In gergness/srvyr: 'dplyr'-Like Syntax for Summary Statistics of Survey Data

This vignette is for testing only and excluded from the R build.

Indexing vs not indexing database objects on order variable

Setup

Difficult to run more than once, so there may be some variance. However, I see that in general, survey package databases are about 2X faster than srvyr. Adding the index takes a slight penalty for SQLite, not for MonetDB (might just be ignoring it). MonetDBLite is slightly faster than SQLite.

library(survey)
library(srvyr)
library(dplyr)
library(microbenchmark)
library(tidyr)
library(ggplot2)
library(scales)
library(MonetDBLite)
library(RSQLite)

micro_times <- 10

set.seed(1)

num_obs <- c(1e5)
num_vars <- 10
num_strata <- 10


data <- data.frame(strata = sample(paste0("st", seq_len(num_strata)), num_obs, replace = TRUE), 
                   probs = runif(num_obs))
data[, c(3:(2 + num_vars))] <- runif(num_obs * num_vars) + rep(seq_len(num_vars), each = num_obs)
data$uid <- order(data$probs) # Random order of uid
# data$uid <- seq_len(nrow(data)) # Sorted order of uid
names(data) <- tolower(names(data))

# Database setup
monetdb_dir <- paste0(tempdir(), "/monetdb")
dir.create(monetdb_dir)
monetdb <- src_monetdblite(monetdb_dir)

sqlitedb_dir <- paste0(tempdir(), "/sqlite/")
dir.create(sqlitedb_dir)
sqlitedb <- src_sqlite(paste0(sqlitedb_dir, "db-file"), create = TRUE)


svys <- list()
setup_times <- 
  microbenchmark(
    `svy_local setup` = {
      svys[["svy_local"]] <- 
        svydesign(~1, strata = ~strata, probs = ~probs, data = data)
    },
    `srvyr_local setup` = {
      svys[["srvyr_local"]] <- 
        data %>% as_survey_design(strata = strata, probs = probs)
    },
    `svy_monetdb setup` = {
      svy_monetdb_data <- copy_to(monetdb, data, name = "svy", temporary = FALSE)

      svys[["svy_monetdb"]] <- svydesign(~1, strata = ~strata, probs = ~probs, data = "svy", 
                        dbtype = "MonetDBLite", dbname = monetdb_dir)
    },
    `srvyr_monetdb_no_index setup` = {
      srvyr_monetdb_no_index_data <- copy_to(monetdb, data, name = "srvyr_no_index", temporary = FALSE)

      svys[["srvyr_monetdb_no_index"]] <- 
        srvyr_monetdb_no_index_data %>% as_survey_design(strata = strata, probs = probs, 
                                                         uid = uid)
    },
    `srvyr_monetdb_index setup` = {
      srvyr_monetdb_index_data <- copy_to(monetdb, data, name = "srvyr_index", temporary = FALSE)
      index_status <- DBI::dbSendQuery(monetdb$con, "CREATE INDEX uid ON srvyr_index (uid)")

      svys[["srvyr_monetdb_index"]] <- 
        srvyr_monetdb_index_data %>% as_survey_design(strata = strata, probs = probs, 
                                                      uid = uid)    
      }, 
    `svy_sqlitedb setup` = {
      svy_sqlitedb_data <- copy_to(sqlitedb, data, name = "svy", temporary = FALSE)

      svys[["svy_sqlitedb"]] <- svydesign(~1, strata = ~strata, probs = ~probs, data = "svy", 
                        dbtype = "SQLite", dbname = paste0(sqlitedb_dir, "/db-file"))
    },
    `srvyr_sqlitedb_no_index setup` = {
      srvyr_sqlitedb_no_index_data <- copy_to(sqlitedb, data, name = "srvyr_no_index", temporary = FALSE)

      svys[["srvyr_sqlitedb_no_index"]] <- 
        srvyr_sqlitedb_no_index_data %>% as_survey_design(strata = strata, probs = probs, 
                                                         uid = uid)
    },
    `srvyr_sqlitedb_index setup` = {
      srvyr_sqlitedb_index_data <- copy_to(sqlitedb, data, name = "srvyr_index", temporary = FALSE)
      index_status <- DBI::dbSendQuery(sqlitedb$con, "CREATE INDEX uid ON srvyr_index (uid)")
      #index_status <- DBI::dbSendQuery(sqlitedb$con, "CREATE INDEX uid_v3 ON srvyr_index (uid, v3)")

      svys[["srvyr_sqlitedb_index"]] <- 
        srvyr_sqlitedb_index_data %>% as_survey_design(strata = strata, probs = probs, 
                                                      uid = uid)    
      },

    times = 1, unit = "s")

setup_times

Ungrouped Operations

Survey's implementation is faster (nearly as fast local). Srvyr's a fair amount slower. Index only matters if you include all of the variables of interest in the index...

mean_function <- function(svy) {
  if (inherits(svy, "tbl_svy")) {
    svy %>% summarize(x = survey_mean(v3))
  } else {
    svymean(~v3, svy)
  }
}

mean_times <- microbenchmark(
  `svy_local mean` = mean_function(svys[["svy_local"]]), 
  `srvyr_local mean` = mean_function(svys[["srvyr_local"]]), 
  `svy_monetdb mean` = mean_function(svys[["svy_monetdb"]]), 
  `srvyr_monetdb_no_index mean` = mean_function(svys[["srvyr_monetdb_no_index"]]), 
  `srvyr_monetdb_index mean` = mean_function(svys[["srvyr_monetdb_index"]]), 
  `svy_sqlitedb mean` = mean_function(svys[["svy_sqlitedb"]]), 
  `srvyr_sqlitedb_no_index mean` = mean_function(svys[["srvyr_sqlitedb_no_index"]]), 
  `srvyr_sqlitedb_index mean` = mean_function(svys[["srvyr_sqlitedb_index"]]), 
  times = micro_times, unit = "s")

mean_times

Grouped Operations

mean_function <- function(svy) {
  if (inherits(svy, "tbl_svy")) {
    svy %>% group_by(strata) %>% summarize(x = survey_mean(v3))
  } else {
    svymean(~v3, svy, byvar = ~strata)
  }
}

grouped_mean_times <- microbenchmark(
  `svy_local mean` = mean_function(svys[["svy_local"]]), 
  `srvyr_local mean` = mean_function(svys[["srvyr_local"]]), 
  `svy_monetdb mean` = mean_function(svys[["svy_monetdb"]]), 
  `srvyr_monetdb_no_index mean` = mean_function(svys[["srvyr_monetdb_no_index"]]), 
  `srvyr_monetdb_index mean` = mean_function(svys[["srvyr_monetdb_index"]]), 
  `svy_sqlitedb mean` = mean_function(svys[["svy_sqlitedb"]]), 
  `srvyr_sqlitedb_no_index mean` = mean_function(svys[["srvyr_sqlitedb_no_index"]]), 
  `srvyr_sqlitedb_index mean` = mean_function(svys[["srvyr_sqlitedb_index"]]), 
  times = micro_times, unit = "s")

grouped_mean_times