)
    Pt = ifelse(date >= intervention_date,
                as.numeric(difftime(date, intervention_date, units = "days")) / 30.44,
                0)
  ) %>%
  arrange(date)

# Display sample of prepared data
sample_indices <- c(
  which(self_emp_its$date == as.Date("2007-01-01")),  # Before intervention
  which(self_emp_its$date == as.Date("2008-08-01")),  # Just before
  which(self_emp_its$date == as.Date("2008-09-01")),  # Intervention point
  which(self_emp_its$date == as.Date("2008-10-01")),  # Just after
  which(self_emp_its$date == as.Date("2010-01-01"))   # Well after
)

kable(self_emp_its[sample_indices, ],
      caption = "Sample of Prepared ITS Data",
      digits = 2,
      col.names = c("Date", "Self-Employed (000s)", "Time (Xt)",
                    "Intervention (Zt)", "Post-Intervention Time (Pt)"))
Sample of Prepared ITS Data
Date Self-Employed (000s) Time (Xt) Intervention (Zt) Post-Intervention Time (Pt)
709 2007-01-01 9672 707.95 0 0.00
728 2008-08-01 9321 726.94 0 0.00
729 2008-09-01 9163 727.96 1 0.00
730 2008-10-01 8908 728.94 1 0.99
745 2010-01-01 8956 743.96 1 16.00

Variable Descriptions:

  • Y (self_employed): Outcome variable - number of self-employed persons in thousands
  • Xt: Sequential time variable representing months since January 1948 (captures pre-existing trend)
  • Zt: Binary intervention indicator (0 before September 2008, 1 after) - tests for immediate level change
  • Pt: Time since intervention in months (0 before crisis, then 1, 2, 3…) - tests for slope change post-intervention

Step 4: Fit the ITS Model

Show Code
# Fit interrupted time series regression model
its_model <- lm(self_employed ~ Xt + Zt + Pt, data = self_emp_its)

# Display results
summary(its_model)

Call:
lm(formula = self_employed ~ Xt + Zt + Pt, data = self_emp_its)

Residuals:
    Min      1Q  Median      3Q     Max 
-1485.6  -215.5   130.6   410.4  1446.8 

Coefficients:
             Estimate Std. Error t value Pr(>|t|)    
(Intercept) 4874.3303    46.4320 104.978  < 2e-16 ***
Xt             6.3433     0.1106  57.355  < 2e-16 ***
Zt          -845.6173    99.0745  -8.535  < 2e-16 ***
Pt            -4.2762     0.7537  -5.673 1.87e-08 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 627.1 on 928 degrees of freedom
Multiple R-squared:  0.827, Adjusted R-squared:  0.8264 
F-statistic:  1479 on 3 and 928 DF,  p-value: < 2.2e-16
Show Code
# Coefficient table
coef_table <- data.frame(
  Coefficient = names(coef(its_model)),
  Estimate = coef(its_model),
  Std_Error = summary(its_model)$coefficients[, "Std. Error"],
  t_value = summary(its_model)$coefficients[, "t value"],
  p_value = summary(its_model)$coefficients[, "Pr(>|t|)"]
)

kable(coef_table,
      caption = "ITS Model Coefficients",
      digits = 4,
      row.names = FALSE)
ITS Model Coefficients
Coefficient Estimate Std_Error t_value p_value
(Intercept) 4874.3303 46.4320 104.9779 0
Xt 6.3433 0.1106 57.3555 0
Zt -845.6173 99.0745 -8.5352 0
Pt -4.2762 0.7537 -5.6733 0

Model Equation:

\[ \text{Self-Employment}_t = \beta_0 + \beta_1 X_t + \beta_2 Z_t + \beta_3 P_t + \epsilon_t \]

Where: - \(\beta_0\) (Intercept): Baseline self-employment level at start of series - \(\beta_1\) (Xt): Pre-intervention trend (slope before 2008) - \(\beta_2\) (Zt): Immediate level change at intervention (jump/drop in September 2008) - \(\beta_3\) (Pt): Change in slope post-intervention (how trend changed after crisis)

Show Code
# Generate predictions from model
self_emp_its$predicted <- predict(its_model, newdata = self_emp_its)

# Plot actual vs predicted with trend lines
ggplot(self_emp_its, aes(x = date)) +
  geom_line(aes(y = self_employed, color = "Actual"), linewidth = 0.8) +
  geom_line(aes(y = predicted, color = "Predicted (ITS Model)"), linewidth = 0.8) +
  geom_vline(xintercept = as.numeric(intervention_date),
             color = "red", linetype = "dashed", linewidth = 1) +
  annotate("text", x = intervention_date, y = max(self_emp_its$self_employed) * 0.95,
           label = "2008 Crisis", color = "red", hjust = -0.1) +
  labs(
    title = "Actual vs Predicted Self-Employment (ITS Model)",
    subtitle = "Model captures pre- and post-intervention trends",
    x = "Year",
    y = "Self-Employed Persons (Thousands)",
    color = "Series"
  ) +
  theme_minimal() +
  scale_color_manual(values = c("Actual" = "steelblue", "Predicted (ITS Model)" = "orange"))

