About

Jonathan P. Caulkins is the H. Guyford Stever University Professor of Operations Research and Public Policy at Carnegie Mellon University's Heinz College. He specializes in systems analysis of problems related to drugs, crime, terror, violence, and prevention. His work in these areas has earned him several awards, including the David Kershaw Award from the Association of Public Policy Analysis and Management, a Robert Wood Johnson Health Investigator Award, and the INFORMS President's Award. Caulkins has contributed to the understanding of issues surrounding cannabis legalization and co-authored the second edition of his book Marijuana Legalization: What Everyone Needs to Know (Oxford University Press). His research interests also include optimal control, reputation and brand management, human trafficking, and black markets. He has taught quantitative decision-making across four continents to students from 50 countries, ranging from undergraduates to Ph.D. candidates and executives. Caulkins has authored or co-authored more than 140 refereed articles and 10 books. He has held leadership roles such as past Co-director of RAND's Drug Policy Research Center and founding Director of RAND's Pittsburgh office, and continues to collaborate with RAND on various projects. His academic background includes a Bachelor of Science and Master of Science in systems science from Washington University, and an S.M. in electrical engineering and computer science along with a doctorate in operations research from MIT.

Research topics

• Computer Security
• Medicine
• Economics
• Computer Science
• Psychology
• Environmental health
• Econometrics
• Economy
• Geography
• Psychiatry
• Mathematical optimization
• Macroeconomics
• Economic growth
• Mathematics

Selected publications

• Long-Run Trends in Self-Reported Heroin Initiation

  SSRN Electronic Journal · 2026-01-01

  preprintOpen access1st authorCorresponding
  PublisherDOI

• Did the illicit fentanyl trade experience a supply shock?

  Science · 2026-01-08 · 10 citations

  article

  A synthesis of government and social media data suggests a disruption, possibly tied to events in China.

  PublisherDOI

• Historical trends in self‐reported US heroin initiation

  Addiction · 2026-02-24

  articleOpen access1st authorCorresponding

  BACKGROUND AND AIMS: Illegal opioids create challenges for public health and safety. There is imperfect understanding of when use of illegally manufactured opioids increased. This paper examined data on self-reported year of first heroin use in the United States. DESIGN: Secondary analysis of general population survey data. SETTING: United States. PARTICIPANTS: 1 708 720 across 38 surveys from 1979 to 2023. MEASUREMENTS: The U.S. National Survey on Drug Use and Health and its predecessors ask respondents about their year of first use for various substances. In any single survey, few report initiating heroin in a specific year, but combining multiple surveys improves precision, enabling the plotting of time trends. Those who initiate can be broken down by age and by whether they report having misused pain relievers before they first used heroin. FINDINGS: Self-reported heroin initiation appears to have been rare before the late 1960s, although data for those years are sparse. It rose sharply to a peak in 1972, fell by almost half, and remained stable from 1978 to 1994. It rose by about 75% in the late 1990s and a further 85% by the early 2010s. The proportion reporting misuse of prescription opioids before first using heroin increased from about one-third before 1990 to one-half by 2000 and 80% by 2010. The proportion who were over age 40 at the time of first heroin use increased from nearly 0 before 1990 to 10% in 2012 and about 40% in 2020-2021. CONCLUSIONS: In the United States, heroin initiation that is self-reported to have occurred after 2000 differs in magnitude and character from that from the late 1970s through mid-1990s. Changes began before dates commonly associated with restrictions on opioid prescribing. This appears consistent with a view that "trading down" from prescription opioid misuse to consumption of illegally manufactured opioids did not only occur after implementation of policies to reduce opioid prescribing.

