21 Novembre 2026

Psicologia Punto One | Psicologia e Psicologi nel Web

Menu

Menu

SCHEDA SCIENTIFICA2026Altra ricerca

Fentanyl purity and overdose decline: A reexamination of geographic trends.

Nabarun Dasgupta, Adams Sibley, Paula Gildner, Katherine Gora Combs, Lori A Post, Sam Tobias, Alex H Kral, Rosalie Liccardo Pacula

Science advances · 2026

Abstract originale

Drug overdose deaths in the United States reached record levels during the unregulated ("illicit") fentanyl era before recently declining. A plausible hypothesis is that a sudden drop in fentanyl purity ("supply shock") beginning in 2023 caused the downturn. We evaluated this hypothesis by replicating a published analysis with regional overdose data, using transfer function models accounting for time trends and autocorrelation; negative controls tested for spurious correlation. Replicating the original paper, we extracted fentanyl purity data from a Drug Enforcement Administration (DEA) figure; overdose mortality used standard vital statistics sources. We found rising national fentanyl purity did not track fatal fentanyl overdose rates in most United States regions, showing only modest association in the West. When both purity and mortality later declined, the observed associations were also seen with unrelated macroeconomic indicators that shared the same time pattern, raising concerns about spurious correlation in the original study. Canadian analyses also revealed variation in both strength and direction of association between fentanyl purity and overdose deaths at subprovince level. Differential selection bias in law enforcement purity assessment may have compromised longitudinal and geographic generalizability of DEA data, notably stimulated reporting for higher-purity wholesale samples in 2023-2024. National fentanyl purity alone does not provide sufficient explanation for the recent decline in overdose deaths. Reductions in localized purity may influence localized overdose decreases, but alternative causes must be considered nationally, including drug supply changes, public health interventions, demographic shifts, behavior changes, post-pandemic recovery, cartel dynamics, and fewer initiates to illicit opioid use.

Parole chiave

FentanylDrug OverdoseHumansUnited StatesCanadaGeography
Leggi il testo integrale originale

Testo estratto dalla pubblicazione: sono esclusi immagini, media, supplementi e parti soggette a diritti distinti.

Fentanyl purity and overdose decline: A reexamination of geographic trends

Abstract

Drug overdose deaths in the United States reached record levels during the unregulated (“illicit”) fentanyl era before recently declining. A plausible hypothesis is that a sudden drop in fentanyl purity (“supply shock”) beginning in 2023 caused the downturn. We evaluated this hypothesis by replicating a published analysis with regional overdose data, using transfer function models accounting for time trends and autocorrelation; negative controls tested for spurious correlation. Replicating the original paper, we extracted fentanyl purity data from a Drug Enforcement Administration (DEA) figure; overdose mortality used standard vital statistics sources. We found rising national fentanyl purity did not track fatal fentanyl overdose rates in most United States regions, showing only modest association in the West. When both purity and mortality later declined, the observed associations were also seen with unrelated macroeconomic indicators that shared the same time pattern, raising concerns about spurious correlation in the original study. Canadian analyses also revealed variation in both strength and direction of association between fentanyl purity and overdose deaths at subprovince level. Differential selection bias in law enforcement purity assessment may have compromised longitudinal and geographic generalizability of DEA data, notably stimulated reporting for higher-purity wholesale samples in 2023–2024. National fentanyl purity alone does not provide sufficient explanation for the recent decline in overdose deaths. Reductions in localized purity may influence localized overdose decreases, but alternative causes must be considered nationally, including drug supply changes, public health interventions, demographic shifts, behavior changes, post-pandemic recovery, cartel dynamics, and fewer initiates to illicit opioid use.

INTRODUCTION

Over the past half-decade, the United States experienced record drug overdose mortality, driven primarily by unregulated fentanyl, followed by an unprecedented decline (1–5). As of July 2025, all states showed substantial declines from peak overdose of at least 19%, with some states approaching 50% decline from peak (6). Vangelov et al. (7) proposed that a fentanyl “supply shock” beginning in May 2023, precipitated by Chinese government actions, was the primary driver of this decline.

Statistically, the conjecture hinges on their observation that, starting March to July 2023, powder purity appeared to fall from around 25 to 11% by the end of 2024. And, that fentanyl overdose deaths declined by “more than half” from a national peak in May 2023 during the same time. They offer a correlation statistic (r = 0.37) for the strength of the association from January 2019 to October 2024 (7).

Others have proposed a similar hypothesis in Canada (8). Given regional variation in overdose rates (9, 10) and subgroup trends (11, 12), we tested whether the “supply shock” hypothesis (7) holds at subnational levels in North America. We were motivated by the fact that overdose death rates by state peaked as early as 2021, with some declining gradually (6). Broadly summarizing previous drug cycles, rapid drops in overdose are believed to more likely represent drug supply disruption, such as that which occurred following the heroin drought in Australia (13). Gradual declines suggest structural factors (14) including interventions, behavior change, and long-run demographic shifts. This distinction has important policy priority-setting implications, informing resource allocation and intervention design. If a purity-only hypothesis can explain the recent decline in overdose, the role of interventions, behavior change, and demographics becomes less salient. Conversely, if purity can only explain a limited portion of overdose declines, the importance of interventions is heightened. Therefore, testing the robustness of the supply shock hypothesis (7) has considerable implications for public health, implementation science, and government policy.

Population studies of overdose increase

Our main test was whether overdose fatalities increased in tandem with fentanyl purity before decline. We analyzed two periods: Block 1 (July 2022 to September 2023), when purity rose, and Block 2 (October 2023 to September 2024), when purity declined. Prior studies reported positive associations between fentanyl supply and overdose mortality at state/province and national levels (15–18), although some appeared inflated by autocorrelation (17).

Quantifying the exact nature of the association is benefitted by understanding geographic variation in drug purity (19), variations in where and when fentanyl initially emerged (20–22), adulterants (22–25), and route of administration (26). A previous study in British Columbia found routine twofold fluctuation in fentanyl purity over 3 years (27). Even within cities, retail fentanyl purity varies widely; in Los Angeles (28), expected “fentanyl” samples ranged from <1 to 697 mg of active ingredient per gram, with a median purity of 10.0%.

To test the robustness of the supply shock hypothesis, we replicated the Vangelov et al. (7) study at a subnational level. We also used negative controls using macroeconomic indicators to evaluate the presence of shared time trend structures, thereby assessing the possibility of spurious correlation.

Methods limitations of original analysis

Trend and autocorrelation

Both the purity and mortality series showed autocorrelation and nonlinearity, so omnibus correlation statistics would be inflated and violate independence assumptions. The original analysis did not specify which correlation coefficient was used to describe the strength of association, the associated assumptions, nor how statistical significance was assessed, limiting replicability. We therefore used autoregressive integrated moving average (ARIMA)–based transfer function models (TFMs) (29) and two negative controls to isolate associations beyond shared secular time trends.

Aggregation bias

The published analysis (7) was conducted at the aggregated national level, despite local variability in both fentanyl purity and overdose (30). The explicit assumption of an exogenous-origin supply shock was that both death rates and fentanyl purity were nationally distributed uniformly. The unregulated drug supply can be highly regionalized, and many states had declining overdose rates (20) years before the onset of the purported supply shock. We were concerned that the independent (purity) and dependent (fatal overdose rate) variables could be subject to aggregation bias in the form of Simpson’s paradox (31), wherein the aggregate national rate increases (despite declines in some states) because the national rate is a weighted composite that can be driven upward by larger or faster-growing increases in a subset of states, reversing or obscuring regional trends. Therefore, using a common-input, differential-response design, we compared overdose mortality rates by four US Census Regions to the national fentanyl purity time series, focusing on two temporally adjacent periods of increasing and decreasing fentanyl purity comprising the crux of the supply shock hypothesis.

