Incarceration Data

Sources

• Vera Institute Incarceration Trends
• Prison Policy Initiative: The Geography of Mass Incarceration
• Wisconsin Population Health Institute's County Health Rankings & Roadmaps

Step 1 - Cleaning the county health rankings data

There are various steps involved in cleaning the data from the county health rankings:

1. There are two sources of data that County Health Rankings make available which I term their cross-sectional estimates and their longitudinal estimates. The cross-sectional estimates are data tables they publish annually, and they make no attempt to link together these year-by-year estimates. Each row corresponds to a specific county + variable estimate (i.e., there is no time component). The longitudinal estimates are a more traditional longitudinal data table where each row represents a unique county + year + variable estimate.

   1. For the cross-sectional estimates, I download the cross-sectional estimates for each year going all the way back to first release (2010). I then combine each of the cross-sectional estimates to create a longitudinal data table. One can look at the specific documentation + code to see how I harmonize the variables across the years.

   2. For the longitudinal estimates, I download each year's longitudinal estimates and compare present estimates to past estimates to ensure none of the values have changed. The only issue I have experienced with this is that the values between 2023 and 2024 changed for school funding adequacy.

2. The next step is to combine the cross-sectional and longitudinal estimates into one data table. It will occasionally be the case that the cross-sectional and longitudinal data tables will have different estimates for the same variable + county + year (or year range). A choice must be made as to which estimate should be used. I investigate this (see the code + report for more details), and I do not think the estimates are different enough to the point where it would substantively affect results + conclusions. Based on my analysis and judgment, I think using the longitudinal estimates in cases where the estimates do not match is the best course of action.

3. The next step has to do with the fact that for some variable + county + year (or year range) observations, there will be multiple, different estimates. The County Health Rankings & Road-maps data tables will sometimes use the previous years' data as part of the current year's release, and it seems as if those estimates change sometimes for whatever reason. A choice must be made as to which estimate to use. As detailed in the code + report, the estimates (thankfully) are not too different so I do not think results + conclusions would change substantively. Based on my analysis and judgment, I think using the estimates from the most recent release of data are the best estimates to use. The logic being that more recent estimates would be more up-to-date and accurate.

4. The final step involves cleaning up variable + county + year observations with multiple, different estimates. This is different from the above problem. In this step, we are trying to associate each observation with a specific year. One can choose to use the starting year or the end year for each variable as the year of that observation. For variables which only range across one year, the choice does not matter. For other variables which represent the estimate across a range of years (e.g., 2002 - 2004) the choice can matter (e.g., whether one chooses 2002 or 2004 as the year this estimate represents). So what will happen is we will get variables with the same starting year or ending year, and they will have different estimates associated with them. One will have to pick which estimate to use. As one may see in a somewhat running theme here, I look into how different the estimates are (see code + report), and I find the estimates are relatively similar. I do not think conclusions would change too much if one used one set of estimates vs. another. Based on my analysis and judgment, I think using the estimates associated with the smallest time range would be the best estimates to use. Meaning if I had a set of estimates for a specific county for a specific variable from 2002 - 2004 and then 2002 - 2003, and I was using the starting year as the representative year, I would use the estimate from 2002 -2003. Note that this specific observation would not present as a problem if we were to use the ending year as the year of observation because these estimates have different ending years.

   1. One additional complication is whether one uses the start year or the end year of a year range to represent the year for any given observation. For modeling purposes, if a variable is included as an independent variable (i.e., a predictor variable or explanatory variable), one must use the end year of its range to signify the year of the observation. This is to avoid situations where you accidentally include an observation that occurred after the outcome took place and could not logically affect it. For example, if you are modeling the effect of child mortality on unemployment, you would want to use the end year for child mortality (e.g., for observations taking place from 2002 - 2008, these observations would count for 2008) when modeling its effect on unemployment (for the year 2008). If one used the starting year for child mortality (i.e., 2002), then one would wind up using data from future years that logically could not have affected the unemployment rate in 2002.

   2. If a variable is included as a dependent variable (i.e., an *outcome variable), then it is up to the researcher if they want to use the starting year or ending year. I personally believe it makes most the sense to use the ending year, but I will also be using start year as a robustness check.

