Introduction to tidystats

2020-09-21

Why use tidystats?

The tidystats package is designed to address two problems common in scientific research: incomplete and incorrect statistics reporting. The problem of incomplete statistics reporting likely stems from a fundamental trade-off between wanting to be comprehensive on the one hand and providing a clear narrative on the other. The problem of incorrect statistics reporting is likely caused by manually copy-pasting statistical output from the output window into a text editor. tidystats addresses these two problems by enabling researchers to combine their statistical analyses into a single file, from which a subset of the analyses can then be reported using a Microsoft Word add-in.

How to use tidystats?

tidystats is designed to easily fit in your data analysis workflow. In fact, tidystats can simply be tacked on at the end of a data analysis session, with only one minor requirement. This requirement is that all analyses are stored in a variable. For example, if you run a regression analysis using the lm() function, you store the result of that analysis in a variable: model1 <- lm(extra ~ group, data = sleep). By storing each analysis in a variable, you can later add each analysis to a list using the add_stats() function from tidystats. Once all the analyses are gathered together, you save the analyses to a .json file using the write_stats() function. This .json file can then be read by a Word add-in to report your analyses in Word, or shared with others and read into R to extract statistics from your analyses (e.g., for meta-analyses).

An example

Below follows an example of a few analyses conducted on the quote_source data contained within the tidystats package. The data is from a large-scaled replication of Lorge & Curtiss (1936). More details can be found in the paper of the replication effort (Klein et al., 2014). In short, participants saw the following quote:

“I have sworn to only live free, even if I find bitter the taste of death.”

The quote was attributed to either George Washington, a liked individual, or Osama Bin Laden, a disliked individual. Participants were asked to what extent they agree with the quote, on a 9-point Likert scale ranging from 1 (disagreement) to 9 (agreement).

We start with a bit of setup.

# Load packages
library(tidystats)
library(dplyr)

# Load data
data <- tidystats::quote_source

The main hypothesis is that people will like the quote more when it is attributed to George Washington compared to Osama Bin Laden. We test this hypothesis by first looking at the descriptives and then by conducting a t-test.

descriptives <- data %>%
  group_by(source) %>%
  describe_data(response, short = TRUE)
descriptives
variable source N M SD
response Bin Laden 3083 5.23 2.11
response Washington 3242 5.93 2.21
t_test <- t.test(response ~ source, data = data)
t_test
#> 
#>  Welch Two Sample t-test
#> 
#> data:  response by source
#> t = -13, df = 6323, p-value <2e-16
#> alternative hypothesis: true difference in means is not equal to 0
#> 95 percent confidence interval:
#>  -0.802 -0.589
#> sample estimates:
#>  mean in group Bin Laden mean in group Washington 
#>                     5.23                     5.93

Participants appear to rate the quote a bit more positively when it is attributed to George Washington.

We can also perform some additional tests. For instance, does it matter if the participant is from the US? And does age matter? To answer these questions, we can perform interaction tests using lm().

The interaction with the participant being from the U.S. or not:

lm_us_or_not <- lm(response ~ source * us_or_international, data = data)
summary(lm_us_or_not)
#> 
#> Call:
#> lm(formula = response ~ source * us_or_international, data = data)
#> 
#> Residuals:
#>    Min     1Q Median     3Q    Max 
#> -5.005 -1.228 -0.228  1.772  3.772 
#> 
#> Coefficients:
#>                                                   Estimate Std. Error t value
#> (Intercept)                                         5.2278     0.0437  119.50
#> sourceWashington                                    0.7769     0.0613   12.67
#> us_or_internationalinternational                    0.0210     0.0955    0.22
#> sourceWashington:us_or_internationalinternational  -0.3717     0.1323   -2.81
#>                                                   Pr(>|t|)    
#> (Intercept)                                         <2e-16 ***
#> sourceWashington                                    <2e-16 ***
#> us_or_internationalinternational                     0.826    
#> sourceWashington:us_or_internationalinternational    0.005 ** 
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> Residual standard error: 2.16 on 6321 degrees of freedom
#>   (18 observations deleted due to missingness)
#> Multiple R-squared:  0.0275, Adjusted R-squared:  0.027 
#> F-statistic: 59.5 on 3 and 6321 DF,  p-value: <2e-16

The interaction is significant, so it appears to matter whether the participant is from the U.S. or not. In fact, participants from the U.S. show a stronger effect than those from outside the U.S.

The interaction with the participant’s age:

lm_age <- lm(response ~ source * age, data = data)
summary(lm_age)
#> 
#> Call:
#> lm(formula = response ~ source * age, data = data)
#> 
#> Residuals:
#>    Min     1Q Median     3Q    Max 
#> -5.743 -1.202 -0.152  1.793  3.883 
#> 
#> Coefficients:
#>                      Estimate Std. Error t value Pr(>|t|)    
#> (Intercept)           4.93370    0.09589   51.45  < 2e-16 ***
#> sourceWashington      0.55737    0.13558    4.11    4e-05 ***
#> age                   0.01147    0.00336    3.41  0.00065 ***
#> sourceWashington:age  0.00545    0.00478    1.14  0.25433    
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> Residual standard error: 2.16 on 6308 degrees of freedom
#>   (31 observations deleted due to missingness)
#> Multiple R-squared:  0.0308, Adjusted R-squared:  0.0304 
#> F-statistic: 66.9 on 3 and 6308 DF,  p-value: <2e-16

No significant interaction effect, so we do not have evidence for age changing the size of the effect.

Let’s say these are the analyses we want to save the output of and report later. This is where tidystats comes in. The steps to perform are to first create an empty list and then to use the add_stats() function to add analyses to the list. This is why we stored each analysis into a variable. The add_stats() function takes an analysis, extracts the statistics, and adds the result to a list. Optionally, you can add additional information about each analysis, such as whether it was preregistered, whether it was a primary, secondary, or exploratory analysis, or simply add some notes.

# Create an empty list to store the analyses in
results <- list()

# Add the analyses
results <- results %>%
  add_stats(t_test, preregistered = TRUE, type = "primary", 
    notes = "A t-test comparing the effect of source on the quote rating.") %>%
  add_stats(lm_us_or_not, preregistered = FALSE, type = "exploratory", 
    notes = "Interaction effect with being from the U.S. or not.") %>%
  add_stats(lm_age)

You can see that I added quite some information to the first and second analysis. This is recommended because it is easy to forget which analysis is which; and you might accidentally report the wrong analysis if you have many of them. It’s also nice to add some documentation so that others who are not as familiar with your data can also better understand each analysis.

To save these analyses to a file, you can use the write_stats() function.

write_stats(results, "lorge-curtiss-1936-replication.json")

Note the file extension: .json. These types of files are simply text files, but in a format that is machine-readable (unfortunately, not very human-readable). This file can be used to share with others so that they can read it back into R and extract statistics (e.g., for meta-analyses) or by you to report the statistics in Word.

References

Lorge, I., & Curtiss, C. C. (1936). Prestige, suggestion, and attitudes. The Journal of Social Psychology, 7, 386-402. https://doi.org/10.1080/00224545.1936.9919891

Klein, R.A. et al. (2014) Investigating Variation in Replicability: A “Many Labs” Replication Project. Social Psychology, 45(3), 142-152. https://dx.doi.org/10.1027/1864-9335/a000178