🖂 Back of the Envelope 📉

https://acm9q.shinyapps.io/Back_of_the_Envelope/

Can we build a ~~basic point-and-click regression analysis tool~~ replacement to general purpose statistical analysis tools with R shiny?

The goal:

Build a general purpose regression tool, incorporating extensions to regression such as heteroskedasticity-robust standard errors, clustered standard errors, multilevel modeling, and logistic regression.

With sufficient on-your-own data preparation, this tool should be sufficient for basic regression analyses for beginner-to-intermediate level social-science use. I have no plans to implement latent-variable / structural equation modeling at this time, but path analysis could be in the distant future.

To Do Items

1.0 To Do List:

fix issue with missing points whenever there are residuals? what's that about??
- something like, for each variable selected, select_if(is_extant)? this should already be working in the back end. Annoying.
- https://tidyr.tidyverse.org/reference/drop_na.html
- This is dropping too much data--needs to only drop the chosen variables

Current Problem:

Honestly

The key may to take all the model output objects that are generated and to standardize their output into my own bespoke formatting, and then render that into everything else... But it sure does sound like a pain.

Status of what doesn't work:

Added Variable Plots (needs broom::augment() for lm_robust) DeclareDesign/estimatr#377
see this response; DeclareDesign/estimatr#377 (comment)
Plot Residuals (needs broom::augment() for lm_robust) DeclareDesign/estimatr#377
model diagnostics: (needs broom::augment() for lm_robust)
- QQ plot
- resid v fitted DeclareDesign/estimatr#377
- histogram of residuals DeclareDesign/estimatr#377
- cook's distance (https://stackoverflow.com/questions/62107571/how-to-manually-calculate-cooks-distance)
{report}

The Plan:

1.1 To Do List

#ui

radioButtons('format', h5('Document format'), c('PDF', 'HTML', 'Word'), inline = TRUE),
downloadButton('downloadReport'),

#server
   output$downloadReport <- downloadHandler(
      filename = function() {
         paste('my-report', sep = '.', switch(
            input$format, PDF = 'pdf', HTML = 'html', Word = 'docx'
         ))
      },
      content = function(file) {
         src <- normalizePath('report.Rmd')
         owd <- setwd(tempdir())
         on.exit(setwd(owd))
         file.copy(src, 'report.Rmd')
         
         library(rmarkdown)
         out <- render('report.Rmd', switch(
            input$format,
            PDF = pdf_document(), HTML = html_document(), Word = word_document()
         ))
         file.rename(out, file)
      })

Downloadable plots:

Example:

plot_server <- function(id, df, vbl, threshhold = NULL) {
  
  moduleServer(id, function(input, output, session) {
    
    plot <- reactive({viz_monthly(df(), vbl, threshhold)})
    output$plot <- renderPlot({plot()})
    output$dnld <- downloadHandler(
      filename = function() {paste0(vbl, '.png')},
      content = function(file) {ggsave(file, plot())}
    )
    
  })
}

1.2 To Do List

The Deep Future To Do List

Binary Outcome Mixed Effects
Multiple Imputation with Chained Equations
specify which variable is the ID variable; then allow users to plot Id variables instead of points.

User Feedback:

~~For the Correlation table, you may want to rotate your x-axis labels 45 or 90 degrees. Getting a lot of overlap for files with > 20 factors~~
Is there a way that you can override or modify the error messages? Instead of "contact the app author", maybe provide a URL to a message board or email?
- MAKE ERRORS GREAT AGAIN
~~my big suggestion is just to clarify and restructure the flow of the user interface. I think my user preference is that I'd want to upload a dataset, look around in it, and then decide on a model~~
- so maybe separate the Upload and Model pieces entirely. Then restructure the left-hand nav to be something like Upload > View Data Set > Descriptive Statistics > Correlation Table > Model > Summary > Plots > Diagnostics
- ~~some suggestions in there to make the nav header more descriptive, and put summary before plots just so we get the immediate output of the model.~~ with this, might even be worthwhile to put the model + summary on the same page actually
my last thought would be that it might be cool to allow for dplyr-style filtering of the uploaded dataset - I think it would be relatively straightforward, but also legit if you don't want to include that functionality as it could also be an enormous pain in the ass to try and catch edge-cases
- (in re dplyr style filtering: editable data tables are now possible, but that's a feature I have planned to work on after I squash all the inital bugs. it's gonna go: squash bugs, ~~added variable plots~~, finish outliers, HLM, instrumental variables, THEN data processing)
ooh, before I forget: I might also have a disclaimer or something re: what you do with the uploaded datasets. could spook some people when you actually have people using it for not just testing purposes

With a sufficient amount of effort, this could actually be a pretty good tool for multilevel linear modeling, ~~provided I can figure out what exactly the error was with mice's MCMC imputation.~~ imputation is gonna have to wait for v1.2 at the absolute earliest.

Domain issues

Better names:

www.backoftheenvelope.com (is currently taken) www.envelope.fyi www.envelo.pe

material:

Reading Material

Watching Material

https://rstudio.com/resources/webinars/testing-shiny-applications-with-shinytest-shiny-developers-now-have-tools-for-automated-testing-of-complete-applications/ https://rstudio.com/resources/webinars/introducing-shiny-gadgets-interactive-tools/ https://rstudio.com/resources/webinars/interactive-graphics-with-shiny/ https://rstudio.com/resources/webinars/help-me-help-you-creating-reproducible-examples/ https://rstudio.com/resources/webinars/scaling-shiny-apps-with-asynchronous-programming/

it's been done:

but not with user-input data: https://rich.shinyapps.io/regression/ (~~use this for model work~~) bruh use this for downloadable reports work

McCartneyAC/Back_of_the_Envelope