  PublisherDOI

• Possible effects of cheap fentanyl on drug markets, use, and harm: A theoretical analysis

  Global Crime · 2025-09-23 · 2 citations

  article1st author
  PublisherDOI

• HEALing communities study results, questions and implications

  Addiction · 2025-12-04

  editorialOpen access1st authorCorresponding

  The HEALing Communities Study (HCS) was a $350 million 4-year multi-site, community-level, cluster-randomized wait-list controlled trial of evidence-based practices for reducing opioid overdose deaths. It did not produce statistically significant reductions in deaths (its central outcome), treatment uptake or behavioral health service delivery, but it did reduce stigma. This disappointing result should prompt serious reflection within our community. The $350 million HCS was the ‘largest implementation science study ever funded in addiction research’ [1]. The evaluation was rigorous, with 67 communities in four states randomly assigned to the Communities that HEAL (CTH) treatment or to the control group. The stated goal was ‘to reduce opioid overdose deaths by 40% in three years’ predicated on a belief that ‘opioid overdose deaths are largely preventable’ [1]. HCS embodied the field's best wisdom by including (1) community engagement; (2) communication campaigns to increase awareness and demand for evidence based practices (EBPs) and to reduce stigma against people with opioid use disorder (OUD) and against medications for treating opioid use disorder (MOUD); and (3) a requirement that communities implement EBP for (3a) overdose education and naloxone distribution (OEND), (3b) MOUD and (3c) safer prescribing of opioid analgesics that could ‘significantly reduce opioid overdose deaths in a relatively short period of time’ [1]. Regarding the central outcome, there was a statistically not-significant 8% reduction in the opioid overdose death rate (P = 0.30) [2] and the overall overdose death rate (P = 0.26) [3]. Secondary outcomes included a statistically significant 37% decline in deaths from opioids combined with a psychostimulant other than cocaine, and small and not statistically significant reductions in deaths from opioids plus cocaine (6%) and opioids with benzodiazepine (1%). Outcomes for additional aims (e.g. testing the study's conceptually driven framework) are less easily summarized. Related studies found no statistically significant effect on (1) initiation, retention, and linkage to MOUD; (2) the rate of waivered practitioners or active prescribing of buprenorphine; or (3) the rate of individuals receiving behavioral health services reflected in Medicaid claims [4-6]. However, ‘the CTH intervention significantly changed stakeholders' perceived community stigma toward OUD and MOUD’ (P = 0.0007 and P = 0.0066, respectively) [7]. These results challenge confidence that CTH's recipe of community engagement; reducing stigma; and EBP for MOUD, OEND and safer prescribing necessarily produce major changes in death and other health outcomes. Researchers, policy makers and others who had that confidence need to adjust their beliefs or identify reasons why the trial failed. Three main conjectures have been offered for why CTH could have failed even if its approach remains sound. First, the intervention began shortly before coronavirus disease (COVID), hampering deployment. However, 235 EBPs were implemented by the start of the evaluation's comparison year [2], which extended through 30 June 2022. Additionally, implementation proceeded enough to reduce stigma [7], although COVID could have interfered more with healthcare delivery than with stigma reduction efforts. Second, ‘change in the illicit drug market may have reduced the effectiveness of the intervention, because fentanyl became a more prevalent opioid … [and] we do not know whether surges in fentanyl use over time were similar across communities’ [2]. Figure 1 plots the proportion of illegal opioid observations that were fentanyl for counties containing control communities, counties containing intervention communities and other counties in the intervention states (see Caulkins and Giri for more on the data) [8]. Fentanyl's spread was similar in all three. If anything, the spread was greater between baseline and evaluation for control counties, which could have enhanced not attenuated the apparent effectiveness of CTH. Therefore, although CTH may be less effective in the fentanyl era than in previous times, it is not clear that the particulars of fentanyl's spread produced a false negative result in HCS's evaluation of CTH. Third, ‘the HCS timeline and reach of selected EBPs may have been insufficient’ [4]; that is, EBPs may produce change—but not quickly—and/or the dose was too small. Barocas et al.'s [9] analysis of the economic value of resources deployed for interventions may support the latter idea. Although the HCS study budget was nearly $350 million, payments made by HCS to implement CTH totaled only $37.5 million. In-kind resources (e.g. community members' time) and non-HCS financial support added another $26.4 million. Therefore, the total economic value of resources devoted to CTH was $63.8 million or just $1.93 million per intervention site. Further, half of those resources were devoted to community engagement, and more to communications campaigns, leaving only one-third ($668 000 per site) for implementing EBP strategies. In theory, a fourth possibility is that because CTH empowered each community to make their own selections from the EBP list, the communities could have selected unwisely. If they selected options on the weaker end of the evidence-based list, that could have diluted the total impact. For those inclined to adjust beliefs about CTH in light of HCS results, there are many possibilities. To mention four possibilities that I ponder: (1) Perhaps the power of stigma-reduction efforts has been over-sold, since HCS reduced stigma but did not produce the other outcomes. (Note: RCTs evaluating stigma reduction's effects on distal outcomes, like death, are scant compared to RCTs on MOUD.) (2) Maybe CTH has high efficacy in small studies, but does not scale-up well. (3) Perhaps EBP are cost-effective (with only $668 000 invested per site in programming, saving a single life per site would cost-justify them), but do not save enough lives to bend the curve appreciably at a population level. (4) Perhaps EBP could bend the curve, but only with much greater investments than planners designing the HCS presumed it would take. Given the HCS results, I also wonder what did produce the much greater than 8% reduction in deaths that occurred between mid-2023 and mid-2025 for the United States as a whole and also for Canada. I do not advocate one view over another or wish for readers necessarily to agree with me. I do, though, suggest that we collectively engage with this challenge and be open to some change in thinking. If we do not change our minds in some way, can we ask taxpayers for another $350 million study—when that sum could instead fund CTH-scale EBP programming in approximately 500 communities? The public health community often urges people to heed scientific evidence on matters ranging from vaccines to dietary advice to COVID mitigation. Will we heed the HCS study evidence enough to change our minds in some way? Jonathan P. Caulkins: Writing—original draft (lead); writing—review and editing (lead). The author thanks Keith Humphreys, Beau Kilmer, Peter Reuter and two anonymous referees for helpful comments on earlier drafts. None. The data that support Figure 1 were made available by the HIDTA PMP program. They cannot be shared directly by the author but should be available from HIDTA PMP to anyone wishing to replicate the figure.

