Introduction

• Purpose & Goal: Using a combination of data visualization and time series analysis, I wanted to evaluate the trends in opioid-related deaths and analyze the impact of current policies and treatment, for the purpose of synthesizing recommendations for the opioid crisis in Washington state

• Data Sources:

  • Opioid overdose statistics: Washington State Department of Health

  • GIS data: Plotly & TIGIRS (scrapped from publicly available GIS GeoJSON & shape files)

• Modification & Data wrangling at a glance:

  • Renamed columns

  • Joined county & other location data to GIS files (GeoJSON) and location identifiers (FIPS codes)

  • Multiple transformations for time series data specific to libraries

• New libraries

  • GIS wrangling & visualization: rjson & tigirs

  • Time series analysis: CausalImpact & Prophet

• Overview of data limitations:

  • 2003 is missing

  • Total number deaths/overdoses is slightly inconsistent (typically off by 10-20 cases)

  • Changes to WA law resulting in limited access and analysis of data from 2021 to present (WAC-492-300)

“Age Adjusted Rate?”

When comparing populations, it’s important to consider differences in size. If raw counts are used, the data will be skewed towards larger of people. This can be addressed by calculating the rate of an event (e.g., opioid overdoses) by dividing the number of deaths by a group’s population, and multiplying the quotient by 100,000. This produces a crude rate that measures event occurrence by 100,000.

However, populations have different distributions of disease (and risk of disease). Age adjusted rate takes the crude rate and adds an age-specific weight to correct for demographic differences:

\(age\;adjusted\;rate =\;100,000\cdot (\frac{n\;deaths}{population})\)

Therefore, age adjusted rate in this data is interpreted as “the rate of deaths due to drug overdose per 100,000 residents in Washington, adjusted for age per population” (which can be state-wide, county-specific, regional, etc.).

It is important to note that this can skew the data. For example, the adjusted rates for Garfield county in 2005 and 2016 are abnormally high due to Garfield’s small population (< 2400). However, age adjusted rates are still useful because some population & demographic correction is better than none.

Summary of Trends

1. Across all drug classes, overdoses cases are increasing. There was a small decrease in 2009 until 2012, however it has greatly inclined first in 2015, and again during 2020.

2. The majority of overdose cases in WA are opioids which include natural Opiates (Morphine, Codeine) and synthetic Opioids (Heroin, Methadone, Vicodin, Oxycontin, Ultram, Fentanyl, and analogues).

3. Methadone and prescription, non-Fentanyl opioid cases are decreasing

4. Fentanyl overdoses are rapidly increasing

   • Note the difference in prescription & prescription, non-Methadone from 2016 onwards

   • Fentanyl prescribing rates have not increased, however Fentanyl drug reports by law enforcement have become more prevelant, suggesting that Fentanyl and its analogs are being obtained and used illegally (O’Donnel et al., 2017)

5. Psycho-stimulants (Methamphetamines) cases have also risen since the early 2010s.

Opioid Deaths in Washington (2000 - 2020)

Other Drug Deaths in Washington (2000 - 2020)

All Drug Deaths in Washington (2000 - 2020)

Interpreting Impact Plots

Making causal inferences when controlled experiments aren’t possible

In applied stats (research, analysis, etc.), we want to find the (potential) impact of what could have been and compare that to what actually happened. Normally, this is done through controlled experiments, but we can’t always do that due to ethical concerns and other constraints (time, money, ability, etc.). In this case, it would not be ethical to conduct an experiment on a subset of the population because that would mean:

1. We are withholding potentially life-saving treatment/policies to a subset of the participants

2. We are conducting experiments on a vulnerable population (opioid dependence/addiction is comorbid with a variety of physical, psychological, and economical problems).

One way to address this issue is through synthetic conditions. Using time series data, we can evaluate the significance of an event by comparing the data with a prediction that assumes an event did not occur. This prediction is known as the “counter-factual.” In this case, the counter-factual predicts the number or rate of death if an event (e.g., COVID-19) never happened. The counter-factual prediction is then compared with the actual data for causal inference. The model can also be adjusted using covariates - variables which impact the observed value (e.g., Opioid Deaths) but are not directly related.

Interpreting impact plots:

• The black line is the actual data. The blue dotted line is the predicted counterfactual. The blue bands are the confidence interval

• The x-intercept (grey dotted line) is the event and separates the graph into a pre-period (event did not exist) and a post-period (event is present)

Model Assumptions:

• Covariates are independent to the observed data (therefore, they aren’t affected by the event/intervention).

Impact on all Opioid Deaths

Impact on all Synthetic Opioid deaths

Prescription Drug Monitoring Programs

A PDMP is a state-run database which records the prescription and dispensing of controlled substances. PDMPs can improve prescribing behavior, thus reducing the number of opioids prescribed and exposure to opioids. Ultimately, this can reduce the number of deaths. Common changes to PDMP policies include improving the ease of access, requiring physicians to consult a PDMP prior to prescription, and real-time data policies (e.g., requiring pharmacists and physicians to update the PDMP on a daily or weekly basis, rather than a monthly schedule)

Washington’s PDMP, the “Prescription Monitoring Program,” was created in 2007 and implemented on October 7, 2011 (RCW 70.225).