Step 5: Counterfactual Analysis

Show Code
# Generate counterfactual predictions (what would have happened without intervention)
self_emp_its$counterfactual <- predict(its_model,
                                        newdata = mutate(self_emp_its, Zt = 0, Pt = 0))

# Plot actual, predicted, and counterfactual
ggplot(self_emp_its, aes(x = date)) +
  geom_line(aes(y = self_employed, color = "Actual"), linewidth = 0.8) +
  geom_line(aes(y = predicted, color = "Predicted (with intervention)"), linewidth = 0.8) +
  geom_line(aes(y = counterfactual, color = "Counterfactual (no intervention)"),
            linewidth = 0.8, linetype = "dashed") +
  geom_vline(xintercept = as.numeric(intervention_date),
             color = "red", linetype = "dashed", linewidth = 1) +
  annotate("text", x = intervention_date, y = max(self_emp_its$self_employed) * 0.95,
           label = "2008 Crisis", color = "red", hjust = -0.1) +
  labs(
    title = "Self-Employment: Actual vs Counterfactual Outcomes",
    subtitle = "Counterfactual shows what would have happened without the 2008 financial crisis",
    x = "Year",
    y = "Self-Employed Persons (Thousands)",
    color = "Trend",
    caption = "Counterfactual assumes pre-intervention trend continued"
  ) +
  theme_minimal() +
  scale_color_manual(values = c(
    "Actual" = "black",
    "Predicted (with intervention)" = "blue",
    "Counterfactual (no intervention)" = "darkgreen"
  ))

Show Code
# Calculate intervention effect at specific time points
effect_2010 <- self_emp_its %>%
  filter(date == as.Date("2010-01-01")) %>%
  mutate(effect = counterfactual - predicted) %>%
  select(date, actual = self_employed, predicted, counterfactual, effect)

effect_2015 <- self_emp_its %>%
  filter(date == as.Date("2015-01-01")) %>%
  mutate(effect = counterfactual - predicted) %>%
  select(date, actual = self_employed, predicted, counterfactual, effect)

effect_2020 <- self_emp_its %>%
  filter(date == as.Date("2020-01-01")) %>%
  mutate(effect = counterfactual - predicted) %>%
  select(date, actual = self_employed, predicted, counterfactual, effect)

intervention_effects <- bind_rows(effect_2010, effect_2015, effect_2020)

kable(intervention_effects,
      caption = "Intervention Effect at Key Time Points",
      digits = 2,
      col.names = c("Date", "Actual", "Predicted", "Counterfactual",
                    "Lost Self-Employment (000s)"))
Intervention Effect at Key Time Points
Date Actual Predicted Counterfactual Lost Self-Employment (000s)
2010-01-01 8956 8679.42 9593.45 914.03
2015-01-01 8725 8803.42 9973.97 1170.55
2020-01-01 8881 8927.41 10354.48 1427.07

Step 6: Interpretation of Results

1. Immediate Effect (Zt coefficient)

Show Code
# Extract Zt coefficient
zt_coef <- coef(its_model)["Zt"]
zt_pval <- summary(its_model)$coefficients["Zt", "Pr(>|t|)"]

cat("Zt coefficient:", round(zt_coef, 2), "\n")
Zt coefficient: -845.62 
Show Code
cat("p-value:", format(zt_pval, scientific = TRUE), "\n")
p-value: 5.655923e-17 

Interpretation: The Zt coefficient represents the immediate level change in self-employment when the 2008 financial crisis hit.

  • If negative and significant (p < 0.05): There was an immediate drop in self-employment right when the crisis began - likely due to credit freeze making it impossible to start businesses and wealth destruction reducing entrepreneurial capital.

  • If not significant: The crisis’s impact was not instantaneous but rather built up over time (would show in Pt instead).

The magnitude tells us how many thousands of self-employed persons were immediately lost/gained when the intervention occurred.

2. Sustained Effect (Pt coefficient)

Show Code
# Extract Pt coefficient
pt_coef <- coef(its_model)["Pt"]
pt_pval <- summary(its_model)$coefficients["Pt", "Pr(>|t|)"]

cat("Pt coefficient:", round(pt_coef, 4), "\n")
Pt coefficient: -4.2762 
Show Code
cat("p-value:", format(pt_pval, scientific = TRUE), "\n")
p-value: 1.871137e-08 

Interpretation: The Pt coefficient represents the change in trend slope after the intervention.

  • Pre-crisis trend slope: β1 (Xt coefficient) - the monthly growth rate before 2008
  • Post-crisis trend slope: β1 + β3 (Xt + Pt) - the monthly growth rate after 2008

If Pt is negative and significant: The growth trend became less steep or even reversed - self-employment stopped growing as fast as it did pre-crisis. This suggests a sustained structural change in entrepreneurial activity.