  PublisherOA PDFDOI

• Improving Opioid Use Estimates Through Multiple Data Sources

  JAMA Health Forum · 2025-05-09 · 1 citations

  articleOpen access1st authorCorresponding
  PublisherDOI

• Historical and contextual variation in daily opioid consumption rates: implications for supply control, service delivery, and research

  The American Journal of Drug and Alcohol Abuse · 2025-07-11 · 1 citations

  review1st authorCorresponding

  The adaptability of consumption has several potential implications. Expansions in supply could have greater effects on quantity consumed than on prevalence. Treatment protocols and overdose prevention strategies may need to adjust for higher baseline consumption. Furthermore, assumptions about health harms from long-term use may need revisiting if they are predicated on lower, historical consumption intensities. These findings are caveated by limitations in reporting of data and variations in methodologies. Hence, greater investments in monitoring consumption intensities are warranted.

  PublisherDOI

• Changes in self‐reported cannabis use in the United States from 1979 to 2022

  Addiction · 2024-05-22 · 113 citations

  articleOpen access1st authorCorresponding

  BACKGROUND AND AIMS: Multiple countries are considering revising cannabis policies. This study aimed to measure long-term trends in cannabis use in the United States and compare them with alcohol use. DESIGN AND SETTING: Secondary analysis of United States general population survey data. PARTICIPANTS: The national surveys had a total of 1 641 041 participants across 27 surveys from 1979 to 2022. MEASUREMENTS: Rates of use reported to the US National Survey on Drug Use and Health and its predecessors are described, as are trends in days of use reported. Four milepost years are contrasted: 1979 (first available data and end of relatively liberal policies of the 1970s), 1992 (end of 12 years of conservative Reagan-Bush era policies), 2008 (last year before the Justice Department signaled explicit federal non-interference with state-level legalizations) and 2022 (most recent data available). FINDINGS: Reported cannabis use declined to a nadir in 1992, with partial recovery through 2008, and substantial increases since then, particularly for measures of more intensive use. Between 2008 and 2022, the per capita rate of reporting past-year use increased by 120%, and days of use reported per capita increased by 218% (in absolute terms from the annual equivalent of 2.3 to 8.1 billion days per year). From 1992 to 2022, there was a 15-fold increase in the per capita rate of reporting daily or near daily use. Whereas the 1992 survey recorded 10 times as many daily or near daily alcohol as cannabis users (8.9 vs. 0.9 M), the 2022 survey, for the first time, recorded more daily and near daily users of cannabis than alcohol (17.7 vs. 14.7 M). Far more people drink, but high-frequency drinking is less common. In 2022, the median drinker reported drinking on 4-5 days in the past month, versus 15-16 days in the past month for cannabis. In 2022, past-month cannabis consumers were almost four times as likely to report daily or near daily use (42.3% vs. 10.9%) and 7.4 times more likely to report daily use (28.2% vs. 3.8%). CONCLUSIONS: Long-term trends in cannabis use in the United States parallel corresponding changes in cannabis policy, with declines during periods of greater restriction and growth during periods of policy liberalization. A growing share of cannabis consumers report daily or near daily use, and their numbers now exceed the number of daily and near daily drinkers.