RESULTS

Descriptive findings

Figure 1 shows that national fentanyl overdose mortality rose sharply beginning in spring 2020, remained elevated through mid-2021, and began declining in mid-2023. Drug Enforcement Administration (DEA)–reported powder purity (text S1) followed a different trajectory, rising mainly from 2022 through 2023 before declining in 2024.

Raw Pearson correlations between national purity and mortality were negative for the Northeast (raw r = −0.47), Midwest (raw r = −0.54), and South (raw r = −0.27), leading to the improbable appearance that overdoses declined while fentanyl purity increased. The West (raw r = +0.79) did see correlation. Hence, accounting for autocorrelation and geographic stratification is essential to interpreting the impact of the purported supply shock.

At the start of Block 1 (July 2022; Fig. 2), the DEA-reported “national” fentanyl powder purity was 12.8% (Table 1), doubling to 25.8% by September 2023. Fentanyl overdose mortality rates decreased substantially during that same time period, with the greatest rate drop in the Midwest (Δ−45.4%), whereas only increasing slightly (9%) in the West.

Table 1.Fentanyl powder purity and overdose rates during two blocks of purity change.

Block 1 (July 2022 to September 2023)

DEA fentanyl powder purity* Monthly fentanyl overdose death rate (per 100,000 population)†

Region July 2022 September 2023 Percent change July 2022 September 2023 Percent change

Northeast 12.8% 25.8% +101.6% 2.46 1.96 −20.3%

Midwest 12.8% 25.8% +101.6% 2.93 1.60 −45.4%

South 12.8% 25.8% +101.6% 1.86 1.58 −15.1%

West 12.8% 25.8% +101.6% 1.55 1.69 +9.0%

Block 2 (October 2023 to September 2024)

DEA fentanyl powder purity* Monthly fentanyl overdose death rate (per 100,000 population)†

Region October 2023 September 2024 Percent change October 2023 September 2024 Percent change

Northeast 24.7% 11.2% −54.7% 2.09 1.07 −48.8%

Midwest 24.7% 11.2% −54.7% 1.58 0.83 −47.5%

South 24.7% 11.2% −54.7% 1.55 0.83 −46.5%

West 24.7% 11.2% −54.7% 1.63 0.91 −44.2%

*The purity series is identical across four panels in each block, reflecting its status as a single national estimate from Drug Enforcement Administration (DEA) without geographic stratification, corresponding to the common-input, differential-response design.

†Drug purity data were extracted from a figure in a US Drug Enforcement Administration report. Overdose metrics were processed by OD Pulse at Northwestern University using US Centers for Disease Control and Prevention vital statistics data.

During Block 2 fentanyl purity declined (Δ−54.6%) from 24.7% in October 2023 to 11.2% in September 2024 (Fig. 2). Monthly fentanyl overdose death rates declined, 44.2 to 48.8% across Census Regions during Block 2.

National TFMs

We first replicated Vangelov et al.’s (7) analysis using autoregressive-corrected models for the sake of direct comparison to the original publication. (This analysis was not included in the preregistration but requested during the manuscript review process.) When the TFM was applied to national overdose mortality against the national purity series, the full-period gain was statistically significant but weak (gain = +0.026 deaths per 100,000 per percentage-point change in purity, P = 0.011, TFM R2 = 0.249); however, residual autocorrelation was detected in the full-period model (Ljung-Box P = 0.001), consistent with a structural break between the rising and declining phases and supporting the analytic decision to estimate the two blocks of increase/decline separately.

Dividing the full period into its two structurally distinct blocks reveals that this pooled national association is driven primarily by Block 2, where the gain approached but did not reach conventional significance (gain = +0.022, P = 0.059, TFM R2 = 0.38). The Block 1 gain (the critical test of whether rising purity tracked rising mortality) was not meaningful (gain = +0.034, P = 0.143, R2 = 0.184). This block-stratified result points to the purity-mortality association detected in national data reflecting convergent post-2023 decline trajectories rather than consistent pharmacological coupling across both increase and decline phases. Although these analyses afford a general sense of decoupling of the purity-overdose relationship nationally, they are prelude to the preregistered analyses where regional models were central to testing the supply shock hypothesis.

Regional models: Period of fentanyl purity increase

In Block 1, during the period of purported rising fentanyl purity (July 2022 to September 2023), the transfer function gain (Table 2) was substantively meaningful only in the West (gain = 0.056 deaths per 100,000 per month per percentage-point change in powder fentanyl purity per month, P = 0.10, TFM R2 = 0.22). The Northeast and Midwest showed near-zero gains (0.008 for both, P > 0.80, TFM R2 = 0.005), indicating that month-to-month innovations in the “national” purity series carried no predictive information for mortality rates. The South showed similarly weak association (gain = 0.012, P = 0.59). The discordance between TFM R2 and penalized R2 further confirms that the fentanyl purity series in the Northeast, Midwest, and South (Table 2) make no meaningful information gains on explaining overdose mortality rates.

Table 2.TFM results for the association between fentanyl purity and monthly fentanyl overdose mortality rate by US Census Region.

Gain represents the estimated change in monthly mortality rate per 100,000 associated with a one percentage-point change in national fentanyl purity, after prewhitening both series via an ARIMA(1,1,0) filter derived from the input series (purity). TFM, transfer function model; raw r, Pearson correlation between purity and mortality in levels (unprewhitened); R2, variance explained by the TFM on prewhitened series; penalized R2, adjusted R2 accounting for number of predictors and sample size, providing a more conservative measure of fit that decreases when predictors do not improve explanatory power; SE, standard error.

Region n Gain SE P TFM R2 Penalized R2 Raw r

Block 1 (July 2022 to September 2023)

Northeast 15 +0.008 0.031 0.815 0.005 −0.085 −0.47

Midwest 15 +0.008 0.032 0.810 0.005 −0.085 −0.54

South 15 +0.012 0.021 0.592 0.027 −0.061 −0.27

West 15 +0.056 0.032 0.103 0.223 0.153 +0.79

Block 2 (October 2023 to September 2024)

Northeast 12 +0.027 0.011 0.038* 0.436 0.366 +0.92

Midwest 12 +0.021 0.007 0.019* 0.519 0.459 +0.95

South 12 +0.010 0.006 0.161 0.230 0.134 +0.86

West 12 +0.020 0.015 0.226 0.177 0.074 +0.83

*P < 0.05.

Regional models: Period of fentanyl purity decrease

The purity-overdose association for the West (TFM R2 = 0.18, P = 0.226) was not meaningful in Block 2 (October 2023 to September 2024), despite being plausibly correlated in Block 1. This divergent finding casts doubt on the ability of the purity series to fully explain overdose rates in the West. Stated another way, it is difficult to conclude that overdoses were declining only due to fentanyl purity in Block 2 because overdose rates were decoupled from increasing purity in Block 1. For the South, the purity-mortality relationship remained weak in both time periods, a region that includes southern Appalachian states with entrenched, elevated intergenerational overdose rates.

In Block 2, as both purity and mortality declined, transfer function gains became statistically significant in the Northeast (0.027, P = 0.04) and Midwest (0.021, P = 0.02), with TFM R2 = 0.44 and 0.52 (penalized R2 = 0.37 and 0.46), respectively. This apparent strengthening of the purity-mortality relationship may reflect convergent macro time trends, where both series decline simultaneously alongside broader post-COVID societal indicators, rather than genuine validation of the purity series as a causal predictor of mortality, necessitating the use of negative controls.

Negative controls

To assess spurious correlation, we compared overdose mortality with two macroeconomic indicators (consumer price inflation and consumer sentiment; text S2) that have no direct pharmacologic pathway to overdose death.

In unadjusted time-series analysis (table S3), overdose mortality rates were strongly correlated with both inflation (Pearson raw r = +0.63, R2 = 0.40) and consumer sentiment (raw r = −0.73, R2 = 0.53), associations substantially greater than the r = 0.37 reported in the original publication (7). Raw cross-correlation between overdose mortality rate and each economic indicator (fig. S3, Row B) would have shown “statistically significant” associations at every lad from −12 to +12 months, further evidence that the overdose associations are spurious.

