Construct long-term trends in college education

[dc0sic]

Long-term trends in college education will be necessary for calibrating the college education model. These long-term trends should be disaggregated by sex, race/ethnicity, and age. One potential source of data is CPS ASEC.

[judah-axelrod]

Just pushed a Quarto notebook that uses ipumsr to get the CPS ASEC. Data from 1971-2023 is <9 million observations (or 6.2 million after filtering out missing IDs and education data), and it ran pretty quickly locally. A few notes here:

• The data is on Box here.
• Ethnicity data is not available from 1962-1970 so I began at 1971.
• CPS ID variables are not available for 1971-75 and 1977, so if we want to track the same people over time, I think we'd have to limit this to 1978-2023.
• SCHLCOLL is not available until 1986, but I think the EDUC variable seems to have everything we'd need.

I generated a few counts by year at the bottom - happy to display some bar or scatterplots if there are specific cuts you're interested in seeing!

[dc0sic]

Thank you, Judah! I will review tomorrow, but just to quickly answer your questions:

• We don't need to track the same person over time, so don't worry about ID.
• EDUC is perfect. Our outcome of interest for this exercise is the highest grade/year of college completed. In particular, we are interested in the rates of people whose highest level of educational attainment is high school diploma, some college, and Bachelor's degree. It would be useful to have more granular levels of college attainment (e.g. 1st, year, 2nd year, etc.) but the coding of these values changed in 1992, so we can do that only for later years.
• We want to show these trends as time series by age, sex, and race/ethnicity (Asian, Black, Hispanic, and white). We could use 10-year age categories starting with age 16 and ending with age 55.

[dc0sic]

Everything looks good. I have only one additional comment regarding race categories. Because the main purpose for the estimates in education trends is the calibration of our model, it is important to have the same race categories as we have there: Asian (non-Hispanic), Black (non-Hispanic), Hispanic, and White (non-Hispanic).

[judah-axelrod]

Just re-committed. Flagging that nlsy_lib.R codes AAPI as Other, but I left it separate in the diagnostic plots. I see that AAPI only begins in the late 1980s, so I can collapse it if necessary.

Also added time series plots where you can see proportion of folks in each age, race, or sex group that has achieved each education milestones.

Lastly, I want to flag that there is a huge jump in total sample from 2000 (75K) to 2001 (123K); it doesn't seem to affect the proportions very much but might be worth keeping in mind. More info [here])(https://www.census.gov/library/working-papers/2007/demo/POP-twps0080.html).

[dc0sic]

Good point about the AAPI. I will make this change in nlsy_lib.R.

I have a couple of comments on get_cps_asec.qmd. @judah-axelrod I know you have limited amount of time, so don't feel like you have to complete all of this.

• We should use weights when calculating the distribution of educational attainment.
• No need for plotting the "other" race group.
• In addition to estimating trends by sex, race, and age separately, we also need these categories interacted, so that we can see, for example, trends for Hispanic women ages 26 to 35. However, we should focus here on one level of educational attainment: Bachelor's degree, and plot it on the same plot for the four race groups with faceting by age group and one figure for each sex.
• It would be useful to estimate slopes of the trends in the charts described in the previous bullet point.

[judah-axelrod]

Sounds good! Two follow-up questions:

• Good reminder on the weights. Do you think we'll want replicate weights too, so that we can generate standard errors, or just the individual weight (ASECWT)?
• By estimating slopes, do you mean running regressions? Or something like calculating a tangent line at every point in the series?

[dc0sic]

No need for replicate weights, individual weights are fine.

Regarding slopes, yes, a simple linear regression on year, or just add geom_smooth(method=lm, se=FALSE) (perhaps using dashed lines).

[judah-axelrod]

Updated to include individual weights and new plots of bachelor's by race, faceted by sex and age to get interactions + a smoothing curve. I added another view without the White race group so that we can zoom in on the y axis for the other 3 groups. Also added a regression if seeing specific coefficients is helpful!

Also, I set up the category as "Bachelor's and above" - not sure if a distinction between bachelor's and graduate degrees is important for validation purposes but can tweak categories as needed.

[dc0sic]

Thank you! Sorry if I wasn't clear, but when I said to focus on BS degree I meant to plot only the share of people with BS degree, not to restrict the sample to people with BS degree. We want to see how, within each demographic group, the share of people who graduate from college changes over time.

Regarding trends, at this point we're interested in whether they can be approximated by linear trends and, if so, what their slopes are. So let's do geom_smooth(method='lm', se=FALSE).

Regression coefficients will be useful, but we want a simple regression freq ~ YEAR estimated for each demographic group.

[judah-axelrod]

Sorry about that, that makes sense! Updated to include all of this.

[dc0sic]

That's great, thank you!