In the wake of the Great Recession of 2009, there has been a good deal of focus on employment statistics,
one of the most important metrics policymakers use to gauge the overall strength of the economy.
In the United States, the government measures unemployment using the Current Population Survey (CPS),
which collects demographic and employment information from a wide range of Americans each month.
The observations in the CPS Dataset represent people surveyed in the September 2013 CPS who actually completed a survey.
The data contains following varibales:
-
PeopleInHousehold: The number of people in the interviewee's household.
-
Region: The census region where the interviewee lives.
-
State: The state where the interviewee lives.
-
MetroAreaCode: A code that identifies the metropolitan area in which the interviewee lives (missing if the interviewee does not
live in a metropolitan area). The mapping from codes to names of metropolitan areas is provided in the file
MetroAreaCodes.csv.
-
Age: The age, in years, of the interviewee. 80 represents people aged 80-84, and 85 represents people aged 85 and higher.
-
Married: The marriage status of the interviewee.
-
Sex: The sex of the interviewee.
-
Education: The maximum level of education obtained by the interviewee.
-
Race: The race of the interviewee.
-
Hispanic: Whether the interviewee is of Hispanic ethnicity.
-
CountryOfBirthCode: A code identifying the country of birth of the interviewee. The mapping from codes to names of countries is
provided in the file
CountryCodes.csv.
-
Citizenship: The United States citizenship status of the interviewee.
-
EmploymentStatus: The status of employment of the interviewee.
-
Industry: The industry of employment of the interviewee (only available if they are employed).
We can load the data with, 131302 interviewees are in the dataset:
CPS = read.csv("CPSData.csv")
The output of
orders the levels of a factor variable like Industry from largest to smallest, so we can see that "Educational and health services" is the most common Industry.
would have provided the breakdown across all industries.
New Mexico has the fewest interviewees while California has the largest number of interviewees.
We see that 123,712 of the 131,302 interviewees are citizens of the United States (either native or naturalized).
This is a proportion of 123712/131302= 0.942.
The CPS differentiates between race (with possible values American Indian, Asian, Black, Pacific Islander, White, or Multiracial) and ethnicity.
A number of interviewees are of Hispanic ethnicity, as captured by the Hispanic variable.
The breakdown of race and Hispanic ethnicity can be obtained with:
table(CPS$Race, CPS$Hispanic)
American Indian, Black, White, and Multiracial are the races for which there are at least 250 interviewees in the CPS dataset of Hispanic ethnicity.
Evaluating Missing Values
Often when evaluating a new dataset, we try to identify if there is a pattern in the missing values in the dataset.
We will try to determine if there is a pattern in the missing values of the Married variable.
The function is.na(CPS$Married) returns a vector of TRUE/FALSE values for whether the Married variable is missing.
We can see the breakdown of whether Married is missing based on the reported value of the Region variable with the function
table(CPS$Region, is.na(CPS$Married))
table(CPS$Sex, is.na(CPS$Married))
table(CPS$Age, is.na(CPS$Married))
table(CPS$Citizenship, is.na(CPS$Married))
For each possible value of Region, Sex, and Citizenship, there are both interviewees with missing and non-missing Married values.
However, Married is missing for all interviewees Aged 0-14 and is present for all interviewees aged 15 and older.
This is because the CPS does not ask about marriage status for interviewees 14 and younger.
As mentioned in the variable descriptions, MetroAreaCode is missing if an interviewee does not live in a metropolitan area.
The breakdown of missing MetroAreaCode by State can be obtained with:
table(CPS$State, is.na(CPS$MetroAreaCode))
Alaska and Wyoming have no interviewees living in a metropolitan area,
and the District of Columbia, New Jersey, and Rhode Island have all interviewees living in a metro area.
To evaluate the number of interviewees not living in a metropolitan area, broken down by region, we can run:
table(CPS$Region, is.na(CPS$MetroAreaCode))
We can then compute the proportion of interviewees in each region that live in a non-metropolitan area: 34.8% in the Midwest, 21.6% in the Northeast,
23.8% in the South, and 24.4% in the West.
While we were able to use the table() command to compute the proportion of interviewees from each region not living in a metropolitan area,
it was somewhat tedious
(it involved manually computing the proportion for each region) and isn't something you would want to do if there were a larger number of options.
It turns out there is a less tedious way to compute the proportion of values that are TRUE. The mean() function,
which takes the average of the values passed to it, will treat TRUE as 1 and FALSE as 0, meaning it returns the proportion of values that are true.
For instance, mean(c(TRUE, FALSE, TRUE, TRUE)) returns 0.75.
tapply(is.na(CPS$MetroAreaCode), CPS$State, mean)
It is actually easier to answer this question if the proportions are sorted, which can be accomplished with:
sort(tapply(is.na(CPS$MetroAreaCode), CPS$State, mean))
From this output, we can see that Wisconsin is the state closest to having 30% of its interviewees from a non-metropolitan area
(it has 29.933% non-metropolitan interviewees) and Montana is the state with highest proportion of non-metropolitan interviewees without them
all being non-metropolitan, at 83.608%.
