[16]{.chapter-number}  [Meta-Analysis]{.chapter-title}

16 Meta-Analysis

Bayesian meta-analysis is a statistical method that combines information from multiple studies to estimate the overall effect size of an intervention or variable of interest. Bayesian methods incorporate prior information into the analysis, updating these beliefs with new data to produce posterior distributions. This approach provides a flexible and dynamic means of integrating evidence, allowing for more informed and adaptive decision-making.

Imagine having access to a wealth of findings from various studies that have explored the impact of a specific intervention on your company’s sales. Each study offers an estimate of the effect size, but these estimates might differ due to variations in sample size, study design, and other contributing factors. Bayesian meta-analysis steps in to harmonize these disparate estimates into a unified overall effect size. It adeptly takes into account the relative precision and reliability of each study, ensuring a balanced and informed assessment. Furthermore, with sufficient data, this methodology allows you to delve deeper and estimate the probability that specific study characteristics are driving heterogeneity in the observed impact of the intervention.

16.1 A Simple Example Using Synthetic Data

Let’s ground this concept with a practical scenario. Suppose your objective is to estimate the incremental lift of an intervention, and you have access to multiple studies that have investigated its effects. Faced with a variety of estimates, how would you determine which one to rely on for your decision-making? While you might be tempted to simply choose the number you prefer, mentally combine the estimates, or calculate a simple average, a more rigorous approach is to conduct a Bayesian meta-analysis.

library(dplyr)
library(ggplot2)

studies <- tibble(
  study = LETTERS[1:5],
  lift_hat = c(0.09, 0.06, 0.06, 0.07, 0.04),
  std_err = c(0.01, 0.01, 0.007, 0.007, 0.01)
) %>%
  mutate(
    LB = lift_hat - 1.96 * std_err,
    UB = lift_hat + 1.96 * std_err
  )

library(ggiraph)

tooltip_css <-
  "background-color:gray;color:white;font-style:italic;padding:10px;border-radius:5px;font-size:16px;"


studies <- studies %>%
  mutate(
    tooltip = glue::glue(
      "lift is {scales::percent(lift_hat)} with a 95% confidence interval between {scales::percent(LB, accuracy=1)} and {scales::percent(UB, accuracy=1)}"
    )
  )

my_plot <- ggplot(data = studies, aes(x = study, y = lift_hat)) +
  geom_pointrange_interactive(aes(
    ymin = LB,
    ymax = UB,
    tooltip = tooltip
  )) +
  scale_y_continuous(labels = scales::percent_format(accuracy = 1)) +
  ylab("Lift") +
  xlab("Study") +
  theme_bw(base_size = 20)

girafe(
  code = print(my_plot),
  pointsize = 20,
  width_svg = 7,
  height_svg = 3.5,
  options = list(opts_tooltip(css = tooltip_css))
)

Given these studies, what insights can you glean? Let’s say you need to make a decision based on whether the lift is at least 5%. Would you be able to confidently determine the answer?

To address this, you can fit a straightforward model to the data:

\[ \begin{aligned} y_i &\sim N(\theta_i,s_i) \\ \theta_i &\sim N(\mu, \tau) \\ \mu &\sim N(0, 0.05) \\ \tau &\sim N^+(0, 0.05) \end{aligned} \] where:

$y_i$ is the estimated lift in study $i$
$s_i$ is the standard error of the lift in study $i$
$\theta_i$ is the true lift in study $i$
$\mu$ is the true lift in the population
$\tau$ is the standard deviation in lift

To fit this model you can use im::metaAnalysis():

library(im)

test_meta <- metaAnalysis$new(data = studies, point_estimates = lift_hat,
                              standard_errors = std_err, id = study)

Calculating the probability that the lift is at least 5% becomes remarkably simple:

test_meta$probability(a = 0.05)

The probability that lift is more than 5% is 88%.

To visualize the insights gained from the meta-analysis, you can plot the posterior probability:

# Plot the lift's prior and posterior distributions
test_meta$PlotLift(
  breaks = c(0, 0.01, 0.05, 0.1),
  break_names = c(" < 0", "(0,1%)", "(1%,5%)", "(5%,10%)", "> 10%"),
  display_mode_name = TRUE
)

16.2 A more complex meta-anlysis

Let’s delve deeper into the realm of Bayesian meta-analysis by incorporating study-specific characteristics that can offer richer insights into the factors influencing the impact of interventions. Imagine that each study in your collection comes with valuable metadata, such as the geographical location where the intervention was implemented, the specific modality of the intervention, or any other relevant attribute. By weaving these details into our model, we can uncover whether the impact varies significantly across different locations or modalities.