  PublisherOA PDFDOI

• Riding the waves from epidemic to endemic: Viral mutations, immunological change and policy responses

  Theoretical Population Biology · 2024-02-02 · 1 citations

  articleOpen access

  Nonpharmaceutical interventions (NPI) are an important tool for countering pandemics such as COVID-19. Some are cheap; others disrupt economic, educational, and social activity. The latter force governments to balance the health benefits of reduced infection and death against broader lockdown-induced societal costs. A literature has developed modeling how to optimally adjust lockdown intensity as an epidemic evolves. This paper extends that literature by augmenting the classic SIR model with additional states and flows capturing decay over time in vaccine-conferred immunity, the possibility that mutations create variants that erode immunity, and that protection against infection erodes faster than protecting against severe illness. As in past models, we find that small changes in parameter values can tip the optimal response between very different solutions, but the extensions considered here create new types of solutions. In some instances, it can be optimal to incur perpetual epidemic waves even if the uncontrolled infection prevalence would settle down to a stable intermediate level.

  PublisherDOI

• Improving Estimates of the Prevalence of Opioid Use Disorder in the United States: Revising Keyes et al

  Journal of Addiction Medicine · 2024-09-02 · 7 citations

  article

  OBJECTIVES: The United States faces an ongoing drug overdose crisis, but accurate information on the prevalence of opioid use disorder (OUD) remains limited. A recent analysis by Keyes et al used a multiplier approach with drug poisoning mortality data to estimate OUD prevalence. Although insightful, this approach made stringent and partly inconsistent assumptions in interpreting mortality data, particularly synthetic opioid (SO)-involved and non-opioid-involved mortality. We revise that approach and resulting estimates to resolve inconsistencies and examine several alternative assumptions. METHODS: We examine 4 adjustments to Keyes and colleagues' estimation approach: (A) revising how the equations account for SO effects on mortality, (B) incorporating fentanyl prevalence data to inform estimates of SO lethality, (C) using opioid-involved drug poisoning data to estimate a plausible range for OUD prevalence, and (D) adjusting mortality data to account for underreporting of opioid involvement. RESULTS: Revising the estimation equation and SO lethality effect (adj. A and B) while using Keyes and colleagues' original assumption that people with OUD account for all fatal drug poisonings yields slightly higher estimates, with OUD population reaching 9.3 million in 2016 before declining to 7.6 million by 2019. Using only opioid-involved drug poisoning data (adj. C and D) provides a lower range, peaking at 6.4 million in 2014-2015 and declining to 3.8 million in 2019. CONCLUSIONS: The revised estimation equation presented is feasible and addresses limitations of the earlier method and hence should be used in future estimations. Alternative assumptions around drug poisoning data can also provide a plausible range of estimates for OUD population.

  PublisherDOI

Frequent coauthors

• Beau Kilmer

  107 shared

• Peter Reuter

  100 shared

• Gustav Feichtinger

  Statistics Austria

  89 shared

• Mark A. R.Kleiman

  New York University

  62 shared

• Gernot Tragler

  TU Wien

  56 shared

• Rosalie Liccardo Pacula

  University of Southern California

  54 shared

• Peter M. Kort

  Tilburg University

  43 shared

• Benedikt Fischer

  University of Auckland

  42 shared

Education

• Ph.D., Operations Research

  Carnegie Mellon University

  1986

• M.S., Operations Research

  Carnegie Mellon University

  1983

• B.S., Mathematics

  University of California, Berkeley

  1981

Awards & honors

• David Kershaw Award from the Association of Public Policy An…
• Robert Wood Johnson Health Investigator Award
• INFORMS President's Award