In potterzot/livingwage: Living Wage Employment Analysis

Living Wage

Using Bayesian Structural time-series models to estimate the causal impact of the living wage on employment.

Installation

devtools::install_github("potterzot/livingwage")

Quick Example for Santa Fe

library(CausalImpact)
library(livingwage)
library(tidyr)
library(dplyr)
library(zoo)
data(bls)

#Load the data
sf <- filter(bls, series %in% c("Santa Fe MSA", "Albuquerque MSA")) %>%
  spread(key = series, value = employees)
names(sf)[4:5] <- c("abq", "sf")
dates <- as.Date(paste(sf$year, sf$periodName, "01", sep="-"), "%Y-%B-%d")
sf$date <- dates
dat <- zoo(sf[, c("sf", "abq")], dates)

#Define the pre and post period
#Santa Fe's $8.50 living wage went into effect in June 2004
pre_pd <- as.Date(c("2001-06-01", "2004-06-01"))
post_pd <- as.Date(c("2004-07-01", "2006-07-01"))

#Model
args <- list(nseasons = 4, season.duration = 3)
impact <- CausalImpact(dat, pre_pd, post_pd, model.args = args)
plot(impact)    
#> Warning: Removed 109 rows containing missing values (geom_path).
#> Warning: Removed 109 rows containing missing values (geom_path).
#> Warning: Removed 109 rows containing missing values (geom_path).

Slightly More Complex Example

library(bsts)

#Set the post-period of the treatment location to NA, 
#but first save to a variable
post_pd_response <- sf[sf$date >= post_pd[1] & sf$date <= post_pd[2], "sf"]
sf[sf$date >= post_pd[1], "sf"] <- NA

#model employment as a linear trend
ss <- AddLocalLinearTrend(list(), sf$sf)
mdl <- bsts(sf ~ abq, ss, niter = 1000, data = sf[sf$date <= post_pd[2],])
#> =-=-=-=-= Iteration 0 Tue Nov  3 12:50:17 2015 =-=-=-=-=
#> =-=-=-=-= Iteration 100 Tue Nov  3 12:50:17 2015 =-=-=-=-=
#> =-=-=-=-= Iteration 200 Tue Nov  3 12:50:17 2015 =-=-=-=-=
#> =-=-=-=-= Iteration 300 Tue Nov  3 12:50:17 2015 =-=-=-=-=
#> =-=-=-=-= Iteration 400 Tue Nov  3 12:50:17 2015 =-=-=-=-=
#> =-=-=-=-= Iteration 500 Tue Nov  3 12:50:17 2015 =-=-=-=-=
#> =-=-=-=-= Iteration 600 Tue Nov  3 12:50:17 2015 =-=-=-=-=
#> =-=-=-=-= Iteration 700 Tue Nov  3 12:50:17 2015 =-=-=-=-=
#> =-=-=-=-= Iteration 800 Tue Nov  3 12:50:18 2015 =-=-=-=-=
#> =-=-=-=-= Iteration 900 Tue Nov  3 12:50:18 2015 =-=-=-=-=

impact <- CausalImpact(bsts.model = mdl, 
                       post.period.response = post_pd_response)

plot(impact)

Next Steps

This is just the barest bones of an analysis, so nothing here should be considered conclusive or even interesting. Next steps are: would be to gather data for multiple locations and specify a series of models.

References

CausalImpact 1.0.3, Brodersen et al., Annals of Applied Statistics (2015). https://google.github.io/CausalImpact/.
blsAPI, by Mike Silva. https://github.com/mikeasilva/blsAPI.
reprex, a reproducible example packages by Jennifer Bryan. https://github.com/jennybc/reprex