Discriminating Frequency Tables using Higher Criticism

Alon Kipnis

2019-12-21

This package implements an adpatation of the Higher-Criticism (HC) test to discriminate two frequency tables footnotes1.

The package includes two main functions: - two.sample.pvals – produces a list of P-values, one for each feature in the two tables. - HC.vals – computes the HC score of the P-values.

A third function two.sample.HC combines the two functions above so that the HC score of the two tables is obtained using a single function call.


Example:

#' # Can be used to check similarity of word-frequencies in texts:
#' text1 = "On the day House Democrats opened an impeachment inquiry of
#'    President Trump last week, Pete Buttigieg was being grilled by Iowa 
#'    voters on other subjects: how to loosen the grip of the rich on 
#'    government, how to restore science to policymaking, how to reduce child
#'    poverty. At an event in eastern Iowa, a woman rose to say that her four
#'    adult children were “stuck” in life, unable to afford what she had in 
#'    the 1980s when a $10-an-hour job paid for rent, utilities and an 
#'    annual vacation."
#' text2 = "How can the federal government help our young people that want to do
#'   what’s right and to get to those things that their parents worked so hard for?”
#'   the voter asked. This is the conversation Mr. Buttigieg wants to have. 
#'   Boasting a huge financial war chest but struggling in the polls, Mr. Buttigieg
#'   is now staking his presidential candidacy on Iowa, and particularly on
#'   connecting with rural white voters who want to talk about personal concerns 
#'   more than impeachment. In doing so, Mr. Buttigieg is also trying to show how
#'   Democrats can win back counties that flipped from Barack Obama to Donald
#'   Trump in 2016 — there are more of them in Iowa than any other state — 
#'   by focusing, he said, on “the things that are going to affect folks’
#'   lives in a concrete way."

tb1 = table(strsplit(tolower(text1),' '))
tb2 = table(strsplit(tolower(text2),' '))
pv = two.sample.pvals(tb1,tb2)

print(pv$pv)
> [1] 1.0000 1.0000 0.2304 1.0000 1.0000 1.0000     NA 0.1936     NA

print(pv$Var1)
> go i or say should stay you and not

HC.vals(pv$pv)
> $HC
> 0.323954762194625
> $HC.star
> 0.323954762194625
> $p
> 0.2304
> $p.star
> 0.2304

Example 2

n = 1000  #number of features
N = 10*n  #number of observations
k = 0.1*n #number of perturbed features

seq = seq(1,n)
P = 1 / seq  #sample from a Zipf law distribution
P = P / sum(P)
tb1 = data.frame(Feature = seq(1,n),  # sample 1
      Freq = rmultinom(n = 1, prob = P, size = N))


seq[sample(seq,k)] <- seq[sample(seq,k)] 
Q = 1 / seq 
Q = Q / sum(Q)  

tb2 = data.frame(Feature = seq(1,n), # sample 2
      Freq = rmultinom(n = 1, prob = Q, size = N))

PV = two.sample.pvals(tb1, tb2)  #compute P-values

HC.vals(PV$pv)  # HC test

# can also test using a single function call
two.sample.HC(tb1,tb2) 

  1. See Kipnis A. Higher Criticism for Discriminating Word-Frequency Tables and Testing Authorship (2019)↩︎