Analyses were conducted in R Studio Version (2023.12.0+369)⁴² using the following packages: ‘epiR’⁴³, ‘lmtests’⁴⁴, ‘sandwich’⁴⁶, ‘effsize’⁴⁵, ‘MatchIT’⁴⁶, ‘ggplot2’⁴⁷, ‘patchwork’⁴⁸, ‘viridis’⁴⁹. First, after calculating descriptive statistics to summarize sample characteristics, distributions of MoCA total scores were compared using histograms and density plots, and crude prevalences of MCI were estimated in FDNY and GRC responders using conservative MoCA criteria (18 < MoCA < 23). The prevalence of MCI was also estimated in FDNY responders using different algorithmic routines applied to domain specific measures of cognitive function, yielding multiple operational definitions of MCI, including Jak/Bondi criteria⁵⁰, Petersen criteria⁵¹, and NIA-AA criteria⁵¹. To help visualize heterogeneity, determine clinical significance, and contextualize the magnitude of estimates, forest plots were used to display crude prevalence rates of MCI for WTC-exposed FDNY responders juxtaposed with meta-analytic estimates from different global, community, and clinical populations.

Second, generalized linear models were used to evaluate cohort differences (FDNY vs. GRC) in the prevalence of MCI, before and after controlling for age, sex, self-reported race/ethnicity, level of education, and county of residence. Specifically, robust Poisson regressions using a Huber-White sandwich estimator for variance estimation were used because odds ratios derived from logistic regressions overestimate relative risk ratios when outcomes are common (e.g., >10%)⁵². Similar Poisson regressions were then used to test whether FDNY responders and New York City Police Department (NYPD) officers, a subset of the GRC henceforth called GRC-NYPD, differ in the prevalence of MCI, as members of the FDNY and GRC-NYPD are similar regarding screening criteria for employment (both must complete a written examination and pass medical, social, and psychological tests) but differ in terms of their work experiences and exposures. Finally, generalized linear models were then estimated after excluding responders 65 years and older.

Third, as a sensitivity analysis, propensity score matching was used to estimate the average marginal cohort difference (FDNY vs. GRC) in the prevalence of MCI accounting for potential confounding by the included covariates (age, sex, race/ethnicity, education, and county of residence). 1:1 nearest neighbor matching was conducted without replacement on a propensity score estimated using logistic regression. To estimate the cohort differences and the corresponding standard error, we fit a quasi-Poisson regression with MCI as the outcome and cohort and covariates included as simultaneous predictors. Covariates were still included after matching for “doubly robust” estimation to reduce any remaining imbalance in case the 1:1 matching is not exact⁵³. G-computation in the matched sample was used to estimate the cohort difference, and a cluster-robust variance was used to estimate its standard error with matching stratum membership as the clustering variable.

Fourth, a series of multiple linear regressions with heteroskedasticity-consistent standard errors (identical to Stata’s ‘robust’ option) were used to investigate the associations of domain-specific cognitive, psychological, and physical functional tests with mild-to-severe cognitive impairment, defined as a binary variable (0=not impaired, 1=impaired) using conservative MoCA criteria (total scores < 23) and adjusting for age-related differences in test performance. Finally, raw scores from norm-referenced tests were transformed to age-adjusted scaled scores (either t-scores or z-scores), accompanied by their recommended interpretations (e.g., “average”, “low average”, “borderline”, “impaired”, “moderate”, “severe”, and “profound”) to help determine the severity of impairments across specific domains of cognitive function.

The prevalence of MCI determined using different diagnostic criteria for the FDNY responders and using MoCA total scores for both FDNY and GRC responders are reported in Table 2. In the FDNY cohort, different operational definitions of MCI (Jak/Bondi, Petersen, NIA-AA) yielded similar prevalence estimates for the full sample (range= 51.91% to 58.11%) and for FDNY responders under 65 years old (range=52.57% to 60.32%). When MCI was determined using conservative criteria applied to total MoCA scores, prevalence estimates for FDNY responders were markedly lower (24.78%) than prevalence estimates based on canonical operational definitions of MCI. To contextualize the clinical significance of these findings, crude prevalence estimates of MCI for WTC-exposed FDNY responders based on canonical operational definitions are descriptively juxtaposed with meta-analytic estimates of MCI in Figure 1. Notably, crude prevalence rates were higher for WTC responders compared to meta-analytic estimates from community and clinical populations.

