HOW WE GOT THE STORY

To calculate the changing interest rates and how COVID-19 affected them, the AJC data specialist Nick Thieme undertook a statistical analysis of municipal bond data obtained from the Municipal Securities Rulemaking Board (MSRB), the regulating body in charge of municipal bonds in the U.S. The MSRB provided real-time bond transaction data covering all bond trades in the U.S. between October 2019 and October 2021. Additionally, data on the roughly 3,000 new bond issuances between October 2018 and October 2021 were acquired. These data provided a complete record of all new bonds issued and characteristics of their issuance along with all bonds traded and characteristics of their trading.

The AJC also acquired metadata on issuers and issuances from S&P Global Services, a financial company tasked with maintaining bond records. These data included the jurisdiction covered by a bond, the economic sector a bond pertains to, and financial characteristics of the bond. Demographic data on covered jurisdictions was obtained from the U.S. Census Bureau. Data on new COVID-19 cases over time comes from the AJC’S database of COVID-19 cases. Data on economic conditions over time comes from the Federal Reserve Economic Database (FRED).

Following strategies developed by public and private sector economic analysts, the AJC cleaned the data, and linked the datasets together to create a usable dataset of new bond issuances, metadata about the issuance, economic and demographic data for the region the bond pertains to, and COVID-19 cases in that same area. To determine whether bond coupons were higher before or after March 2020, the AJC used a so-called regression discontinuity design model. These models are considered the best post-treatment methods for causal inference.

Lastly, to estimate the size of the effect of COVID-19 on bond coupons, the AJC built a series of statistical models, increasing in complexity from a simple linear model, to general additive models (GAM), to mixed-effects GAMS, to mixed-effects GAMS with a large degree of freedom. These models were then used to estimate the effect of COVID-19, and all showed almost total agreement.

This methodology was vetted by Michael Lavine, professor emeritus of statistics at the University of Massachusetts Amherst.