It is possible if one were to make all the opposite decisions I made, one might end up with a much different set of estimates than what I have. In other words, even though each data cleaning step does not change the estimates all that much, cumulatively the impact of each data cleaning step together might be moderately large. There are a few things specifically which make me believe this will not be the case. First, the proportion of observations affected by the data cleaning decisions is small (roughly 7% - 8%), and the magnitude of change in the estimates is also very small. Second, there are no cascading data changes. By this I mean that the estimates affected by one set of data cleaning decisions are not affected by any other data cleaning decisions. The set of estimates affected by cross-sectional vs. longitudinal differences are not the estimates affected by multiple observations for the same year + county + variable and vice versa. The cumulative impact of these changes is thus additive and not multiplicative which reduces the chance that these data cleaning decisions might affect results.

Of course, I also create a data table using the complete opposite of each of the data cleaning decisions I have made so I can test and ensure results are not substantively different.

Main Analysis Tables

• longitudinal_late_endyear_min.csv
• longitudinal_late_startyear_min.csv
  • This means we are using the *longitudinal estimates (rather than the cross-sectional), the estimates from the most recent (latest) release, and the estimates associated with the smallest year range.

Robustness Analysis Tables

• cross_early_endyear_max.csv
• cross_early_startyear_max.csv

Step 2 - Selecting variables based on years/counties of availability, reliability and relevance

There are a wide variety of variables made available by the County Health Rankings & Road-maps. However, some of these variables are only of limited utility due to the fact that they are only available for a limited number of years/counties. Additionally, due to the way the variable was collected, it may not be very reliable. Here I will detail the variables we drop and briefly why. If the variable is of particular substantive interest, I will provide details on the source of the variable and if more information could be obtained about it.

Checking for reliability (below variables are dropped)

1. Violent crime: This variable stopped being reported after a certain period of time, and the last few periods of observation are all based on the same years. Additionally, this data comes from the UCR, and there have been some issues in using UCR data to aggregate crime to the county level.
2. Binge drinking: It has been supplanted by the inclusion of excessive drinking.
3. Primary care provider rate: This has been supplanted by the inclusion of primary care providers.
4. Hospice use: The data is only available for one point in time.
5. College degrees: The data is only available for one point in time.
6. Gini coefficient: The data is only available for one point in time.
7. Single parent households: It has been supplanted by the inclusion of children in single parent households.
8. Air pollution particulate days and air pollution ozone days: These data are only available for three points in time, and they have been made supplanted by the inclusion of air pollution particulate matter.
9. Access to healthy foods: It has been supplanted by the inclusion of the food environment index and its component variables.
10. Liquor store density: This data is only available for two points in time.
11. Access to recreational facilities: This data is only available for three points in time. It also has been supplanted by the access to exercise opportunities variable. There is also really no documentation available about this variable so it is unclear how it was constructed.
12. High housing costs: This data is only available for three points in time. Additionally, it is not well documented so it is unclear what it is measuring exactly. It also has been supplanted by severe housing cost burden and the percentage of households with high housing costs. In actuality, this variable may be the same as severe housing cost burden since both come from the ACS.
13. Inadequate social support: This data is only available for two points in time and is poorly documented.
14. Fast food restaurants: This data is only available for two points in time and is poorly documented. It also has been supplanted by the inclusion of the food environment index and its component variables.
15. Could not see a doctor due to cost: This data is only available for three points in time and is poorly documented.
16. Access to parks: This variable is only available at one point in time. It has been supplanted by the inclusion of the access to exercise opportunities variable.
17. Drug poisoning deaths: This variable is only available at two points in time, and there is no documentation available concerning this variable. It has been supplanted by the inclusion of the drug overdose variable.
18. Drinking water violations: This variable is only available at two points in time, and it has been supplanted by the drinking water violations (binary) variable.
19. Children eligible for free lunch: This variable has been supplanted by children eligible for free or reduced price lunch.
20. Drug overdose deaths - modeled: This variable is only available for two points in time, and there is no documentation available about it.
21. COVID19 age-adjusted mortality: This variable is only available for one point in time.
22. Voter turnout: This variable is only available for one point in time.
23. Census participation: This variable is only available for one point in time.
24. Illiteracy: This variable is only available for one point in time.