To achieve this, we can refine our model as follows:

\[ \begin{aligned} y_i & \sim N(\theta_i,s_i) \\ \theta_i & \sim N(\mu_i, \tau) \\ \mu_i & = \mu_0 + X\beta \end{aligned} \]

In this augmented model $X$ is a matrix of study-specific characteristics. By fitting this model, we can not only estimate the overall effect size but also discern whether the impact is significantly higher or lower for specific locations, modalities, or any other characteristic captured in the metadata. This granular understanding allows for more targeted decision-making, enabling you to tailor interventions to specific contexts and maximize their effectiveness.

However, it is crucial to remember that the quality of your meta-analysis hinges on the quality of the data you feed into it. As with any statistical model, the adage “garbage in, garbage out” holds true. If the underlying studies are flawed or biased, the meta-analysis will not magically erase those imperfections. For instance, to conduct even a simple meta-analysis, you ideally need at least five high-quality randomized controlled trials (RCTs) to ensure robust results.

In essence, Bayesian meta-analysis acts as a versatile instrument, empowering you to harmonize diverse sources of evidence, account for heterogeneity, and extract actionable insights from study-level characteristics. As you continue your exploration of causal inference in the tech industry, remember that Bayesian methods offer a robust framework for navigating uncertainty, optimizing interventions, and driving impactful decisions that propel your company forward. However, the success of this endeavor rests on the foundation of sound data and rigorous study design.

Learn more

Gelman et al. (2013) Parallel Experiments in Eight Schools.