On average, MoCA total scores were significantly lower for FDNY responders (mean = 23.83, median = 24) compared to GRC responders (mean = 24.51, median = 25), as indicated by Welch’s t-test (p < .001) and Wilcoxon rank sum test with continuity correction (p < .001), amounting to a small standardized mean difference (Cohen’s d with Hedge’s correction = .24, 95% CI = [.13, .35]). The crude prevalence of MCI was also higher for FDNY than GRC responders when using MoCA criteria (24.78% vs. 18.11%). The observed distributions of MoCA total scores for FDNY and GRC responders that were used to determine MCI are depicted in Figure 2.

Results of Poisson regressions testing cohort differences in MCI are reported in Table 3. MCI prevalence remained significantly higher for WTC-exposed FDNY responders, compared to WTC-exposed responders from the GRC, adjusting for differences in sample characteristics (adjusted RR=1.53, 95% CI = [1.24, 1.88], p < .001), and similarly, when compared to GRC-NYPD law enforcement responders (adjusted RR=1.43, 95% CI = [1.14, 1.80], p = .002). Reported in Table 3, the results of null hypothesis significance tests remained unchanged after excluding responders 65 years and older. Finally, after using propensity score nearest neighbor matching to create covariate-balanced samples of equal size (343 matched pairs, total n = 686), the standardized mean differences for the covariates indicated excellent balance (see supplemental Figure S1). Quasi-Poisson regression estimated the relative risk to be 1.13 (CI 95% = 1.07 – 1.20, p < 0.001), indicating that the average cohort difference (FDNY vs. GRC) in the prevalence of MCI remained statistically significant after propensity score matching on age, sex, race/ethnicity, education, and county of residence.

To examine the domain-specific features of cognitively impaired FDNY responders, Figure 3 displays the age-adjusted correlates of mild-to-severe cognitive impairment diagnosed using MoCA total scores < 23. Specifically, standardized regression coefficients (β) from a series of multivariable linear regressions are plotted with 95% confidence intervals and p-values calculated using heteroskedasticity-consistent standard errors. These coefficients quantify whether responders with cognitive impairment, on average, exhibit deficits in a specific domain of cognitive function, while adjusting for the effects of age. Compared to unimpaired responders, responders with cognitive impairment exhibited deficits in all experimenter-administered tests of cognitive function, including total recall (β = −0.65, p < .001), delayed recall (β = −0.59, p < .001), retention (β = −0.29, p = .038), and recognition (β = −0.49, p < .001) measured by the HVLT, visual memory measured by the BNT (β = −0.39, p = .016), reading comprehension measured by the WRAT-4 Reading (β = −0.41, p = .002), psychomotor speed measured by the TMT-A (β = −0.40, p = .003), choice reaction speed and attentional shifting measured by the TMT-B (β = −0.58, p < .001), processing speed measured by the SDMT (β = −0.34, p = .004), and verbal fluency measured by the COWA (β = −0.73, p = .002).

On average, compared to unimpaired responders, responders with cognitive impairment also exhibited deficits in computerized assessments of episodic memory (β = −0.49, p < .001), working memory (β = −0.30, p = .006), cognitive throughput (β = −0.34, p = .021), and visual learning (β = −0.29, p < .001). Notably, self-reports of subjective cognitive concerns (average CFI scores) were not significantly different when comparing cognitively impaired and unimpaired responders (β = 0.06, p = .679). Similarly, indicators of physical function were not significantly different when comparing cognitively impaired and unimpaired responders, including lower extremity function measured by the SPPB (β = −0.05, p = .666) and maximum grip strength (β = −0.02, p = .864).

The severity of impairments across specific domains of cognition that were measured by norm-referenced tests are depicted in Figure 4 and reported in Table S3. The age-adjusted prevalence rates of moderate (0.58%–3.50%), severe (1.17%–6.12%), and profound (8.75%–10.79%) impairments were high for tests of verbal learning and memory (measured by the HVLT; panels A-D), compared to certain executive functions, specifically verbal fluency and processing speed (measured by the COWA & SDMT; panels E & H), whereby moderate impairments were quite rare (0%–0.29%) and severe and profound impairments were entirely absent. On the other hand, the age-adjusted prevalence rates of moderate (0.87%–1.46%), severe (0%–0.29%), and profound (2.62%–9.62%) impairments were more common for other executive functions, like psychomotor speed and choice reaction speed (measured by the TMT-A & TMT-B). Finally, following NIA-AA diagnostic guidelines for early-onset dementia, 4.96% of WTC-exposed FDNY responders met the criteria for severe cognitive impairment (95% CI = [2.91% to 7.82%]). Furthermore, this prevalence rate remained largely unchanged after excluding FDNY responders over the age of 65 years (Table 2).