Checking for missing values (below variable-year combinations are dropped)

1. Strangely, the living wage data is missing for all county-year observations.
2. Mental health providers is missing from the years 2009 - 2011 and 2013.
3. Diabetes prevalence is nearly all missing for the year 2018.
4. High school graduation is missing for nearly 90% and 75% of all counties respectively in 2008 and 2010. Values for 2014 and 2016 are completely missing.
5. Other primary care providers is missing for the year 2013.
6. Jail admission rate is missing for nearly 95% of counties from 1970 - 1977.
7. Jail discharges is missing for nearly 95% of counties from 1970 - 1977
8. Prison admission rate is missing for nearly 95% of counties from 1970 - 1982.
9. Prison population rate is missing for nearly 95% of counties from 1970 - 1982.
10. % rural is only available in 2000 and 2010.
11. Access to exercise opportunities is missing for 2015.
12. Adult smoking is completely missing for 2013.
13. Child mortality is missing for 2011 and 2014.
14. Dentists is missing for 2008 - 2009.
15. Disconnected youth is completely missing for 2013, 2015, and 2016.
16. Drinking water violations (binary) is completely missing for 2015.
17. Excessive drinking is completely missing for 2013.
18. Health care costs is missing completely in 2008, 2010, and 2016.
19. HIV prevalence is completely missing in 2011 and 2014.
20. Homicides is completely missing in 2009, 2011, and 2014.
21. Infant mortality is completely missing in 2009, 2011, 2014, and 2015.
22. Injury deaths is completely missing in 2011 and 2014.
23. Insufficient sleep is completely missing in 2015.
24. Limited access to healthy foods is completely missing from 2007 - 2009, 2011 - 2014, and 2016.
25. Low birth-weight is completely missing in 2009 and 2015.
26. Median household income is completely missing in 2009.
27. Motor vehicle crash deaths is completely missing in 2009, 2011, and 2014.
28. Poor mental health days is completely missing in 2013.
29. Poor or fair health is completely missing in 2013.
30. Poor physical health days is completely missing in 2013.
31. Premature age-adjusted mortality is completely missing in 2011 and 2014
32. Teen births is completely missing in 2009 and 2015.
33. Uninsured adults is completely missing from 2006 - 2007.

Variables with a moderate to a moderately high amount of missing values (not enough to drop)

1. Alcohol-impaired driving deaths (~8% - 9%).
2. Suicides.
3. Juvenile arrests.
4. Disconnected youth (~35%).
5. Firearm fatalities (~30%).
6. Drug overdose deaths (~50%).
7. Homicides. Note that homicides routinely has missing observations for more than 50% of counties.
8. Residential segregation (Black/White) (~35%) and (White/non-White) (~11%).
9. Infant mortality. Note that infant mortality routinely has missing observations for more than 50% of counties.
10. Child mortality (~35% - 40%).
11. HIV prevalence (~20% - 30%).
12. Adult smoking (only 2008 - 2012) (~10% - 20%).
13. Excessive drinking (only 2009 - 2012) (~15% - 30%).
14. High school graduation (~15%).
15. Injury deaths (~5% - 10%).
16. Low birth-weight (~3% - 7%).
17. Mental health providers (~5% - 15%).
18. Motor vehicle crash deaths (only in 2010, 2012 - 2013, 2015 - 2016) (~10% - 15%).
19. Poor mental health days (only 2008 - 2012) (~5% - 15%).
20. Poor or fair health (only 2008 - 2012) (~12% - 13%).
21. Poor physical health days (only 2008 - 2012) (~5% - 10%).
22. Premature death (~2% - 10%).
23. Primary care physicians (only in 2009) (~6%).
24. Teen births (~3% - 5%).
25. Sexually transmitted infections (sporadically) (~3% - 6%).
26. Jail admission rate and population (DCRP) (~9%).
27. Jail discharge rate, jail admission rate, population rate, and pre-trial rate (~5% - 10%).
28. Prison admission and population rate (~10% - 30%).

Checking for years of availability (all variables dropped)

1. All observations after 2018 are dropped since 2018 is the latest year of availability for the jail incarceration rate. Some variables which only have observations after 2018 are, as a result, also dropped. These variables high school completion, broadband access, school segregation, gender pay gap, childcare cost burden, and child care centers.

2. Suicides and traffic volume each only have one observation prior to 2018 so these variables are dropped.

3. Life expectancy, home ownership, severe housing cost burden, and juvenile arrests are all only available from 2016 and on. The latest year of availability for prison incarceration rates is 2016. As a result, these variables will be dropped when analyzing prison incarceration rates.