Integrating Metropolitan Area Data
Codes like MetroAreaCode and CountryOfBirthCode are a compact way to encode factor variables with text as their possible values, and they are therefore quite common in survey datasets. In fact, all but one of the variables in this dataset were actually stored by a numeric code in the original CPS datafile.
When analyzing a variable stored by a numeric code, we will often want to convert it into the values the codes represent. To do this, we will use a dictionary,
which maps the the code to the actual value of the variable. We have provided dictionaries MetroAreaCodes.csv and CountryCodes.csv,
which respectively map MetroAreaCode and CountryOfBirthCode into their true values.
271 observations (codes for metropolitan areas) are there in MetroAreaMap and 149 observations (codes for countries) are there in CountryMap.
To merge in the metropolitan areas, we want to connect the field MetroAreaCode from the CPS data frame with the field Code in MetroAreaMap.
The following command merges the two data frames on these columns, overwriting the CPS data frame with the result:
CPS = merge(CPS, MetroAreaMap, by.x="MetroAreaCode", by.y="Code", all.x=TRUE)
The first two arguments determine the data frames to be merged (they are called "x" and "y", respectively, in the subsequent parameters to the merge function).
by.x="MetroAreaCode" means we're matching on the MetroAreaCode variable from the "x" data frame (CPS),
while by.y="Code" means we're matching on the Code variable from the "y" data frame (MetroAreaMap).
Finally, all.x=TRUE means we want to keep all rows from the "x" data frame (CPS), even if some of the rows' MetroAreaCode doesn't match any codes in
MetroAreaMap (for those familiar with database terminology, this parameter makes the operation a left outer join instead of an inner join).
we can read that Boston-Cambridge-Quincy, MA-NH has the largest number of interviewees of these options, with 2229.
In order to know which metropolitan area has the highest proportion of interviewees of Hispanic ethnicity, It will be easiest to obtain the maximum by actually using the sorted output:
sort(tapply(CPS$Hispanic, CPS$MetroArea, mean))
As we can see, 96.6% of the interviewees from Laredo, TX, are of Hispanic ethnicity, the highest proportion among metropolitan areas in the United States.
In order to determine the number of metropolitan areas in the United States from which at least 20% of interviewees are Asian,
sort(tapply(CPS$Race == "Asian", CPS$MetroArea, mean))
We can read from the sorted output that Honolulu, HI; San Francisco-Oakland-Fremont, CA; San Jose-Sunnyvale-Santa Clara, CA; and Vallejo-Fairfield, CA had at least 20%
of their interviewees of the Asian race.
To determine which metropolitan area has the smallest proportion of interviewees who have received no high school diploma.
sort(tapply(CPS$Education == "No high school diploma", CPS$MetroArea, mean, na.rm=TRUE))
We can see that Iowa City, IA had 2.9% of interviewees not finish high school, the smallest value of any metropolitan area.
Integrating Country of Birth Data
ust as we did with the metropolitan area information, we can merge in the country of birth information from the CountryMap data frame,
replacing the CPS data frame with the result.
CPS = merge(CPS, CountryMap, by.x="CountryOfBirthCode", by.y="Code", all.x=TRUE)
From summary(CPS), we can read that Country is the name of the added variable, and that it has 176 missing values.
From the summary(CPS) output, or alternately sort(table(CPS$Country)), we see that the top two countries of birth were United States and Mexico,
both of which are in North America. The third highest value, 839, was for the Philippines.
For proportion of the interviewees from the "New York-Northern New Jersey-Long Island, NY-NJ-PA" metropolitan area have a country of birth that is not the United States:
table(CPS$MetroArea == "New York-Northern New Jersey-Long Island, NY-NJ-PA", CPS$Country != "United States")
we can see that 1668 of interviewees from this metropolitan area were born outside the United States and 3736 were born in the United States
(it turns out an additional 5 have a missing country of origin). Therefore, the proportion is 1668/(1668+3736)=0.309.
To obtain the number of TRUE values in a vector of TRUE/FALSE values, you can use the sum() function. For instance, sum(c(TRUE, FALSE, TRUE, TRUE)) is 3.
Therefore, we can obtain counts of people born in a particular country living in a particular metropolitan area with:
sort(tapply(CPS$Country == "India", CPS$MetroArea, sum, na.rm=TRUE))
sort(tapply(CPS$Country == "Brazil", CPS$MetroArea, sum, na.rm=TRUE))
sort(tapply(CPS$Country == "Somalia", CPS$MetroArea, sum, na.rm=TRUE))
sort(tapply(CPS$Country == "Pakistan", CPS$MetroArea, sum, na.rm=TRUE))
We see that New York has the most interviewees born in India (96), Boston has the most born in Brazil (18), and Minneapolis has the most born in Somalia (17). Washington-Arlington-Alexandria,
has the most born in Pakistan (18).