R/app.R

In skranz/MetaStudies: Shiny app and function for Meta studies following Andrews and Kasy (2019).

#' Create and run the MetaStudies shiny app
#'
#' @param csv.file if not NULL a file that will be initially used by the app
#' @param show.cor if TRUE add a tabPanel with correlations between estimates and standard
#'                 errors to get better insights whether the independence assumption
#'                 of Andrews & Kasy (2019) may be violated in the given data set.
#' @param ... additional parameters passed to \code{shinyApp}
MetaStudiesApp = function(csv.file=NULL,show.cor = TRUE,...) {
  res.ui = tagList(
    h2("Model estimates", align = "center"),
    h4("Distribution of true effects, conditional publication proabilities", align = "center"),
    column(12, align="center",
           tableOutput("estimatestable"),
           plotOutput("estplot", width = "70%")
    )
  )

  if (show.cor) {
    lower.ui = tabsetPanel(
      tabPanel("Results",res.ui),
      tabPanel("Specification tests",
        h3("Estimated correlations between estimate and standard deviation in latent distribution"),
        tableOutput("cortable"),
        helpText("
Explanation: One crucial assumption of the Andrews and Kasy (2019) approach is that in the latent distribution without publication bias the estimate and its standard error are statistically independent across tests.

While the latent distribution cannot be observed, we use an inverse probability weighting approach to estimate the correlation inside the latent distribution.
More precisely, we compute a weighted correlation, weighting each observation inversely with its estimated publication probability. Assuming all
assumptions are satisfied, this should yield a consistent estimator of the correlation in the unobserved latent distribution.

We show the correlation of absolute estimates and standard errors both in levels and in logs. To compute confidence intervals for the estimated correlations, you can call manually the function 'bootstrap_specification_tests' in the package MetaStudies (the computation can be quite slow).
")
      )
    )
  } else {
    lower.ui = tagList(hr(), res.ui)
  }


  ui <- fluidPage(
    # titlePanel(h1("Optimal treatment assignment given covariates", align = "center")),
    sidebarLayout(
      sidebarPanel(
        fileInput("file1", "Choose CSV File of estimates and standard errors",
          multiple = FALSE,
          accept = ".csv"
        ),
#          accept = c("text/csv",
#                     "text/comma-separated-values,text/plain",
#                     ".csv")),
        hr(),
        h3("Model parameters"),
        fluidRow(
          column(6, checkboxInput("symmetric", "Symmetric p(.)", value = FALSE),
                    checkboxGroupInput("cutoffs", "Cutoffs for p(.)",
                                          choiceNames = c("1.65", "1.96","2.33"),
                                          choiceValues = c(1.645, 1.960, 2.326),
                                    selected = 1.960)),
          column(6, radioButtons("modelmu", "Model for the distribution of effects",
                                 choices = c("Normal" = "normal",
                                             "Student-t" = "t" #,
                                                #"nonparametric" = 3
                                             ),
                                             selected = "normal"))
        ),
        #hr(),
        actionButton(inputId = "estimatebutton", label = "Estimate model")
      ),

      mainPanel(
        h2("Funnel plot, histogram of z-stats", align = "center"),
        fluidRow(splitLayout(cellWidths = c("50%", "50%"),
                      plotOutput("funnel"),
                      plotOutput("hist")))
      )
    ),
    lower.ui
  )

  server <- function(input, output, session) {
    #object to store data and estimation results
    v = reactiveValues()

    load.data = function(file) {
      metadata=read.table(file,sep=",")
      v$X=metadata[,1]
      v$sigma=metadata[,2]
    }

    observe({
      req(input$file1)
      file = input$file1$datapath
      load.data(file)
    })


    if (!is.null(csv.file)) {
      #v$file = csv.file
      load.data(csv.file)
    }

    # read data, generate funnel plot
    output$funnel=renderPlot({
      req(v$X)
      metastudies_funnel_plot(v$X,v$sigma)
    })

    # generate histogram
    output$hist=renderPlot({
      req(v$X)
      z_histogram(v$X, v$sigma)
    })

    #estimation
    observeEvent(input$estimatebutton,{
      v$cutoffs=as.numeric(unlist(input$cutoffs))
      v$symmetric=input$symmetric
      v$modelmu=input$modelmu
      if (!v$symmetric) v$cutoffs= c(-rev(v$cutoffs), 0 ,v$cutoffs)

      v$ms=metastudies_estimation(v$X,v$sigma,v$cutoffs, v$symmetric, model= v$modelmu)

      if (show.cor) {
        v$cors = metastudy_X_sigma_cors(v$ms)
      }

      output$estplot = renderPlot({
          estimates_plot(v$ms)
      })
    })

    #render estimates to table
    output$estimatestable =  renderTable(rownames=TRUE,hover=TRUE,digits=3,
     {req(v$ms)
      v$ms$est_tab
     })

    if (show.cor) {
      # show correlations for specification tests
      output$cortable =  renderTable(rownames=TRUE,hover=TRUE,digits=3, striped=TRUE,{
        req(v$cors)
        tab = v$cors %>%
          filter(mode=="ipv") %>%
          mutate(
            cor = round(cor,3),
            label = case_when(
              trans=="log" ~ "Logarithms",
              TRUE ~ "Levels"
            ),

            #label = case_when(
            #  mode=="ipv" ~ "Estimated latent distribution",
            #  TRUE ~ paste0("Observed distribution in range ",mode)
            #),
            #ci = case_when(
            #  mode == "ipv" ~ "Not computed",
            #  TRUE ~ paste0("[",round(conf.cor.low,3),", ",round(conf.cor.up,3),"]")
            #)
          ) %>%
          select(`Transformation`=label, `Correlation`=cor)
        tab
      })
    }

  }
  shinyApp(ui = ui, server = server,...)
}