A discussion on branches in the multiverse

Abhraneel Sarma



Branches in a multiverse analysis

In a multiverse analysis, researchers identify sets of defensible analysis choices (e.g., different ways of excluding outliers, different data transformations), implement them all, and then report the outcomes of all analyses resulting from all possible choice combinations. Thus, a multiverse analysis consists of different paths or branches of analysis for the same problem.

For an example of how a branch might arise during analysis, let’s say that you have a variable, \(x\), which can assume any value between 1 and 60. You want to discretise this variable into two ordinal categories — high (if x >= 30) and low (if x < 30). However, another researcher might argue for discretising this variable into three ordinal categories — high (if x >= 40), medium (if 20 <= x < 40), and low (if x < 20).

An important aspect of this package is to make declaring such branches in the analysis more natural and convenient. Each branch is identified by a parameter, which can take two or more options. Thus, the different options are different branches within the multiverse. A branch may be defined for:

In this document we describe different ways of declaring a branch, using different syntactical styles that are popular with R users, including base R syntax and tidyverse syntax.


In this vignette, we make use of multiverse code chunks, a custom engine designed to work with the multiverse package, to implement the multiverse analyses. Please refer to the vignette (vignette("multiverse-in-rmd")) for more details. Users could instead make use of the function which is more suited for a script-style implementation. Please refer to the vignettes (vignette("complete-multiverse-analysis") and vignette("basic-multiverse")) for more details.

Example 1: A simple branch

Let’s say that there is a variable x which can take any of three different values — “low”, “medium”, “high”. We can perform the effect of encoding x as each of these values separately through separate branches. To do this, we first create a multiverse object, M:

M = multiverse()

Then declare x as three possible values within the multiverse using the branch() function.

```{multiverse default-m-1, inside = M}
x <- branch(a_parameter, "low", "medium", "high")

Now, we can inspect the multiverse object to see if all the options were created.

expand(M) %>%
  select(.universe, everything())
## # A tibble: 3 x 5
##   .universe a_parameter .parameter_assignment .code            .results
##       <int> <chr>       <list>                <list>           <list>  
## 1         1 low         <named list [1]>      <named list [1]> <env>   
## 2         2 medium      <named list [1]>      <named list [1]> <env>   
## 3         3 high        <named list [1]>      <named list [1]> <env>

Example 2: A branch within functions in a “pipe”: %>%

Let’s say that we have some data which indicates the amount of time spent by a user in four different conditions which are indexed 1, 2, 3 and 4 (the modality column in the following dataset). We will first load the data and convert the column into factor from integer.

data.userlogs.raw = userlogs %>%
  mutate( modality = factor(modality) ) %>%

We then create the multiverse object for this analysis

M = multiverse()

The data variable duration is approximately normally distributed, however, as is common with such data it has a long right-tail. Before using a regression model, we can choose to either log-transform the data, or assume that it is normally distribution and analyse it. We can perform both analysis in our multiverse by declaring two branches — one where the variable is log-transformed and the other where it is not.

```{multiverse default-m-2, inside = M}
df <- data.userlogs.raw %>%
    select(modality, duration) %>%
    mutate( duration = branch( data_transform, 
        "none" ~ duration,
        "log" ~ log(duration)

The expand shows that there are two rows for the data transform branch which correspond to the two different reasonable alternatives for transforming time spent data.

## # A tibble: 2 x 5
##   .universe data_transform .parameter_assignment .code            .results
##       <int> <chr>          <list>                <list>           <list>  
## 1         1 none           <named list [1]>      <named list [1]> <env>   
## 2         2 log            <named list [1]>      <named list [1]> <env>

If we look at the code column, which is a list of expressions, we see that the variable duration is defined in two separate ways, one for each row of the multiverse table. In other words, for each alternate analysis path, one of these would get executed and that would be the how the duration variable is transformed for that analysis path.

## [[1]]
## [[1]]$`h-2`
## {
##     df <- data.userlogs.raw %>% select(modality, duration) %>% 
##         mutate(duration = duration)
## }
## [[2]]
## [[2]]$`h-2`
## {
##     df <- data.userlogs.raw %>% select(modality, duration) %>% 
##         mutate(duration = log(duration))
## }

Example 3: A branch as a function

M = multiverse()

Say we are not familiar with R’s tidyverse packages and it’s style of writing code. An alternate way of declaring a branch for transforming the duration variable in the analysis would be to use two separate functions for the transformation depending on which universe one is.

```{multiverse default-m-3, inside = M}
duration <- branch(data_trans,
        "log-transformed" ~ log,
        "un-transformed" ~ identity

Then we execute the multiverse and inspect the results:


expand(M) %>%
  mutate(transformed_duration = map(.results, "duration"))
## # A tibble: 2 x 6
##   .universe data_trans   .parameter_assign… .code     .results transformed_dura…
##       <int> <chr>        <list>             <list>    <list>   <list>           
## 1         1 log-transfo… <named list [1]>   <named l… <env>    <dbl [512]>      
## 2         2 un-transfor… <named list [1]>   <named l… <env>    <dbl [512]>

As you can see, this creates, in each universe within the multiverse, a vector named duration which is either un-transformed or log-transformed. However, this suffers from being slightly unintuitive as it is harder to read or write the code. What is actually happening here is that the entire branch(...) gets replaced with either log() or identity() within a given universe.

A more readable pattern might be:

M = multiverse()

```{multiverse default-m-4, inside = M}
duration_transform <- branch(data_trans,
    "log-transformed" ~ log,
    "un-transformed" ~ identity

duration <- duration_transform(data.userlogs.raw$duration)