This aligns with our hypothesis: the 2008 crisis didn’t just cause a temporary dip but fundamentally altered the entrepreneurship landscape through: - Increased regulatory burden (Dodd-Frank) - Risk aversion among millennials who witnessed the crisis - Student debt explosion making entrepreneurship unaffordable - Shift toward gig economy rather than traditional self-employment

3. Counterfactual

What would have happened without the 2008 financial crisis?

The counterfactual line (green dashed) shows self-employment continuing along its pre-2008 growth trajectory. The gap between the counterfactual and actual/predicted lines represents the total cost of the intervention - how many self-employed persons the economy “lost” due to the crisis.

Key observations: - Immediate gap: The vertical distance at the intervention point shows immediate losses - Widening gap: If the gap grows over time, it indicates the crisis had compounding effects - Policy implications: The counterfactual quantifies what recovery efforts need to restore

4. Delayed Effect

Show Code
# Calculate delayed effect 10 years post-intervention (2018)
delayed_effect <- self_emp_its %>%
  filter(date == as.Date("2018-09-01")) %>%  # Exactly 10 years after
  mutate(
    immediate_effect = counterfactual - (counterfactual - Zt * coef(its_model)["Zt"]),
    trend_effect = counterfactual - predicted,
    total_effect = counterfactual - self_employed
  ) %>%
  select(date, immediate_effect, trend_effect, total_effect)

kable(delayed_effect,
      caption = "Decomposition of Delayed Effect (10 years post-intervention)",
      digits = 2,
      col.names = c("Date (10 yrs after)", "Immediate Effect", "Trend Change Effect", "Total Effect (000s)"))
Decomposition of Delayed Effect (10 years post-intervention)
Date (10 yrs after) Immediate Effect Trend Change Effect Total Effect (000s)
2018-09-01 -845.62 1358.65 1347

Delayed Effect Interpretation:

Ten years after the 2008 financial crisis, the intervention’s impact can be decomposed:

  1. Immediate effect still present: The initial shock (Zt) persists unless there’s a recovery mechanism
  2. Accumulated trend change: Slower post-crisis growth compounds over 10 years (120 months × Pt coefficient)
  3. Total delayed effect: The sum of immediate + accumulated trend effects

Visual interpretation from the plot: - At 2018 (10 years post-crisis), the gap between counterfactual and actual is larger than at the intervention point - This widening gap indicates the crisis had long-lasting structural effects on entrepreneurship - The effect is not just a temporary shock that recovers - it’s a permanent shift in the entrepreneurial landscape

Economic significance: If the total effect shows hundreds of thousands of “missing” self-employed persons a decade later, it suggests: - The 2008 crisis permanently damaged entrepreneurial ecosystems - Young people who should have become entrepreneurs in 2008-2018 never did - This supports our “blocked entrepreneurship pathway” hypothesis - student debt and risk aversion created by the crisis continue suppressing entrepreneurial activity years later

Summary of Findings

Show Code
# Create comprehensive summary
xt_coef <- coef(its_model)["Xt"]

summary_df <- data.frame(
  Effect = c("Pre-Crisis Trend", "Immediate Level Change", "Post-Crisis Trend Change", "Post-Crisis Trend (Total)"),
  Coefficient = c("β1 (Xt)", "β2 (Zt)", "β3 (Pt)", "β1 + β3"),
  Value = c(xt_coef, zt_coef, pt_coef, xt_coef + pt_coef),
  Interpretation = c(
    "Monthly growth before 2008",
    "Drop/jump when crisis hit",
    "Change in monthly growth rate",
    "New monthly growth rate after crisis"
  )
)

kable(summary_df,
      caption = "Summary of ITS Model Effects",
      digits = 4,
      row.names = FALSE)
Summary of ITS Model Effects
Effect Coefficient Value Interpretation
Pre-Crisis Trend β1 (Xt) 6.3433 Monthly growth before 2008
Immediate Level Change β2 (Zt) -845.6173 Drop/jump when crisis hit
Post-Crisis Trend Change β3 (Pt) -4.2762 Change in monthly growth rate
Post-Crisis Trend (Total) β1 + β3 2.0670 New monthly growth rate after crisis

Conclusion:

The interrupted time series analysis provides strong empirical evidence that the 2008 financial crisis fundamentally disrupted U.S. entrepreneurship. The crisis didn’t just cause a temporary setback - it produced lasting structural changes in self-employment trends. This finding directly supports our project’s central thesis: economic shocks that increase student debt burdens and risk aversion have long-term consequences for wealth-building pathways, particularly entrepreneurship.

The counterfactual analysis quantifies this cost - showing how many entrepreneurial ventures were “lost” due to the crisis. This has direct policy implications: student debt relief, entrepreneurship support programs, and regulatory reforms could help restore the pre-crisis entrepreneurial dynamism.