4. Math scores, reading scores, and preventable hospital stays (2019 - 2024 release) each only have one observation prior to 2016. Thus, they are dropped for prison incarceration analysis.

Filling in missing values and obtaining older estimates

1. Violent crime: Some researchers have cautioned against creating crime rates at the county level using UCR/NIBRS data. We could try, though.

2. ACS (American Community Survey) data

   1. College degrees and Gini coefficient. County Health Rankings and Road-maps data only collected these variables for one year. We could collect them ourselves.
   2. Disconnected youth: CHRR used ACS Table B14005 to create this variable. We could do something similar to fill in missing values and get data points from earlier time periods.
   3. % not proficient in English: CHRR used ACS Table B16005 to create this variable. It would require a bit of work to completely recreate this variable from that table, though.
   4. No documentation is provided as to what table is used for calculating residential segregation. It might have been calculated manually using demographic compositional variables. We could do the same, but it would require some work. It could be very worthwhile, though, to fill in missing values and obtain data from earlier time points.
   5. High school completion: CHRR used ACS Table S1501 to create this variable. We could easily use this table to fill in missing values and obtain data from earlier time points.
   6. Home ownership: CHRR used ACS Table B25003 to create this variable. We could easily use this table to fill in missing values and obtain data from earlier time points.
   7. Severe housing cost burden: CHRR used ACS tables B25074 and B25095 to create this variable. It might be bit a bit involved, but we could attempt to recreate this variable to fill missing variables and obtain data from earlier time points.
   8. Income inequality: CHRR used ACS Table B19080 to create this variable. We could easily use this table to fill in missing values and obtain data from earlier time points.
   9. Children in single-parent households: CHRR used ACS Table B09005 to create this variable. Geographic coverage is pretty good, but we could get older data potentially.
   10. Some college: CHRR used ACS Table B15001 to create this variable. Geographic coverage is pretty good, but we could get older data potentially.

3. BRFSS (Behavioral Risk Factor Surveillance System) data

   1. Variables include: Inadequate social support, Could not see a doctor due to cost, Poor physical health, Poor mental health, Poor or fair health, Insufficient sleep Excessive drinking, Adult smoking, Diabetes prevalence, Frequent physical distress, and Frequent mental distress. All of these variables have at least some gaps in their years of availability.
   2. Adult obesity is the one variable from this set that actually has pretty geographic coverage and no gaps in yearly availability.
   3. These data all come from the CDC BRFSS program. They have data going back to 2006, but I am not sure of the geographic coverage or if all variables are available every year. It is possible that what is provided by CHRR is the best that we have.

4. High school graduation: Missing values as well as older data can potentially be obtained from the EDFacts data portal. It is possible, though, that what CHRR collected is the best that is available.

5. CDC Wonder/Vital Statistics: Many of the variables have gaps in their temporal coverage and are missing at least some counties consistently (likely due to a small number of incidents). If we go to the source, it is possible we could fill in the gaps temporally and geographically.

   1. Teen births, Premature age-adjusted mortality, Premature deaths, Motor vehicle crash deaths, Low birth-weight, Injury deaths, Infant mortality,Homicides, Child mortality, Alcohol-impaired driving deaths, Suicides, Drug overdose deaths, Life expectancy, and Firearm fatalities.

6. Small Area Health Insurance Estimates (SAHIE)

   1. There are some temporal gaps in uninsured adults which could be potentially filled by going to the source data. Uninsured children and Uninsured have good geographic coverage, but we may be able to get older estimates.

7. . National Center fro HIV, Viral Hepatitis, STD, and Tuberculosis Prevention (NCHHSTP).

   1. HIV prevalence has moderately high rates of geographic missing values and has some temporal gaps. Sexually transmitted infections has much lower rates of missing and no temporal gaps. By going to the CDC Atlas Plus, the data can be obtained from the source to potentially fill in the gaps for these missing values and obtain older values.

8. Population Estimates Program (PEP)

   1. % 65 and older, % American Indian or Alaskan Native, % Asian, % below 18 years of age, % female, % Hispanic, % Native Hawaiian or Other Pacific Islander, % Non-Hispanic Black, % Non-Hispanic White and Total Population all have some temporal gaps. By going to the source, we could fill in these gaps, and also potentially obtain older estimates.

