Predicting gun death rate from gun laws: The balance between life & liberty is measurable

Life, Liberty, and the pursuit of Happiness were deemed to be God-given rights when America’s government was formed. American gun laws balance restricting individual liberty with protecting individual life.

The following are incontrovertible: 1. Gun laws restrict individual freedoms 2. Gun laws save lives. There’s no reason to waste energy debating these facts. Instead, the question is, where will Americans choose to find the balance between life and liberty, for it is a choice.

Leaving aside effects of lobbyists on gun policy or of gun policy on crime, which are not addressed here, States have chosen to adopt gun laws that represent a range of points along the life vs. liberty continuum. The analysis summarized here is not an opinion, but a depiction of a complex set of interacting, measurable factors and outcomes, teased apart using a mathematical algorithm. Numbers in (firearm death rates and gun law counts per category by State and year) …….. numbers out (gun law categories that increase or decrease predicted death rate and by how much).

The result is a sideways tree depicting predicted average annual State gun death rates depending on the number of laws they hold within a category. Each split in the tree shows how some law subcategories increase or decrease the predicted number of gun-related deaths. Starting at the left of the tree, the predicted average firearm death rate per 100,000 State residents is 11.2, but for States with 4–8 laws within the Permitting category, it drops to 7.3. For States with 0–3 Permitting laws, the rate increases to 12.7… and so on.

Particularly timely is the question of whether age restrictions are important. In some cases, States with more Age Restrictions (as defined by the The State Firearm Laws Project codebook) are predicted to have considerably lower death rates. Specifically, in States with more Permitting laws but no Fingerprinting law, the death rate is predicted be cut in half by imposing more age restrictions (4.3 vs. 9.7, 2nd and 3rd boxes from the top on the far right).

The data:

Boston University’s State Firearm Laws Project maintains a database of the status of 133 State firearm law provisions from 1991 — 2017, classified into 14 categories and 49 subcategories by their team of legal researchers, based on intended impact on gun violence. It is a publicly available resource that you can also explore through their interactive maps and charts. I downloaded their dataset of presence/absence of firearm law provisions by State and year.

The CDC’s Underlying Cause of Death database breaks down death certificate cause of death by region and year. I downloaded raw death rate per 100,000 residents per State/year with firearm as cause of death. This includes any death due to firearm whether accidental or not, self-inflicted or not, crime-related or not, but no data about gun type, gender, age, or circumstances.

Since gun laws, alone, as The State Firearm Laws Project cautions, are unlikely to be the sole determinants of firearm death rate, I ran the same analysis including unemployment rate per State/year and then with weighted population density per State/year, neither of which significantly changed results. There may be other factors by State/year that could be informative in a repeat analysis, such as proximity to States with more or fewer gun laws.

Summarizing the quantified effects of balancing liberty with life:

Based on Recursive Partitioning And Regression Tree (RPART) analysis, ten of 49 gun law subcategories were found to significantly predict State gun death rate decreases vs. increases with an average error rate of 1.76 deaths per 100,000. The predicted average annual death rate is 11.2. From there the tree splits into two parts: upper and lower.

The upper section shows that in States with 4–8 Permitting laws (as defined by the The State Firearm Laws Project codebook), the rate drops from 11.2 to 7.3. Of these, States that also have a Fingerprinting law are predicted to see the rate drop to 4.8; otherwise, it goes up to 9.0. If there is no Fingerprinting law, the predicted rate decreases to from 9.0 to 4.3 with 6–7 Age Restriction laws, but with 0–5 such laws it increases to 9.7.

The lower section of the tree starts with States that have 0–3 Permitting laws and an average annual predicted gun death rate rate of 12.7. This can decrease to 9.5 for States with 2–8 Storage laws, otherwise, the rate increases to 13.3. It is predicted to decrease from 13.3 to 10.6 for States with 1–3 (any) laws in the Violent Misdemeanor subcategory; if not, the predicted rate increases to 14.0.

Moving farther down and right, between 2–3 Gun Trafficking laws decreases the predicted rate in these States from 14.0 to 10.8, but for States with 0–1 Gun Trafficking laws, the predicted rate increases to 14.7. Maintaining 4–11 laws in the Domestic Violence Restraining Order subcategory decreases the predicted rate from 14.7 all the way to 9.1, but if these States have 0–3 such laws, the rate increases to 15.2. The predicted rate falls back from 15.2 to 11.4 for these States if there are 2–3 laws in the Preemption subcategory, but increases to 15.3 with 0–1 laws. The rate is predicted to lower a bit to 15.1 if these States have 2–4 laws in the Background Checks subcategory, but if there are 0–1 such laws, the rate increases to a high of 17.6. For those States with a greater number of Background Checks laws, the predicted firearm death rate drops from 15.1 to 13.6 with 1–3 (any) Dealer Licensing subcategory laws and increases to 15.5 with none.

The above three paragraphs are the reason that infographics exist. It might be better to scroll back up and look at the tree. Or it can be compressed thus: 1. Gun laws restrict individual freedoms 2. Gun laws save lives.

Thank you for reading.

I welcome constructive feedback . “Clap” with approval, share on FB, Twitter, Linkedin, Reddit, wherever, or if you have a specific response or question, message me here. I’m also interested in hearing what topics you’d like covered in future posts. If you want the data and R code used for data wrangling, feature engineering, and analysis, please fork or download my associated GitHub repository. It would be great if someone repeated the analysis after adding additional State/year variables that might affect the outcome. If you do, let me know.

Read more about my work on jenny-listman.netlify.com. Feel free to contact me via Twitter @jblistman or LinkedIn.

Notes on data and analysis:

1. All data, code, and notes to replicate this analysis can be found on my GitHub repo. Go for it!
2. Presence or absence of 133 firearm laws across US States from 1991 to 2017 was downloaded from The State Firearm Project: https://www.statefirearmlaws.org/table.html
3. The codebook of laws, their categories, subcategories, and explanations is downloadable from The State Firearm Laws Project’s website: https://www.statefirearmlaws.org/download-codebook.html.
4. CDC cause of death by State/year from firearms data were downloaded from: https://wonder.cdc.gov/ucd-icd10.html using filter criteria: State, Year, cause of injury = Firearm. Or download it from my GitHub repo.
5. Using the rpart package in the R statistical computing language, the input variables consisted of the per State per year total number of laws within each of 49 subcategories as coded by The State Firearm Laws Project and the output variable was annual raw firearm death rate per State per year. Input variables were scaled because law categories varied widely in size. The data were split into train and test sets. Typically, this is done by randomly selecting 20% of the data as a test set and 80% as a train set. However, for this data set, it is likely that there are correlations across States for a given year and correlations across years for a given State, or correlations among years within each small range of years. Therefore, the data set was split by odd or even year into two sets, each balanced for State and removing early/recent year bias. A repeat analysis completed with the typical 20/80 random split gave incredibly similar results.
6. Although I didn’t end up using unemployment or weighted population density per State per year, the datasets are available in my GitHub repo. I spent the time acquiring and putting them into tidy format and other people might want to use them. The 2016 State unemployment data were from US Department of Labor, Bureau of Labor Statistics. The unemployment rate data up to 2015 came from Iowa State University, Iowa Community Indicators Program under “Additional data: Download historical annual series (1980-current) in Excel”. They had already compiled the US Dept of Labor data for those years so I didn’t have to. The weighted population density data were obtained using code from Dan Goldstein on this post in Decision Science News.