Skip to contents

Hypothesis tests

Source: R/toolbox.R
hypothesis_test.Rd

Mirrors the C# HypothesisTests static class: thirteen one- and two-sample parametric and nonparametric hypothesis tests, reached through the shared hypothesis toolbox group. Every method but "f_models" returns the 2-sided p-value of its test statistic; "f_models" (the F-test comparing two nested regression models) additionally returns the F statistic itself.

Usage

hypothesis_test(
  x = NULL,
  y = NULL,
  method = "jarque_bera",
  population_mean = 0,
  lag_max = NULL,
  index = NULL,
  sse_restricted = NULL,
  sse_full = NULL,
  df_restricted = NULL,
  df_full = NULL
)

Arguments

x

numeric vector: the sample (the two-sample methods' first sample, or the response series for "linear_trend"). Ignored for "f_models".

y

numeric vector, the second sample. Required for the two-sample methods listed above, and rejected (must be NULL) for every other method – earlier versions silently discarded a y supplied to a one-sample method instead of raising.

method

one of "one_sample_t", "equal_variance_t", "unequal_variance_t", "paired_t", "f", "f_models", "jarque_bera", "wald_wolfowitz", "ljung_box", "mann_whitney", "mann_kendall", "linear_trend", "unimodality".

population_mean

the hypothesized mean for "one_sample_t". Default 0.

lag_max

the max lag for "ljung_box". Default NULL (use the C# default rule).

index

the index (x-axis) vector for "linear_trend". Default NULL, meaning seq_along(x).

sse_restricted, sse_full, df_restricted, df_full

the four "f_models" inputs: the restricted and full models' sum of squared errors and degrees of freedom.

Value

a named numeric vector: p_value for every method but "f_models", which returns c(f_statistic =, p_value =).

Details

Argument use by method, and the C# guard each one inherits:

  • "one_sample_t": x, population_mean (default 0). Needs at least 2 observations.

  • "equal_variance_t" / "unequal_variance_t": x, y. equal_variance_t needs a combined length of at least 3; unequal_variance_t has no length guard (upstream has none either).

  • "paired_t": x, y, which must be the same length.

  • "f": x, y, each needing at least 2 observations.

  • "f_models": sse_restricted, sse_full, df_restricted, df_full (all required; x and y are ignored). df_restricted must differ from df_full, and df_full must be positive.

  • "jarque_bera" / "wald_wolfowitz": x. No length guard.

  • "ljung_box": x, lag_max (default NULL, meaning floor(min(10 * log10(length(x)), length(x) - 1)), the C# default rule).

  • "mann_whitney": x, y. x must be no longer than y, each must have more than 3 observations, and the combined length must exceed 20.

  • "mann_kendall": x. Needs at least 10 observations.

  • "linear_trend": x (the sample), index (default seq_along(x), i.e. 1:length(x) – a VALUE the regression is fit against, not an index into x). index and x must be the same length.

  • "unimodality": x. Needs at least 10 observations. Fits a 1-component and a 2-component Gaussian mixture model (both at the hard-coded seed 12345, so the result is deterministic) and returns the p-value of the likelihood-ratio statistic against a chi-square with 3 degrees of freedom, so a SMALL p-value is evidence against unimodality. If either mixture fit fails numerically the result is NaN rather than an error, matching upstream.

Examples

hypothesis_test(c(4, 5, 5, 6, 9, 12, 13, 14, 14, 19, 22, 24, 25), method = "jarque_bera")
#>  p_value 
#> 0.592128