Monte Carlo Simulation and Power Analysis

Introduction

Before running an event study, it is important to understand whether your research design has enough statistical power to detect an effect of a given magnitude. The simulate_event_study() function provides Monte Carlo simulation to estimate:

Power: The probability of rejecting the null hypothesis when the alternative is true
Size: The actual rejection rate under the null (should be close to the nominal \alpha)

Basic Usage

library(EventStudy)

# Size analysis: no abnormal return injected
size_result <- simulate_event_study(
  n_events = 20,
  event_window = c(-5, 5),
  estimation_window_length = 120,
  abnormal_return = 0,        # H0 is true
  n_simulations = 1000,
  seed = 42
)
print(size_result)
#> Event Study Simulation
#>   N events:        20
#>   Event window:   [ -5 , 5 ]
#>   Abnormal return: 0
#>   Test statistic:  CSectT
#>   Alpha:           0.05
#>   N simulations:   1000
#>   Power (day 0):   0.048

The power at day 0 should be close to the nominal significance level (0.05) when no abnormal return is injected.

Power Analysis

To estimate power, inject a non-zero abnormal return:

# Power analysis: 2% abnormal return on event day
power_result <- simulate_event_study(
  n_events = 20,
  event_window = c(-5, 5),
  estimation_window_length = 120,
  abnormal_return = 0.02,     # 2% effect
  n_simulations = 1000,
  seed = 42
)
print(power_result)
#> Power (day 0):   0.723

Varying Effect Size

Compare power across different abnormal return magnitudes:

effects <- c(0, 0.005, 0.01, 0.02, 0.05)
power_curve <- sapply(effects, function(ar) {
  res <- simulate_event_study(
    n_events = 20,
    abnormal_return = ar,
    n_simulations = 500,
    seed = 42
  )
  res$power
})

plot(effects, power_curve, type = "b",
     xlab = "Abnormal Return", ylab = "Power",
     main = "Power Curve")
abline(h = 0.05, lty = 2, col = "grey")

Varying Sample Size

sample_sizes <- c(5, 10, 20, 50, 100)
power_by_n <- sapply(sample_sizes, function(n) {
  res <- simulate_event_study(
    n_events = n,
    abnormal_return = 0.01,
    n_simulations = 500,
    seed = 42
  )
  res$power
})

plot(sample_sizes, power_by_n, type = "b",
     xlab = "Number of Events", ylab = "Power",
     main = "Power vs Sample Size (AR = 1%)")

Customizing the DGP

You can control the data-generating process parameters:

result <- simulate_event_study(
  n_events = 20,
  abnormal_return = 0.01,
  n_simulations = 500,
  dgp_params = list(
    alpha = 0.0001,         # drift
    beta = 1.2,             # market exposure
    sigma_firm = 0.015,     # idiosyncratic volatility
    sigma_market = 0.02     # market volatility
  ),
  seed = 42
)

Higher idiosyncratic volatility makes it harder to detect abnormal returns, reducing power.

Different Test Statistics

Compare the power of different test statistics:

tests <- c("CSectT", "PatellZ", "BMP", "SignT", "RankT")
power_by_test <- sapply(tests, function(ts) {
  res <- simulate_event_study(
    n_events = 20,
    abnormal_return = 0.01,
    test_statistic = ts,
    n_simulations = 500,
    seed = 42
  )
  res$power
})

barplot(power_by_test, main = "Power by Test Statistic (AR = 1%)",
        ylab = "Power", las = 2)
abline(h = 0.05, lty = 2, col = "grey")

Rejection by Day

The rejection_by_day component shows the rejection rate at each event-window day:

result <- simulate_event_study(
  n_events = 20,
  abnormal_return = 0.02,
  n_simulations = 500,
  seed = 42
)

result$rejection_by_day
#> # A tibble: 11 x 2
#>    relative_index rejection_rate
#>             <int>          <dbl>
#>  1             -5         0.048
#>  2             -4         0.052
#>  ...
#>  6              0         0.723   <-- event day
#>  ...

Only the event day (relative_index = 0) should show elevated rejection. Other days should be near the nominal level.

Seed Reproducibility

Results are fully reproducible when a seed is set:

r1 <- simulate_event_study(n_events = 10, n_simulations = 100, seed = 42)
r2 <- simulate_event_study(n_events = 10, n_simulations = 100, seed = 42)
identical(r1$power, r2$power)  # TRUE

Simon Mueller

2026-09-06