Hey @erhla --

In the process of replicating the COD, PRD, and PRB statistics generated via the make_report function I noticed that all three IAAO stats were slightly different from what you'd get if you were to manually compute them. I believe bootstrapping is a good solution for estimating errors but unsure if that is the best choice for estimating the actual IAAO performance indices. Or alternatively perhaps what you can do is provide an option in the make_reports fucntion to generate the statistics without bootstrapping.

As a comparison this is what I got manually computing them without bootstrapping vs using make_report:

• COD 40.4 (manual) vs 34.55 (make_report)
• PRD 1.12 (manual) vs 0.843, 1.058, 1.312, 1.177, 1.027 (make_report .. this was least stable)
• PRB -0.1071028 (manual) vs -0.088 (make_report)

Below is the function I used modeled directly after your code w/o bootstrapping (referencing this: https://github.com/cmf-uchicago/cmfproperty/blob/master/R/iaao_stats.R#L1-L33)

#' @param data dataframe of data
#' @param assessment_value_col string of assessment_value_col name
#' @param sale_price_col string of sale_price_col name
#' @return dataframe 
gen_iaao_stats <- function(data, assessment_value_col, sale_price_col) { 
  df_iaao <- data %>% 
    rename_at(vars(c(assessment_value_col, sale_price_col)), function(x) c('assessment_value','sale_price')) %>% 
    mutate(av_ratio = assessment_value/sale_price,
           log2 = log(2),
           count = 1) %>%
    mutate(cod = 100 * sum(abs(av_ratio - stats::median(av_ratio)))/(n() * stats::median(av_ratio)),
           prd = mean(av_ratio, na.rm = TRUE)/stats::weighted.mean(av_ratio, sale_price, na.rm = TRUE),
           prb_value = 0.50 * (assessment_value/median(av_ratio)) + 0.50 * sale_price,
           prb_ln_value = log(prb_value)/log(2),
           prb_pct_diff = (av_ratio - median(av_ratio))/median(av_ratio)
    )
  prb_model <- linear_reg() %>% 
    fit(prb_pct_diff ~ prb_ln_value,  data = df_iaao) %>%
    tidy(., conf.int = TRUE) 
  iaao_out <- list('cod' = c(unique(df_iaao$cod)) ,
                   'prd' = c(unique(df_iaao$prd)) ,
                   'prb' = c(prb_model %>% filter(term == 'prb_ln_value') %>% select(estimate) %>% distinct() %>% pull())) %>%
    as.data.frame()
  return(iaao_out)
}

test <- gen_iaao_stats(data = data_input, assessment_value_col = 'ASSESSEDVALUE', sale_price_col = 'sale_price')

Hi @erhla, I recently used cmfproperty to analyze sales data for St. Louis County, Mo., for the period 2010 to 2023.

Everything worked well, but I am puzzled by one aspect of the final report that was generated.

The diagram for the Coefficient of Dispersion in our report shows a very large confidence interval -- far bigger than the CI on the COD chart for St. Louis County on the CMF's 2020 report online. (So big that it's changing the Y-axis scale)

Here's a comparison:

On CMF website	In our new report

None of the other diagrams in our new report have this issue. All the other diagrams' confidence intervals are very tightly aligned to the data lines, similar to those in the 2020 CMF report.

Anyway, I'm happy to send the source sales data we used (it's a 17mb CSV), or to share my code.

README

• Add links to https://erhla.github.io/cmfproperty/
• Reduce to simple example

pkgdown

• Add tab for example report & example report w/ diagnostics
• Update reference page to group functions into data processing, regressivity tests, plots, and report
• Link vignettes to each other when appropriate

vignettes

• preprocessing---more detailed step-by-step example of how to get data in correct format
• methods---example of output applied for each function & latex equations behind functions (include arm's length/inflation adj) rename to Methods?
• list of figures---example of output applied for each function
• Get Started---motivation and overview

plots

• split plots up into different sections

regressivity_tests

• update documentation to refer to Methods