Multiverse case study #4

In this document we re-implement the specification curve analysis by Simonsohn et al. [http://dx.doi.org/10.2139/ssrn.2694998] using the Multiverse library.

Introduction

The specification curve analysis is in principle similar to a multiverse analysis, where all alternate specifications of a particular analysis asking the same research question are explored. In their study, Simonsohn et al. explore the robustness of the analysis by Jung et al. [https://doi.org/10.1073/pnas.1402786111], which investigated whether hurricanes with female sounding names are more deadlier than hurricanes with more male sounding names. We first begin by loading the dataset which is provided by the library. We then rename some of the variables and perform some data transformations which standardises some of the variables (mean = 0 and standard deviation = 1).

data("hurricane")

# read and process data
hurricane_data <- hurricane %>%
    # rename some variables
    rename(
        year = Year,
        name = Name,
        dam = NDAM,
        death = alldeaths,
        female = Gender_MF,
        masfem = MasFem,
        category = Category,
        pressure = Minpressure_Updated_2014,
        wind = HighestWindSpeed
    ) %>%
    # create new variables
    mutate(
        post = ifelse(year>1979, 1, 0),
        zcat = as.numeric(scale(category)),
        zpressure = -scale(pressure),
        zwind = as.numeric(scale(wind)),
        z3 = as.numeric((zpressure + zcat + zwind) / 3)
    )

Original analysis

We then illustrate an implementation of the original analysis by Jung et al. [https://doi.org/10.1073/pnas.1402786111]. The original analysis used a negative binomial model, which is suitable for overdispersed count data. Due to some issues with model fit with the MASS::glm.nb function (see Note 3: https://github.com/uwdata/boba/tree/master/example/hurricane), we instead use the simpler poisson regression model which will ensure that none of the models fail while fitting.

In the original analysis, Jung et al. exclude two hurricanes which caused the highest number of deaths (Katrina and Audrey) as outliers. They transform the variable used the interactions between the 11-point femininity rating and both damages and zpressure respectively, as seen below:

df <- hurricane_data %>%
    filter( name != "Katrina" & name != "Audrey" )

fit <- glm(death ~ masfem * dam + masfem * zpressure, data = df, family = "poisson")

Multiverse Analysis

To implement a multiverse analysis, we first need to create the multiverse object:

M <- multiverse()

Excluding outliers

In their implementation, Simonsohn et al. describe a principled method of excluding outliers based on extreme observations of death and damages. The consider it reasonable to exclude up two most extreme hurricanes in terms of death, and upto three most extreme hurricanes in terms of damages. This space of decisions is implemented using multiverse as follows:

Note

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.

```{multiverse default-m-1, inside = M}
df <- hurricane_data %>%
    filter(branch(death_outliers, 
        "no_exclusion" ~ TRUE,
        "most_extreme_deaths" ~ name != "Katrina",
        "most_extreme_two_deaths" ~ ! (name %in% c("Katrina", "Audrey"))
    )) %>%
    filter(branch(damage_outliers,
        "no_exclusion" ~ TRUE,
        "most_extreme_one_damage" ~ ! (name %in% c("Sandy")),
        "most_extreme_two_damage" ~ ! (name %in% c("Sandy", "Andrew")),
        "most_extreme_three_damage" ~ ! (name %in% c("Sandy", "Andrew", "Donna"))
    ))
```

Identifying independent variables

The next decision involves identifying the appropriate dependent variable for the primary effect — how do we operationalise femininity of the name of a hurricane. Simonsohn et al. identify two distinct ways. First, using the 11 point scale that was used in the original analysis; or second using a binary scale.

The other decision involved is whether or not to transform damages, another independent variable. damages follow a long tailed, positive only valued distribution.