A strength of our study was the administration of an extensive battery of objective neurocognitive assessments, including tests of global and domain-specific neurocognitive function. This enabled the diagnosis of MCI using different algorithmically determined operational definitions in WTC-exposed FDNY responders, while simultaneously enabling a thorough domain-specific characterization of neurocognitive deficits among impaired responders. Another strength was our ability to directly compare differently exposed WTC groups – FDNY vs. GRC and FDNY vs. GRC-NYPD. A third strength was finding similar MCI prevalence rates across secondary operational definitions of MCI including Jak/Bondi criteria, Petersen criteria, and NIA-AA criteria, while estimates of MCI prevalence were even higher when using traditional MoCA criteria but noticeably lower when using more conservative MoCA criteria, as employed in past studies of MCI in GRC responders².

There are several limitations. First, the present study was cross-sectional and, consequently, provides no information about the rate of change in cognitive decline, potential reversion of MCI over time, long-term incidence, or prognosis. Related, despite following NIA-AA guidelines for the diagnosis of severe cognitive impairment³⁹, the lack of repeated measures precludes determination of the “continuing decline” criteria for major NCD. Second, although randomly selected, participants voluntarily chose whether to enroll in the study, introducing the potential for self-selection bias. Third, comparisons to non-WTC responders were descriptive due to potential comparability issues regarding operational definitions of MCI and differences in statistical methodologies. Fourth, apart from the MoCA, GRC-responders did not complete experimenter administered neurocognitive assessments and, therefore, could not be diagnosed using a wide variety of well-established algorithmic routines like FDNY-responders, limiting the number of directly comparable estimates across the FDNY and GRC. However, the MoCA was administered to both FDNY and GRC responders, enabling comparison of prevalence estimates of MCI across these cohorts. Fifth, like all non-experimental studies, cause-effect relationships cannot be discerned. However, such high prevalence rates of MCI among WTC-exposed responders, even higher than clinical samples of patients with serious neurological and medical conditions, is consistent with pathogenic occupational exposures endured during the response efforts to the WTC attacks. Similarly, such high prevalence rates of severe-to-profound impairments across specific domains of cognition (~3% to 16%), as well as a high prevalence rate of severe impairments and functional limitations indicative of a possible major NCD (~5%) are unlikely to be explained by common geriatric disorders, as these prevalence rates were considerably higher than what would be expected in the general population under 65 years of age (e.g., 0.01% to 0.50% prevalence for early onset dementia)^16–20. Nevertheless, future analyses will investigate the relationship between level of WTC exposures and cognitive dysfunction. Sixth, no genetic testing was done among WTC-exposed FDNY responders, so we cannot rule out the possibility that APOE e4 or other genetic factors might contribute to the high rates of cognitive impairment observed in the present study. However, this seems unlikely as previous research in GRC responders has shown the link between WTC exposure severity and cumulative hazard of cognitive impairment is not explained by polygenic risk for Alzheimer’s Disease⁷⁰. Seventh, FDNY responders were predominately White and male which is reflected in this study’s source population. Consequently, the generalizability of findings to more diverse populations might be limited. Finally, though cognitive symptoms are one of the most common functional indicators of neurodegenerative disease, the present study did not validate findings using neuroimaging or other biomarkers. Future work should seek to corroborate the present findings, not only using longitudinal studies of cognition, but also using targeted neuroimaging and serological studies.

In the first study to examine the prevalence of MCI and possible major NCD in WTC-exposed-FDNY-responders, we identified a slightly higher level of MCI compared to WTC-exposed-GRC-responders in propensity score analyses that controlled for differences in sample characteristics and demographic factors. Crude prevalence rates of MCI based on canonical criteria (Jak/Bondi, Petersen, and NIA-AA) were also higher than recent meta-analytic estimates from different community and clinical populations, and rates of severe cognitive impairment were higher than expected in the general population based on meta-analyses of early onset dementia. Cognitive symptoms were widespread, with deficits that spanned domains of cognitive function, including verbal learning, memory, psychomotor speed, and choice reaction speed. Such high rates of cognitive impairments are concerning. Researchers and clinicians focused on screening and prevention efforts should give occupational and environmental exposures serious consideration as potential pathogenic risk factors for cognitive impairment. Future research is needed to examine longitudinal rates of cognitive decline in FDNY-responders and other cohorts with significant occupational and environmental exposures, as well as the impact of comorbidities and treatments on rate of decline. Following occupational exposures, clinicians should include screening and treatment planning for mild-to-severe cognitive impairment, and this should be a high priority for public health efforts moving forward.