9. Juvenile arrests data comes from Easy Access to State and County Juvenile Court Case Counts (EZACO). It seems like the data is relatively easily obtainable from them, and it goes back all the way to 1997.

10. Children eligible for free and reduced price lunch data comes from the NCES Common Core of Data and Private School Survey. One can get results going all the way back to 1986 which is quite impressive. I am not sure what the geographic coverage will be. It may be worth investigating to obtain this data.

    1. School segregation data comes from the National Center on Education Statistics (NCES) so it is possible we could gather this data ourselves. I am not sure, though, how relevant it is for our analysis. We would be more interested in residential segregation. Additionally, I think researchers at CHRR created this measure themselves so it is not a variable we can simply pull from the NCES.

11. Child care cost burden combines data from the Small Area Income Poverty Estimates (SAIPE) Census program with data from the Living Wage Institute. Specifically, it uses data from the Living Wage Institute's living wage measure, and it extracts the the childcare cost component from the living wage. It then divides by the average healthcare costs by median income estimates from the SAIPE to determine childcare cost burden. Much like the living wage variable, in order to get access to past years of data, one needs to request this data from the Living Wage Institute.

    1. To access living wage data, one needs to submit a request to the Living Wage Institute.
    2. I am not sure how relevant child care cost burden and living wage are for our research, though.

12. Child care center data comes from the Homeland Infrastructure Foundation-Level Data (HIFLD) project. It does not appear as if they make the data available longitudinally, though. On the other hand, I am not certain how useful this variable would be for our research.

13. Small Area Income & Poverty Estimates

    1. Going to the source, we can fill in missing values for median household income and also obtain older estimates of the data.
    2. Children in poverty has good geographic and temporal coverage. The only benefit would be potentially being able to obtain older estimates.

14. Math and reading scores are available from the Educational Opportunity Project at Stanford University. They have data going back to 1991 (not sure of the geographic coverage), but this could be really powerful data to collect.

15. Percentage of households with high housing costs data comes from the Comprehensive Housing Affordability Strategy maintained by HUD using specially tabulated ACS data from the Census. The data goes back to at least the 2006 ACS so it may be worthwhile to invest some time obtaining this data.

16. Social associations: This data comes from the County Business Patterns dataset, and it has data going back to 1987. It could be useful to collect this data.

17. Local Area Unemployment Statistics (LAUS). Unemployment has good geographic and temporal. The only benefit would be that we could potentially obtain older estimates of unemployment.

18. School funding adequacy comes from the School Finance Indicators Database. It may be worth it to use the data from the source instead of the estimates from CHRR. The data are outputs from a model and as a result, the team behind the database continually develop new estimates for past years. Additionally, the longitudinal estimates from 2023 for school funding adequacy do not match the estimates from 2024. Furthermore, neither the 2023 or 2024 data match what is published on the website. This leads me to believe the estimates from the CHRR are from an older model.

19. Air pollution and drinking water violations come from the CDC's National Environmental Public Health Tracking Network (NEPHTN). I am not so familiar with environmental data so I can only make educated guesses as to which measures from the CDC match what CHRR did.

20. Access to exercise opportunities is a specially created variable by the CHRR team.

21. Food and diet environment

    1. Food insecurity comes from the Map the Meal Gap. Food insecurity has good geographic coverage and no temporal gaps. The website dashboard only has data going back to 2017, but one can request the data which might have earlier time points. I am not sure how relevant this variable would is for our research, though.
    2. Limited access to healthy foods comes from the USDA's Food Access Research Atlas. Unfortunately, it does not look like there are many time points available for this dataset.
    3. Food environment index is a little confusing. I believe it is a custom created variable by CHRR using data from Map the Meal gap and the Food Environment Atlas and possibly also the Food Access Research Atlas.

Relevance

The main goal of the research project is create plausible estimates of the prison incarceration rate on childhood and teenager health + socioeconomic outcomes at the community level. This may also extend into trying to create estimates of the effect of incarceration on residential demographics and residential instability at the community level. As a result, there are going to be some variables which are just not very relevant to that modeling task, and they will be dropped.

These variables include: preventable hospital stays (2010 - 2018 measure), diabetes monitoring, mammography screening (ages 67 - 69), driving alone to work, long commute (driving alone), mammography screening (ages 65 - 74), and flu vaccinations.