We implement these two decisions in our multiverse as follows:

```{multiverse label = default-m-2, inside = M}
df <- df %>%
    mutate(
        femininity = branch(femininity_calculation,
          "masfem" ~ masfem,
          "female" ~ female
        ),
        damage = branch(damage_transform,
          "no_transform" ~ identity(dam),
          "log_transform" ~ log(dam)
        )
    )
```

Declaring alternative specifications of regression model

The next step is to fit the model. We can use either a log-linear model or a poisson model for the step. Both are reasonable alternatives for this dataset. We also have to make a choice on whether we want to include an interaction between femininity and damage. This results in the following specification:

broom::tidy(fit)
## # A tibble: 7 x 5
##   term              estimate std.error statistic  p.value
##   <chr>                <dbl>     <dbl>     <dbl>    <dbl>
## 1 (Intercept)        -4.05     0.463     -8.75   2.11e-18
## 2 femininity          0.0289   0.537      0.0538 9.57e- 1
## 3 damage              0.851    0.0536    15.9    7.97e-57
## 4 z3                 -0.745    0.105     -7.10   1.29e-12
## 5 femininity:damage  -0.0338   0.0603    -0.560  5.75e- 1
## 6 femininity:z3       0.852    0.109      7.79   6.87e-15
## 7 damage:post        -0.0687   0.00601  -11.4    3.33e-30
```{multiverse label = default-m-3, inside = M}
fit <- glm(branch(model, "linear" ~ log(death + 1), "poisson" ~ death) ~ 
          branch(main_interaction,
              "no" ~ femininity + damage,
              "yes" ~ femininity * damage
          ) + branch(other_predictors,
              "none" ~ NULL,
              "pressure" %when% (main_interaction == "yes") ~ femininity * zpressure,
              "wind" %when% (main_interaction == "yes") ~ femininity * zwind,
              "category" %when% (main_interaction == "yes") ~ femininity * zcat,
              "all" %when% (main_interaction == "yes") ~ femininity * z3,
              "all_no_interaction" %when% (main_interaction == "no") ~ z3
          ) + branch(covariates, "1" ~ NULL, "2" ~ year:damage, "3" ~ post:damage), 
          family = branch(model, "linear" ~ "gaussian", "poisson" ~ "poisson"),  
          data = df)

broom::tidy(fit)
```

Once we have implemented the analysis model in our multiverse, the corresponding step will be applied to each analysis path. To interpret the results, we first estimate a prediction interval corresponding to each analysis path.

```{multiverse label = default-m-4, inside = M}
pred <- predict(fit, se.fit = TRUE, type = "response")

pred2expectation <- function(mu, sigma) {
    branch(model, "linear" ~ exp(mu + sigma^2/2) - 1, "poisson" ~ mu)
}

disagg_fit <- df  %>%
    mutate(
        fitted = pred$fit,                                # add fitted predictions and standard errors to dataframe
        se.fit = pred$se.fit,
        deg_f = df.residual(fit),                         # get degrees of freedom
        sigma = sigma(fit),                               # get residual standard deviation
        se.residual = sqrt(sum(residuals(fit)^2) / deg_f) # get residual standard errors
    )

# aggregate fitted effect of female storm name
expectation <- disagg_fit %>%
    mutate(expected_deaths = pred2expectation(fitted, sigma)) %>% 
    group_by(female) %>%
    summarise(mean_deaths = mean(expected_deaths), .groups = "drop_last") %>% 
    compare_levels(mean_deaths, by = female)
```

Execution and Results

After we’ve specified our multiverse analysis, we would like to execute the entire multiverse, and view the results. Below, we plot the mean difference point estimate for expected deaths when a hurricane has a more feminine name, for each unique analysis path. We find that based on these arbitrary specifications of the multiverse, there is perhaps no relation between femininity of the name of a hurricane and the number of deaths that it causes, as some models predict a lower number of deaths, and some predict much higher.

execute_multiverse(M)

mean_deaths <- multiverse::expand(M) %>%
  extract_variables(expectation) %>%
  unnest(expectation)

mean_deaths %>%
  arrange(mean_deaths) %>%
  mutate(.id = 1:nrow(.)) %>%
  ggplot(aes(y = mean_deaths, x = .id)) +
  geom_point() +
  theme_minimal() +
  labs(x = "universe", y = "Mean difference in expected deaths")