Following TFM adjustment for autocorrelation and time trend, both associations reversed direction entirely: the overdose-inflation correlation shifted from rraw = +0.63 to rTFM = −0.34, and the overdose-sentiment correlation shifted from rraw = −0.73 to rTFM = +0.32. Sign reversals of this magnitude are diagnostic in TFMs; when prewhitening transforms a strong positive correlation into a negative one, the original association is, by definition (29), an artifact of shared temporal structure rather than causal underlying relationship. Fentanyl purity follows the same post-COVID arc as these macroeconomic indicators, rendering fentanyl purity structurally indistinguishable from unrelated series. The negative control analyses reveal that the hypothesized purity-overdose associations cannot be interpreted as causal and are more likely spurious.

Comparison with Canada

As Vangelov et al. (7) did, we also compared our findings to British Columbia, where Tobias et al. (15) used geographically resolved fentanyl concentration data from community drug checking services to model the same fentanyl purity-mortality relationship, using generalized additive mixed models stratified by n = 16, subprovince health districts. Each one percentage-point increase in fentanyl concentration was associated with a +0.072 gain in drug-related mortality per 100,000 population (P = 0.029), an effect size roughly 30% larger than our strongest Block 1 gain (West, 0.056), and six- to ninefold larger than gains in the other three regions.

Critically, the British Columbia study also found substantial regional heterogeneity in the strength of association, with gain estimates ranging from −0.024 to +0.171 across subprovince health service delivery areas, indicating that, even with geographically specific data, the purity-mortality relationship varies highly by local context. The attenuation of our effect sizes relative to those reported by Tobias et al. (15) is consistent with measurement error introduced by geographic mismatch between the exposure and outcome.

In Toronto, opioid mortality peaked in late 2020 through early 2021 and subsequently entered a sustained downward trajectory, albeit with a prolonged period of fluctuation before sharper declines observed in 2024. In contrast, mean fentanyl purity remained elevated through 2023 and early 2024 before declining (32). This lack of temporal alignment suggests that reductions in average potency may have contributed to later stages of the trajectory but are unlikely to explain the initial downward inflection in mortality.

DISCUSSION

Our primary test for assessing the supply shock hypothesis (7) was to evaluate whether increases in the fentanyl purity were associated with increases in overdose mortality. Without proper time-series adjustment for trend and autocorrelation, correlations during the period of purity increase would have found a negative association in most of the country (raw Pearson r = −0.54 in the Midwest, raw r = −0.47 in the Northeast), suggesting that higher purity coincided with fewer deaths paradoxically. Previous analyses (15, 18, 30, 33, 34) have shown a strong positive correlation between fentanyl purity and overdose mortality; if we could not replicate even this established association in Block 1, it seems the purity series published in the National Drug Threat Assessment should be used cautiously in making causal inference. The DEA cautioned (35) against exactly this interpretation due to temporary sampling directives in 2023–2024 that focused on higher-purity wholesale samples (800 g or more), detailed in text S3.

In time-series analysis, when multiple social indicators move in the same direction, associations can arise spuriously. Our negative control analyses show that Block 2 overdose rate associations are compatible with shared post-COVID time trend structure rather than a specific pharmacologic effect of fentanyl purity. Contemporaneous declines in overdose mortality from psychostimulants and non-opioid substances (36) lend further credence to considering broader societal time trends as influencing all-substance overdose mortality declines.

After correcting for assumption violations, the national purity series was meaningfully associated with overdose increases only in the West during Block 1, and even there modestly. In the Northeast and Midwest, the same series explained essentially none of the month-to-month variation during the rise phase and then became only mildly associated during the decline phase. Together, these findings support two conclusions: The purity series is not nationally representative, and structural factors likely contributed materially to both the rise and fall in overdose mortality.

The South further illustrates this mismatch: Despite heavy representation in DEA field-lab sampling (44% of all powder samples in 2023; text S3), the purity series showed weak association with overdose in both blocks. This likely reflects the fact that several southern states had already entered decline (6) years before the purported 2023 supply shock. Some local associations may still exist, but they would require geographically resolved data to interpret credibly. A single national purity series regressed against national mortality is insufficient to support causal claims about a fixed-time-point supply shock in a large, heterogeneous country.

These findings have practical policy implications. If no clear national supply shock explains overdose downturns, greater attention should be paid to local structural factors that changed over the same period, including naloxone diffusion, treatment expansion, recovery supports, demographic shifts, and timing of fentanyl displacement of heroin (1). We do not foreclose the possibility that fentanyl purity declined in specific locations around 2023, including settings where fentanyl was displaced by sedatives or numbing agents (37, 38). However, a simplistic supply shock is not supported by the evidence.

We also uncovered that the purity series used in the DEA figure does not align with other DEA reporting. For example, in 2024, the DEA reported (39) the national powder fentanyl purity of 11.3%, whereas the extracted data from the figure had it as 15.2%. In 2023, the reported average DEA field lab powder purity was 19.7% and DEA’s Fentanyl Profiling Program purity was 33.5% (35), whereas figure-extracted data had it as 23.3%. These values were substantially higher than contemporaneous retail-level estimates from Los Angeles (28), raising concern about sampling bias. We emphasize the need for caution when using poorly documented law enforcement–derived data in policy analyses, and sampling methodology and datasets should be explicitly disclosed as part of governmental reports.

Localized purity impacts

Despite our analysis not being able to support a supply shock hypothesis at national or regional levels, we contend that local variation in fentanyl supply is still likely to be real and consequential. An arising question is what the appropriate level of geographic granularity would be required to interrogate the purity-overdose association with fidelity. This question is consistent with our original suspicion of Simpson’s paradox, whereby aggregate national fentanyl purity and overdose rates mask local variation.

Returning to Tobias et al. (15), the most geographically resolved evidence from British Columbia reinforces the need to precisely measure local fentanyl concentration at the retail level. Point-of-care drug-checking samples are arguably the closest available approximation to true street-level exposure, in contrast to wholesale law enforcement seizures. Median fentanyl concentration varied across districts in 2023, from 6.0% on Central Vancouver Island to 18.2% in North Shore/Coast Garibaldi, a more than threefold difference. Tobias et al. (15) also note that areas with more urban communities carried higher fentanyl concentrations than rural ones, but the distribution was not uniform.

With such wide heterogeneity in the exposure, it is not unexpected that the concentration-mortality association reported by Tobias et al. (15) was likewise heterogeneous. A random-slope specification was statistically required to model the association, isolating and allowing each subprovince health district to have its own trajectory. Even despite this, the resulting model’s explanatory power ranged from R2 = 0 to 58% across areas. Because the strongest slopes appeared in both a dense urban setting (Vancouver, β = 0.169) and a largely rural one (Northern Interior, β = 0.171), the relationship is shaped by local context in ways that do not reduce to a simple urban-rural dichotomy. This contingency is critical, pointing beyond simple pharmacological explanation and requiring consideration of local structural factors such as market maturity, time since fentanyl emergence, and hyperlocal drug distribution networks. The assumption of uniformity in the single national supply shock hypothesis ignores this evidentiary nuance.

In the British Columbia study, estimate precision was greatest in populous districts with dense drug checking coverage and least stable where sampling was sparse, so the apparent strength of a purity-mortality association depends on observing the local drug supply accurately. The national series we analyzed is built on sampling that is profoundly uneven across US states and regions. As detailed in text S3, 44% of 2023 DEA powder exhibits originated in the South (n = 3370) but only 14% in the West (n = 1117), of which California alone accounted for 59% (n = 657). By contrast, states like Alaska contributed only n = 18 samples. Placing Vangelov et al. (7) in the context of our analysis, we found that the West where the series tracked mortality most closely relies on the most geographically sparse sampling base. In contrast, the South, which dominates the national purity sample, showed weak association with overdose in both blocks. A “national” fentanyl purity series that obscures such unequal representation cannot be read as equally informative nationwide, reflecting sampling artifact as much as any true local effect. Resolving this question will require localized, adequately powered exposure data, such as those generated by drug checking interventions (40).