Next, we execute and inspect the results. Again, this alternate calculation also leads to the same result.


expand(M) %>%
  mutate( transformed_duration = map(.results, "duration" ) ) %>% 
  unnest(c(transformed_duration)) %>%
## # A tibble: 6 x 6
##   .universe data_trans   .parameter_assign… .code     .results transformed_dura…
##       <int> <chr>        <list>             <list>    <list>               <dbl>
## 1         1 log-transfo… <named list [1]>   <named l… <env>                 4.25
## 2         1 log-transfo… <named list [1]>   <named l… <env>                 4.80
## 3         1 log-transfo… <named list [1]>   <named l… <env>                 3.89
## 4         1 log-transfo… <named list [1]>   <named l… <env>                 3.75
## 5         1 log-transfo… <named list [1]>   <named l… <env>                 4.01
## 6         1 log-transfo… <named list [1]>   <named l… <env>                 3.41

Example 4: Parameters can be reused

If you want to use a parameter multiple times, or just want to define it upfront, you can assign it to a variable and use the variable in your code. For example, in the next example, we declare three different branches for data transformation, data modeling and data presentation respectively. We also define any necessary functions along the way.

samplemean <- function(x, d) {return(mean(x[d]))}

M = multiverse()

t_test_ci <- function(x, y) {
    broom::tidy(t.test(x, conf.level = y))

bootstrapped_ci <- function(x, y) {
    boot(data = x, statistic = samplemean, R = 5000) %>%
        broom::tidy( conf.int = TRUE, conf.method = "bca", conf.level =  y) %>%
        rename(estimate = statistic)
```{multiverse default-m-5, inside = M}
data_trans <- branch(data_trans,
      "log-transformed" ~ log,
      "un-transformed" ~ identity

  fit <- branch(bootstrap,
        "t-dist" ~ t_test_ci,
        "bca" ~ bootstrapped_ci
  ci <- branch( ci, 0.5, 0.68, 0.95, 0.99 )
  df <- data.userlogs.raw  %>%
    mutate( duration = data_trans(duration) ) %>% 
    mutate( modality.f = factor(modality) ) %>%
    group_by(subject, modality.f, modalityname) %>% 
    summarise( duration = mean(duration), .groups = "drop" ) %>%
    group_by(modality.f) %>%
    rename( value = duration ) %>%
    summarise( data = list(value), .groups = "keep" )
  df <- df %>%
    mutate( fit = map(data, ~ fit(.x, ci) ) ) %>%
    select(-data) %>%
    unnest( cols = c(fit) )

We then write the analysis and use those variables to indicate branches in our analysis within the multiverse.


To see the results of using branches, we plot a part of this multiverse. We see the different transformations (no transformations in the right column and log-transform in the left column). Each row shows the different data presentation branches by presenting different confidence levels.

expand(M) %>%
  mutate(df = map(.results, "df" ) ) %>%
  unnest( cols = c(df) )  %>%
  mutate( modality.f = recode( modality.f, 
                 "1" = "physical-touch",
                 "2" = "no-touch",
                 "3" = "virtual-prop",
                 "4" = "virtual-mouse"
  ) ) %>%
    estimate = ifelse(data_trans == "log", exp(estimate), estimate),
    conf.low = ifelse(data_trans == "log", exp(conf.low), conf.low),
    conf.high = ifelse(data_trans == "log", exp(conf.high), conf.high)
  ) %>%
  arrange(ci, desc(data_trans), desc(bootstrap)) %>%
  filter( bootstrap == "bca" ) %>%
  ggplot() + 
  #geom_vline( xintercept = 0,  colour = '#979797' ) +
  geom_point( aes(x = estimate, y = modality.f)) +
  geom_errorbarh( aes(xmin = conf.low, xmax = conf.high, y = modality.f), height = 0) +
  facet_grid(ci ~ data_trans, scales = "free")

Example 5: Declaring continuous variables in branch

So far we have only looked at examples where we declare parameters within a branch which takes discrete values. In most cases, it makes sense to declare them separately and if the number of values for a discrete variable is not too high, sufficiently easy and concise to declare them in this manner. However, this does not apply to continuous variables as declaring all possible values would be painstakingly difficult. Hence we provide a alternate way of declaring branches where the parameter takes a continuous sequence of values.

M = multiverse()
```{multiverse default-m-6, inside = M}
set.seed(branch(seed, .options = 1:25))

x1 <- rnorm(25)
x2 <- rnorm(25)
y <- x1 + x2 + runif(25)

m <- lm(y ~ x1+ x2)

intervals <- broom::tidy(m, conf.int = TRUE)

We then plot the estimates from the multiverse to see if we obtain the correct results:

expand(M) %>%
  mutate(df = map(.results, "intervals" ) ) %>%
  unnest( cols = c(df) ) %>%
  unnest(seed) %>%  # won't be necessary once issue #34 is fixed
  ggplot(aes(x = .universe, y = estimate, ymin = conf.low, ymax = conf.high), alpha = 0.5) +
  geom_pointrange() +
  facet_grid( . ~ term )  +
  scale_x_continuous( breaks = seq(0, 25, by = 5) ) +
  labs(x = "Universe (i.e. the seed parameter)", y = "Model Parameter Values", title = "Estimates with 95% confidence intervals")

Note: in this vignette we used the script-oriented inside() function for implementing the multiverse. However, we can implement the exact same multiverse in RMarkdown using the multiverse-code-block for more interactive programming. To implement this using a multiverse-code-block, we can simply place the code passed into the inside function (the second argument) inside a code block of type multiverse, provide it with the appropriate labels and multiverse object, and execute it. See (multiverse-in-rmd) and (branch) for more details and examples.