Wage analysis

Author

EPI data team

Note: Users will need to install epiextractr for this example. Refer to EPI packages for R for installation instructions.

library(tidyverse)
library(epiextractr)

Defining the universe for wage analysis

For standard definition (i.e., no special exclusions or inclusions), filter to age >= 16, emp == 1, and cow1 <= 5. This is non-self-employed, non-self-incorporated, employed workers at or over 16 years old. 

  • Note: self-employed and self-incorporated workers do not have wages in the CPS ORG, so this filter is somewhat duplicative when analyzing wage data. It’s necessary for Basic data.

  • Note: this is not, of course, always the target universe (e.g., EPOPs). Use discretion when defining the sample.

cps_org <- load_org(2020:2024, "year", "age", "statefips", "wage", "emp", "union", "orgwgt", "cow1") |> 
    filter(age >= 16, emp == 1, cow1 <= 5)

Inflation adjusting

Use chained CPI c_cpi_u for analysis; and chained extended c_cpi_u_extended for analysis pre-2000 (package: realtalk). 

Median wages

Use averaged_median in the epidatatools package, rather than binipolate. Note, quantiles_n and quantiles_w default to the following so are not strictly necessary to include in the code (although may help others understand the process). 

wages_gender <- cps_org |>  
    summarise(  
        wage_median = averaged_median(  
            x = realwage,  
            w = orgwgt/12,  
            quantiles_n = 9L,  
            quantiles_w = c(1:4, 5, 4:1)),  
        n=n(),  
        .by=c(female, year)
    )

Wage premiums & regressions 

See GitHub repo here: https://github.com/Economic/equal_pay_day  

EPI’s standard wage regression uses the logarithm of real wages as the dependent variable and includes race (wbho), gender (female), education (educ), age, and age squared as the regressors (right-hand side variables). We also include fixed effects for marital status (married) and state (statefips).  If you are using multi-year data you must use year fixed effects as well. Note that using log nominal or log real wages will make no difference to the results if you use year fixed effects. In R, these models are commonly estimated using the feols function from the fixest package. See the EPI data library code for an example. 

Note: Because the dependent variable is in logs, the coefficients need to be converted from natural log changes to percentage changes. Specifically, a one-unit change in a regressor is associated with a [(exp(β) − 1) × 100%]  change in wages, holding other factors constant. For example, in 2025 the coefficient on female in the conventional EPI wage regression is −0.203424. This implies that women’s wages are approximately [(exp(−0.203424) − 1) × 100 ≈ −18.4%] different from men’s wages, after controlling for race, education, age, marital status, and state. 

library(fixest) 

wage_reg <- cps_org |>  
 mutate(  
    log_realwage = log(realwage),   
    age_2 = age^2 
) |>  
feols(  
    log_realwage ~   
    i(female, ref = "0") +  
    age + age_2 |  
    educ + wbho + 
    married + statefips, 
    weights = ~orgwgt  
)

Imputed wage filtering 

There may be instances where you want to remove imputed wages allocated by the BLS. This is especially common when dealing with union vs non-union wage comparisons, as the BLS does not account for union status when calculating imputed wages. Because of 1) the way that the a_earnhour and a_weeklypay variables are coded and 2) that filter() default drops N/A values, simply coding for filter(a_earnhour != 1) (i.e., dropping imputed wages) will also remove all non-hourly workers. You do not want this. Use this code to preserve salaried workers with self-reported wages: 

If you want to keep imputed wages in the sample for analysis: 

mutate(
    wage_imputed = case_when( 
    paidhre == 1 & a_earnhour == 1 ~ 1, 
    paidhre == 0 & a_weekpay == 1 ~ 1, 
    .default = 0 
    )
)

You can always filter to wage_imputed = 0 if you want to remove the imputed wages. 

If you want to filter out imputed wages from the start:  

filter_out(a_earnhour == 1 & paidhre == 1 | a_weekpay == 1 & paidhre == 0)

These should result in the same samples for non-imputed wages. You can check this by running crosstab(data, paidhre, wage_imputed) and crosstab(data, paidhre, a_earnhour).

Back to top