Alternative causal hypotheses

The basis for the supply shock hypothesis is the pharmacological expectation that higher fentanyl concentrations should increase risk of respiratory depression and fatal overdose at the individual level. However, comparing opioid-naïve humans to those with opioid tolerance from chronic use, experimental studies have shown fentanyl to require a 4.3-fold difference in concentration to achieve the same 50% ventilatory depression in both populations (41). Therefore, population-level heterogeneity in the relationship between purity and overdose may be conditional on local levels of opioid tolerance, similar to the protective effect of steady-state pharmacokinetics during methadone maintenance (42). We foregrounded the opioid-tolerance hypothesis first not because it outranks other explanations but because it is the mechanism most directly faithful to the pharmacological premise of the supply shock hypothesis proposed by Vangelov et al. (7). Beyond this purity consideration, we posit five nonexclusive mechanisms that may explain the decline in overdose.

Drug supply changes

Two sequential waves of illicit drug supply changes are suspected to have broadly modified overdose risk. First, the replacement of heroin with fentanyl, spreading east to west in the United States, catalyzed an unprecedented acceleration in overdose deaths (20, 36), with state-by-state (43) trajectories subsequently returning toward baseline in a consistent multiyear cascade. Second, in places where fentanyl had appeared earliest, the unregulated drug supply then became adulterated with strong sedatives and numbing agents, accompanied by reports of growing user dissatisfaction (44), conditions that can suppress consumption, both independent of and as a reaction to fentanyl concentration (45, 46). Adulteration of fentanyl with xylazine has been shown to result in compensatory protective behavior to use less drugs or less frequently (44) and less severe overdose in animal models (47) and medical practice (48). Upstream causes of drug supply fluctuations are poorly understood, but dynamics within and between organized criminal organizations (49) cannot be discounted, with external pressure exerted from interdiction, outside of the United States and Canada, as well as at ports of entry.

Demographic shifts

Second, demographic shifts were evident during the study period, which cannot be explained by a simple national supply shock hypothesis. The population of people who use opioids has been shifting in recent years, with an established cohort aging through middle age and fewer initiates entering use (36, 50), structurally reducing the size of the at-risk population over time. During the study period, as overdose deaths reached peak levels in the United States, the greatest growth (51, 52) in fentanyl mortality was observed in African American and Native American men. A single national supply shock hypothesis cannot cohesively explain these shifting overdose demographics without exploring causal mediators.

Behavior changes

Within drug using networks, behavioral adaptation, using smaller amounts, using less often, using test strips, shifting away from injection toward smoking and snorting (53, 54), and adopting safer drug use practices have been shown to alter overdose risk (55, 56). The aforementioned length of time elapsed because the emergence of unregulated fentanyl in the local drug supply may afford people time to adapt to the different potency (relative to heroin) while causing an initial rapid rise in overdose deaths (43) that diminish the pool of current illicit drug users, in turn creating intergenerational negative feedback loops (57, 58) that discourage continued illicit drug use and depress the number of illicit opioid initiates. A single time-point supply shock does not measure or consider these as mediators.

Public health interventions

Community-based and clinical interventions expanded substantially over the study period, particularly since the height of the coronavirus pandemic (COVID), including naloxone distribution (43) and broadened access to evidence-based treatment (1, 59). In addition, distribution of opioid settlement funds, reduced stigma for seeking treatment among young adults and youth (60), and policy changes (61) including telehealth and Medicaid expansion (62), further created the conditions for fundamental reductions in overdose risk by creating off-ramps for problematic substance use. Although these interventions may take time to make meaningful impact on overdose rates, the simultaneous implementation of these many courses of action could deliver faster results than each would be expected to produce in isolation.

Societal postpandemic recovery

There is mounting evidence (63) that overdose declines parallel a post-COVID pandemic recovery, with declining (socially positive) trends seen in broader societal indicators, including violent crime (64, 65), motor vehicle fatalities (66), and suicide (67). These downward trajectories are similar to the macroeconomic series examined in our negative control analysis. Our finding that the purity-mortality association rides on shared postpandemic temporal structure with unrelated phenomena suggests that broad societal trends cannot be entirely disentangled from the decline in overdose deaths, although their causal connection remains methodologically challenging to establish.

Limitations

Neither our study nor others (7, 18, 19, 30, 68) accounted for myriad interventions implemented to reduce overdose mortality at a local level. These overlapping efforts are notoriously difficult to measure prospectively over large areas, as we previously (69) demonstrated.

We acknowledge that the analytic blocks were short for transfer function modeling. That constraint was imposed by the brief rise-then-fall structure of the purity series; pooling across the structural break would have violated stationarity assumptions. Bayesian Information Criterion (BIC) confirmed that no lagged transfer terms were warranted. The consistency of null findings across regions in Block 1 supports the robustness of the central result despite limited power. Time-varying underascertainment of fentanyl’s involvement in overdose, especially in jurisdictions where postmortem toxicology practices changed over time, may have affected regional mortality rates (70).

The purity series was manually digitized from a figure and may contain measurement error. We did not model adulterants such as xylazine, medetomidine, or bromazolam, which may alter both overdose risk and compensatory behavior; fentanyl purity alone is an incomplete measure of mortality risk in the contemporary synthetic drug supply.

We acknowledge that systematically collected subnational purity data would have been a boon to this analysis, and the common input-differential response design is not inherently causal. Without more granular governmental or drug checking data, direct causal modeling was not possible, and our intent was to evaluate the hypothesis posed by Vangelov et al. (7) through replication.

For data derived from drug checking services in British Columbia, we also note that different health service areas had varying levels of sampling volume (15). We are not aware of any precedent to accurately draw a fully representative sample of the unregulated drug supply over large geographic areas (40).

We also did not examine demographic shifts in overdose mortality, including age and race/ethnicity, which may be central to explaining the recent decline. A fundamental causal explanation for the decline in overdose deaths is likely fewer initiates to illicit opioid use (14). These limitations should be addressed in future studies.

Methodological and ethical implications

Methodologically, national time series should not be entered into regression models without addressing autocorrelation, time trends, and subnational variation. Also, enforcement-derived data should not be used without adequate documentation of sampling.

We do not dispute that drug supply characteristics and overdose mortality are linked. Our point is narrower: This national purity series (7) does not provide causal proof. In the current synthetic-drug era, “concentration” may be a more useful concept than “purity,” because mortality risk depends on a changing mixture of fentanyl, sedatives, novel opioids, and other co-occurring substances. A decline in fentanyl purity alone may not correspond to a decline in total overdose risk.

No statistical model can substitute for the obligation to provide humane, effective care to reduce overdose deaths. The geopolitical framing of the supply shock hypothesis risks implying that interdiction was the sole causative factor, but the recent history of drug supply changes, interventions, demographic shifts, and behavior changes require us to consider more complex and interwoven explanations. A purely reductionist approach demeans the considerable efforts enacted across the United States and Canada to reduce overdose deaths, and we entreat researchers to engage in pursuing analysis of multicause possibilities.

MATERIALS AND METHODS

Data sources

Fentanyl purity

The DEA publicly reports purity quantification from regional field labs where law enforcement seizures (ostensibly at or near street retail level) are quantified for prosecutorial evidence. In addition, a smaller number of intentionally selected samples is sent to the DEA’s Fentanyl Profiling Program (FPP). In 2023, the focus of both shifted to higher-volume (“wholesale”) drug seizures of greater than 800 g (35), with the average powder purity via DEA field labs as 19.7% and selected higher-volume samples via the FPP as 33.5%. Therefore, the data used by Vangelov et al. (7) from the National Drug Threat Assessment may represent a sampling bubble in 2023–2024, where higher-purity samples were comingled with retail samples. We recreated the purity series used by Vangelov et al. (7), extracting data points from the fentanyl powder figure in the 2025 DEA National Drug Threat Assessment (71) using code-assisted manual coordinate extraction to handle pixel-level anti-aliasing (text S1). We were able to replicate (fig. S1) the fentanyl powder purity time series with fidelity: Lin’s concordance correlation coefficient for reproducibility was 0.999.