--- title: "Meta-Analysis" share: permalink: "https://book.martinez.fyi/meta-analysis.html" description: "Business Data Science: What Does it Mean to Be Data-Driven?" linkedin: true email: true mastodon: true --- Bayesian meta-analysis is a statistical method that combines information from multiple studies to estimate the overall effect size of an intervention or variable of interest. Bayesian methods incorporate prior information into the analysis, updating these beliefs with new data to produce posterior distributions. This approach provides a flexible and dynamic means of integrating evidence, allowing for more informed and adaptive decision-making. Imagine having access to a wealth of findings from various studies that have explored the impact of a specific intervention on your company's sales. Each study offers an estimate of the effect size, but these estimates might differ due to variations in sample size, study design, and other contributing factors. Bayesian meta-analysis steps in to harmonize these disparate estimates into a unified overall effect size. It adeptly takes into account the relative precision and reliability of each study, ensuring a balanced and informed assessment. Furthermore, with sufficient data, this methodology allows you to delve deeper and estimate the probability that specific study characteristics are driving heterogeneity in the observed impact of the intervention. ## A Simple Example Using Synthetic Data Let's ground this concept with a practical scenario. Suppose your objective is to estimate the incremental lift of an intervention, and you have access to multiple studies that have investigated its effects. Faced with a variety of estimates, how would you determine which one to rely on for your decision-making? While you might be tempted to simply choose the number you prefer, mentally combine the estimates, or calculate a simple average, a more rigorous approach is to conduct a Bayesian meta-analysis. ```{r basic_plot, warning=FALSE, message=FALSE} library(dplyr) library(ggplot2) studies <- tibble( study = LETTERS[1:5], lift_hat = c(0.09, 0.06, 0.06, 0.07, 0.04), std_err = c(0.01, 0.01, 0.007, 0.007, 0.01) ) %>% mutate( LB = lift_hat - 1.96 * std_err, UB = lift_hat + 1.96 * std_err ) ``` ::: {.content-visible when-format="pdf"} ```{r static} ggplot(data = studies, aes(x = study, y = lift_hat)) + geom_pointrange(aes( ymin = LB, ymax = UB )) + scale_y_continuous(labels = scales::percent_format(accuracy = 1)) + ylab("Lift") + xlab("Study") + theme_bw(base_size = 20) ``` ::: ::: {.content-visible when-format="html"} ```{r interactive, fig.height=7, fig.width=14, message=FALSE, warning=FALSE} #| eval: !expr knitr::is_html_output() library(ggiraph) tooltip_css <- "background-color:gray;color:white;font-style:italic;padding:10px;border-radius:5px;font-size:16px;" studies <- studies %>% mutate( tooltip = glue::glue( "lift is {scales::percent(lift_hat)} with a 95% confidence interval between {scales::percent(LB, accuracy=1)} and {scales::percent(UB, accuracy=1)}" ) ) my_plot <- ggplot(data = studies, aes(x = study, y = lift_hat)) + geom_pointrange_interactive(aes( ymin = LB, ymax = UB, tooltip = tooltip )) + scale_y_continuous(labels = scales::percent_format(accuracy = 1)) + ylab("Lift") + xlab("Study") + theme_bw(base_size = 20) girafe( code = print(my_plot), pointsize = 20, width_svg = 7, height_svg = 3.5, options = list(opts_tooltip(css = tooltip_css)) ) ``` ::: Given these studies, what insights can you glean? Let's say you need to make a decision based on whether the lift is at least 5%. Would you be able to confidently determine the answer? To address this, you can fit a straightforward model to the data: $$ \begin{aligned} y_i &\sim N(\theta_i,s_i) \\ \theta_i &\sim N(\mu, \tau) \\ \mu &\sim N(0, 0.05) \\ \tau &\sim N^+(0, 0.05) \end{aligned} $$ where: - $y_i$ is the estimated lift in study $i$ - $s_i$ is the standard error of the lift in study $i$ - $\theta_i$ is the true lift in study $i$ - $\mu$ is the true lift in the population - $\tau$ is the standard deviation in lift To fit this model you can use `im::metaAnalysis()`: ```{r meta-analysis, warning=FALSE, results = "hide"} library(im) test_meta <- metaAnalysis$new(data = studies, point_estimates = lift_hat, standard_errors = std_err, id = study) ``` Calculating the probability that the lift is at least 5% becomes remarkably simple: ```{r probability} test_meta$probability(a = 0.05) ``` ::: {.content-visible when-format="pdf"} To visualize the insights gained from the meta-analysis, you can plot the posterior by calling `test_meta$PlotLift`. ::: ::: {.content-visible when-format="html"} To visualize the insights gained from the meta-analysis, you can plot the posterior probability: ```{r} #| eval: !expr knitr::is_html_output() # Plot the lift's prior and posterior distributions test_meta$PlotLift( breaks = c(0, 0.01, 0.05, 0.1), break_names = c(" < 0", "(0,1%)", "(1%,5%)", "(5%,10%)", "> 10%"), display_mode_name = TRUE ) ``` ::: ## A more complex meta-anlysis Let's delve deeper into the realm of Bayesian meta-analysis by incorporating study-specific characteristics that can offer richer insights into the factors influencing the impact of interventions. Imagine that each study in your collection comes with valuable metadata, such as the geographical location where the intervention was implemented, the specific modality of the intervention, or any other relevant attribute. By weaving these details into our model, we can uncover whether the impact varies significantly across different locations or modalities. To achieve this, we can refine our model as follows: $$ \begin{aligned} y_i & \sim N(\theta_i,s_i) \\ \theta_i & \sim N(\mu_i, \tau) \\ \mu_i & = \mu_0 + X\beta \end{aligned} $$ In this augmented model $X$ is a matrix of study-specific characteristics. By fitting this model, we can not only estimate the overall effect size but also discern whether the impact is significantly higher or lower for specific locations, modalities, or any other characteristic captured in the metadata. This granular understanding allows for more targeted decision-making, enabling you to tailor interventions to specific contexts and maximize their effectiveness. However, it is crucial to remember that the quality of your meta-analysis hinges on the quality of the data you feed into it. As with any statistical model, the adage "garbage in, garbage out" holds true. If the underlying studies are flawed or biased, the meta-analysis will not magically erase those imperfections. For instance, to conduct even a simple meta-analysis, you ideally need at least five high-quality randomized controlled trials (RCTs) to ensure robust results. In essence, Bayesian meta-analysis acts as a versatile instrument, empowering you to harmonize diverse sources of evidence, account for heterogeneity, and extract actionable insights from study-level characteristics. As you continue your exploration of causal inference in the tech industry, remember that Bayesian methods offer a robust framework for navigating uncertainty, optimizing interventions, and driving impactful decisions that propel your company forward. However, the success of this endeavor rests on the foundation of sound data and rigorous study design. ::: {.callout-tip} ## Learn more @gelman2013bayesian Parallel Experiments in Eight Schools. :::