Creating Custom Test Statistics

In EventStudy: Event Study Analysis

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

Introduction

The EventStudy package supports two types of test statistics:

1. Single-event statistics -- Computed for each event individually (e.g., AR t-test, CAR t-test)
2. Multi-event statistics -- Computed across all events in a group (e.g., Cross-Sectional t-test, Patell Z, Sign test)

Both types inherit from TestStatisticBase and implement a compute() method. This vignette walks through creating custom test statistics of each type.

The TestStatisticBase Interface

Every test statistic must:

1. Inherit from TestStatisticBase
2. Set a name field (short code used as column name in results)
3. Implement compute(data_tbl, model) returning a tibble

For single-event tests, data_tbl contains one event's data and model is the fitted model for that event. For multi-event tests, data_tbl contains all events in a group and model is a nested tibble of all models.

Example: Wilcoxon Signed-Rank Test (Single Event)

A non-parametric test for whether abnormal returns in the event window are symmetrically distributed around zero:

library(EventStudy)

WilcoxonARTest <- R6::R6Class("WilcoxonARTest",
  inherit = TestStatisticBase,
  public = list(
    name = "WilcoxonAR",

    compute = function(data_tbl, model) {
      event_data <- data_tbl %>%
        dplyr::filter(event_window == 1)

      ar <- event_data$abnormal_returns

      # Wilcoxon signed-rank test against 0
      test_result <- wilcox.test(ar, mu = 0, conf.int = TRUE)

      tibble::tibble(
        n_obs        = length(ar),
        V_statistic  = test_result$statistic,
        p_value      = test_result$p.value,
        median_ar    = median(ar, na.rm = TRUE),
        conf_low     = test_result$conf.int[1],
        conf_high    = test_result$conf.int[2]
      )
    }
  )
)

Example: Kolmogorov-Smirnov Test (Multi-Event)

A non-parametric test comparing the distribution of AARs against a normal distribution:

KSTest <- R6::R6Class("KSTest",
  inherit = TestStatisticBase,
  public = list(
    name = "KS",

    compute = function(data_tbl, model) {
      aar_stats <- data_tbl %>%
        dplyr::filter(event_window == 1) %>%
        dplyr::group_by(relative_index) %>%
        dplyr::summarise(
          aar        = mean(abnormal_returns, na.rm = TRUE),
          n_events   = dplyr::n(),
          n_pos      = sum(abnormal_returns > 0, na.rm = TRUE),
          n_neg      = sum(abnormal_returns <= 0, na.rm = TRUE),
          .groups    = "drop"
        ) %>%
        dplyr::mutate(caar = cumsum(aar))

      # KS test of AARs against normal
      ks <- ks.test(aar_stats$aar, "pnorm",
                    mean = 0, sd = sd(aar_stats$aar))

      aar_stats %>%
        dplyr::mutate(
          ks_D      = ks$statistic,
          ks_pvalue = ks$p.value
        )
    }
  )
)

Registering Custom Statistics

Add your test statistic to a statistics set:

# Single-event: add to SingleEventStatisticsSet
single_stats <- SingleEventStatisticsSet$new()
single_stats$add_test(WilcoxonARTest$new())

# Multi-event: add to MultiEventStatisticsSet
multi_stats <- MultiEventStatisticsSet$new()
multi_stats$add_test(KSTest$new())

# Or create a custom set from scratch
custom_single <- StatisticsSetBase$new(
  tests = list(ARTTest$new(), CARTTest$new(), WilcoxonARTest$new())
)

Using Custom Statistics in the Pipeline

params <- ParameterSet$new(
  single_event_statistics = single_stats,
  multi_event_statistics = multi_stats
)

task <- EventStudyTask$new(firm_data, index_data, request_data)
task <- run_event_study(task, params)

# Results appear as columns in the task data
task$data_tbl$WilcoxonAR  # single-event results
task$aar_caar_tbl$KS       # multi-event results

Built-in Test Statistics

The package includes the following built-in test statistics:

Single-event: - ARTTest -- Abnormal Return t-test - CARTTest -- Cumulative Abnormal Return t-test - BHARTTest -- Buy-and-Hold Abnormal Return t-test

Multi-event: - CSectTTest -- Cross-Sectional t-test (AAR/CAAR) - PatellZTest -- Patell standardized residual test - SignTest -- Simple sign test - GeneralizedSignTest -- Cowan (1992) generalized sign test - RankTest -- Corrado (1989) rank test - BMPTest -- Boehmer, Musumeci & Poulsen (1991) test - CalendarTimePortfolioTest -- Calendar-time portfolio approach

Key Points

• Inherit from TestStatisticBase
• Set a unique name field (becomes the column name in results)
• Return a tibble from compute()
• For multi-event stats, include relative_index, aar, caar, and car_window columns for compatibility with plotting functions
• Use add_test() to add to existing statistics sets

EventStudy documentation built on Oct. 31, 2026, 1:07 a.m.