CODEWITHSUNDEEP
rdplyrstatisticst-test
Thomas Philips
Thomas Philips

Reputation: 1089

A subtle problem with t-tests over multiple columns

I have a dataframe with responses to multiple questions (reproducible example with 2 questions below)

set.seed(1)
df <- data.frame (
          UserId = c(rep("A", 4), rep("B", 4), rep("C", 4), rep("D", 4)),
          Sex = c(rep("Female", 8), rep("Male", 4), rep("No_Response", 4)),
          Answer_Date = as.Date(c("1990-01-01", "1990-02-01", "1990-03-01", "1990-04-01",
                                  "1991-02-01", "1991-03-01", "1991-04-01", "1991-05-01",
                                  "1992-03-01", "1992-04-01", "1992-05-01", "1992-06-01",
                                  "1993-07-10", "1992-08-10", "1993-09-10", "1993-10-10")),
          Q1 = sample(1:10, 16, replace = TRUE),
          Q2 = sample(1:10, 16, replace = TRUE)
      ) %>%
      group_by(UserId) %>%
      mutate(First_Answer_Date = min(Answer_Date)) %>%
      mutate(Last_Answer_Date  = max(Answer_Date)) %>%
      ungroup()

Following the suggestion in

https://sebastiansauer.github.io/multiple-t-tests-with-dplyr/

I run t-tests for Q1 and Q2 against the null hypothesis that the true mean is 0:

questions <- c("Q1", "Q2")
df %>%
  select(questions, Sex) %>%
  filter(Sex != "No_Response") %>%
  gather(key = variable, value = value, -Sex) %>%
  group_by(Sex, variable) %>%
  summarize(value = list(value)) %>%
  spread(Sex, value) %>%
  group_by(variable) %>%
  mutate( p_Female = t.test(unlist(Female))$p.value,
          p_Male   = t.test(unlist(Male)  )$p.value,
          t_Female = t.test(unlist(Female))$statistic,
          t_Male   = t.test(unlist(Male)  )$statistic) %>%
  mutate( Female = length(unlist(Female)),
          Male   = length(unlist(Male))
  )

which gives me

# A tibble: 2 x 7
# Groups:   variable [2]
  variable Female  Male  p_Female p__Male t_Female t_Male
  <chr>     <int> <int>     <dbl>   <dbl>    <dbl>  <dbl>
1 Q1            8     4 0.0000501 0.00137     8.78  11.6 
2 Q2            8     4 0.00217   0.0115      4.71   5.55

All good so far. My troubles start when I want to do the t-test only on the First_Answer_Date.

df %>%
  filter(Answer_Date == First_Answer_Date) %>%
  select(questions, Sex) %>%
  filter(Sex != "No_Response") %>%

    # A tibble: 3 x 3
         Q1    Q2 Sex   
      <int> <int> <chr> 
    1     9     5 Female
    2     2     5 Female
    3     1     9 Male

Now, there is only one response from a Male and two from a Female, and on Q2, both Female respondents have the same answer. If I rerun my t-test code

df %>%
  filter(Answer_Date == First_Answer_Date) %>%
  select(questions, Sex) %>%
  filter(Sex != "No_Response") %>%
  gather(key = variable, value = value, -Sex) %>%
  group_by(Sex, variable) %>%
  summarize(value = list(value)) %>%
  spread(Sex, value) %>%
  group_by(variable) %>%
  mutate( p_Female = t.test(unlist(Female))$p.value,
          p__Male = t.test(unlist(Male))$p.value,
          t_Female = t.test(unlist(Female))$statistic,
          t_Male = t.test(unlist(Male))$statistic) %>%
  mutate( Female = length(unlist(Female)),
          Male   = length(unlist(Male)))

Error: Problem with `mutate()` input `p_Female`.
x data are essentially constant
i Input `p_Female` is `t.test(unlist(Female))$p.value`.
i The error occurred in group 2: variable = "Q2".

The error message I get is logical, but this is a situation that I am likely to encounter in practice - some subsets can be of size 1 or 0, all respondents to some questions are likely to give the same answer etc. etc. How can I make the code degrade gracefully, just putting a blank or NA in those cells in its output tibble where no answer can be computed for one reason or another?

Sincerely

Thomas Philips

Upvotes: 0

Views: 41

Answers (1)

Ronak Shah
Ronak Shah

Reputation: 389115

Perhaps, you can use tryCatch to handle the error :

library(dplyr)
library(tidyr)

df %>%
  filter(Answer_Date == First_Answer_Date) %>%
  select(questions, Sex) %>%
  filter(Sex != "No_Response") %>%
  pivot_longer(cols = -Sex, names_to = "variable") %>%
  group_by(Sex, variable) %>%
  summarize(value = list(value)) %>%
  pivot_wider(names_from = Sex, values_from = value) %>%
  group_by(variable) %>%
  mutate( p_Female = tryCatch(t.test(unlist(Female))$p.value, error = function(e) return(NA)),
          p_Male   = tryCatch(t.test(unlist(Male) )$p.value, error = function(e) return(NA)),
          t_Female = tryCatch(t.test(unlist(Female))$statistic, error = function(e) return(NA)),
          t_Male   = tryCatch(t.test(unlist(Male))$statistic,error = function(e) return(NA))) %>%
  ungroup %>%
  mutate( Female = lengths(Female),
          Male   = lengths(Male))

Upvotes: 1

Related Questions