Missing Values Exercises¶
One of the defining features of
pandas is the use of indices for data alignment. Like many features in
pandas, it can make live very easy, but if you aren’t careful, it can also lead to problems. This is especially true because indices lead to behavior that is very different from what one sees in other languages and library (like
julia). So let’s spend a little timing practicing interacting with indices (and missing values)!
Today, we will be using the ACS data we used during out first
pandas exercise to examine the US income distribution, and how it varies by race. Note that because the US income distribution has a very small number of people with extremely high incomes, and the ACS is just a sample of Americans, the far right tail of the distribution will not be very well estimated. However, this data should suffice for helping to understand wealth inequality in the United States.
To begin, load the ACS Data we used in our first pandas exercise. That data can be found here. We’ll be working with
Let’s begin by calculating the mean US incomes from this data (recall that income is stored in the
Hmmm… That doesn’t look right. The average American is definitely not earning 1.7 million dollars a year. Let’s look at the values of
value_counts(). Do you see a problem?
value_counts() with the argument
normalize=True to see proportions of the sample that report each value instead of the count of people in each category. What percentage of our sample has an income of 9,999,999? What percentage has an income of 0?
As we discussed before, the ACS uses a value of 9999999 to denote that income information is not available for someone. The problem with using this kind of “sentinel value” is that pandas doesn’t understand that this is supposed to denote missing data, and so when it averages the variable, it doesn’t know to ignore 9999999.
To help out
pandas, use the
replace command to replace all values of 9999999 with
Now that we’ve properly labeled our missing data as
np.nan, let’s calculate the average US income once more.
OK, now we’ve been able to get a reasonable average income number. As we can see, a major advantage of using
np.nan is that
pandas knows that
np.nan observations should just be ignored when we are calculating means.
But it’s not enough to just get rid of the people who had
inctot values of 9999999. We also need to know why those values were missing. Suppose, for example, that the value of 9999999 was used for anyone who made more than 100,000 dollars: if we just dropped those people, then our estimate of average income wouldn’t mean much, would it?
So let’s make sure we understand why data is missing for some people. If you recall from our last exercise, it seemed to be the case that most of the people who had incomes of 9999999 were children. Let’s make sure that’s true by looking at the distribution of the variable
age for people for whom
inctot is missing (i.e. subset the data to people with
inctot missing, then look at the values of
Then do the opposite: look at the distribution of the
age variable for people who whom
inctot is not missing.
Can you determine when 9999999 was being used? Is it ok we’re excluding those people from our analysis?
Note: In this data, Python doesn’t understand
age is a number; it thinks it is a string because the original data has categories like “90 (90+ in 1980 and 1990)” and “less than 1 year old”. So you can’t just use
max(). We’ll discuss converting string variables into numbers in a future class.
Great, so now we know why those people had missing data, and we’re ok with excluding them.
But as we previously noted, there are also a lot of observations of zero income in our data, and it’s not clear that we want everyone with a zero-income should be included in this average, since those may be people who are retired, or in school.
Let’s limit our attention to people who are currently working. We can do this using
empstat. Remember you can use
value_counts() to see what values of
empstat are in the data!
Now let’s estimate the racial income gap in the United States. What is the average salary for employed Black Americans, and what is the average salary for employed White Americans? In percentage terms, how much more does the average White American make than the average Black American?
Note: these values are not quite accurate estimates. As we’ll discuss in later lessons, to get completely accurate estimates from the ACS we have to take into account how people were selected to be interviewed. But you get pretty good estimates in most cases even without weights – your estimate of the racial wage gap without weights is within 5% of the corrected value.
Note: This is actually an underestimate of the wage gap. The US Census treats hispanic respondents as a sub-category of “white”, so in pooling what most Americans think of as “White” respondents (but which Census thinks of as “White, Non-Hispanic”) with Hispanic respondents (who tend to earn less), we get an underestimate of the average white salary in the US.
While all ethnic distinctions are socially constructed, and so on some level these distinctions are all deeply problematic, this coding is inconsistent with what most Americans think of when they hear the term “White”, which is though of as a category that is distinct from being Hispanic or Latino (categories which are also usually conflated in American popular discussion). With that in mind, most researchers working with US Census data split “White” into “White, Hispanic” and “White, Non-Hispanic” using
Want more practice?¶
(1) As noted above, these estimates are not actually quite correct because we aren’t using survey weights. To calculate a weighted average that takes into account survey weights, you need to use the following formula:
(As you can see, when \(weight_i\) is constant for all observations, this just simplifies to our normal formula for mean values. It is only when weights vary across individuals that weights must be explicitly addressed).
In this data, weights are stored in the variable
perwt, which is the number of people for which each observation is a stand-in (the inverse of that observations sampling probability).
Using the formula, re-calculate the weighted average income for both populations.
(2) Now calculate the weighted average income gap between non-hispanic White Americans and Black Americans.
Absolutely positively need the solutions?¶
Don’t use this link until you’ve really, really spent time struggling with your code! Doing so only results in you cheating yourself.