Overdose data

Monthly drug overdose mortality counts were obtained from OD Pulse, an overdose surveillance data platform maintained by Northwestern University (36). OD Pulse preprocesses mortality data from the CDC Wide-ranging Online Data for Epidemiologic Research (WONDER). Overdose deaths were defined X40-X44, X60-X64, X85, or Y10-Y14 (36, 72) and coded as T40.4 (synthetic opioids, predominantly unregulated fentanyl and analogs). Monthly death counts were aggregated by the four US Census regions (Northeast, Midwest, South, and West) and converted to population-adjusted mortality rates per 100,000 using Census Bureau intercensal population estimates.

Statistical analysis: TFMs

We specified a common-input, differential-response design: The same national purity series served as the exposure for all four Census Regions, with region-specific monthly overdose mortality rates per 100,000 as the output. The input series was monthly fentanyl powder purity (percent) derived from the extracted DEA figure, which lacked geographic stratification. This design directly tests the assumption underlying Vangelov et al.’s analysis (7), that a single national purity measure is equally informative across the country, by using the original study’s own premise as the testable hypothesis. If the DEA-reported national series were truly representative of local supply conditions, we would expect comparable transfer function gains across regions. Conversely, divergent results constitute evidence of geographic mismatch between the national measure and local supply.

To assess the temporal relationship between fentanyl powder purity and overdose mortality rates across US Census Regions, we estimated TFMs following the procedure established by Box-Jenkins (29, 73) and applied in infectious disease surveillance (74) and other health fields (75).

ARIMA-based prewhitening

TFMs were applied to the primary analysis and for the negative controls. We first fit (“prewhitened”) an ARIMA time-series model that captures both trend and month-to-month momentum in the purity series, with residuals representing each month’s value that could not have been predicted from the series’ own history. Applying the identical filter to the mortality series removed shared temporal structure and autoregressive dependence, with surviving correlation representing the contemporaneous relationship between changes in purity and changes in mortality. This conservative procedure preserves the genuine purity-mortality relationship and eliminates shared time trend (29).

We selected ARIMA(1,1,0) as the prewhitening filter because the purity series exhibits both a clear trend (addressed by the differencing term, d = 1) and modest persistence in month-to-month changes (addressed by a single autoregressive term, P = 1), whereas the short series length (12–15 months per cell) precludes more complex specifications without overfitting. The BIC (76) comparison of lag-0 month versus lag-0,1 transfer function specifications confirmed that no additional complexity was warranted in any of the eight region-block cells. The same autoregressive-differencing filter was then applied to the mortality series, yielding a pair of prewhitened series free of shared autocorrelation structure. The same process was applied to the negative controls.

Time block structure

On the basis of visual time-series inspection, the study period (July 2022 to September 2024; 27 months) was divided into two analytic blocks to reflect an apparent structural purity shift: Block 1 (July 2022 to September 2023, n = 15 months), characterized by rising fentanyl purity, and Block 2 (October 2023 to September 2024, n = 12 months), characterized by declining purity. Models were estimated separately for each region-block combination (eight cells). Model fit was assessed using adjusted coefficient of determination (R2), penalized for model complexity (77).

ARIMA diagnostics

For each Region-Block pair (n = 8), ARIMA fit was assessed for normality (Jarque-Bera) and serial correlation (Ljung-Box) (78). Visual inspection included residuals over time, residual autocorrelation function (ACF), and residual partial autocorrelation function (PACF) (with 6-month lags), quantile-quantile (QQ) plot of residuals, and plot of observed versus expected purity values over time. The .ipynb notebook provided in the Dryad repository contains every diagnostic plot.

TFM regression model outputs

We estimated impulse-response weights by regressing the filtered mortality series on the contemporaneous filtered purity series (or negative controls) via ordinary least squares. Impulse-response weights estimate magnitude of the mortality response to a one-unit “impulse” (innovation) in fentanyl purity at each month lag. In our lag-0 specification, a single weight represents how much of an “innovation” in purity passes through to mortality within the same month.

The primary interpretive metric was transfer function gain, defined as the sum of impulse-response weights, representing the expected change in monthly mortality rate per 100,000 population associated with a one percentage-point change in fentanyl purity. We report gain estimates with standard errors and significance levels for contemporaneous effect.

Assessing model fit via penalized R2

We assessed model fit in the prewhitened domain using the coefficient of determination (R2). R2 was computed as 1 − (SSR/SST), where SSR is the residual sum of squares from the prewhitened regression and SST is the total sum of squares around the mean of the prewhitened outcome. Because R2 increases monotonically with additional predictors, we also report the adjusted R2 (77), which penalizes model complexity as penalized R2 = 1 − [(1 − R2)(n − 1)/(n − p − 1)], where n is the number of observations used in the prewhitened regression and p is the number of predictors (excluding the intercept). In our primary specification with a single predictor and relatively small monthly blocks, R2 provides an intuitive measure of variance explained, whereas penalized R2 was preferred for comparing specifications and guarding against overfitting, particularly when exploring lagged terms that increase the ARIMA p component. We present both metrics but rely on penalized R2 for model selection across transfer structures; close agreement between R2 and penalized R2 indicates a parsimonious model with genuine explanatory power, whereas divergence suggests that apparent in-sample fit may be inflated by model complexity relative to sample size.

Modeling assumptions

Time-series methods, including the TFM, assume stationarity (that the statistical properties of a series, its mean, variance, and autocorrelation remain stable over time), and violating this assumption risks detecting spurious associations driven by shared trends rather than genuine short-run coupling. Both the purity and mortality series are clearly nonstationary over the study period (purity rises then falls, mortality declines), but the differencing component (d = 1) of our ARIMA(1,1,0) prewhitening filter explicitly addresses this by transforming both series into their month-to-month changes before estimating the transfer function, effectively removing the nonstationary trend regardless of whether it is deterministic or stochastic. The additional step of dividing the study period into two analytic blocks provides further protection as any residual nonstationarity from the structural shift between the rising and declining phases is isolated rather than allowed to span the full series.

Negative controls

Any set of indicators that share a structurally similar time trajectory (rise through 2021–2022, decline thereafter) should produce a large, statistically significant association regardless of causality. We conducted a negative control cross-correlation analysis using two common macroeconomic indicators (text S2) with no plausible pharmacological pathway to overdose mortality: the Consumer Price Index (CPI), analyzed as the 12-month percent change in the US Bureau of Labor Statistics all-items index for all urban consumers (CPIAUCSL) (79), a standard measure of consumer price inflation, and the University of Michigan Consumer Sentiment Index (MCSI) (80) (a survey-based measure of household economic confidence). For each negative control, we computed cross-correlation functions (CCFs) on unadjusted time series and then applied the same TFM process, using prewhitened residuals obtained by fitting auto-ARIMA models.

Ethics statement

This study used public, geographically aggregated, anonymized data and was not eligible for human subjects review, also because we used only data from decedents. The study was preregistered at the Open Science Framework: https://osf.io/h6fdw/.

Artificial intelligence

Generative artificial intelligence (AI) was used to create initial drafts of Python code, deployed within Deepnote.com notebooks. Instead of being one proprietary standalone foundation model, Deepnote AI functions as a virtual collaborator that relies on ensemble third-party generative AI models; specifically for code generation, Codeium Windsurf was deployed. Generative AI was also used to generate initial drafts of R code for ARIMA models, using Claude Opus 4.7 (Anthropic). All code was reviewed and modified by humans, and final models were run in traditional (non-AI) integrated development environments. Citations were identified and manually inserted using the non-AI software (Zotero, version 9), but Claude was used to find errors in final formatting for citations under nongenerative conditions to conform to journal standards.

Acknowledgments

We acknowledge the intellectual contribution of Vangelov et al. (7) to open this discussion and for innovative methods that we replicated. We thank the peer reviewers who provided material feedback to improve the manuscript. N.D. is a noncompensated board member of a nonprofit organization (Remedy Alliance/For The People) that bulk distributes the opioid overdose antidote naloxone to nongovernmental organizations and public health departments. We are grateful to our colleagues at our institutions who provide support services for preprocessing data, as well as accounting and administrative support.

Funding:

This work was supported by the NIH/NIDA [1R21DA058583, Leveraging State Drug Overdose Data to Build a Comprehensive Case Level National Dataset to Inform Prevention and Mitigation Strategies (to L.A.P.)], NIH/NIDA Opioid Policy Tools and Information Center of Excellence [grant no. 2P50DA046351 (to R.L.P.)], John D. and Catherine T. MacArthur Foundation Fellowship 2025-2030 (to N.D.), and Foundation for Opioid Response Efforts (to N.D.).

Author contributions:

Conceptualization: N.D., P.G., A.H.K., L.A.P., R.L.P., and A.S. Data curation: N.D. Formal analysis: N.D. Funding acquisition: N.D. Investigation: N.D., R.L.P., and A.S. Methodology: K.G.C., N.D., L.A.P., R.L.P., and A.S. Project administration: N.D. and P.G. Resources: N.D. and P.G. Software: N.D. Supervision: N.D. Validation: N.D., P.G., and R.L.P. Visualization: N.D. and P.G. Writing—original draft: N.D., R.L.P., and A.S. Writing—review and editing: K.G.C., N.D., P.G., A.H.K., L.A.P., R.L.P., A.S., and S.T.

Competing interests:

The authors declare that they have no competing interests.

Data, code, and materials availability:

All data and code needed to evaluate and reproduce the results in the paper are present in the paper and/or the Supplementary Materials. This study did not generate new materials. Code, datasets, codebook, and output in the form of Jupyter .ipynb notebooks are available from the Dryad repository: https://doi.org/10.5061/dryad.zkh1893rq.

Supplementary Materials

This PDF file includes:

REFERENCES

1.Dowell D., Nataraj N., Rikard M., Park J., Zhang K., Baldwin G.,
Why have overdose deaths decreased? Widespread fentanyl saturation and decreased drug use among key drivers. Lancet Reg. Health Am.
51,
101226 (2025).

2.D. R. Holtgrave, R. LaBelle, “Policies and programs to accelerate declines in U.S. drug fatalities,” Health Affairs Forefront, 14 April (2025); 10.1377/forefront.20250411.182531.

3.Shin J., Ledingham E. M., McKenzie M., McKee H., McCray E., Turbidez J., Reichley N., St John K., Hallowell B. D.,
Trends from the Rhode Island Harm Reduction Surveillance System: 2021-2024. R. I. Med. J.
109,
25–30 (2026).

4.Liu S., Casillas S., Ussery E., Sheppard M., Vivolo-Kantor A.,
Trends of all opioid-, heroin-, and fentanyl-involved nonfatal overdoses in the United States, October 2020–April 2025. Drug Alcohol Depend.
276,
112886 (2025).

5.J. R. Friedman, J. J. Palamar, D. Ciccarone, T. L. Gaines, A. Borquez, C. L. Shover, S. A. Strathdee, Charting the decline of the fourth wave: US overdose deaths by race, geography, and substance involvement. medRxiv 2026.01.25.26344769 [Preprint] (2026). 10.64898/2026.01.25.26344769.

6.National Center for Health Statistics, “VSRR Provisional Drug Overdose Death Counts” (National Center for Health Statistics, 2026); https://data.cdc.gov/National-Center-for-Health-Statistics/VSRR-Provisional-Drug-Overdose-Death-Counts/xkb8-kh2a/about_data.

7.Vangelov K., Humphreys K., Caulkins J. P., Pollack H., Pardo B., Reuter P.,
Did the illicit fentanyl trade experience a supply shock?
Science
391,
134–136 (2026).

8.Fischer B., Hall W.,
Recent decreases in drug overdose deaths: Assessing evidence for supply-related drivers and implications for health-oriented drug policy in Canada. Prev. Med.
209,
108607 (2026).

9.Monnat S. M.,
Demographic and geographic variation in fatal drug overdoses in the United States, 1999-2020. Ann. Am. Acad. Pol. Soc. Sci.
703,
50–78 (2022).

10.Post L. A., Lundberg A., Moss C. B., Brandt C. A., Quan I., Han L., Mason M.,
Geographic trends in opioid overdoses in the US from 1999 to 2020. JAMA Netw. Open
5,
e2223631 (2022).

11.Mason M., Soliman R., Kim H. S., Post L. A.,
Disparities by sex and race and ethnicity in death rates due to opioid overdose among adults 55 years or older, 1999 to 2019. JAMA Netw. Open
5,
e2142982 (2022).

12.Powell D., Han D., Grooms J., Pacula R. L.,
Geography and fentanyl: Explaining the disproportionate rise in opioid overdose deaths among Black Americans. Health Aff.
44,
1050–1059 (2025).

13.Weatherburn D., Jones C., Freeman K., Makkai T.,
Supply control and harm reduction: Lessons from the Australian heroin drought. Addiction
98,
83–91 (2003).

14.Stringfellow E. J., Lim T. Y., Hasgul Z., Jalali M. S.,
Structural drivers of the drop in opioid overdose deaths in the US. JAMA Psychiatry
83,
101–103 (2026).

15.Tobias S., Bird J., Guzman S. M., Palis H., Crabtree A., Hein J. E., Wood E., Ti L.,
Temporal and regional associations between fentanyl concentrations in the unregulated drug supply and drug-related mortality in British Columbia, Canada. Int. J. Drug Policy
150,
105200 (2026).

16.Hempstead K., Yildirim E. O.,
Supply-side response to declining heroin purity: Fentanyl overdose episode in New Jersey. Health Econ.
23,
688–705 (2014).

17.Baldwin N., Gray R., Goel A., Wood E., Buxton J. A., Rieb L. M.,
Fentanyl and heroin contained in seized illicit drugs and overdose-related deaths in British Columbia, Canada: An observational analysis. Drug Alcohol Depend.
185,
322–327 (2018).

18.Kennedy M. C., Dong H., Tobias S., Buxton J. A., Lysyshyn M., Tupper K. W., Ti L.,
Fentanyl concentration in drug checking samples and risk of overdose death in Vancouver, Canada. Am. J. Prev. Med.
66,
10–17 (2024).

19.Kleinman R. A.,
Fentanyl, carfentanil and other fentanyl analogues in Canada's illicit opioid supply: A cross-sectional study. Drug Alcohol Depend. Rep.
12,
100240 (2024).

20.Zoorob M. J., Park J. N., Kral A. H., Lambdin B. H., Del Pozo B.,
Drug decriminalization, fentanyl, and fatal overdoses in Oregon. JAMA Netw. Open
7,
e2431612 (2024).

21.B. Pardo, J. Taylor, J. P. Caulkins, B. Kilmer, P. Reuter, B. D. Stein, “The future of fentanyl and other synthetic opioids” (Research Report, RAND Corporation, 2019).

22.Kilmer B., Pardo B., Pujol T. A., Caulkins J. P.,
Rapid changes in illegally manufactured fentanyl products and prices in the United States. Addiction
117,
2745–2749 (2022).

23.Zoorob M.,
Fentanyl shock: The changing geography of overdose in the United States. Int. J. Drug Policy
70,
40–46 (2019).

24.Lim T. Y., Dong H., Stringfellow E., Hasgul Z., Park J., Glos L., Kazemi R., Jalali M. S.,
Temporal and spatial trends of fentanyl co-occurrence in the illicit drug supply in the United States: A serial cross-sectional analysis. Lancet Reg. Health Am.
39,
100898 (2024).

25.Gladden R. M., O'Donnell J., Mattson C. L., Seth P.,
Changes in opioid-involved overdose deaths by opioid type and presence of benzodiazepines, cocaine, and methamphetamine—25 states, July-December 2017 to January-June 2018. Morb. Mortal. Wkly. Rep.
68,
737–744 (2019).

26.M. Godvin, J. R. Friedman, C. A. Molina, A. J. Koncsol, R. Romero, D. N. Juurlink, C. L. Shover, Estimating the daily milligrams of morphine equivalent of illicit fentanyl use in Los Angeles: Clinical and epidemiological implications. medRxiv 2025.10.07.25337514 [Preprint] (2026). 10.1101/2025.10.07.25337514.

27.Tobias S., Grant C. J., Laing R., Arredondo J., Lysyshyn M., Buxton J., Tupper K. W., Wood E., Ti L.,
Time-series analysis of fentanyl concentration in the unregulated opioid drug supply in a Canadian setting. Am. J. Epidemiol.
191,
241–247 (2022).

28.Shover C. L., Koncsol A. J., Godvin M. E., Goodman-Meza D., Pardo B., Poimboeuf M., Molina C. A., Romero R., Feng J., Friedman J. R.,
High variation in purity of consumer-level illicit fentanyl samples in Los Angeles, September 2023–April 2025. Int. J. Drug Policy
145,
104977 (2025).

29.G. E. P. Box, G. M. Jenkins, G. C. Reinsel, Time Series Analysis: Forecasting and Control, Wiley Series in Probability and Statistics (Wiley, ed. 4, 2008).

30.Dahlen A., Lei F., Agyabeng K., Chen R., Johnson C. E., Amaro G., Spinnler J., Khezri M., Pagán J. A., Healton C., Farhat T. M.,
Proportion of fentanyl reports in illicit drug seizures and opioid mortality. JAMA Health Forum
7,
e256286 (2026).

31.Glynn R. J., Knight E. L., Levin R., Avorn J.,
Paradoxical relations of drug treatment with mortality in older persons. Epidemiology
12,
682–689 (2001).

32.Toronto Public Health Surveillance & Epidemiology Unit, Toronto Overdose Information System, http://public.tableau.com/app/profile/tphseu/viz/TOISDashboard_Final/ParamedicResponse [accessed 11 March 2026].

33.Zibbell J. E., Aldridge A. P., Cauchon D., DeFiore-Hyrmer J., Conway K. P.,
Association of law enforcement seizures of heroin, fentanyl, and carfentanil with opioid overdose deaths in Ohio, 2014-2017. JAMA Netw. Open
2,
e1914666 (2019).

34.Park J. N., Rashidi E., Foti K., Zoorob M., Sherman S., Alexander G. C.,
Fentanyl and fentanyl analogs in the illicit stimulant supply: Results from U.S. drug seizure data, 2011-2016. Drug Alcohol Depend.
218,
108416 (2021).

35.Drug Enforcement Administration, Office of Forensic Sciences, Special Testing and Research Laboratory, “Annual Fentanyl Report CY 2023” (PRB# 2025-017, Drug Enforcement Administration, 2025); www.dea.gov/documents/2025/2025-09/2025-09-22/cy-2023-annual-fentanyl-report.

36.Post L. A., Ciccarone D., Unick G. J., D'Onofrio G., Kwon S., Lundberg A. L., Sharma S., Mason M.,
Decline in US drug overdose deaths by region, substance, and demographics. JAMA Netw. Open
8,
e2514997 (2025).

37.US Centers for Disease Control and Prevention, “Medetomidine in the U.S. illegal fentanyl supply increasing risk for overdose and severe withdrawal syndrome—Health Advisory” (2026); www.cdc.gov/han/php/notices/han00527.html.

38.Walton S. E., Stang B. N., Kacinko S., Papsun D. M., Logan B. K., Krotulski A. J.,
Medetomidine quantitation and enantiomer differentiation in biological specimens collected after fatal and non-fatal opioid overdoses. J. Anal. Toxicol.
49,
551–558 (2025).

39.A. Winokur, DEA Emerging Trends (2025); https://orsprogram.org/wp-content/uploads/2025/12/May-2025-ORS-TAT-Webinar-Slides.pdf.

40.Tobias S., Bird J., Law M. R., Buxton J. A., Ti L.,
How population-level drug supply data are integrated into health outcomes research: A scoping review. Epidemiol. Rev.
48,
mxag010 (2026).

41.Algera M. H., Olofsen E., Moss L., Dobbins R. L., Niesters M., Van Velzen M., Groeneveld G. J., Heuberger J., Laffont C. M., Dahan A.,
Tolerance to opioid-induced respiratory depression in chronic high-dose opioid users: A model-based comparison with opioid-naïve individuals. Clin. Pharmacol. Ther.
109,
637–645 (2021).

42.Dyer K. R., Foster D. J., White J. M., Somogyi A. A., Menelaou A., Bochner F.,
Steady-state pharmacokinetics and pharmacodynamics in methadone maintenance patients: Comparison of those who do and do not experience withdrawal and concentration-effect relationships. Clin. Pharmacol. Ther.
65,
685–694 (1999).

43.Newman S. T., McNamara I. A., Campbell K. D., Park B., Carpenter R. W., Blanchard B. A., Thater P., Green L., Winograd R. P.,
An interrupted time series analysis of fentanyl, naloxone, and opioid-involved deaths in five counties in Eastern Missouri. J. Subst. Use Addict. Treat.
169,
209564 (2025).

44.Sibley A. L., Miller C. W., Joniak-Grant E., Bell A., Visnich M., Alsum S., Dasgupta N.,
A brick to a bundle: A qualitative study of behavioral responses to xylazine adulteration. Int. J. Drug Policy
145,
105017 (2025).

45.W. H. Eger, M. Plesons, T. S. Bartholomew, A. R. Bazzi, M. H. Hauschild, C. C. McElrath, C. Owens, D. W. Forrest, H. E. Tookes, E. L. Crable, Protective or potentially harmful? Altering drug consumption behaviors in response to xylazine adulteration. Research Square [Preprint] (2024); 10.21203/rs.3.rs-4810429/v1.

46.Eger W. H., Plesons M., Bartholomew T. S., Bazzi A. R., Hauschild M. H., McElrath C. C., Owens C., Forrest D. W., Tookes H. E., Crable E. L.,
Syringe services program staff and participant perspectives on changing drug consumption behaviors in response to xylazine adulteration. Harm Reduct. J.
21,
162 (2024).

47.Voronkov M., Cernea M., Stefanut C., Nikonov G., Milevich G., Abernethy J.,
Does kappa agonism improve reversal of 'tranq-dope' overdose? Evidence from a rodent model. Pharmaceuticals
19,
846 (2026).

48.Love J. S., Levine M., Aldy K., Brent J., Krotulski A. J., Logan B. K., Vargas-Torres C., Walton S. E., Amaducci A., Calello D., Hendrickson R., Hughes A., Kurt A., Judge B., Pizon A., Schwarz E., Shulman J., Wiegand T., Wax P., Manini A. F.,
Opioid overdoses involving xylazine in emergency department patients: A multicenter study. Clin. Toxicol.
61,
173–180 (2023).

49.InSight Crime, Local Rules, Global Lessons: How Criminal Governance Shapes Fentanyl Markets in Northern Mexico ( 2026); https://insightcrime.org/wp-content/uploads/2026/03/Local-Rules-Global-Lessons_-How-Criminal-Governance-Shapes-Fentanyl-Markets-on-Mexicos-Northern-Border.pdf.

50.England K., Luo L., Verdery A. M., Monnat S. M.,
Shifting cohort patterns in the use of drugs with elevated overdose risk in the United States. Soc. Probl.
72,
1239–1259 (2025).

51.Smith M. K., Planalp C., Bennis S. L., Stately A., Nelson I., Martin J., Evans P.,
Widening racial disparities in the U.S. overdose epidemic. Am. J. Prev. Med.
68,
745–753 (2025).

52.Cadet K., Smith B. D., Martins S. S.,
Intersectional racial and sex disparities in unintentional overdose mortality. JAMA Netw. Open
8,
e252728 (2025).

53.Karandinos G., Baggett T. P., Ciccarone D.,
Understanding the health impacts of the rapid shift to drug smoking in the US. JAMA Intern. Med.
186,
639–640 (2026).

54.Megerian C. E., Bair L., Smith J., Browne E. N., Wenger L. D., Guzman L., Kral A. H., Lambdin B. H.,
Health risks associated with smoking versus injecting fentanyl among people who use drugs in California. Drug Alcohol Depend.
255,
111053 (2024).

55.McCormick K. A., Samora J., Claborn K. R., Steiker L. K. H., DiNitto D. M.,
A systematic review of macro-, meso, and micro-level harm reduction interventions addressing the U.S. opioid overdose epidemic. Drugs
32,
1–14 (2025).

56.Dayton L., Mazhnaya A., Schneider K. E., Kong X., Winiker A., Davey-Rothwell M., Tobin K. E., Latkin C. A.,
Trends in overdose experiences and prevention behaviors among people who use opioids in Baltimore, MD, 2017-2019. Drug Alcohol Depend.
221,
108650 (2021).

57.Jones C. M., Zhang K., Han B., Guy G. P., Losby J., Einstein E. B., Delphin-Rittmon M., Volkow N. D., Compton W. M.,
Estimated number of children who lost a parent to drug overdose in the US from 2011 to 2021. JAMA Psychiatry
81,
789–796 (2024).

58.Verdery A. M., Ryan-Claytor C., Smith-Greenaway E., Sarkar N., Livings M.,
More than 1.4 million US children have lost a family member to drug overdose. Am. J. Public Health
114,
1394–1397 (2024).

59.Nunes J. C., Costa G. P. A., De Aquino J. P., Adinoff B.,
Expanding access to buprenorphine and methadone: Global perspectives and policy recommendations. Subst. Use Addict. J.
47,
595–610 (2026).

60.Crockett M. A., Núñez D., Martínez P., Borghero F., Campos S., Langer Á. I., Carrasco J., Martínez V.,
Interventions to reduce mental health stigma in young people: A systematic review and meta-analysis. JAMA Netw. Open
8,
e2454730 (2025).

61.Krawczyk N., Rivera B. D., Levin E., Dooling B. C. E.,
Synthesising evidence of the effects of COVID-19 regulatory changes on methadone treatment for opioid use disorder: Implications for policy. Lancet Public Health
8,
e238–e246 (2023).

62.Siegal N., Gupta S., Miles J., Samples H., Simon K., Aalsma M. C., Crystal S.,
Medicaid expansion and buprenorphine dispensing in early vs recent expansion states. JAMA Netw. Open
9,
e2559803 (2026).

63.Friedman J. R., Palamar J. J., Ciccarone D., Gaines T. L., Borquez A., Shover C. L., Strathdee S. A.,
US overdose mortality saw first drop below the Jalal-Burke exponential growth curve in 2024. Int. J. Drug Policy
153,
105320 (2026).

64.Grobman B., Liu M., Mansur A., Lu C. Y.,
Trends and disparities in firearm homicides in the USA, 2018-2024. J. Gen. Intern. Med.
41,
1694–1697 (2026).

65.Cullip M., Rosenthal J. S., Dmetrichuk J. M.,
Trending to baseline: A long-term look at the COVID-19 pandemic effects on non-natural deaths in Ontario. Forensic Sci. Int.
387,
113031 (2026).

66.Pressley J. C., Aziz Z., Pawlowski E., Hines L., Roberts A., Guzman J., Bauer M.,
Using a Safe System framework to examine the roadway mortality increase pre-COVID-19 and in the COVID-19 era in New York State. Int. J. Environ. Res. Public Health
22,
61 (2025).

67.Patel V. R., Liu M., Jena A. B.,
Suicide mortality among adolescents and young adults after launch of a suicide and crisis lifeline. JAMA
335,
1721–1723 (2026).

68.Cano M., Daniulaityte R., Marsiglia F.,
Xylazine in overdose deaths and forensic drug reports in US states, 2019-2022. JAMA Netw. Open
7,
e2350630 (2024).

69.Alexandridis A. A., McCort A., Ringwalt C. L., Sachdeva N., Sanford C., Marshall S. W., Mack K., Dasgupta N.,
A statewide evaluation of seven strategies to reduce opioid overdose in North Carolina. Inj. Prev.
24,
48–54 (2018).

70.D'Errico S.,
Commentary. Fentanyl-related death and the underreporting risk. J. Forensic Leg. Med.
60,
35–37 (2018).

71.US Drug Enforcement Administration, 2025 National Drug Threat Assessment (DEA-DCT-DIR-007-25, 2025); www.dea.gov/sites/default/files/2025-07/2025NationalDrugThreatAssessment.pdf.

72.Slavova S., Freeman P. R., Rock P., Brancato C., Hargrove S., Liford M., Quesinberry D., Walsh S. L.,
Changing trends in drug overdose mortality in Kentucky: An examination of race and ethnicity, age, and contributing drugs, 2016-2020. Public Health Rep.
138,
131–139 (2023).

73.G. E. P. Box, G. M. Jenkins, Time Series Analysis: Forecasting and Control, Holden-Day Series in Time Series Analysis and Digital Processing (Holden-Day, 1979).

74.Monnet D. L., López-Lozano J. M., Campillos P., Burgos A., Yagüe A., Gonzalo N.,
Making sense of antimicrobial use and resistance surveillance data: Application of ARIMA and transfer function models. Clin. Microbiol. Infect.
7,
29–36 (2001).

75.Catalano R. A., Dooley D., Jackson R.,
Selecting a time-series strategy. Psychol. Bull.
94,
506–523 (1983).

76.Schwarz G.,
Estimating the dimension of a model. Ann. Stat.
6,
461–464 (1978).

77.N. R. Draper, H. Smith, Applied Regression Analysis, Wiley Series in Probability and Statistics (Wiley, ed. 3, 1998).

78.Beard E., Marsden J., Brown J., Tombor I., Stapleton J., Michie S., West R.,
Understanding and using time series analyses in addiction research. Addiction
114,
1866–1884 (2019).

79.U.S. Bureau of Labor Statistics, “Consumer Price Index for All Urban Consumers: All Items in U.S. City Average [CPIAUCSL]” (retrieved from FRED, Federal Reserve Bank of St. Louis, 2026); http://fred.stlouisfed.org/series/CPIAUCSL.

80.University of Michigan, “University of Michigan: Consumer Sentiment [UMCSENT]” (retrieved from FRED, Federal Reserve Bank of St. Louis, 2026); http://fred.stlouisfed.org/series/UMCSENT.

Associated Data

Supplementary Materials

Data Availability Statement

All data and code needed to evaluate and reproduce the results in the paper are present in the paper and/or the Supplementary Materials. This study did not generate new materials. Code, datasets, codebook, and output in the form of Jupyter .ipynb notebooks are available from the Dryad repository: https://doi.org/10.5061/dryad.zkh1893rq.

Come citare questo studio

Nabarun Dasgupta, Adams Sibley, Paula Gildner, Katherine Gora Combs, Lori A Post, Sam Tobias, Alex H Kral, Rosalie Liccardo Pacula. (2026). Fentanyl purity and overdose decline: A reexamination of geographic trends.. Science advances. DOI: 10.1126/sciadv.aei7742

Consulta la pubblicazione originale ↗